Sai Initiative

Sai is Project Shohin's return to its original objective: build the strongest practical model near four billion parameters. This repository is the live scratchpad and implementation surface for that effort.

Nothing is called an improvement until it beats the unchanged parent and an equal-compute control on real, source-disjoint benchmarks.

Data precedes architecture. Sai first earns a trustworthy learning sequence: verified source bytes, quality and duplication controls, explicit prerequisite coverage, gradual difficulty, later rehearsal of fundamentals, and source-disjoint evaluation. A sophisticated mixer cannot compensate for bad examples or for teaching dependent concepts before their primitives. A 4B run may now be used as the decisive architecture experiment once its data is qualified, but the architecture is not called successful until that checkpoint beats the unchanged and equal-compute baselines on real benchmarks.

Sai treats data admission as three separate scientific questions:

1. Quality: is each example accurate, coherent, useful, legally and operationally admissible, and free of benchmark contamination, spam, answer farms, and high-confidence duplication?
2. Coverage: does the mixture teach the English, mathematics, code, science, technology, reasoning, and communication capabilities we intend to measure, without allowing one abundant source to crowd out the rest?
3. Pedagogy: are primitives taught before dependent concepts, are new ideas composed gradually, and are foundations rehearsed after difficulty rises?

These gates cannot substitute for one another. Clean advanced text is still a bad first lesson when its prerequisites are absent. A readable document is not necessarily foundational. A source with a high upstream quality score is not automatically a Sai training source. The target is not merely an easy-to-hard sort; it is a measured semantic learning progression whose exact records and ordering beat a same-record order control before scaling.

Sai data constitution

Data decisions precede tokenizer, architecture, and scale decisions.

The latest open-recipe evidence is reconciled in docs/SAI_2026_DATA_RESEARCH_SYNTHESIS.md. Sai uses progressive composition with continuous broad rehearsal, retains low-dose foundational code/math/science/technical exposure early, and measures specialist upsampling, reasoning mid-training, and long-context adaptation as separate factors rather than copying another model's ratios.

The corpus objective is stated more directly in docs/SAI_POLYMATH_DATA_THESIS.md. Web is a candidate reservoir rather than a protected percentage. The final pool must cover human expression, literature, arts, history, philosophy, society, everyday life, reference knowledge, mathematics, science, engineering, and software, and explicitly reward accurate cross-domain bridges. Selection is driven by marginal concept, style, and capability coverage under a spiraled prerequisite curriculum--not by copying another model's 75--85% web ratio.

The executable replacement for dataset concatenation is described in docs/SAI_DATA_COMPILER.md. It treats global raw sources as reality anchors, chooses preservation and English-translation policy, derives grounded representations, protects form-bearing human expression, constructs verified cross-domain and procedural reasoning, and leaves final sampling to a coverage- and model-responsive curriculum controller.

The first large book reservoir is now pinned in docs/SAI_INSTITUTIONAL_BOOKS_COMPILER.md. Its 983,004 Harvard Library volumes are screened metadata-first rather than blindly downloaded. Sai separates archive facts from model inferences, measures linguistic/conceptual/reasoning complexity independently, builds cited prerequisite edges, and uses a spiral schedule that never stops rehearsing fundamentals. Valuable non-English technical works route to English; literature must use a reputable human translation or separately labeled literal and literary synthetic translations. Dataset terms and per-volume rights evidence remain independent admission gates.

The first metadata-only book queue now contains 10,000 duplicate-safe volumes across 772 language×subject cells: 9,409 non-English and 591 English. This is a translation-discovery workload, not a fixed training ratio. Authenticated access has now replayed the first exact enriched-text shard and built a separate 185-volume compiler pilot. The gated text remains local under the pinned IDI terms; it is not redistributed through GitHub or Hugging Face.

The durable data catalog is Godlydonuts/Sai. It separates upstream source references, model judgments, verified representations, curriculum artifacts, and final training shards. The registry never treats a downloaded dataset as training-ready and records reference-only sources without copying bytes when their terms prohibit redistribution.

Hugging Face may display an inherited Parquet split named train or test. Those names describe the upstream file layout, not Sai admission state. A row under an upstream train split can still be corrupt, contextless, duplicated, non-English, contaminated, or rights-blocked; an upstream test split is never silently treated as Sai evaluation data. Only a future hash-receipted verified/ → curriculum/ → training/ release may be consumed by training.

Materialized Hugging Face source-lake checkpoint

The first physical source-lake boundary is now complete. At exact Hugging Face dataset head cc8576fbb3f949bdaf59049a150c1fa1d35f47c3, Sai has 13,974 byte-identical LFS source shards containing 8,802,247,613,960 bytes (8.802247614 TB; 8.0055975686 TiB). This exceeds the current 8.5 TB decimal registry-capacity target by 302,247,613,960 bytes and the earlier 8 TiB binary target by 6,154,591,752 bytes. Every included destination object was replayed against the size and SHA-256 of its pinned upstream object; the replay found zero identity mismatches.

Materialized source family	Files	Bytes	Snapshot status
PleIAs Common Corpus	10,000	4,489,486,652,558	Complete selected data snapshot; per-row rights required
FinePDFs	1,250	3,082,436,502,565	Deterministic partial snapshot at storage boundary
Common Pile, 31 families	845	540,438,290,489	Complete selected source snapshots
UltraData-Math L1	1,485	366,622,518,811	Complete source snapshot
Nemotron specialized reasoning	219	244,286,609,368	Complete source snapshot
Nemotron specialized v1.2	90	53,621,158,028	Complete source snapshot
FineMath-4plus	64	18,365,184,633	Complete source snapshot
Nemotron Legal v1	21	6,990,697,508	Complete source snapshot
Total	13,974	8,802,247,613,960	8.5 TB and 8 TiB physical targets met

Hugging Face accepted five deterministic FinePDFs copy batches and then returned a public-storage-quota 403 before batch six. No partial sixth batch was committed. Sai freezes the accepted head rather than silently changing the selection. The file-level manifest contains 13,974 rows and has SHA-256 56dc0d512db07aa26533decce8efd86bcb1b705355cd40069fdf9cb6311b5665; the aggregate receipt is 0715eefc3c3bda8ee800fc4c80155df461055da3bbf2a473ad6c93cf93bea9d8. Both live under artifacts/sai_hf_materialized_source_lake_20260825_r1, and the executable remote verifier is src/sai/data/hf_materialized_source_lake.py. The receipt, updated dataset card, and a four-part text publication of the manifest were remotely replayed at Hugging Face evidence head 122b98e0aa38130b0165ed3166cc7a569c3cddf0. Ordered concatenation of the four remote parts reproduced all 13,974 rows and the exact manifest SHA-256. The parts descriptor has canonical receipt 3287b272bd78bd12fca1b9a580928a8b0556815ecc404f57b3cb6ee401632c7a.

This is a major storage and custody milestone, not a training-readiness claim. These are compressed source candidates, not eight trillion tokens and not a final mixture. Global rights adjudication, English routing/translation, benchmark decontamination, exact and semantic deduplication, Hermes quality and pedagogy compilation, tokenizer accounting, curriculum assignment, and final Parquet publication remain fail-closed. Stokes's 10 Gbps connection is reserved for transformations and verification that cannot use zero-download Hugging Face server-side copies; its network link is not the present bottleneck.

The physical lake is now joined to the exact reservoir rights inventory in a 38-source fail-closed admission matrix under canonical receipt e70b65ebec4d451be5d4a7094fe798e1154019a0db79cf64d99ec1ff6ee26ab6. It accounts for every one of the 13,974 materialized files and all 8,802,247,613,960 bytes. Of those bytes, 5,027,859,142,584 require per-row license evidence, 3,100,801,687,198 have recognized declarations whose obligations still must be applied, and 673,586,784,178 require source-terms resolution. Every source separately exposes incomplete language/translation, decontamination, exact and semantic deduplication, full-population Hermes, representation, prerequisite, and spiral-curriculum gates. Physical custody can therefore never be mistaken for silent admission.

Source-lake retention is evidence-based, not permanent. A file is deleted when the whole exact object is proven unusable, or after every retained row has been published in a byte-hash-verified filtered replacement. Mixed shards are not deleted merely because an audit finds bad rows: those row identities are excluded immediately, then the raw shard is reclaimed only after its good rows have replacement custody. Each material deletion must record the remote path, object SHA-256, bytes removed, decision evidence, replacement path and hash when applicable, deletion commit, and whether recovery remains possible upstream. Provenance manifests, benchmark-boundary versions, and other replay evidence are retained even when they are small; deleting an extra pointer to the same LFS object is forbidden when it saves no object bytes but breaks historical replay. Transient acquisition caches are a separate boundary: a pinned, upstream-recoverable cache may be reclaimed without asserting whole-source unusability when it was never authoritative Sai custody, no active or admitted derivative depends on it, and the source-safe audit and recovery coordinates are already durable.

Full FineMath-4plus census

Sai has now scanned the entire materialized FineMath-4plus snapshot rather than extrapolating from an audit sample. Sixty-four independent Stokes CPU jobs hash-verified and scanned 6,699,493 rows, 34,126,971,204 UTF-8 text bytes, and 9,573,187,002 upstream tokens. All jobs exited successfully; the dependency-bound global aggregate completed in 15 seconds.

The global content census found zero byte-exact duplicate rows and only seven normalized duplicate rows after NFKC, case folding, and whitespace normalization. Upstream marks every row en, but its own language confidence places 349,840 rows below 0.50 and another 591,870 between 0.50 and 0.70. Only 2,339,612 rows contain a nonzero math-extraction feature, demonstrating that the source name and provider quality score cannot replace direct content qualification.

Three nested, non-admission measurement profiles quantify the available quality/length headroom:

Profile	Rows	UTF-8 text bytes	Upstream tokens
Broad: English confidence ≥0.70, score ≥3, 64–32k tokens	5,734,795	29,622,108,077	7,941,391,804
Core: English confidence ≥0.80, score ≥4, 128–32k tokens	4,879,600	26,493,021,099	6,949,987,393
Elite: English confidence ≥0.90, score ≥5, 128–32k tokens	250,540	1,570,280,996	402,534,176

The complete source-safe publication receipt is bb578f5e969e8d15d96ae40ae3511d4dd6d2d9c42e834e5c641204719d53e4c2. All 64 shard receipts and the aggregate replayed byte-for-byte at Hugging Face commit db79c6bb4e7752aee2de8ce2414fcf5ef709e5c1. These profiles remain training_ready=false: official-boundary decontamination, semantic/subdocument deduplication, content-density analysis, Hermes judgment, rights obligations, and final curriculum assignment are still required.

Conservative FineMath candidate qualification

The next population-wide pass intersected score 5, English confidence ≥0.90, explicit found_math=true, 128–32,767 upstream tokens, at least 512 UTF-8 bytes, and a valid source host. From all 6,699,493 rows it retained 56,654 mechanical candidates, 434,498,432 text bytes, and 117,445,585 upstream tokens. This is a candidate filter, not a quality-admission shortcut.

Exact matching against the frozen 27,979,728-word-shingle and 475,804-code- shingle official benchmark boundary then retained 52,277 rows and rejected 4,377. The cause-complete receipt reports 1,422 rows with word overlap, 3,290 with eligible-code overlap, and 335 with both; the candidate population had zero normalized-exact duplicates. A separate 32-CPU replay independently reproduced receipt b0ba86aaa60dddfdfae6653882d489fde1ecf3ab0f043d9a3954bdd38e191277 and output SHA-256 c61a840375572bca1a9872d50d99c66ee0452f8f9be4aacd89c1cd7af5d84a7a.

After a contextless-MCQ-answer-key failure mode was identified in the broader source lake, a source-agnostic high-precision filter was added and replayed over all 52,277 survivors. It found zero instances matching the strict bare-key signature in this conservative FineMath subset, so it changed no document identities. That does not absolve the wider lake: completed Hermès judgments already flag 36 answer-farm rows, 149 SEO/content-farm rows, and 97 corrupted rows among 2,243 audited documents. These are hard admission exclusions. The post-filter 512-row semantic audit population is frozen at receipt de365117bee119d224196a3a712518a2814214e130580bf643ef261c56327e1b.

Source-agnostic mechanical quality gate

The answer-key incident is now covered by a reusable gate rather than a FineMath-only exception. Before semantic admission, the gate detects bare MCQ keys, scored answer sheets without problem statements, embedded control-byte corruption, Unicode replacement-character corruption, repeated-character gibberish, contextless link/markup/structured fragments, and heavily duplicated boilerplate. Hard-reject, context-review, and cleanup-review routes take precedence over a mechanical pass. A pass never implies semantic quality, rights clearance, decontamination, global deduplication, or training readiness.

The code-expanded exact replay covers 10,371 distinct candidate rows across 13 current populations, with zero candidate-identity overlap and zero exact source-content overlap between those populations. It routed 10,360 to mechanical pass, held 9 for duplicated-boilerplate cleanup, routed 1 short contextless bibliographic form to context review, and hard-rejected 1 contextless Cambridge physics mark-scheme row. The catalog-form detector was added after a live Hermès audit exposed a 258-byte field list that named a valuable book but contained no book content; replay over all 10,371 identities found exactly that one new nonpass. That row independently triggered both the scored-answer-sheet detector and 136 embedded backspace controls. An initially broader scoring-marker detector was rejected during development because it falsely matched citations, array indexes, PEPs, and papers; those cases are regression fixtures in the final policy.

The current source-safe publication is artifacts/sai_source_mechanical_quality_gate_publication_20260826_r3.json, with canonical receipt df34d6507032269351df3d841032e068de5ff986dcbcb7d5f92f212e98e82385 and policy SHA-256 436ea538156447a7188a15404764302c7b3290b3a06c12677d316f265ccc6c80. The decision streams retain identities and measurements but no source text and are excluded from Git history. The preceding 12-population r2 evidence remains byte-replayed at immutable Hugging Face commit 4ab25974f7b7f40d5ef0bfe2dd8eedfe267831fc; the r3 publication adds the exact 2,048-row OpenCoder population and explicit source-content duplicate accounting. All 28 r3 evidence files plus the dataset card were downloaded back byte-identically from Hugging Face commit 444b1c482ff7e510d68f7e7115f1bf1d2087c936. The authorized Stokes evidence root contains the same evidence plus the dataset card as 29 byte-matched files under manifest SHA-256 929c6b46f7e4de7ca17c5fc360337465d33ec62ba284fbd5ca052c7d61a73c89. Every nonpass row is barred from direct admission, and all 10,371 rows remain training_ready=false.

Cross-domain connection compiler

Hermès has already proposed cross-domain metadata for 2,105 of 2,243 completed source judgments: 6,751 assignments, 730 distinct directed domain-pair labels, and 22,345 distinct concept labels. These are bridge proposals, not finished training examples.

Sai now converts that metadata into paired-source development work. A frozen 512-pair population uses 1,381 high-confidence English anchors across 290 directed bridge labels, caps each label at eight pairs and each source anchor at two, and requires different candidate and content identities. The paired synthesis contract requires exact evidence from both anchors, a conceptual explanation, worked transfer problem, counterexample, analogy limits, prerequisite map, and verification questions. Generated rows remain training_ready=false until independent claim verification, benchmark decontamination, deduplication, and transfer ablations close. The development proposal receipt is 0ef933e21252d060a4691cc9f5c63441bd40f4a796d5952db6651298c3c133e5. Generation now has a dependency-bound aggregate stage that requires exactly one receipt per pair and all 64 shard summaries, replays every normalized judgment, strips literal source quotes from derived candidates, and preserves only their SHA-256 bindings. Even after generation closes, the aggregate remains training_ready=false pending independent claim and transfer verification.

The first complete generation campaign closed on 2026-08-26 with 512/512 pair receipts and 64/64 shard summaries. It produced 512 unique derived candidates spanning all 290 frozen directed bridge labels from 1,381 source-disjoint anchors, using 3,488,388 prompt tokens and 1,169,045 completion tokens. Independent replay confirmed that no literal source quotes remain in the derived candidate file. Its candidate SHA-256 is d46f7e2c637b71085876cf180fff572095030e2eecf553ff4875aaee13e96bde and its canonical aggregate receipt is f2eaccffa188fe6ced475f7544006c863f0f9d3979031e35d984a12a6b0566e5. The exact aggregate and its 512-row verification population were copied and byte-replayed under the authorized Stokes evidence root. This is a completed generation result, not a quality result: the independently requested verifier is active, and every generated row remains training_ready=false.

A second, dependency-staged Hermès pass now verifies all 512 synthesized bridges against both restored exact anchors. It must cover every generated claim with a byte-exact quote and separately judge the shared structure, substantive domain link, worked transfer solution, counterexample, and analogy limits. Retention requires every check to pass; otherwise the bridge enters an explicit revision or rejection lane. This verifier is a separate request but uses the same model family, so its aggregate truthfully records independent_model_family_verification_complete=false, strips the private anchor text, and still requires decontamination, global deduplication, and transfer ablation before any bridge can become training-ready.

Prerequisite-edge compiler

Sai is also converting Hermès's document-level prerequisites_assumed and concepts_taught fields into graph-verification work. It does not equate co-occurrence with a prerequisite. After both active compiler populations close, the builder requires each proposed direction to recur in at least two distinct candidate and content identities, applies the same conservative source-quality floor, balances selection across domains, and freezes 192 repeated-evidence edges. A separate Hermès request then compares every edge against two or three exact source documents and distinguishes strict prerequisites, helpful foundations, co-taught nonedges, and unsupported directions with byte-exact quotes. Even positive same-family decisions remain graph candidates until independent verification and acyclic graph construction close; every route is training_ready=false.

The complete same-family verification closed on 2026-08-26 with 192/192 edge receipts and 64/64 shard summaries. It classified only 28 proposed directions as strict prerequisites and 4 as helpful foundations, while 120 were merely co-taught nonedges and 40 were unsupported. This is an important negative filter: repeated document-level co-occurrence overpredicted directional prerequisites for 160/192 proposals, so those edges are excluded instead of being allowed to distort the spiral curriculum. All output records retain only source-evidence hashes. The aggregate canonical receipt and file SHA-256 are 81546db8ddf82bb85c45ed4dd083f266116fd55b7084c75f4cdc12ad82846e82 and dcca78e7b017536781880532280ee5746601f212f5c9f04d08cf9d423feb4c28. The exact source-text-free aggregate is byte-replayed under the authorized Stokes evidence root. The 32 positive candidates still require independent model-family verification and acyclic graph construction, so none is yet training-ready.

All Hermès compiler-style workers now acquire the same persistent logical-shard lock before replaying or creating receipts. This permits dependency-prestaged single-process fan-out over disjoint bridge, book, prerequisite, representation, and verification shards without duplicate model calls. The local capacity fan-out begins only when its exact input receipt exists, preserves already-complete shard summaries, and retries only unresolved identities when the provider rate-limits a request.

Provider admission is now bounded across independent processes, not merely inside each worker. A live replay of the 50 most recently completed receipts at implementation time contained 62 HTTP-429 retries in addition to 50 valid responses, demonstrating that 32 simultaneous client attempts exceeded useful provider capacity. A subsequent 10-row frontier shard still exhausted all five retries on four rows at a 16-slot ceiling. The next accepted-capacity search measured 4.59 completed rows/minute at eight slots and 4.29 at sixteen. A prospective ten-slot probe then sustained 4.90 completed rows/minute over its first ten-minute comparison window, while HTTP-429 outcomes per completion fell from about 0.64 in the eight-slot cohort to 0.55. Subsequent loopback workers therefore share ten OS-locked request slots across every compiler, bridge, prerequisite, book, and verifier process. The limiter changes only request timing: candidate identities, prompts, model, temperature, reasoning effort, and receipt hashes remain governed by the same contracts. Non-loopback endpoints are unaffected, and new receipts record the applied shared limit explicitly.

A final bounded twelve-slot probe bracketed the optimum rather than assuming that ten was maximal. It held at most eleven simultaneous requests, produced zero valid probe receipts in 3m48s, and coincided with repeated five-attempt 429 exhaustions in the independently running frontier shard. Only the two probe lanes were terminated; healthy production workers and every completed receipt were preserved. Twelve was therefore rejected and the measured ten-slot limit remains active.

Transient HTTP retries are now deterministically staggered by candidate identity instead of waking every independent worker on the same 1/2/4/8-second boundaries. The exponential backoff and 30-second ceiling remain intact; only the transient-HTTP retry timing receives a stable 1.000--2.000 multiplier. This breaks provider-side retry herds without changing prompts, judgments, candidate assignments, request hashes, or the content of any accepted training record. Every new receipt names the retry-timing policy, while completed receipts remain immutable and replayable.

The resumed 1,024-row byte-weighted teacher population has now closed across FinePDFs, FineWeb-Edu, SmolLM, FineMath, Dolma, and OpenWebMath. Hermès returned 818 retain, 163 review, and 43 reject verdicts, but conservative routing sent 511/1,024 identities to quarantine, 176 to cleanup review, 116 to translation review, 89 to factual-grounding review, 23 to rights hold, and only 78 to representation verification. This is a coverage screen, not a source- wide yield estimate. It blocks bulk admission for all six sources: FinePDFs alone sent 324/596 sampled rows to quarantine, while OpenWebMath sent 20/35 to rights hold. Aggregate receipt 14f09f39cad9e8e7b0c4032deb3b1589d4ecbfc0f7ce4591beca43eb872fb784 and decision receipt 6c28bc37575e6b96e0857d942f91a903ca7f93f88e926dbacd875bd986313075 are byte-matched under weighted-reservoir-audit/20260826-r1 in the authorized Stokes evidence root.

All 511 quarantine routes are also sealed in a text-free dataset-exclusion manifest with 511 unique source-row, candidate-identity, and content hashes. Every record has dataset_materialization_allowed=false; rejected source text is absent. The manifest SHA-256 is 4811e01af6474a41322620bebffb781d8934c99a70da9f86f7a8209695a36185 and its canonical receipt is a4f279db7d7609453307bf8f52acc28db5fb3a6fd12ba2e7a8183c3ada313bab. It is durable at weighted-reservoir-audit-quarantine-exclusions/20260826-r1. Mixed raw files remain evidence-only until each salvageable row has clean replacement custody; the quarantined identities cannot re-enter a later dataset materialization. The README, aggregate, decision, manifest, and receipt were force-downloaded and byte-replayed from Hugging Face dataset commit e38aea8688b6e1e5ec6b9cad4f23444d220a19a0.

The four current exact-replay quarantine manifests are now merged into one fail-closed materialization registry: 12 UltraData-Math identities, 244 frontier-source identities, 511 weighted-reservoir identities, and an explicit zero-row Institutional Books exclusion manifest produce 767/767 unique candidate and content hashes with no collisions. The empty book manifest is a sealed audited input, not an omitted source. Every registry row denies materialization and carries no source text. The registry SHA-256 is 93d4eb890c61a7eb742b00fbf922c5ce5814a134b6fb0a6e164c70ba0fd4b180 and its canonical receipt is f89d1da4e9b56e4112a19c3f03c5c3ab3297fdfb79540077ad7540980666d45e. Both files are byte-matched under quarantine-exclusion-registry/20260826-r2 in the Stokes evidence root. Future materializers must join against this deny registry before emitting a training candidate; new sealed audit manifests extend it deterministically.

The second resumed population contains 1,007 benchmark-clean PubMed full texts. Its workers reuse the shared provider cap and automatically seal a complete aggregate and source-work decision. These rows broaden future teacher distillation evidence; they do not authorize bulk source admission.

A global Hermès teacher census is now dependency-staged behind all 13 exact source aggregates. It will refuse to run unless the aggregate candidate-file hashes form a one-to-one match with the 10,371-row mechanical-quality publication, every source population is complete, and the global verdict, language, style, and conservative-route partitions each cover every row exactly once. Reservoir and Institutional Books schemas are normalized without collapsing their distinct curriculum taxonomies. The resulting publication contains only counts, usage, source hashes, and aggregate receipts; it cannot promote teacher opinions into verified admission. This turns the completed Hermès work into an exact global gap map for scalable triage, representation verification, translation, and curriculum allocation rather than another pile of disconnected per-source reports.

Host-diverse code-web teacher expansion

The live teacher census exposed a material code shortage: among the first 1,692 high-confidence English anchors, only 16 came from code-repository sources and only 13 had code style. NVIDIA's newer pretraining-code releases remain manually gated for the current account, so Sai does not count their metadata as code content. Instead, an exact public OpenCoder code-web shard is now pinned at revision 9e8e48e…c06f3. The complete 286,437,437-byte Parquet member replayed its published LFS SHA-256 and all 197,882 rows.

The code audit considered 162,487 rows between 512 bytes and 512 KiB, found the same number of unique content hashes, and froze an 8,192-row host-diverse screen. The source-agnostic mechanical gate rejected two rows and the official benchmark boundary rejected 39, leaving 8,151 clean screen candidates. A final bottom-hash selection froze 2,048 unique documents across 1,922 web hosts, with no host contributing more than two rows. Candidate SHA-256 is 3cf1a97021a22f8a2dbab932c0bbf58ac724bd49b03c679aa61d447126e46182; population receipt is 53abfd09fb2bc71b17dba5b922c1eaa2c7752cb216654e1557b442701937e7c9. The dataset card's MIT declaration is bound but does not establish rights for every underlying web document, so per-row rights provenance remains false. Hermès compilation is dependency-staged and every row remains training_ready=false. The source-safe receipt and updated dataset card were downloaded back byte-identically from Hugging Face commit 861b793e68504f4a7df6b4e6ade4ce6322454300. The same two files are mirrored under the authorized Stokes evidence root; their SHA-256 values are e6570108…105affb and 410f8040…03ba87.

The code-web population now has a prospective 276-row promotion screen: the exact identity buckets 64..71 and 96..103 under the frozen 128-shard partition. This boundary was fixed after 47 receipts exposed a low initial yield but before any screen result existed. The full 2,048-row audit is promoted only if the screen routes at least 30% of rows to representation verification, quarantines no more than 25%, assigns computer_science to at least 60%, and achieves mean educational-value and technical-depth scores of at least 2.5/4 each. The wrappers pause without canceling an in-flight request as soon as all 16 screen-shard summaries exist. Failure reallocates scarce Hermès capacity to a better code source; it does not relabel the failed web material or weaken the thresholds.

sai-evaluate-opencoder-promotion-screen now makes that boundary executable. It pins the population and receipt hashes above, replays all 16 shard summaries and all 276 compiler receipts, computes every threshold with exact integer comparisons, and emits a signed, source-text-free pass/fail receipt. The local dependency runner invokes it as soon as the screen closes, so a failed screen releases Hermès capacity without a manual scoring gap. The evaluator cannot weaken a threshold, add a row, call a model, or turn a teacher judgment into training admission.

The eight-trillion data program

Sai is now executing two related but deliberately separate programs:

1. Hash-bound source-candidate reservoirs. Two immutable inventories now reference 21.537 physical TiB of reality anchors and candidate material. They provide breadth and filtering headroom; neither is itself a training set.
2. A prospective 8T-token training curriculum. This is the maximum-horizon schedule for turning qualified material into a moving-center spiral. It is a curriculum contract, not authorization to train and not a claim that eight trillion accepted tokens already exist.

Conflating these two quantities would hide the most important work. Raw bytes become training tokens only after rights checks, normalization, exact and near-duplicate removal, benchmark decontamination, quality judgment, concept and prerequisite annotation, grounded transformation, and final replay. A source can be excellent and still require a different representation or a later curriculum position.

Quality-core shift: approximately 2TB, not an 8TB trophy

The current optimization target is a verified approximately 2TB quality core, not preservation of the largest possible raw mirror. Two terabytes is not a padding floor: if a byte does not contribute reliable knowledge, human expression, executable procedure, grounded reasoning, or useful curriculum coverage, volume alone cannot admit it. The earlier 8TB reservoir remains a historical source-candidate checkpoint and recovery index, not the desired training corpus.

The first bulk action under this policy removed the current FinePDFs mirror from Godlydonuts/Sai: exactly 1,250 files and 3,082,436,502,565 bytes. The weighted audit had routed 324/596 sampled FinePDFs rows to quarantine and only a small minority to representation verification, so keeping the entire mirror as a volume anchor was contrary to the measured goal. Every deleted path, LFS SHA-256, Xet hash, and Git blob identity was sealed before removal; the exact upstream revision remains HuggingFaceFW/finepdfs@220bac3acbf07789502c621d2d33952f51ac7f86, and repository history also remains recoverable.

The removal plan has canonical receipt 3c4d686a86c0c11e8eba1e8b32bfc3b39cfb7d496f387a2a9bb68b424b2380cd. It was published and byte-replayed at dataset commit d396d5f522518665c506e84411e4ef2d5e7e5682. Deletion commit b99ba469952341100fd9a1780944729431b7fbf8 contains zero files under the planned prefix. The verified removal receipt is 507b38a20b00d4fa4f303bbe33afb204c021171c2bf13bdadad07fd04ca3d552 and was byte-replayed at dataset commit 6eeb4b868151f5b0ea4b9d815efe12083c2a7a9a. The same plan and receipt are hash-matched in the authorized Stokes evidence root.

After removal, the current Hugging Face sources/ tree contains 12,834 files and 5,719,814,783,273 bytes; its 12,724 payload files contain 5,719,811,111,395 bytes. The next large decision is PleIAs. It will be reduced by measured collection, language, rights, damage, and conservative-route strata after its 1,024-row audit closes. No bulk PleIAs collection is promoted merely to reach 2TB.

The exact candidate envelope is now frozen at 2,000,000,000,000 maximum bytes. Outside PleIAs, the post-removal lake contains 36 source components, 2,724 payload files, and 1,230,324,458,837 raw candidate bytes. Therefore, if every one of those nonbulk bytes survived all gates, PleIAs could contribute at most 769,675,541,163 bytes; this is a provisional ceiling, not an admission. The envelope's canonical receipt is 1151bdcdb37e31f5793d9401cfed70f22ba8d23e544342e1b3b693c9ae749cf2 and its file SHA-256 is f7b4f32b7d87953f0848726429e7882f83c17940f560792d27eee850a733df7d. It was byte-replayed at dataset commit 2c7390534cee3a545d63e1298dd4811dcfbacf0c and copied byte-identically to the authorized Stokes evidence root. The envelope explicitly permits a final core smaller than 2TB and grants no training admission.

Exact source-reservoir checkpoint

Reservoir v2 was sealed on 2026-08-23 from exact Hugging Face revisions. It contains 16,001 files and 8,796,890,808,426 bytes (8.0007255823 TiB), exceeding the exact 8 TiB target by 797,786,218 bytes. The selection includes every file from the specialist sources and only the minimum deterministic, path-ordered FineWeb-Edu prefix needed to cross the target.

Source	Exact revision	Files	Selected bytes	Function
FinePDFs	220bac3acbf07789502c621d2d33952f51ac7f86	3,573	5,375,642,953,643	Global PDF reality anchors
Institutional Books enriched text	92fcdf938eb87edfe0fbf09d4f692fa3d8bc9bcd	4,916	870,263,633,412	Books, human expression, and historical knowledge
FineMath	e92b25a616738fe95dc186b64dfb19f9c8525594	288	149,447,371,427	Mathematical reality anchors
Dolma 3 mix-150B	afa92bfb22366821c5e6cd427cdd036b34b713ef	6,081	110,586,325,507	Broad multidomain reality anchors
SmolLM corpus	3ba9d605774198c5868892d7a8deda78031a781f	338	672,430,417,560	Curated educational web and synthetic textbooks
OpenWebMath	fde8ef8de2300f5e778f56261843dab89f230815	114	27,431,041,597	Mathematical exposition
FineWeb-Edu deterministic fill	87f09149ef4734204d70ed1d046ddc9ca3f2b8f9	691	1,591,089,065,280	Broad educational-web coverage after specialists

The manifest is 7,808,445 bytes with file SHA-256 36d41d579511af2479281b3199f242d5d039ec8a261c2e3d37b07354ea6d7ccf and ordered-row SHA-256 3e1e95121ef44b0c87d430bbd46356cd978ede8719328a7416923a95a64c4e18. The receipt SHA-256 is 38e777da2a81e90919d4404d00d2a8e17531e8e0aa1405424ec645e4cdaddf44. The manifest and receipt were replayed byte-for-byte after publication in Hugging Face dataset commit 8199183b8064ffd0c0b3748bdb40a90a10da2b23. Every source revision was resolved exactly, an object from every source was access-probed, and every selected file is bound to its upstream LFS SHA-256.

The v2 correction came from actual schema replay: SmolLM's two python-edu files contain blob identifiers, repository names, paths, lengths, and scores, but no code text. Sai excludes those metadata-only bytes instead of counting them as code training material. Code coverage remains present through Dolma's content-bearing Stack-Edu shards; it is not fabricated from an index.

This checkpoint does not claim that 8 TiB is locally downloaded, unique, licensed as one combined corpus, quality-approved, translated, decontaminated, or training-ready. Institutional Books remains authenticated, reference-only material under its pinned early-access terms. The Stack v2 and manual-gated sources are excluded until their exact bytes can be accessed lawfully and reproducibly. Source inclusion also does not establish a final training percentage: origin is metadata, not an epistemic function or a fixed mixture allocation.

Modern-source augmentation checkpoint

The first semantic results exposed an unavoidable arithmetic fact: an 8 TiB raw reservoir cannot yield 8 TiB of finished data after rejecting damaged, duplicated, unsafe, low-value, or rights-incompatible material. Sai therefore sealed a second, source-only frontier inventory on 2026-08-24. Frontier v3 contains 26,599 files and 14,883,185,490,335 physical bytes (13.5361783490 TiB) from modern, independently curated source families. Together, the two reservoirs reference 23,680,076,298,761 physical bytes (21.5369039313 TiB) before cross-reservoir deduplication and quality compilation.

Those bytes are now independently accounted under source-safe conversion ledger release r7 receipt b638e13a2e7cc74118071f14e29e6dbe0f0a0234f3d150644dc281a2dfe04c47. The ledger hash-verifies both reservoir manifests, all six immutable audit populations containing 2,103 rows, corrected rights-inventory v2, and both exact bounded text-payload probes. It also binds the first complete source census: 2,504,679 pinned arXiv abstracts with 2,458,156 mechanically eligible rows and 2,380,856,330 mechanically eligible text bytes. Duplicate audit, probe, or full-census receipts are rejected. Its current funnel is deliberately blunt: 21.5369 TiB referenced candidates, 2,103 acquired audit rows, 17,638,716,209 mechanically useful bytes measured in nine bounded members, two completed source pilots containing 3,290 bounded near-deduplicated rows, one complete source census, and 0 training-ready bytes. The bounded measurement is not extrapolated to the reservoir. Of the candidate bytes, 7,899,196,133,417 require declared-license obligation handling, 5,027,859,142,584 require per-row license evidence, and 10,753,021,022,760 require source-terms resolution. Cross-inventory overlap and full-reservoir text-payload yield remains unresolved, so the candidate-byte sum cannot be used as a training-data claim. The path-portable text-free r7 receipt has file SHA-256 9fb929317de7de7aecdee12b97b91150e99ffe3876ca92c49bd32655a68089bc and replayed byte-for-byte from Hugging Face dataset commit 288fe22adf52f8b5430cfa6834c039c45004cbbc; earlier ledger releases remain immutable historical evidence.

Candidate slice	Exact revision	Files	Physical bytes	Intended comparison
Ultra-FineWeb current English L2 (2026-08-20 slice)	02c85641e3d19a854be2e09139c25adaa9518063	6,000	477,974,475,357	Newest model-selected English web
Ultra-FineWeb benchmark-validated English L2	02c85641e3d19a854be2e09139c25adaa9518063	2,048	2,661,358,122,836	Earlier L2 generation with published proxy evidence
FineWeb2-HQ multilingual	c0c06e94fd3a44ae9e802b2b0fc533817601eb5e	5,891	6,042,406,965,380	Twenty-language high-value translation discovery
Nemotron specialized reasoning	9ed3718b5f2ae29074c5e34e64115432b7c4320f	219	244,286,609,368	RQA, InfiniByte, math textbooks, and scientific coding
UltraData-Math L1 selected slice	fe10db8efd35597fd7fcff8ff576b5ec4ea5ff87	1,485	366,622,518,811	Filtered and deduplicated mathematical source material
PleIAs Common Corpus	307910e4c5d040d6f318e6edf2a2b97849155771	10,000	4,489,486,652,558	Traceable open-license and public-domain global reality anchors
Nemotron Specialized v1.2	807afc1fa65c441d46ebc7d9b95295a35499a527	90	53,621,158,028	Fact seeking, generative tasks, moral scenarios, and multiple choice
Nemotron Legal v1	3d91d58a5c0c46fe9944300ec46719f97a385b13	21	6,990,697,508	Primary law and legal reasoning
Common Pile filtered collection	31 exact repository revisions	845	540,438,290,489	Courts, government, patents, science, books, education, code, reference, and culture

UltraData Math L2/L3 quality audit

Sai does not treat a provider's tier name as evidence of quality. On 2026-08-25 it therefore acquired a deterministic 160-row screen from the exact openbmb/UltraData-Math revision fe10db8efd35597fd7fcff8ff576b5ec4ea5ff87: 32 rows from the 33.7B-token L2 quality-selected tier and 32 rows from each of the four L3 synthetic formats (conversation, multi-style, question-answer, and textbook-exercise). Every official dataset-server response returned that exact revision in X-Revision; 20 response hashes and 160 source identities are sealed under population receipt d62c4bf9711135f8d3d2aabbbeca4891c7e02eba3ec799974d3dbc65197cd138.

The active official-public benchmark boundary rejected 12/160 rows before quality scoring: 3/32 L2, 0/32 conversation, and 3/32 in each remaining L3 format. The benchmark-disjoint population contains 148 rows under receipt 54cfe887c504bb1415fc5fd803eaea7438cc40a4ee0f7882b279f07e31d3afe3. Hermes compilation is dependency-staged behind the two already-running source populations, so these 121.7B published upstream tokens remain source candidates, not accepted or training-ready Sai tokens. Raw sampled text and individual contamination decisions remain local. The source-safe publication envelope, population receipt, and 20 batch-custody receipts were uploaded and replayed byte-for-byte in Hugging Face dataset commit 28227ed9ba5a22887c2a0bb3bee20502e0982253.

The 148-row Hermès audit completed on 2026-08-26 with exact 148/148 receipt and 128/128 shard-summary custody. It returned 129 retain and 19 review verdicts, but verdicts are not admissions: conservative row routing assigned 68 to deterministic cleanup, 31 to source-bound transformation review, 27 to factual-grounding review, 12 to quarantine, 8 to rights hold, and only 2 directly to representation verification. The source decision is therefore targeted_recovery_and_verification, with bulk admission false. The aggregate file SHA-256 and canonical receipt are 9bbad154f8fc5f30c8cc3655b62655cd5d7b88eb1d296d523cd39cab6117e206 and 336817928a03afdf5efbba0fe4d7838b9337ae17b09dfa0c8f75160c76b605b0; the decision receipt is e3944ddd66e88d1c1d87327a5fae25f357d0c945eb8d026c1c2d779945825af1. Both source-safe files were byte-replayed into the authorized Stokes evidence root. The 12 quarantined and 8 rights-held rows are barred from dataset-facing materialization; because this was a coverage screen rather than a source-wide acceptance estimate, it does not justify deleting whole upstream parents. The 12 exact quarantine identities are now sealed in a reusable text-free exclusion manifest with SHA-256 327c0a9437c521075e59fa24bbe26e8aa627b563652535b6c310965140a42c5f and canonical receipt 4ddfb6632c4a9c6a91797a7d491d468e54e54b8656a5e3af3d8fc55bebadee17. Future materializers can therefore delete those rows by identity without retaining their source text or silently discarding neighboring good rows.

Complete Python Enhancement Proposal census

The 32-row source-disjoint Common Pile confirmation recovered the Python Enhancement Proposal lane as a narrow but unusually traceable reality anchor: 32/32 rows received retain judgments, 27/32 were routed to representation verification, 2/32 were quarantined, all 32 carried the recognized LicenseRef-Public-Domain declaration, and none overlapped the active benchmark boundary. That evidence authorizes a complete filtered census of one pinned parent; it does not authorize source-wide quality or training.

Sai downloaded the exact 3,723,467-byte common-pile/python_enhancement_proposals_filtered parent at revision 582170907dd303c207770fceacd38e6abf133edc and verified compressed SHA-256 4bb61eded5168ac7f0059a92ed242577c67e4fced8c0d019c84bfaca5596c791. The exhaustive pass scanned 655 rows, removed 36 identities already used in audit populations and one mechanically short row, then screened all 618 remaining rows against the official-public benchmark boundary. It rejected 50 overlapping rows, near-deduplicated the 568 clean rows into 567 unique candidates, and joined every survivor to an exact public-domain attribution record. The compressed parent was removed after the one-host census.

All 567 survivors now form a create-only Hermès compiler population with 8,822,685 excerpt bytes and complete text-free lineage. Compilation is dependency-staged after the already-running UltraData tier audit so the proxy is not overloaded and no identity is scored twice. The census receipt is c1f18f641a31672cda7d2b10caf60769df766aa7edea62418d1089645319c92b, the compiler-population receipt is 255e9aa09ec8d2f00c01db05b8eabb6bd06d9f93c77d59b1ed6e8ce2caf7a5ba, and the source-safe publication envelope is 3eeb07c28d575542d87a670396452a155f0c97aaaeaa15f19d526f210568168e. Source text, individual decontamination decisions, and machine-local paths remain unpublished. The 567 candidates remain non-training-ready until Hermès quality compilation and representation verification close. The three source-safe receipts were downloaded back and replayed byte-for-byte from Hugging Face dataset commit 756d941130a01fabb042178bf94a67b230a64e4c.

CC0 arXiv temporal screen

Sai next expanded scientific reality anchors using the exact 2,504,679-row common-pile/arxiv_abstracts_filtered snapshot at revision dc1ceab4755eb037ec61e49cf1350dab7ceee6e7. The 1,128,382,223 compressed source bytes expand to 3,473,188,609 reported in-memory bytes. A 32-row source-disjoint confirmation had already retained 32/32 abstracts, routed 29/32 to representation verification, found one quarantine row, and observed zero benchmark overlap. That result justified a broad screen, not bulk source admission.

The active r2 screen partitions the complete ordered snapshot into 32 equal temporal strata, chooses one deterministic SHA-256 window per stratum, and selects 32 source-disjoint rows from each window. All 32 official dataset-server responses returned the exact pinned revision. The resulting 1,024-row population is disjoint from 36 earlier arXiv audit identities. The official-public benchmark boundary rejected one row for one eligible code shingle and zero word shingles, leaving 1,023 rows. An exhaustive 522,753-pair exact/high-confidence near-duplicate replay found zero pairs.

Every sampled row carries the same exact CC0 declaration. The independent declaration audit recognized CC0-1.0 on 1,024/1,024 rows with zero rights holds, attribution obligations, or share-alike obligations, while explicitly not claiming source-provenance or legal clearance. The source-safe publication receipt is 0014f665fbbd09c691d03c8964fd7841bd8932e8ff4947c25e9b0b98eaeecdca. Source text and individual contamination decisions remain local. The 1,023 survivors are staged for Hermès after the existing PEP compiler closes; neither the screen nor the 2.5-million-row parent is training-ready or authorized for bulk ingestion. All five source-safe evidence files were downloaded back and replayed byte-for-byte from Hugging Face dataset commit 5047dee73c4acbdc0f2f1abf044ff5049d4d59e9.

The complete text-free parent census then streamed both exact gzip parents, verified all 2,504,679 rows and 1,128,382,223 compressed bytes, and removed each parent before acquiring the next. It replayed all 1,060/1,060 protected locator and full-text identities across the earlier audit populations. The embedded provenance is valid and strictly monotonic for every row; parent 0 contains 1,654 upstream source-position gaps, explaining why physical gzip lines and embedded provenance differ for 1,062,885 rows without constituting lineage failures.

After excluding the protected audit rows and 45,463 mechanically short rows, the census measured 2,458,156 unique eligible rows and 2,380,856,330 eligible UTF-8 text bytes. This is a complete mechanical ceiling, not a quality, benchmark-disjoint, near-deduplicated, curriculum-ready, or training-ready population. The census receipt is 507561b16269da59bfe5f85ab9ae64e4f9b8b88d815078812d06f843e0cf2708 and its source-safe publication envelope is 00cc4bcd19ec550adef2f323e57b81746db7d52c07a6555fdeda16d86bfa52a3. The census, publication, and r7 ledger receipts were force-downloaded and replayed exactly from Hugging Face dataset commit 288fe22adf52f8b5430cfa6834c039c45004cbbc.

The corrected reservoir rights inventory now binds 46 source lanes, 45 exact repository revisions, 42,600 files, and all 23,680,076,298,761 candidate bytes. Five lanes have an exact manifest declaration with obligations, 31 Common Pile lanes require per-row evidence, and ten require source-terms resolution. A permissive wrapper card can no longer override manifest labels such as “with upstream terms” or “generator terms.” StackV2 HTML has no README.md in its exact pinned tree; that absence is recorded rather than substituted from another revision. Corrected receipt 8e72391081af17323aa1e1b8d0480ddbe70dcb232006e6cf37ed7228d34d3d80 contains no source text and establishes no legal clearance. Its remote bytes replayed exactly in Hugging Face commit b7b60404ab737b9fd1e44740f6f781dc8d56da38. The earlier 11/31/4 routing receipt is retained as superseded audit history.

The 16,200,072-byte v3 manifest has SHA-256 0a59e8a24208f8593f806b919a65a3e3e64d911936f137286235c31627f56ebd and ordered-row SHA-256 510ddd35f23a00474d1ba5e6468f65bfd44e82cf33b98c8ed9870e85bf744819. Its canonical receipt is 5c3423c8d473a6155f6c402deeef298f95e89cc3769c968ced6da79f24d488a1 and the receipt file SHA-256 is 3a4a3169a8fbb75bb805dc8726c2e15e96bd21a0958f1b9bf1e2585156b09468. Every selected revision is exact, every file is upstream-LFS-hash-bound, and at least one selected object per slice passed a byte-access probe. The source-safe manifest and receipt were uploaded and re-downloaded byte-for-byte in Hugging Face dataset commit 65729b3e32cc5f86aa440cb2ff2f6e3bc8d64611.

The Common Pile expansion deliberately supplies epistemic functions missing from a web-heavy reservoir: case law, regulations, Hansard, US government publications, patents, arXiv and PubMed, biodiversity archives, public-domain and open-access books, open textbooks, Stack-Edu code, technical discussion, Wikipedia-family reference material, and spoken explanation. Sai uses the filtered releases as source candidates, retains each component's exact revision, and still requires per-record rights and quality review.

Five deterministic coverage populations now bind 1,879 source candidates across the original, weighted, frontier, Common Pile, and v3-expansion screens. The cross-population exact-content replay found zero duplicate pairs; its receipt is e31954f5bd2b220004c6b19c0dd35949052f74a464f0ad009af476e2f6dff0be. This is an exact result for the screened candidates, not a claim about the full 21.537 TiB reservoirs and not a near-duplicate result. The source-safe lineage, population receipts, duplicate reports, and combined report were uploaded and re-downloaded byte-for-byte in Hugging Face dataset commit de17529bd3ba9ea67355c26985b70350e6b8377f. Raw candidates and evidence-bearing compiler judgments were deliberately not redistributed through that public commit.

The first complete 128-row compiler screen materially changed the acquisition strategy. Although Hermes marked 98 rows retain, the independent conservative routing layer sent only 22 rows (17.2%) directly to representation verification; the other 106 still require quarantine, rights, cleanup, factual-grounding, or translation work. This is exactly why model preference is never an admission decision.

Original-reservoir source	Screen rows	Representation verification	Main measured obstacle	Current work decision
FineWeb-Edu fill	24	9 (37.5%)	7 factual, 5 cleanup, 3 quarantine	Priority targeted verification
Dolma 3 mix-150B	24	8 (33.3%)	8 quarantine, 5 cleanup, 3 factual	Targeted recovery and verification
FineMath	16	3 (18.8%)	7 cleanup, 5 factual	Targeted recovery and verification
FinePDFs	40	2 (5.0%)	21 quarantine, 9 translation, 6 cleanup	Bulk expansion paused
SmolLM corpus	16	0	9 factual, 5 cleanup, 2 quarantine	Targeted recovery and verification
OpenWebMath	8	0	7 rights holds, 1 quarantine	Rights-blocked pending resolution

These are exact descriptive results for a coverage-first screen, not estimated full-source acceptance rates. They are still decisive for resource allocation: Sai will not build its center of gravity around raw FinePDF volume, will not silently count unresolved OpenWebMath bytes, and will spend verification effort first where the screen found recoverable signal. The aggregate receipt is c0706f92535aded29c679fff5c35798a6380c01b58dc9bdf95ffd155f9a76359; the deterministic source-work ledger receipt is 7cd1a6b040eaa00a40eb37f2578045780815931d6f712a43d5bd33848a4e250e.

The 31-source Common Pile breadth audit has also completed all 124/124 compiler judgments under aggregate receipt d79749882b8e306e87997a2e0f13bd558e0bef268356b696e6d140eab656bd22. At four rows per source it is a discovery screen, not an acceptance-rate estimate. ArXiv Abstracts and Public Domain Review each routed 4/4 rows to representation verification with no quarantine; Python Enhancement Proposals and StackExchange each routed 3/4 there with one cleanup review. Those are high-priority candidates for a larger source-specific confirmation screen. Wikiteam routed 4/4 rows to quarantine, while arXiv Papers, peS2o, and Wikimedia each routed 3/4 there, so bulk expansion of those representations is paused.

No compiler-only result can bypass an independent gate. USGPO illustrates why: Hermes routed 4/4 rows to representation verification, but the corrected exact word boundary independently found benchmark overlap in 3/4 rows. It is not a priority clean lane. The complete aggregate and conservative work ledger were published and byte-replayed in Hugging Face commit 90a87727f9b5e88b0268153001f19d47c091101d. Raw candidates and evidence-bearing judgments were not redistributed.

The next Common Pile gate is now frozen rather than selected by intuition. Receipt 350e96f2c1bbffa473eb7801fcd43548b03141754622c1ba0cd55a1e7bb9e625 combines the completed compiler aggregate with the independent v2 contamination screen. Promotion requires at least four observed rows, at least 50% representation-verification routing, zero quarantine or rights routes, and zero benchmark-overlap rows. It selects ArXiv Abstracts, GitHub Archive, LibreTexts, Pressbooks, Public Domain Review, Python Enhancement Proposals, and StackExchange for a 224-row confirmation. Confirmation rows must be exact-row and exact-content disjoint from discovery; acquisition uses a different pinned parent whenever one exists and otherwise reuses the only pinned parent with fail-closed discovery-line and content-hash exclusions. This is confirmation workload selection, not training admission. The executable plan was byte-replayed from Hugging Face commit 77cb201f68dab8f447f3d3a6e81b63a9ee4407f5. The earlier receipt a48d9860193460e037c095f5483eb18b4b5199ec6b7be05eba8c6ebcfe562676 and dataset commit 6618216352dbecfae8e3c92eef53d4e14e1e24f1 are retained as superseded audit history: their universal different-parent requirement is infeasible for selected sources with only one pinned parent.

The v2 confirmation population is now sealed under receipt 40e72050e1c5a44d0e7618413d6e731de23232be9982f8f4be5d13eada44b6a5: 224/224 rows, exactly 32 from each selected source, three different-parent lanes, and four single-parent lanes with enforced discovery exclusions. It verified 2,637,343,362 compressed parent bytes while holding at most one parent locally. An independent duplicate replay covered all 60,378 possible pairs across the 124-row discovery and 224-row confirmation populations and found zero byte-exact or normalized-token duplicates (receipt 6fd6b8491a627021d2cfd2db75c6eb8b495bcd48044254452583037eff2f8785). The corrected benchmark boundary found 223/224 clean rows: six lanes were 32/32 clean, while one GitHub Archive row contained two exact word-shingle hits. Therefore GitHub Archive cannot receive blanket source promotion, and no lane is training-ready until the remaining compiler, rights, full-corpus deduplication, and transformation gates close. Screen receipt: 02fa2ead3bd14689fb6f46bf7eaca4f1518342aea8e3c08393d44aac1eb9acba.

The bounded production-pilot path is now executable but evidence-locked. sai.data.confirmation_promotion combines the confirmation compiler aggregate, corrected benchmark screen, discovery/confirmation duplicate report, and exact population receipt. A source must have at least 32 confirmation rows, at least 50% representation-verification routing, zero quarantine, zero rights holds, zero benchmark-contaminated rows, zero exact/normalized duplicate pairs, and exact identity/content disjointness. A pass authorizes only a bounded streaming pilot, never bulk ingestion or training.

sai.data.common_pile_streaming_pilot consumes that promotion receipt. It chooses the smallest hash-pinned parent not used by discovery or confirmation when available, downloads only that parent, verifies the full compressed hash, excludes every audit line and content identity, and chooses deterministic bottom-k rows in a text-free first pass. A second pass replays the exact rows, writes provenance-complete raw candidates, applies the corrected official benchmark boundary and exact normalized deduplication, and then runs an exhaustive bounded near-duplicate join over every surviving pilot pair. The join uses exact SHA-256 identities of five-word shingles, the frozen reservoir Jaccard/containment thresholds, deterministic canonical survivors, and a source-safe receipt; it does not claim that cross-source global deduplication is complete. The pilot then seals receipts and removes the downloaded parent. Documents outside 200 bytes to 128 KiB are counted rather than silently truncated. Pilot rows still remain non-training data until full rights, cross-source deduplication, and representation verification close.

Every surviving pilot document also receives a text-free attribution companion record. The replay reconstructs the exact upstream repository revision, source file, and row index from the raw population; reclassifies the original license declaration; verifies that its canonical license matches the cleaned document; and preserves attribution/share-alike obligations. This closes internal lineage loss without claiming external provenance verification or legal clearance, both of which remain explicit open fields.

sai.data.cross_source_pilot_duplicates is the next no-idle-gap gate. Once at least two source pilots exist, it preserves a deterministic per-source floor, fills unused capacity by a global bottom-k key, and exhaustively compares every unordered pair in the resulting sample with the same exact sparse shingle join. Its receipt distinguishes cross-source duplicate components and says whether the sample happened to cover every pilot row. It never upgrades a sample result into full-reservoir deduplication or training admission.

That gate has now completed over the entire bounded pilot population. The Pressbooks lane selected 2,000 rows, rejected 42 against the public benchmark boundary, and removed ten more through the within-pilot near-duplicate filter, leaving 1,948. Public Domain Review had 1,353 eligible rows, rejected ten at the benchmark boundary, and removed one near duplicate, leaving 1,342. The combined cross-source pass covered all 3,290 survivors and all 5,410,405 unordered pairs logically; it found zero additional duplicate groups and dropped zero rows. Its receipt is f489ab77ec8c8cd930e8b7b7dfafb17e36f4c8936d3d8f6f7630ce33421729ba. This closes the cross-source gate for these two bounded populations only. It does not establish reservoir-wide deduplication, rights verification, representation verification, or training readiness.

The two pilot receipts, nested filter receipts, text-free attribution manifests, cross-source receipt, and conversion ledger r6 were all replayed byte-for-byte from Hugging Face dataset commit f5ec9e07e987f008c52a29b31922c2e361c8472a. Raw and transformed source text was deliberately not published.

All 3,290 cross-source survivors are now joined back to their exact source revision, file, row, declaration, decontamination evidence, and pilot receipt in a compiler population. It contains 1,948 Pressbooks candidates and 1,342 Public Domain Review candidates under receipt b87e7c864fec79de60dc90576777b347dfd66bdbe40af40c42db9f91f422a372. The candidate file SHA-256 is 9e2da348914c6c175bcb9d2fe3272caaa891923bdc27f211118b8e849f33f93a; the text-free lineage SHA-256 is b68ff632fa6c5fda2677a289569679387d29157771a1f382851e0ef983fe76a0. Hermes compilation is running across 128 immutable identity shards with resumable create-only receipts. The compiler can recommend or reject representations, but its output remains a judgment rather than independent verification or admission.

The post-audit expansion decision is now implemented in sai.data.common_pile_full_source_promotion. It cannot run on a partial audit: it replays all 3,290 compiler receipts, all 128 shard summaries, the combined cross-source survivor population, and both bounded pilot receipts. Promotion is per source and requires at least 1,024 bounded rows, at least 85% retain, at most 5% reject, at most 15% quarantine, at most 2.5% rights hold, and source-specific educational-value, reliability, and coherence floors. A pass authorizes only full-source candidate materialization. It does not authorize raw-source admission, training, or the 4B run. This makes the expansion path automatic at audit closure while keeping bad rows and weak sources out of the compiled stream.

sai.data.common_pile_full_source_candidates consumes only that completed per-source promotion. It downloads one hash-pinned parent, excludes all prior audit identities, proves every mechanically eligible row is covered, applies the corrected official benchmark boundary, rejects high-confidence contextless answer keys, runs external-memory normalized exact deduplication, and writes exact attribution custody. The compressed parent is removed after the scan. Every dropped row is absent from the final candidate file, while the raw and intermediate files remain available for replay until replacement custody and later cleanup close. Global near/semantic deduplication, representation verification, rights clearance, training readiness, and the 4B run remain false. A no-duplicate watcher is staged to run this path for Pressbooks immediately if the final source-specific gate passes.

Only the source-safe receipt and text-free lineage were published and remotely replayed in Hugging Face dataset commit bb34c47c1cf77f3bb9b3603ccdfa8c61ac6d2caf. The evidence-bearing candidate text remains local.

The original compressed parents were then replayed once more to recover the source metadata that a language-model judgment cannot reconstruct. Every one of the 3,290 retained identities now has a text-free binding to its original parent row, native ID, declared license, bibliographic metadata, and source URLs. Pressbooks contributes 1,948 complete records and 2,844 unique chapter or book URLs; Public Domain Review contributes 1,342 complete records and 1,342 unique article URLs. The exact metadata-manifest SHA-256 values are 253e433031600aae7a5b1155122d2f93afd276f8e61d8c61663a5b6b87dfef40 and d4228cc643ae4797b294cc1b697ee8f97c61082d7f00cbf7469f7af57e8eb4d1. Both compressed parents were size/hash verified and removed after replay.

This closes internal metadata lineage, not external rights provenance. The pinned Common Pile source cards explicitly warn that inaccurate metadata or license laundering can mislabel documents. Live source-page and work-level rights verification therefore remains mandatory. The text-free manifests and receipts were uploaded and byte-replayed in Hugging Face dataset commit 35efe5b49e62a44dbd430f2c238116acbc571e82.

A bounded live page probe has now measured that next boundary without storing source HTML. It froze 901 unique Pressbooks work pages, 258 retained Public Domain Review essay pages, and the official Public Domain Review reuse-policy page that covers the 1,084 retained collection/conjecture records. Those 1,160 targets cover every one of the 3,290 pilot records exactly. Public Domain Review returned HTTP 200 plus the expected CC BY-SA evidence on all 259 targets, covering 1,342 records. Pressbooks returned 193 HTTP 200 responses; 182 contained the expected declaration, covering 377 records. Its remaining targets produced 632 HTTP 403, four HTTP 401, two HTTP 404, and 70 exhausted transport retries. Across both sources, matching evidence was observed for 1,719 records. No response was truncated and no source page body was written to disk.

This is temporal page evidence, not legal clearance. In particular, a license string appearing on a page does not prove that it governs every retained text span, while a blocked page cannot be treated as negative rights evidence. The receipt therefore deliberately keeps rights_provenance_verified=false, legal_clearance_established=false, and training_ready=false for the full population. Receipt 2483d76d4c596541044cd45eda8c73ad0b6539c5e5767f6f3529865f8ee5b5de and results SHA-256 5f85c155881a820153136bfb2c4c774cd9f16a468f6fb1a0be207893db1966e8 were uploaded and byte-replayed from Hugging Face dataset commit be6becc19c2e2e1bccfa12640b9f5ca4368da43c.

The probe has also been joined back to every exact identity as a fail-closed adjudication queue. It routes 377 records to Pressbooks book/section scope review, 258 to PDR essay license/exception review, and 1,084 to PDR policy and embedded-third-party-material review. The unresolved Pressbooks population is separated into 52 records whose HTTP 200 page lacked the expected canonical pattern, 1,300 behind HTTP 401/403, two on missing pages, and 217 attached to exhausted transport outcomes. Access controls were not bypassed and no route is an automated legal decision. Queue receipt afc84f09628ca9153c626d1f527f715b23bb967a644a19ca3f719db5825fe7c3 and queue SHA-256 8273f615b97e13a7cd7815078acc4b60666a1874e2b32355eed507ab1e638335 were uploaded and byte-replayed from Hugging Face dataset commit e1a2f00a121cbfec417cabe657111e5cb6a2de30.

Public Domain Review has now received a stricter, per-identity scope replay. Inspection of the pinned Common Pile collector at commit 9457f04a14cb2355ab00023420369d46ffd4a395 found that its permissive-license checker was defined but not applied in record construction, while quoted blockquote/q material could enter the extracted text. The audit therefore re-fetched every one of the 1,342 frozen PDR pages, reproduced the pinned selector geometry, required the applicable official-policy or page-specific CC BY-SA evidence, and constructed a quotation-excluded hash without retaining HTML, source text, or scoped text.

The first immutable replay (20260825-r1) exposed a live license-footer class collision and is retained only as superseded audit history. The corrected active replay (20260825-r2) deterministically excludes the exact license UI before comparison. It accounts for all 1,342 identities: 1,253 exactly replay the frozen candidate, 85 require source-page drift review, and four returned unavailable responses. Across the inspected pages, 961 quotation elements containing 446,625 codepoints were excluded from the scoped hashes. The active receipt is 779eeef0a192dcd73744f68aa47af46305c5a45391601a244a87be3cdbf0f40a; the results SHA-256 is 3d9b8793a2696dfc072ab226262c577ceb76212a22794fcd312d8e254a7d6271. Both r1 and r2 source-safe evidence were force-downloaded from and replayed byte-for-byte at Hugging Face dataset commit c2dbb5dfe68a85b06e85c5d1962162d12a62c68f. This measures scope evidence only: every row remains non-cleared and non-training-ready.

The active scope was then materialized into actual candidate data rather than left as a hash-only plan. A third live replay reproduced all 1,253 eligible pages without drift and emitted 5,919,449 UTF-8 text bytes across 995 Collections, 243 Essays, and 15 Conjectures. It excluded 883 quoted elements containing 412,039 codepoints; the other 89 pilot identities remain absent. Because quotation deletion creates new token adjacencies, the exact transformed text was re-screened against the pinned official benchmark boundary. All 1,253 rows remained clean with zero word or eligible-code shingle hits. Receipt 52484c5f8b22d79b231e71d2d03962fd10ea18b29c6740c02b86afd25ebd7741 binds the scoped candidates; receipt 9a051d33874a8515938d072914dbe3888e7cde52ed7eff7754c12d7efd528097 binds the post-transformation screen. This is a replayable open candidate population, not content-quality verification or training admission. The card, candidate bytes, receipts, and text-free decisions were force-downloaded and replayed from Hugging Face dataset commit 6885a18a0a98eb10c3d5d0e73ad276dd49a99a0d.

The next PDR compiler stage is now executable but has not been launched. Once the complete 3,290-row compiler population seals, it will join the clean PDR texts to their exact content and rights lanes, retain only identities routed to representation verification, and freeze at most six compiler-requested derivative types per source. The generation contract requires one exact source citation for every representation, preserves CC BY-SA attribution and share-alike obligations, and treats prerequisite edges and cross-domain connections as unverified candidates. Generated text is emitted separately from source text, with source citations represented by hashes in the candidate corpus. It remains nontraining until post-generation benchmark screening, global deduplication, source-claim verification, and independent representation verification complete. No representation generation or model training is authorized by the code-only preparation.

A single high-throughput verification pass is also dependency-staged after the post-generation screen. It compares every generated representation with its exact source in a fresh request, requires literal evidence from both texts, and routes outputs into retain, revise, or reject lanes using strict entailment, factual-fidelity, uncertainty, cultural-specificity, prose-quality, and copying checks. Because the verifier uses the same model family as the generator, a retained row records same_model_family_verification_complete=true while keeping independent_model_family_verification_complete=false, representation_verified=false, and training_ready=false. This gives Sai a fast quality filter without mislabeling same-family agreement as independent truth.

Rights are independently fail-closed. The exact pinned Hugging Face cards for all seven confirmation candidates currently expose no top-level license field; source-specific READMEs instead describe their collection policy and, for several sources, point to per-document license metadata. The 224 confirmation rows contain concrete CC0, Public Domain, CC BY, CC BY-SA, Apache-2.0, MIT, BSD-2-Clause, WTFPL, and two unversioned “GNU Free Documentation License” declarations. sai.data.license_policy canonicalizes only exact recognized aliases and attaches attribution/share-alike obligations. The unversioned GFDL label and every unknown value enter rights_hold; they are excluded from pilot selection. A recognized declaration still records source_provenance_verified=false and legal_clearance_established=false. The sealed 224-row audit found 222 recognized declarations and two rights holds, both in LibreTexts. Its receipt is 357414811d687921225830732feae6f45508707f126c01cf7b01624eaed0df40; the text-free artifact was remotely byte-replayed in Hugging Face commit e6b1210f26a7fb7e06e45c193131aa71d2c574df. The v2 promotion decision now requires this independent rights receipt in addition to compiler, contamination, duplicate, and disjointness evidence. Forward conversion uses exact-declaration policy schema v2: the complete alias table is hash-bound, and the observed reservoir declarations ODC-By-1.0 and CC-BY-2.0 are recognized with attribution obligations. This does not retroactively change the immutable 224-row audit receipt.

The official public-benchmark contamination boundary is executable, but its first code-shingle policy has been superseded. It projects 18,235 rows from MMLU-Pro, HumanEval+, MBPP+, CorrectBench, LiveCodeBench release v6, LongBench Pro, LiveBench 2024-11-25, IFEval, and MuSR without retaining benchmark text. The strictly ordered binary indexes contain 27,979,728 unique 13-token word shingles (895,351,296 bytes) and 1,907,051 unique 8-token code shingles (61,025,632 bytes). Those r1 artifacts passed their byte-level replay, but the code index also admitted punctuation-only windows. Its receipt, 073bb9f8a9ab9954ed3913b2414ff718e8f86a5020b2eb1feb18069cd75510f1, and non-reversible index are mirrored in Hugging Face commit ad178281de02625f043359a89070e905944452b9. It remains immutable audit history but is not an active admission gate. RULER remains an explicit gap until its generator is pinned to Sai's exact tokenizer and length geometry. Building the boundary makes contamination testing possible; it does not retroactively decontaminate any source bytes.

The active v2 boundary keeps the exact r1 word index byte-for-byte and admits an exact 8-token code window only when it has at least four alphanumeric-bearing tokens, three distinct alphanumeric-bearing tokens, and 16 total characters. The corrected code index contains 475,804 unique shingles (15,225,728 bytes), down from 1,907,051. Its receipt is 9fee65cb9f99813407ea4d5e4c35b4bc0bb7659c1720342f0f50bd1a8c237667; the receipt file SHA-256 is cd985016d5a301b4a1d17e9ee0f5290edda956f0434ba7293475cc187037d20a. Both indexes passed an independent hash, byte-count, and strict-order replay.

All five v2 population screens are complete. They found 69/1,879 flags: 42 rows with word overlap and 27 additional eligible code-only rows. The population counts are 6/128 original, 26/1,024 weighted, 28/512 frontier, 7/124 Common Pile, and 2/91 frontier expansion. Nemotron specialized reasoning is 25/96, comprising five word-overlap rows and 20 additional code-only rows. Neither an individual flag nor a clean screen licenses bulk source admission or rejection. The boundary and all five source-safe screen receipts were uploaded and byte-replayed in Hugging Face commit 43ae57ee4981c78ae23c111436b1fc9b6aa27023.

The 91-row modern-source expansion compiler pass is now complete under aggregate receipt afd82b43ac66f3a485d167b97f79fccc75bc67026c94e182d846a6923f9dea23. The result again separates a model's retain verdict from actual source readiness: Hermes retained 60/91 rows, but conservative routing sent only 14/91 to representation verification. Nemotron Legal contributed 8/21 such rows and is the only priority targeted-verification lane; Nemotron Specialized v1.2 sent 22/30 rows to factual-grounding review and none directly to representation verification; PleIAs Common Corpus sent 16/40 to quarantine, 11/40 to cleanup, 6/40 to translation, and 6/40 to representation verification. Independent v2 benchmark screening found both contaminated rows in PleIAs (2/40), while the Nemotron Legal and Specialized samples were clean. These are coverage-screen results, not source-wide yield estimates. The aggregate and fail-closed work ledger were remotely byte-replayed in Hugging Face commit 2a085eacf1479293e3c369d7eaa8e476d7f84054.

Because PleIAs is the lake's largest unresolved component, a new 1,024-parent screen has now completed across all ten common_corpus_* partitions. It selected 102–103 parents per partition by deterministic SHA-256 rank, excluded all 40 parents from the earlier screen, and fully verified 460,098,704,855 parent bytes without claiming statistical representativeness. Eight identity- disjoint 128-parent shards each opened exactly one parent at a time, verified the full pinned byte count and SHA-256 on disk, iterated the selected row group in 16-record batches, persisted one deterministic usable row, and removed the parent before continuing. All 1,024 parent identities are unique; the aggregate candidate and lineage SHA-256 values are 4f4bbb3c87b0f5e65f7490cf469a04dda03b11e5bc863d0404c6f4fd4f90fd32 and 566404ea466227d61db98b0cc3664c26c99564b6e70efb4e866caa32c106fe4d; the canonical acquisition receipt is 06cce088ccdc2c58f89d0e467961f4b2a648b5448775f8ed333445af672ed1e3.

The corrected official-public benchmark boundary then found 33 contaminated rows and 991 clean rows, with 846 exact word-shingle and 28 eligible code- shingle overlaps. Contamination spans every partition and is highest in the sampled partition 8 lane at 7/102 rows. The source-safe screen receipt is bc9f207c328c3d8ea8387d0c4692f2fdae216706eb00761e906fc8e3b0f17988. The source-safe lineage, acquisition receipt, and screen receipt were force-downloaded and byte-replayed from Hugging Face dataset commit 4741fcd1da4462733d463475c276bd87d4ab7d5d. Hermès source-quality judgment is now staged against the immutable audit population, but source-wide yield, clean full-source materialization, and training admission remain false.

The larger 512-row frontier-source compiler has now closed with exact identity and receipt coverage. Hermès returned 348 retain, 125 review, and 39 reject verdicts, but conservative routing sent only 25/512 (4.88%) to representation verification and 244/512 (47.66%) to quarantine. FineWeb2-HQ and both Ultra-FineWeb snapshots are therefore bulk-paused; Nemotron specialized reasoning and UltraData-Math L1 remain targeted-recovery and verification sources. The aggregate replayed 512 valid outcomes, 210 transient HTTP errors, and 177 repaired or retried invalid model outputs. Aggregate receipt 9c2d0e49d062c1886f9809b8852c4c48c38f41b40bea210631d1cc0f7236c6de and decision receipt 7656214825b6f66984c007f00a6f089c5d2c43791af6b775c466db56d952a48d were remotely byte-replayed in Hugging Face commit ced4fa4db0a90b8804aa0b42ba98e01597920433. The same three source-safe files are byte-matched under the authorized Stokes evidence root at frontier-source-audit/20260825-r1. All 244 quarantined identities are now additionally sealed as a text-free dataset-exclusion manifest. Its SHA-256 is 9ccb9c64c125d907d8bdfa46d01dcc2dbf5c6e1cffc6adeee3fa6995400615ab and its canonical receipt is 50803c3adb2b2dc758344542106735b9a0b2e9c403ef629e3d9ed872191ff256. Those rows can no longer re-enter Sai through a later bulk materializer; the manifest deletes identities from admission without treating a mixed source parent as uniformly bad. These measured routes are neither whole-source yield estimates nor admission; they prevent dataset branding from silently substituting for content quality.

The frozen OpenCoder code-web promotion screen has also closed with exact 276/276 identity coverage across 16 preselected logical shards. It passed only computer-science coverage: 130/276 (47.10%) rows route to quarantine, only 20/276 (7.25%) route to representation verification, educational value is 2.195/4, and technical depth is 1.659/4. The exact replay therefore records stop_full_audit_and_reallocate_hermes_capacity, not a full-audit promotion. All post-screen OpenCoder workers were stopped. Its 286,437,437-byte local Hugging Face acquisition-cache blob and snapshot symlink were deleted, reclaiming the full physical blob while retaining the population, compiler receipts, shard summaries, and source-safe decision. The raw object was never uploaded to Godlydonuts/Sai and remains recoverable from pinned upstream revision 9e8e48e666c226294d6f9e6c2e13f2c84c1c06f3; this bounded screen does not claim every upstream row is unusable. Screen receipt 29a7ceed9841f99213d4087a40e0107277a07793b490760ee800242bcad7be70 and cache-reclamation receipt a45fa68a77018c1b58900b58aa703ad295c249183528ec8495816fe95c6ac172 were remotely byte-replayed in Hugging Face commit f6151be578e7e353af45152426a55681a27eae80 and copied byte-identically to the authorized Stokes evidence root at opencoder-promotion-screen/20260826-r1.

The same cache policy removed a second, unrelated object without waiting for a semantic source audit: a partial willdepueoai/parameter-golf snapshot holding ten FineWeb training binaries and one validation binary already encoded by an unrelated 1,024-token benchmark tokenizer. The 2,124,321,260-byte cache was not original text, appeared nowhere in Sai code or live processes, contained only 11 of the upstream manifest's 196 binaries, and had zero paths in Godlydonuts/Sai. Every local file hash was verified before the exact cache root was quarantined and deleted. It is not locally recoverable, but is exactly re-downloadable from upstream revision a85b0e6035c3c94bc23685a07c81a8f3bf89db80. Reclamation receipt 612423e30092478571c9a43eae23d0271d8278eaa816e92f90a3d605ae1a91fe was remotely byte-replayed in Hugging Face commit 40d688182f3e6a65b3b09a96eeb33ee9e48a441e and mirrored byte-identically to Stokes under parameter-golf-cache-reclamation/20260826-r1.

An unused ProsusAI/finbert cache was also removed after proving zero Sai-code, live-process, and open-handle references. It contained two redundant formats of a narrow financial-sentiment classifier, not a Sai data, tokenizer, semantic deduplication, or training dependency. All 13 files and six symlinks were hash-manifested before 876,191,371 bytes were deleted. The cache is not locally recoverable; main revision 4556d13015211d73dccd3fdd39d39232506f3e43 and safetensors revision 7db323f79b751944bcfa66298ec06977e4518306 remain pinned upstream. Receipt bc82ae86511c507edf8128f8e5ced567e58b2e2d300128258da2b02bab7b2117 was remotely replayed in Hugging Face commit 985b5bae55c694825d3ba4cfaffa01774d04287c and mirrored to Stokes under finbert-cache-reclamation/20260826-r1. The similarly sized all-mpnet-base-v2 cache is deliberately retained because it remains directly useful for semantic deduplication.

The r1 286/1,879 overall and 77/96 Nemotron conclusions are retracted because they were materially inflated by nonsubstantive code windows. Those screens remain immutable evidence of the discovered policy failure in Hugging Face dataset commit 5dc89bfeceadf56663a8f00c479f5d41d5229671. They are explicitly superseded and are not active contamination decisions.

PleIAs Common Corpus adds a distinct 2.27T-token, traceable open corpus rather than another opaque web mixture. Its rows expose collection, open-status, license, language, creator, and date metadata alongside text. Valuable non-English material is a translation-discovery pool, not automatic English training data; literary form and cultural context remain protected from indiscriminate rewriting.

These are physical source-object bytes, not measured text-payload bytes. FineWeb2-HQ, for example, includes large embedding columns, so treating its repository size as English training text would be materially false. The two Ultra-FineWeb generations may overlap, and every web-derived source may overlap the original reservoir. None of those bytes are counted as unique or training-ready until text-column extraction and global semantic deduplication complete.

Two exact text-payload probes now measure that distinction instead of guessing it. One member per source was selected by a frozen SHA-256 rank before size or content inspection. Eight selected members fit the first probe's 4 GiB parent cap. FinePDF's independently selected 4.84 GB member was blocked rather than replaced by a conveniently smaller shard, then measured in a second prospective probe with a 6 GiB cap. Every measured member was fully downloaded, matched to its pinned size and SHA-256, streamed one at a time, and deleted afterward. “Useful” below means only the mechanical 200 B–128 KiB size window; it is not a quality, rights, uniqueness, or admission judgment.

Source	Exact physical bytes	Text UTF-8 bytes	Useful UTF-8 bytes	Useful/physical
FineWeb-Edu fill	2,378,402,603	3,832,560,263	3,741,009,274	1.572908×
Dolma 3 mix-150B	5,349,719	62,584,337	59,977,943	11.211419×
FineMath	733,726,864	1,199,827,734	1,165,858,399	1.588954×
SmolLM corpus	2,391,060,328	3,915,215,823	3,822,727,059	1.598758×
PleIAs Common Corpus	430,252,575	795,883,158	750,836,358	1.745106×
FineWeb2-HQ multilingual	1,203,684,113	447,018,551	384,827,189	0.319707×
Ultra-FineWeb current L2	82,007,099	138,259,448	136,403,753	1.663316×
Ultra-FineWeb earlier L2	1,298,592,398	2,240,142,912	2,190,701,659	1.686981×
FinePDFs	4,836,418,450	9,944,850,928	5,386,374,575	1.113711×

Across only the nine exact measured members, 13,359,494,149 physical bytes contained 22,576,343,154 text bytes and 17,638,716,209 mechanically useful bytes. Ratios above one are expected for compressed members. FinePDF contained 397,166 useful rows, 4,821 short rows, and 12,013 rows above 128 KiB. Those long documents are a structure-aware segmentation queue, not automatic rejects. The roughly 35× spread between the observed useful/physical ratios proves that repository size is not a defensible acquisition objective, but these bounded member probes are not source-wide yield estimates and cannot be extrapolated. The first plan receipt 4f5312f7d9ae86b3fbe8998c7e780c7238eae9394fc767fdbedad2affbacc66c and measurement receipt 1d550e0abc513c5b4e61f0ce5890155bfff01bcdbd2a6896f9a078c26952f848 were remotely replayed in Hugging Face commit fecd9d596c18dd63ab6ea7a89dda7b2544eca4a1. The FinePDF plan receipt 325382746db5836ccffa12ea437fcfdfaf12ee0f29e469ac47cf0e43c0559017 and measurement receipt 5947564751b941b18d8a025abd3451c2e81cfa6e6357c0cad28213561d372919 were remotely replayed in Hugging Face commit e15ca127c695d2d42df04e15738e56525f0bb3ce.

Sai now has a create-only long-document recovery path rather than a truncation policy. sai-segment-long-documents splits only over-budget raw documents, preferring paragraph, line, sentence, clause, and word boundaries before a lossless Unicode-character fallback. It never normalizes or rewrites source text, proves exact parent reconstruction, issues collision-free child row identities, and emits a text-free segment-lineage manifest. Segment geometry is carried through benchmark decontamination and the attribution manifest. This is preparatory compiler infrastructure, not a FinePDF admission result: segmented documents still require contamination screening, global deduplication, rights verification, representation verification, and curriculum placement.

The first global deduplication layer is also executable as sai-deduplicate-global-exact. It external-sorts compact text-free indexes in bounded fan-in passes, groups documents by NFKC/casefold/whitespace-normalized SHA-256, selects the minimum immutable document identity, and replays every apparent collision against the full normalized source text before dropping it. Outputs are deterministic across input order, temporary indexes are removed, and the duplicate manifest contains identities and byte locators rather than source text. This closes normalized exact duplicates only; scalable semantic near-duplicate filtering remains a separate unresolved gate.

The second exact layer is now executable as sai-deduplicate-subdocuments. It follows the August 2026 frequency/length-aware method: natural-boundary segmentation with short forward merges, normalized global exact counting, the explicit T(C,L) copy budget, document-identity ordering, whole-boundary-document retention, and deletion only for sufficiently long contiguous candidate runs. It uses bounded external-sort fan-in, replays every indexed occurrence against the immutable source before trusting a hash, recalculates identities for changed documents, and writes a text-free parent-to-output transformation manifest. Numeric template normalization is restricted to natural-language chunks; fenced code is indivisible and exact, while full code-domain documents currently fail closed as indivisible. This is deliberately safer than pretending a language-agnostic brace parser preserves every programming language.

The first complete-parent execution of this layer is now closed over the 54,509 Pressbooks and Public Domain Review candidates. It indexed 4,650,337 chunks, found 494,414 duplicate groups and 943,565 duplicate occurrences, removed 90,691 chunks / 11,216,449 characters, modified 9,573 documents, and fully removed two duplicate-only documents. The adaptive output contains 54,507 documents under canonical receipt 9d3ee20c4d5d0732589c3baea55414752e0184bb4d81a06a3b71f6894bff1e8e; the text-free 19,949-record transformation manifest has SHA-256 fed1c0767589b5b596adf8008c0d8a273c58e502421424cacd00ad9d5f75ff5f. All temporary indexes were removed. This is a deterministic transformed candidate, not evidence that the adaptive policy improves a model.

An exact post-transform deletion join then removed all 11 Pressbooks source rows that the sealed 3,290-row Hermès compiler pass had hard-rejected. Those judgments include ten weak-grounding flags, eight duplicated-boilerplate flags, and five personal-or-secret-data flags. The join also removed attribution for the two upstream fully deduplicated documents, leaving exactly 54,496 candidate and attribution rows with identical source-row sets. Candidate, attribution, and text-free exclusion-manifest SHA-256 values are 02bd40ea6a7b9a5710861e4284a09981101512f341e91e549c75effc0a76faa8, 9cb7af6399bc527865dee98fc623691d827222155fecc011946b54cfa7f8011a, and d07030784790a8799d039c275428b1ca25947e62eafeb2f7c1de12cffa99a474; the canonical exclusion receipt is 89756c4dbd45b772889bbd813138583fcc9a73850305a2202baefcdbef18df43. All four outputs were byte-verified in durable Stokes custody before the local unfiltered candidate was permanently unlinked. The four small source-safe receipts/ledgers were force-downloaded and byte-replayed from Hugging Face dataset commit 10c6b42c61cf9eac463014416e659109d7e639f4. The clean 534,981,699-byte candidate and 43,758,121-byte attribution files could not be added to that commit because Hugging Face returned an explicit public storage-quota 403 before creating any large-file commit. Their exact local and Stokes custody remains valid; this is a publication-capacity blocker, not a missing-data condition.

The implementation does not turn the paper's result into a Sai result. Corpus promotion still requires an identical-token, identical-compute, source-disjoint comparison of unchanged, keep-one, and adaptive retention. The CLI freezes both executable transformed arms with --retention-policy keep_one_control and the default adaptive_frequency_length; the immutable input is the unchanged arm. Semantic near-duplicates remain a separate measured gate, and every receipt keeps training_ready=false and four_b_training_authorized=false.

NVIDIA's organic/translated Nemotron-CC v2.1, CC-Code v1, and Code v2 repositories were investigated but are not counted: metadata is visible while the current user token receives HTTP 403 on the actual gated objects. The public Nemotron specialized-reasoning slices are counted and will still face the same quality, novelty, generator-lineage, contamination, and benchmark gates as every other synthetic source. In particular, Sai does not import the 2.1T-token medium-high synthetic-rephrase bulk merely to inflate volume. Nemotron Pretraining Code v3 is also excluded from byte counts because its published Parquet schema contains repository, path, language, and commit metadata but no code text. Those locators may support a future rights-aware source fetch, but an index is not training data.

Moving-center spiral for an 8T-token run

The prospective schedule moves its center of gravity from foundations to synthesis and expertise while preserving both tails. Expert material begins in the first token stage, and foundational rehearsal remains present through the last 400B tokens.

Stage	Token interval	Stage tokens	Foundation	Intermediate	Advanced	Expert	Minimum cross-domain material
Foundation	0–2.0T	2.0T	55%	30%	12%	3%	1%
Expansion	2.0–4.8T	2.8T	25%	45%	23%	7%	4%
Depth	4.8–6.8T	2.0T	12%	28%	42%	18%	10%
Synthesis	6.8–7.6T	0.8T	10%	20%	35%	35%	30%
Annealing	7.6–8.0T	0.4T	10%	18%	30%	42%	20%

These percentages are the current prospective difficulty-band allocation, not a frozen source mixture. Within each stage, the compiler and proxy experiments must still discover which concept, style, source, and reasoning regions buy the largest source-disjoint marginal capability gain without causing retention loss. Annealing uses the highest-value mixture supported by that evidence; it does not simply become “all expert data.”

The policy is executable in sai.data.eight_trillion_spiral. It binds exactly 8,000,000,000,000 tokens, contiguous stage boundaries, exact per-band token allocations, nonzero early expertise, nonzero late foundations, and receipt SHA-256 ffa85e065bb7a3895af55bfa9ffdb7f65e236d05722e1ca8abea6161bf259bd2. Both training_authorized and four_b_training_authorized remain false in the artifact: a long-horizon policy cannot authorize a run whose accepted stream does not yet exist.

Synthetic data is a bridge compiler, not a prose factory

Sai's synthetic advantage is intended to be knowledge composition. The model should not merely know biology and information theory independently; it should learn when and how their structures connect. Candidate pairings include biology × information theory, music × Fourier analysis, law × logic, architecture × structural engineering, history × economics, literature × psychology, computer systems × thermodynamics, chemistry × quantum mechanics, and art × geometry. The pairing list is not a quota and novelty is not assumed merely because two domain labels appear in one prompt.

Every admitted synthetic bridge must carry:

• at least two genuinely distinct domain identities;
• exact source anchors and immutable source hashes;
• the concepts and prerequisite edges required to understand the bridge;
• evidence that prerequisites were taught earlier or are explicitly rehearsed;
• a relationship that is not a paraphrase or a superficial word collision;
• an independently solved answer or deterministic verifier;
• translation and transformation lineage; and
• benchmark-overlap evidence sufficient to keep evaluation prompts out of training.

Generic ungrounded generation is forbidden. The compiler should prefer representations whose truth can be checked: executable code and tests, symbolic solvers, simulations, constraint systems, multiple independent solutions, cited synthesis from primary sources, and contradiction-seeking review. A model-generated explanation can improve presentation, but it cannot create a reality anchor or verify its own unsupported claim.

The cross-domain generation loop is therefore:

qualified reality anchors + concept/prerequisite graph
    -> identify distant but structurally meaningful concept pairs
    -> generate a task, derivation, explanation, or worked example
    -> solve independently or execute a deterministic verifier
    -> reject unsupported, duplicate, stylistically collapsed, or contaminated work
    -> assign difficulty on linguistic, conceptual, and reasoning axes
    -> place into the spiral only after prerequisite and retention checks

Gradient-space or capability-gap targeting can later choose underrepresented bridges, but it may not weaken those grounding requirements. Synthetic volume is never the target by itself; verified marginal learning is.

Hermes and Institutional Books operating state

Hermes is the compiler workforce, not an oracle whose output is accepted by default. The current Institutional Books program starts from 983,004 Harvard Library volumes totaling 242,051,626,500 upstream tokens. A metadata-first 10,000-volume review queue spans 115 languages and 772 language×subject cells; 9,409 rows are non-English translation-discovery candidates and 591 are English controls. This is coverage-first sampling, not a desired final language ratio.

The first authenticated enriched-text pilot contains 185 candidates from 200 source rows: 164 English and 21 non-English, with 14 OCR rejects and one token bound reject. Hermes has now completed the first production-schema judgment for an advanced 1917 geology volume. It retained the source as a historical scientific anchor, marked outdated claims as a risk, extracted 34 concepts, 11 prerequisites, and six evidence-backed edges, and left the raw archive source explicitly non-training-ready. The successful request used 12,314 prompt and 1,454 completion tokens after one invalid first response and one schema-bound repair. This single result proves the worker and repair path function; it does not estimate corpus-wide acceptance quality.

The first three completed book receipts exposed a throughput issue: valid outputs required two to five provider calls because generic retry text did not spell out the book schema's most common failures. Future shard processes now receive deterministic book-specific correction hints for exact excerpt quotes, concept-edge evidence, English translation disposition, enum fields, risk keys, domains, and representation labels. The strict schema is unchanged; the correction only reduces wasted invalid retries. Existing healthy book workers are not interrupted and automatically pick up the new code when their next immutable shard process starts.

The complete bounded book pilot has now closed: 185/185 candidate receipts cover 182 nonempty immutable hash shards, with 182 retain, three review, and zero reject model verdicts. The compiler identified 5,121 unique concept labels and 1,129 unique explicit prerequisite-edge claims. The curriculum distribution is four basic, 61 intermediate, 108 advanced, and 12 expert rows; the source-language distribution is 173 English, eight Czech, three French, and one German. The model requested 2,123,954 prompt tokens and emitted 207,638 completion tokens. These are compiler measurements, not admissions: OCR, historical-context, factual-grounding, deduplication, translation, and representation-verification work remains separately routed, and the aggregate retains training_ready=false.

The aggregate receipt is 31a20de0a616b61ac1c5f5fbc22c36fdecb0575237b346f3a4b1252909315d78 and its file SHA-256 is 7fbd1ebb7ab85f6dc48abaa15a1098f1e0b4b793c399709b974dbff021596336. No row reached the conservative quarantine route, so the source's exclusion manifest is intentionally empty and hash-sealed, with receipt 57aead38995182a329188aedf5318720afa806e63a1dea30ef4a763059f3eccb. The aggregate, empty exclusion evidence, and four-source registry are byte-matched in the authorized Stokes evidence root and were force-downloaded and byte-replayed from Hugging Face dataset commit 0e26ff13821ae4cca9c64c380aecefddaa265c98.

Independent frontier-model review capacity

Sai now has a separate, fail-closed review worker for exact provider/model pairs. Its receipts bind the candidate, rubric, request, response model, endpoint, token accounting, and repair attempts; they cannot be silently merged into the primary Hermès stream. Live qualification confirmed Google Gemini 3.1/3.5 Flash Lite and Gemma 4 26B-A4B/31B, Groq GPT-OSS 120B and Qwen 3.6-27B, and Cohere Command A Plus/Reasoning. NVIDIA's independent grounded bridge verifier is separately pinned to nvidia/nemotron-3-ultra-550b-a55b and cannot mark a bridge or training row ready by itself.

The first matched one-row check produced schema-valid judgments from Hermès, Gemini 3.1 Flash Lite, and Groq GPT-OSS 120B. All three independently selected retain, grounding, and biology_medicine; Hermès and Gemini also agreed on the duplicated_boilerplate risk, while GPT-OSS did not. This verifies cross-family execution and exposes a real disagreement; one row is not an accuracy estimate. Provider qualification also failed closed where appropriate: Cerebras returned a billing gate, Groq Qwen repeatedly returned 429, Google Gemma and both Cohere models failed the strict JSON contract in their first pilot, and none of those lanes was scaled. Gemini 3.5 completed eight of nine rows before one strict enum failure. Deterministic enum-specific repair hints were added without weakening the schema, and the full repository suite passes 1,078 tests.

A larger calibration screen now freezes 45 rows across PleIAs (16), PEP (14), and PubMed (15), stratified into clean retain, cleanup-risk retain, severe-risk retain, and non-retain cells. Gemini 3.1 and 3.5 covered all 45 identities and agreed with each other on 45/45 verdicts and 37/45 conservative routes. Each agreed with the primary Hermès verdict on 36/45 rows, but only 17/45 and 14/45 primary routes. Most importantly, Hermès marked nine rows non-retain while both Gemini models retained all nine. Across Hermès and both Gemini lanes, only 12/45 routes were unanimous; 11 of those 12 came from the clean-retain controls.

Nemotron Ultra then targeted the 12 clean controls and nine disputed non-retains. It returned 18/21 valid signed judgments; three remained explicit non-JSON endpoint failures after all retries. Nemotron retained all 11 covered clean controls and all seven covered non-retains, but routed ten of those 18 rows to cleanup, grounding, translation, or rights work rather than direct representation verification. All four available judges agreed on eight clean routes and zero non-retain routes. This is actionable calibration evidence: single-model risk labels are useful for triage, but cannot justify irreversible deletion or automatic admission. Future row deletion requires deterministic evidence or adjudicated agreement; collection-level pruning remains eligible when the complete source audit demonstrates persistently poor yield.

The Gemini consensus receipt is 99a5e35c7d8f42ef1a5961670c99df22cd2ccf3e67bcb53cbb15b4615bbb52cd. The Nemotron target-coverage receipt is 3b3e5c38407f5d83586f5f0c6c95725987439736842d68dc80777f42403180a5, and the combined Gemini/Nemotron comparison receipt is 8085928cdfa68daf5c439d57912ce2a11d429d1421147e33412591ce5103a5c2. All source-safe evidence was byte-replayed at dataset commit bee9b4d2d619e8bc8edcfd5a97c79fcc0c4ba5f3 and copied byte-identically to the authorized Stokes evidence root. Sampled source text was not published.

The previously reported contextless physics answer sheet is also bound to an exact source location: FineMath-3plus row 50 of upstream train-00076-of-00128.parquet, content SHA-256 5f3c24824cb8c10cfeb19ff6242966541ed2a274fd0bdd6f88418b296d718482. Its hard-reject decision remains active. That mixed 503,424,134-byte upstream shard is not present in Godlydonuts/Sai, so there is no bad published file to delete; deleting the whole mixed shard would also discard unrelated good rows.

A separate reservoir-wide coverage audit now freezes 128 generic compiler candidates across six content-bearing source families: 40 FinePDFs rows (16 English and 24 named non-English language strata), 16 FineMath rows across all four quality/subset bands, 24 Dolma rows spanning reference, papers, math, Python/Rust/Java, and 18 PDF/web topics, 16 SmolLM educational-web and Cosmopedia rows, eight OpenWebMath rows, and 24 FineWeb-Edu rows across six crawl years. Institutional Books remains on its richer book-specific schema and is not flattened into this generic audit.

The population file contains 128 unique identities and 1,059,279 bytes, with SHA-256 d3f24cd2855400a00e16f0bcf6dca63190a0f6653ee64031d77af9b35e83d823. Its 138,940-byte lineage file has SHA-256 dc4cab1b3a0cea23b12841afd830970e588f3e0a17a52cd5d50849e8dbe8207b. Twenty-four compressed Dolma parent files were fully downloaded and matched to their upstream SHA-256; 104 Parquet parents were read by exact-revision range requests and remain bound to the upstream LFS hash without falsely claiming a local whole-file rehash. The population receipt is 4c5a179bb6863969850cc7d70133650211a445eb124df6dd28948053cb817ed4. This is a coverage-first diagnostic, not a statistically weighted corpus acceptance estimate.

The exact/near-duplicate replay compared all 8,128 unordered pairs in this 128-row audit population. It found zero byte-identical, normalized-token identical, five-word-shingle Jaccard, or five-word-shingle containment matches at the frozen conservative thresholds. The report's canonical receipt is ecd131b92a708d0cf002004b62a4c69e86b208a96afa2fea2631ca54e511fc2d and its file SHA-256 is 0cd6e93fa4008728a409de82de29fda9e5f9ce5960316c91981d6f62b523568c. This establishes that the diagnostic excerpts are not obvious copies of one another; it does not establish that the 8 TiB reservoir is globally deduplicated.

Compiler judgments also pass through deterministic conservative routing. A model verdict of retain is quarantined for personal/secret data, held for rights ambiguity, sent to factual-grounding review for weak reliability, sent to translation review for non-English material, and sent through cleanup and transformation review before representation verification. Even representation_verification is not training admission. This prevents a confident semantic reviewer from silently overriding objective data gates.

Hermes routing depends on content type:

• Preserve high-value English literature, rhetoric, letters, essays, and other form-bearing expression instead of flattening it into generic model prose.
• For non-English technical and factual work, create English representations with exact source and translation lineage, while preserving source metadata.
• For non-English literature, prefer a reputable admissible human translation. If none exists, keep separately labeled literal and literary synthetic translations and do not represent either as the original voice.
• For papers, standards, reference works, and technical books, preserve the source and derive multiple grounded representations such as prerequisite maps, concise references, textbook explanations, worked examples, FAQs, and misconception/correction pairs.
• Reject or quarantine OCR damage, bibliographic ambiguity, missing rights evidence, duplicative editions, unsupported factual claims, benchmark material, and synthetic voice collapse.

The exact Common Pile confirmation is now complete: all 224 source-disjoint rows have compiler receipts and all 32 shards have summaries. Hermes returned 207 retain, 11 review, and six reject verdicts, but deterministic routing is deliberately stricter: 141 rows require representation verification, 49 cleanup review, 15 factual-grounding review, 17 quarantine, and two transformation review. The sample spans seven sources; 187 rows identify at least one cross-domain bridge. The run consumed 958,783 model tokens and 48 rows required bounded schema repair. These are diagnostic results, not an acceptance rate or training admission.

Only common_pile_pressbooks and common_pile_public_domain_review cleared the frozen zero-quarantine, zero-rights-hold, zero-benchmark-contamination, identity/content-disjoint, and minimum-representation-verification checks for a bounded streaming pilot. The other five sources remain held, and neither promoted source has bulk ingestion or training authorization. Source-safe aggregate, decision, and promotion receipts were uploaded and byte-replayed in Hugging Face dataset commit 44fbdd30cedc89ac908057929468d3162651d645.

The complete bounded compiler gate has now closed across all 3,290 rows and 128 immutable shards. Hermès returned 3,163 retain, 116 review, and 11 reject verdicts, while deterministic routing held 240 rows in quarantine and 21 on rights review. Pressbooks retained 1,829/1,948 rows with 11 rejects; Public Domain Review retained 1,334/1,342 with zero rejects. Both sources pass the frozen threshold for full-source candidate materialization only. This is not verification or training admission: global near/semantic deduplication, rights resolution, representation verification, and final curriculum custody remain open. The aggregate and source decision have canonical receipts ca9884af6ec7e5ef8f2b39a7fcbbe8892423be0e53b75006fcbf987f7ae76484 and 99f96fcf15a4bdd69fdf451c1220d544c7d008c7682abdfc303450b4797075b2. All three source-safe release files were force-downloaded and replayed byte-for-byte from Hugging Face dataset commit 7a447380b42cf631581a1604b249accecbb153bc.

The resulting complete-parent Pressbooks pass scanned all 54,455 rows. It excluded 36 prior-audit rows, 35 short rows, 130 oversized rows, and 1,088 official-boundary overlaps, leaving 53,166 unique candidate and source-row identities. The final candidate SHA-256 is 85256ecea000b6aa0a2b1e638a61af87362e6d7effc87e0722a0ac6e994da2d7; its exact attribution manifest SHA-256 is 59fb344c48fa68eabf740faba647387afe305399c64d0d9ce7e3e0aea67d6119; and the canonical run receipt is eb7e822323ec3cd928cd5b7775809207ed610634ce0c123ce77546b23065ab1e. After final candidate custody was hash-verified on Stokes, three redundant recoverable intermediates were deleted locally, reclaiming 1,645,579,101 bytes. No mixed Hugging Face source shard was deleted. The three source-safe receipt files were force-downloaded and replayed byte-for-byte from dataset commit 207d24434b073d552003d11154ab43bfe2c1bdb0.

The companion Public Domain Review pass scanned all 1,406 parent rows, excluded 36 prior-audit and 17 short rows, and removed ten official-boundary overlaps. It leaves 1,343 unique candidate/source-row identities. Its final candidate and attribution SHA-256 values are 1ac38f4a6dde7e05011da644fa7e8db7acd5a0db3432eb51d259874bd5fe80a2 and 81401c05a387dbcbeef157c892912fab055d589b7d049d1d79ee72b687b7811e; the canonical run receipt is 1214c429cdd9251998c0c947660344e63d0384d513ff2c8c1c07d896b8e03cac. After durable candidate custody, 22,406,095 bytes of redundant local intermediates were reclaimed. The source-safe PDR receipts replay byte-for-byte from dataset commit 2360c039136873ef3b4a653bf642425a78fe440a.

Cross-source normalized exact deduplication then covered all 54,509 surviving Pressbooks and PDR candidates. It found zero duplicate groups and zero drops. The canonical run and source-safe publication receipts are 1be527cf0a814b7586a350757c9f63c90ee43844cae67ddb124929c650a397b4 and 7e69e12d131e4cc78d6956d0f7793a418c1940c202ea49feb028943d40576de2. Because its 546,324,994-byte combined output was only a deterministic reordered copy of two already durable inputs, it was reclaimed after evidence custody. The later unfiltered subdocument candidate was also moved to the separate Stokes evidence root before its 535,008,085-byte local copy was permanently unlinked; the clean 54,496-row replacement remains local and durable. Across all evidence-backed cleanup to date, 6,036,268,343 local bytes have now been removed. The source-safe exact-dedup receipts replay from dataset commit 526e801fae1fb01a4f9ced8f260c6a2ef51c7823.

The next operational work is to expand sustainable, stratified compiler lanes; build cross-source exact and semantic duplicate families; populate the concept prerequisite graph; create verified English translations and grounded representations; and measure accepted bytes and tokens by domain, culture, style, complexity axis, and epistemic function. Only the accepted, replayable output of those steps can become a curriculum shard.

Evidence and status vocabulary

To keep progress legible, Sai uses these states consistently:

• Referenced: exact upstream repository, revision, path, size, and object hash are known.
• Locally present: bytes were actually fetched and their content hash was verified.
• Candidate: the source passed mechanical parsing into a compiler queue.
• Judged: a model or human produced a schema-valid assessment with a sealed request/response receipt.
• Verified: rights, content, transformations, evidence, and contamination checks passed independently.
• Curriculum-ready: concepts, prerequisites, difficulty axes, duplicate family, and spiral placement are complete.
• Training-ready: the final packed stream replays exactly and every required gate is closed.

The current truthful status is: the 8 TiB reservoir is referenced and hash-bound; the 8T-token spiral is prospective; one production book, 224 Common Pile confirmation rows, and both bounded source pilots are judged; the two pilots contain 3,290 benchmark-screened, near-deduplicated rows and have passed the source-specific full-candidate-materialization gate without becoming training-ready; the new UltraData Math L2/L3 screen has 148/160 benchmark-disjoint rows awaiting Hermes compilation; a complete PEP parent census has 567 benchmark-disjoint, near-deduplicated survivors dependency-staged for Hermès compilation; a 1,024-row CC0 arXiv temporal screen has 1,023 benchmark-disjoint survivors, zero near-duplicate pairs, and complete declaration coverage awaiting Hermès; the complete arXiv parent census measures 2,458,156 mechanically eligible unique rows and 2,380,856,330 text bytes while leaving every downstream quality, contamination, deduplication, and training gate open; a live source-page probe observed matching declarations for 1,719 pilot rows without establishing governing scope; and the reservoir as a whole is not training-ready.

Every training population must pass these gates in order:

1. Source truth: reopen exact source bytes; reject corruption, spam, benchmark overlap, unsupported claims, and high-confidence duplicates.
2. Semantic foundations: bind each lesson to the concepts it teaches and assumes. A dependent concept cannot appear confidently before its declared prerequisites have accumulated enough independent prior exposure.
3. Learnable progression: move from grounding to composition, reasoning, and specialization. Surface complexity may break ties, but it cannot overrule the semantic prerequisite graph.
4. Rehearsal and retention: keep foundational material in every later phase. Evaluate the same phase-stratified held-out population at prospectively fixed phase boundaries so acquisition and forgetting are both visible.
5. Matched falsification: compare the proposed order with the identical record multiset under a frozen order control, with the same tokenizer, initialization, optimizer, token budget, and observation schedule. Require held-out likelihood and real source-disjoint capability evidence; any prerequisite-phase regression vetoes promotion.

Specialized data is therefore not automatically "better" data. It becomes useful only after the learner has the language, symbols, primitives, and compositions needed to extract its signal. Sai will not compensate for a bad curriculum by adding parameters, inference-time reasoning, or architectural machinery.

The executable milestone evaluator now enforces the retention part of this constitution for future matched runs. It scores the same phase-stratified development population at initialization, every prospectively declared phase boundary, and termination, then compares curriculum and order-control acquisition and forgetting phase by phase. The current live 500M-token run was frozen before milestone snapshots were added; it retains valid final phase-stratified evidence but will not be misrepresented as a full dynamic learning-curve experiment.

Execution status: on 2026-08-22 the user authorized Sai to proceed directly to the 4B architecture once the data is ready. Small-scale architecture tournaments are no longer a launch prerequisite. The remaining scientific launch gate is the exact data artifact: canonical unique documents, provenance and licensing decisions, benchmark decontamination, frozen exposure weights, semantic-prerequisite curriculum order, tokenizer qualification, and a replayed packed stream. A bounded one-update 4B execution canary remains a technical requirement so the full job does not fail on memory, kernel, or checkpointing mechanics. Neither the authorization nor a successful training run is evidence of improvement; only matched real-benchmark results can establish that claim.

Data-first reset — 2026-08-22

• Data quality and presentation order are now the primary Sai admission gate. The 30-file FineWeb-Edu prefix remains useful raw material, but passing its upstream educational score, basic hygiene filter, and benchmark decontamination is not sufficient to make it a training curriculum.
• Audit found that the previous 500M-token freezer preserved upstream file/row order. It had no prerequisite progression, difficulty strata, domain pacing, or near-duplicate-cluster gate. Dependency-held stream job 769226 and GQA launcher 769687 were therefore cancelled before execution: both recorded zero elapsed time, zero restarts, no node, no child job, and no scientific output. Parallel decontamination recovery 769626 continues because its benchmark-disjoint admitted corpus is still required input evidence.
• sai.data.curriculum now defines a create-only four-band, four-phase data contract. It performs a second quality pass, removes high-confidence five-word-shingle near duplicates, measures every document's surface difficulty, excludes specialization from grounding, backloads advanced material, and rehearses foundational material in every later phase. Its validator reopens source/decontamination evidence and replays every output row, band, duplicate decision, phase mean, and identity fingerprint.
• Recovery job 769626 completed the exact twenty-boundary decontamination pass without retries. It published a 9,541,423,202-byte admitted corpus with SHA-256 5e234e56d3df101c668beb49d4582740d6bc3fbe723448231c73a6d4e6e57dda; its receipt file is SHA-256 a396590cc253fb208c276fb004c53b5dd50bca651811c9f1c315cd69c7c8cb5f. Curriculum builder 769780 atomically published a 9,469,603,720-byte curriculum and receipt before entering its original single-core replay. It scanned 2,153,160 documents, admitted 2,125,835, rejected 27,325, and emitted every admitted identity exactly once. The admitted bands contain 33,220 foundation, 360,044 composition, 1,660,568 reasoning, and 72,003 specialization documents. Phase mean difficulty rises monotonically from 0.23614 to 0.27084, 0.29141, and 0.30431; grounding contains no specialization and later phases retain foundation rehearsal. Independent eight-worker replay 770001 completed in 1,351 seconds with zero restarts, empty stderr, exact receipt SHA-256 b575bebb18509c13848ac34146ac9b8a7d4be54e83e4d66f43100eb717545c1b, and status qualified. The redundant single-core replay was then cancelled; no published artifact was removed or changed.
• A second audit found that the legacy development stream was produced from an earlier corpus slice and could not prove disjointness from the new admitted population. Dependency-held freeze/control jobs 769787 and 769788 were cancelled with zero elapsed time, zero restarts, and no node before they could publish a stream. sai.data.curriculum_split now performs the train/development separation only after the global quality and high-confidence near-duplicate pass. It preserves the four phase boundaries in training, assigns every accepted identity exactly once by a frozen hash modulus, and replays the full source against both outputs. New training and development token streams must bind that split receipt; the legacy development stream is not admissible for the curriculum experiment.
• A valid curriculum receipt is still not proof that the order helps. Before using it for a larger model, Sai will compare it against a same-document deterministic order control at small scale, with matched tokenizer, initialization, optimizer, updates, and compute. Held-out NLL/UTF-8 byte and source-disjoint real capability must improve or remain nonnegative without a domain regression. Tokenizer, filtering, and ordering remain separate factors.
• The current four-phase schedule remains explicitly a surface-complexity falsification experiment, not the final semantic curriculum. It cannot prove that a concept's prerequisites were learned. The next data gate must bind an acyclic concept graph, evidence spans for every exposure, minimum prior coverage before dependent exposure, first-exposure positions, phase-local rehearsal, and unresolved violations. This is the executable version of the rule that a model should encounter the concepts of yellow and blue before it is expected to learn their composition into green.
• A separate model-centric scheduler is now executable for web-scale pacing. sai-score-learnability first compares a predeclared earlier milestone with the terminal state from the same independent probe trajectory on an exact-record-disjoint target stream. Its immutable receipt binds both model states, the target and probe streams, tokenizer, runtime, and every normalized loss row. sai-build-learnability-curriculum then consumes that receipt and a prospectively frozen weak/strong-checkpoint policy. It reorders the identical token-and-boundary record multiset into ready, developing, challenging, and stretch bands, preserves ready-record rehearsal in every later phase, and hash-randomizes within each phase so continuous loss order is not a hidden factor. The full permutation is replayable and treatment checkpoints and terminal benchmark feedback are prohibited. This is a future matched factor, not a semantic prerequisite claim; no qualifying production score population or training run exists yet. Sai therefore treats difficulty as two independent axes: model-relative learnability and audited semantic prerequisites. Neither can silently substitute for the other.
• sai-compose-semantic-learnability is the production composition boundary. It replays the complete semantic taxonomy, curriculum, annotations, and progression report; requires zero skipped or truncated semantic documents; and locks every packed record to its audited semantic phase. Weak/strong learnability evidence may only determine bands and order inside that phase. An advanced low-loss record therefore cannot jump ahead of missing prerequisites. The output retains the identical token-and-boundary multiset and remains unauthorized for training or 4B until a matched comparison wins.
• The semantic gate now freezes exactly 120 review documents: eight from each of the 15 pedagogically valid phase/band cells before any labels are produced. grounding:specialization is excluded because qualified progression requires that cell to remain empty; demanding examples there would contradict the curriculum being audited. A separate replay compares the prospective annotator with an independently identified reviewer, validates every cited evidence span against immutable source text, and computes concept-set disagreement directly. The taxonomy cannot be built unless at least 100 documents were reviewed and disagreement is at most the prospectively frozen five-percent ceiling; callers can no longer satisfy this gate with an unattached arithmetic-only receipt.
• The first held-out selector attempt, CPU job 770519, correctly failed before publishing an artifact because its legacy 16-cell geometry demanded examples from grounding:specialization, a stratum that the qualified curriculum requires to be empty. The corrected selector was pushed at commit 89fda9b76d3898e10a150b25557d7b14768be7d3 and executed as CPU job 770526. It completed in 151 seconds on evc21, with zero restarts and empty stderr. Independent replay verified exactly 120 immutable rows, eight in every one of the 15 valid phase/band strata, and no grounding-specialization row. The population file SHA-256 is 85a7e804b0622f85b2f45c3edf5a37a1a200fdd3e9c833a8c65edf09a804cce8; its canonical receipt SHA-256 is 34ca4ca64acccaf3bc1ae04152156ea879efed47d85dcfdac3242ef5ee8b171a. The population is selected but deliberately unreviewed, and its receipt keeps both training and 4B authorization false. Independent blinded annotation and disagreement review remain mandatory before this sample can qualify the semantic taxonomy or any training curriculum.
• sai-build-prerequisite-blind-review closes the remaining label-leakage boundary before that annotation begins. It shuffles the 120 rows by a salted review identity and exposes only immutable text and the requested evidence format. Curriculum phase, surface band, source identity, and original order are held in a separately sealed key. Reviewers therefore cannot infer the answer from the schedule they are auditing, and no packet can authorize training or 4B scaling. CPU job 770533 executed the exact pushed implementation at commit d798e4edc48d06ff112c95d44872b0d123ad01ab, completed in 224 seconds on evc21 with zero restarts, and replayed the packet twice. The blinded packet SHA-256 is ae4ebe9721f2b2e156cb72aba9f46f73396f0cdcecd0480754784b32c0efba2d, the withheld key SHA-256 is f2986e52cedec2a5db8802c43505539d76b17210d7e1fac026e79ef2aa8aa3fa, and the canonical receipt SHA-256 is 6a07e0e1fa8e5832f2b196cd65e1547f563a47d0f896278fcbb364a3626e347d. Independent inspection confirmed all 120 packet rows omit phase, surface band, source, original index, and document identity; the separate key retains all mappings. The packet is now ready for two independently identified blind reviews, but contains no completed labels and authorizes no training. The same tool now compiles a frozen blind response only after reopening the sealed key: it validates the candidate concept vocabulary and every cited character span, restores phase and document identity in canonical population order, and emits a separate lineage receipt. A single compiled review remains explicitly insufficient; qualification still requires two independent identities and a passing disagreement replay. The replay now additionally requires each independent side to label at least 100 of the 120 documents with evidence-backed concepts. Two empty annotation files therefore fail even if their nominal disagreement is zero.
• Source-disjoint split 770039 is now running from that independently qualified receipt with four ordered workers across both receipt validation and exact split reconstruction, plus a four-hour fail-closed wall limit. Earlier split 770024 was cancelled after 542 seconds, before creating any output, when byte-equivalence tests proved the ordered parallel reconstruction. Development stream 770040, update-aligned 500M-token training stream 770041, exact sequence-multiset order control 770042, and matched GQA launcher 770050 are dependency-staged. Launcher 770043 was canceled before allocation (zero elapsed and zero outputs) when the frozen runtime was advanced solely to give terminal replay four hours. The first three curriculum boundaries occur exactly after optimizer updates 191, 429, and 715, so no gradient accumulation window mixes adjacent difficulty phases.
• Split construction has atomically published all 2,125,835 admitted identities exactly once: 2,104,726 training documents in 9,375,399,692 bytes (SHA-256 6a43417411f886336632f2ad1abbf539504043f28180cdc4a6fa792e7de6241b) and 21,109 development documents in 94,204,028 bytes (SHA-256 56ebbf200bae4ce21c454bd80e91328ab9a486798c83a0b729477f57e0122289). Its qualified receipt self-hash is 0580b683427ecae5943cc0a706e9fb31686c46051283b154e1aebc09b78eb0aa; job 770039 completed the independent full replay at 0:0 with zero restarts. Its stdout contains the same qualified receipt identity for build and replay, and stderr is empty.
• A downstream audit caught a validation confound before allocation: freezing the first 1,024 sequences of the phase-ordered development file would test almost only grounding data. The development freezer now requires exactly 256 sequences from each of grounding, integration, reasoning, and specialization and binds those four token/byte strata in the stream receipt. Original jobs 770040 and 770050 were canceled at zero elapsed with no node, logs, or outputs. Training-stream job 770041 completed at 0:0 in 5,386 seconds with zero restarts and empty stderr. Its 60 shards contain exactly 244,140 packed sequences and 499,998,720 tokens. The four emitted phase budgets are 48,896 / 60,928 / 73,216 / 61,100 sequences, and their boundaries match the declared optimizer updates. Receipt file SHA-256 is 8de9780d4b5b873e668260ee2423c1536912163f9fb6696597663ac0c1e026b1; ordered-stream identity is c4c271f38b55ab277c7660719e3d36bc485063d440a3745b8f4d532545d51636. Dependent job 770042 completed the exact sequence-multiset order control at 0:0 with zero restarts. It contains the identical 244,140 packed records, 499,998,720 tokens, boundary masks, tokenizer, source, and admitted UTF-8 bytes, but applies frozen permutation seed 2026082201 with zero fixed points. Its receipt file is SHA-256 a8481d767a6e468a5c46a69331ebe33fca80ff3d382bfd014b6e7ffa62c05ad0; ordered-stream identity is 0d40a828e83b7cda52fcf77489dbe2a223761fe70f8972f2d3e66297d4439513.
• The likelihood evaluator and terminal order comparator now retain all four development strata separately. A lower aggregate NLL cannot pass if grounding, integration, reasoning, or specialization regresses against the exact-order control. This closes the remaining possibility that easier rows could hide curriculum damage to advanced material. Development job 770086 exposed a zero-work runtime packaging defect: eight historical executable bits had been stripped, so the runtime correctly failed its clean-tree check before Python or data access. Its dependent launcher 770088 canceled without allocation. Both output roots remained absent. Replacement job 770105 completed at 0:0 with zero restarts and empty stderr against clean immutable runtime 78f9d7b. It published exactly 1,024 packed sequences: 256 and 524,288 tokens from each of grounding, integration, reasoning, and specialization. The stream receipt file is SHA-256 6ae403c18f683fa3ddd7536989c38f7c5c4c14a3b448993e020cf84cabb8eb9a; its ordered-stream identity is 232f68b380db1cbfa75aeda8c8bb3a878f9afe1551528b6efcbccbb4c6e6a34a. Its exact train/development split receipt and development source hashes match the independently qualified lineage. Launcher 770106 then failed before submission because its export omitted the literal _stream_ component of the completed training-stream path. It ran for four seconds on CPU, emitted no output, and submitted no GPU child; continuation 770127 was dependency- cancelled without work. Recovery launcher 770136 uses the corrected exact path and is currently replaying the split and all three streams with zero restarts. It has not yet submitted a GPU child. Continuation 770137 remains dependency-held. The trainer, evaluator, comparator, data bytes, model geometry, seed, optimizer, and evaluation rows are unchanged. Pre-allocation runtime audit then found that the prior 100M-token GQA arm required 9,905 seconds; a linear 500M-token projection is 49,525 seconds and the predeclared 25% safety margin is 61,907 seconds. The original eight-hour child limit was therefore deterministically insufficient. Future launcher bytes request 18 hours for each full arm. A create-only live guard will apply the same limit to the two still-pending children immediately after 770136 publishes their dispatch, before either training arm starts; it changes no data, model, seed, optimizer, update, or evaluation identity.
• Source-disjoint MMLU-Pro and MuSR evaluation now admits curriculum-derived streams only through the exact completed lineage from the benchmark-audited decontaminated source, through the qualified curriculum and split receipts, to the frozen train stream. Direct-source evaluation remains unchanged; missing, partial, parent-drifted, or train-drifted lineage fails before benchmark GPU submission.
• Population refresh now publishes the same canonical aggregate contract as the original population builder. The immutable 12,032-row MMLU-Pro and 756-row MuSR sources are reconverted against the current decontamination receipt; a refresh-only schema can no longer strand an otherwise valid evaluation before submission. Recovery job 770126 completed at 0:0 in 60 seconds with zero restarts and empty stderr. Its canonical aggregate file is SHA-256 dbbbeb2904a7d6d9c5e9fdc06017b97cdf5befaa70c0dba3622bd179386f43f1 with self-hash a6df423a114c2611ce6b3af16df1303278225c354fd9bd14e6a9db5d36a93f68.
• The real-development decision is frozen before scores exist. Both matched checkpoints must complete all 12,032 MMLU-Pro rows and all 756 MuSR rows. Curriculum order is retained only with nonnegative deltas on both boards, a positive unweighted macro, a strictly positive paired 95% bootstrap lower bound, and no domain regression worse than one percentage point. This development-only receipt cannot authorize architecture promotion or 4B.
• The positive-NLL handoff is now executable rather than manual. A CPU-only stager reopens the exact order receipt and both checkpoints, then submits two matched fan-outs: eight independent one-H100 MMLU-Pro shards plus one MuSR job for each arm. Two CPU merges and a terminal CPU comparator bind all 18 H100 jobs to COMPLETED|0:0|0 with zero restarts before applying the frozen benchmark decision. Partial submission is cancelled, while a negative NLL decision submits no benchmark work.
• A dependency-held continuation reads the live training dispatch and stages that handoff against the exact comparison and canonical population jobs. This removes the manual gap between a clean NLL decision and real-board measurement without reserving an idle GPU or pre-authorizing a favorable result. Job 770127 is held on launcher 770106; it requests no GPU and submits the real-board graph only through the frozen positive-NLL condition.
• Architecture promotion remains behind this boundary. Two matched 100M GQA jobs now measure curriculum ordering versus a deterministic order control on the identical record multiset; they are a data-order falsification, not an architecture result. The 4B prohibition remains unchanged.
• A deterministic forty-document qualitative audit of the exact development population confirms that the frozen surface score is not semantic pedagogy: grounding can include electrostatics and engineering, reasoning can include introductory rocks or biographies, and specialization can include basic traffic-light explanations or mythology. The exact sample rule, source hashes, observations, and resulting 4B-data prohibition are recorded in docs/SAI_CURRICULUM_QUALITATIVE_AUDIT_20260822.md. The current comparison remains an order falsification experiment; it cannot waive a later semantic- prerequisite or source-mixture gate.
• The first public math-source audit rejected blind use of FineMath 4plus. Exact shard 5d0b2611...1fe5 contains 104,680 unique rows, but only 34.7583% set found_math=true; direct evidence includes incoherent score-5 algebra with essay-service links plus answer-farm, SEO, gambling, and commercial-homework material. docs/SAI_FINEMATH_SHARD_AUDIT_20260822.md freezes the input and findings. FineMath remains a candidate only after a new Sai quality, provenance, deduplication, and decontamination filter.
• FineMath filter V1 then applied a prospectively frozen high-precision policy to all 104,680 rows and accepted zero. The dominant cause was a miscalibrated upstream language-confidence floor of 0.98: only 367 total rows reached it, while 3,114 rows passed every non-language criterion. V1 remains an immutable filter_empty_no_candidate result; it was not relaxed after the outcome. docs/SAI_FINEMATH_FILTER_V1_RESULT_20260822.md binds the receipt, rejection counts, and exact funnel. The next prospective step is a blind human-review ladder at no language floor, 0.90, and 0.95—not training.
• That prospective ladder has now frozen 3,114 non-language-qualified candidates and a 192-row blind packet: exactly 64 rows from each of <0.90, 0.90–0.95, and >=0.95. The packet hides score and stratum; its separate key remains closed until labels are complete. Receipt a17dcf57…d6d0 authorizes no training.
• An offline workspace now exposes only the 192 row identities and texts, with resumable evidence-backed labels and no external requests, URLs, language scores, strata, or hidden key. The post-review decision was frozen before labels: two independent reviewers, at least 90% consensus acceptance and an 80% Wilson 95% lower bound in every included stratum, then the lowest passing floor among none, 0.90, and 0.95. If the >=0.95 stratum fails, FineMath is rejected rather than retrospectively relaxed. Any selected rows remain non-training candidates pending global deduplication, benchmark decontamination, provenance replay, and semantic prerequisite placement. docs/SAI_FINEMATH_HUMAN_REVIEW_WORKSPACE.md freezes this boundary.
• Bulk code admission now begins with an executable Stack-Edu metadata audit, not a download. It pins revision eeec5caa…814c, rejects every unlicensed or mixed-unallowlisted row, and measures quality score, encoding, length, provenance, and duplicate identities before content retrieval. Metadata can only nominate blobs for a later content audit; missing current opt-out replay, secret scanning, legal review, deduplication, or benchmark decontamination keeps training authorization false. The complete pinned five-shard Python population measured 25,286,019 rows, but only 514,566 (2.0349%) survived this preliminary filter; 20,722,635 were marked no_license, and 18,536 more were marked permissive without a detected license. The independently replayed complete-language evidence is recorded in docs/SAI_STACK_EDU_PYTHON_LANGUAGE_AUDIT_20260822.md.
• The complete Stack-Edu candidate identities are now frozen separately from source content. CPU job 770639 completed in 3,100 seconds with exit 0:0, zero restarts, and empty stderr. It froze 514,566 unique blob identities spanning 127,672 repositories and 514,559 unique repository/path pairs; the candidate JSONL SHA-256 is 7429c9d4…07c5 and canonical receipt identity is ec6b0aa9…0762. Exact execution, hash, and population evidence is in docs/SAI_STACK_EDU_CANDIDATE_IDENTITY_AGGREGATE_20260822.md. This remains metadata-only and authorizes no content retention or training. Current-release alignment is executable. The new sai-stack-v2-current-python-snapshot-v1 boundary requires the complete Python metadata shard set from bigcode/the-stack-v2 revision e565caa3…90e47 (v2.2.0, opt-outs enacted through 2026-07-29) plus the exact dataset card and self-hashed access evidence. The freezer queries the authenticated Hub API at that commit and verifies every local file against its remote Git or LFS identity; caller-chosen local hashes are insufficient. Alignment retains an old candidate only when the exact (repo_name, path, blob_id) still exists and the current row is permissive, non-vendor, and non-generated. Missing rows are treated as removed. This closes current opt-out drift only; content-byte verification, attribution, secret/PII/malware scanning, global exact and near deduplication, benchmark decontamination, and semantic curriculum review all remain mandatory. See docs/SAI_STACK_V2_CURRENT_ALIGNMENT_CONTRACT.md.
• Authorized acquired code must then pass the separate exact-content bundle verifier. It requires a sealed contiguous bundle and ordered index, recomputes the Git/SWH blob SHA-1 over blob <length>\0<bytes>, checks independent SHA-256 and declared length, proves strict UTF-8 round trips, and rejects any gap, overlap, missing row, or trailing byte. A valid byte receipt still keeps training and 4B authorization false; quality, secrets, global duplication, contamination, and semantic placement remain downstream gates.
• Verified bytes now feed a separate bounded safety/quality findings pass. High-confidence private-key and credential formats plus invalid control bytes are vetoes; personal-email, high-entropy/JWT-like strings, generated markers, extreme repetition/minification, and Python-version parse failures require review. A row with no bounded finding is still only a candidate because this scanner cannot prove the absence of novel secrets, malware, dependency hazards, or subtle benchmark-derived code. The exact policy and limitations are frozen in docs/SAI_STACK_EDU_CONTENT_SAFETY_CONTRACT.md.
• A separate create-only selector now resolves those bounded findings without pretending they constitute source admission. High-confidence rejects cannot be overridden; every manual-review row requires one exact hashed adjudication; bounded-clean rows remain candidates. The selected population retains training and 4B authorization false until accuracy, usefulness, duplication, contamination, semantic placement, and matched source-addition evidence pass. Its contract is docs/SAI_STACK_EDU_SAFETY_SELECTION_CONTRACT.md.
• An authored programming-curriculum candidate now preserves 111 Rust Book chapters and 16 CPython tutorial chapters at exact pinned revisions, in publisher order with byte-exact code and license evidence. Its 127-row candidate receipt is 80de7bef…08e; it authorizes no training. Python's tutorial explicitly assumes prior programming knowledge, so every Python row requires programming_foundations instead of being mislabeled as grounding. The authored sequence is a prospective pedagogical spine, not a complete corpus; semantic review, global deduplication, decontamination, source-addition controls, and identical-document order controls remain mandatory. Exact evidence is in docs/SAI_AUTHORED_CURRICULUM_CANDIDATE_20260822.md. All 127 rows are now frozen in a salted blind-review packet (f052ff87…b906) that hides our provisional order/stage key until independent concept labels and evidence spans are complete. The two-reviewer adjudicator now verifies all evidence spans and preserves separate concept, prerequisite, quality, admission, and defect disagreement; no completed labels or PASS exist yet.
• The final 4B mixture boundary can no longer be satisfied by plausible-looking 64-character hashes. The new relocation-safe v3 validator reopens every source manifest, selection policy, license decision, quality audit, decontamination receipt, and pedagogical progression receipt; validates exact file bytes plus canonical receipt schema/status/self-hash; and rejects links, missing evidence, and re-signed drift. Each receipt must also carry its exact role-specific positive decision; a generic status: qualified cannot admit a source. Each decision must also name the exact source-manifest hash it covers, preventing cross-source receipt reuse. There is no structure-only v3 mode; every validation reopens the evidence root. No v3 mixture passes yet.
• The source-addition gate is now executable rather than a prospective table. sai-compare-source-addition requires equal training tokens and compute, identical model/initialization/optimizer/tokenizer/development evidence, and distinct replayed source qualifications. It compares target-normalized and UTF-8-byte-normalized held-out likelihood in every development stratum; any stratum regression vetoes the source. A likelihood pass still retains nothing until real source-disjoint benchmark confirmation completes. The NLL receipt now binds each terminal checkpoint plus manifest using the exact bundle identity consumed by evaluation. sai-confirm-source-addition-benchmarks then requires paired complete MMLU-Pro and MuSR development evidence, exact checkpoint lineage, a positive 95-percent paired macro lower bound, no negative benchmark delta, and no domain regression below one percentage point. Only that terminal receipt can retain a source, and it still cannot promote an architecture or authorize 4B training.
• The matched curriculum-order experiment is now live. CPU launcher 770136 completed at 0:0 after replaying the split and all three streams and published dispatch SHA-256 4d682933…7ee. Its create-only wall-time receipt 895fd3e3…2cc extended both full arms to the predeclared 18-hour bound before release. Exact-geometry canary 770153 completed on one H100 in 294 seconds, with zero restarts and empty stderr; its terminal run receipt is f484fcf8…68b1. Curriculum job 770154 and identical-record deterministic- order control 770155 then started independently on evc22 and evc24. Both train exactly 499,998,720 tokens with the same model, initialization, optimizer, seed, update count, tokenizer, and admitted record multiset. Both independently published their first durable checkpoints at optimizer step 10: exactly 2,560 sequences each. Their checkpoint manifests bind the same code (16439922…410), configuration (a34e96f8…f11), environment (778d1372…f29), and model (ef67fa5d…574) identities; only the ordered- stream and resulting run identities differ. The curriculum checkpoint hash is 02fe57b4…e2d, and the control checkpoint hash is 1fef6028…996.
• Continuation 770137 failed in five CPU seconds before submitting work because Slurm would not accept a new dependency on completed population job 770126 after it aged out of the live controller table. The population and split jobs remain exactly COMPLETED|0:0|0 in accounting. A later audit found that pending comparator 770156 would have published a raw checkpoint-file hash where the benchmark evaluator requires the canonical checkpoint plus manifest bundle hash. Pending benchmark stage 770159 inherited that deterministic lineage mismatch. Both jobs had zero elapsed time and zero restarts, created no outputs, and were cancelled without touching live training. The exact scientific inputs and destinations are now staged behind corrected CPU comparator 770503 and benchmark stage 770505, using sealed, clean runtime commit c4b3b011ddee4d5e8aad3b60a30219b799c5b686. Comparator 770503 still depends only on training jobs 770154 and 770155; stage 770505 depends only on 770503 and will independently reopen all completed population, split, stream, checkpoint, and manifest evidence before it may submit benchmark work. No GPU job was duplicated and no scientific identity, record order, token budget, model byte, or evaluation population changed. Source now avoids both the aged-dependency and checkpoint-identity failures in future continuations.
• Semantic annotation policy v2 and prerequisite taxonomy v3 now require every accepted positive concept label to bind at least 16 exact source Unicode codepoints. A bare term mention therefore cannot establish that a prerequisite was taught. This is only a minimum evidence guard; independent annotation review, prerequisite order, concept-density, and later-rehearsal gates remain conjunctive. No semantic curriculum has passed yet.
• Semantic audit selection can now run against the already-qualified, source-disjoint development split. The selector preserves the same frozen salt, four phases, four surface bands, and eight documents per stratum while re-reading 94 MB rather than repeatedly replaying the 9.5 GB training curriculum. Its 120 selected documents remain an unreviewed audit packet; the speedup changes no label, threshold, or training byte.
• The authored-curriculum prerequisite lane now has two exact blind candidate reviewers. CPU context jobs 770444 (Qwen3.5-9B) and 770445 (SmolLM3-3B) completed with zero restarts and verified all 127 prompts against the same hidden-key-free packet. Qwen observed 966–11,620 input tokens and SmolLM3 952–11,098, both below the frozen 24,576-token ceiling. Independent one-H100 review jobs 770450 and 770451 both loaded their models but failed on row zero after exhausting three invalid structured-response attempts; no label, draft, or review receipt was published, and comparison 770471 never ran. The repaired runner now preserves every rejected response, constrains output complexity, canonicalizes list order and unique whitespace-equivalent source quotes, and retains every original evidence threshold. Fresh collision-safe jobs 770735 and 770736 preserved exact failures: Qwen repeatedly placed evidence-backed taught concepts in both semantic roles; SmolLM3 completed two candidate rows before repeatedly recommending admit with an empty taught set. Neither published a complete result and comparison 770738 never ran. Pushed commit a14e6eae…d387 now gives explicit quoted taught evidence precedence over an ungrounded duplicate assumption and conservatively maps empty-taught admit to revise. All evidence thresholds remain unchanged. Fresh jobs 770761 and 770762 each completed three replayable rows before failing on row 3; comparison 770763 never ran. Their immutable failure artifacts show only redundant nested quality fields and unsupported, ambiguous, or sub-minimum evidence strings. Pushed commit 2aaf0f3…af5bd9 conservatively discards those unsupported strings, retains only unchanged unique literal 16-codepoint evidence, and never upgrades a recommendation. Fresh collision-safe jobs 770785 (Qwen) and 770786 (SmolLM3) are staged independently, with CPU comparison 770787 dependent on both. They may rank cross-family disagreements for human attention, but model output still cannot qualify labels. Both reviewers crossed the former row-3 boundary but terminated at rows 4 and 5 respectively: Qwen could not ground defect quotes, while SmolLM3 exceeded evidence bounds and then returned non-JSON text. Comparison 770787 never ran. This ends parser iteration rather than weakening the evidence contract; the packet now proceeds only through the frozen two-human review path. sai-build-authored-review-workspace now makes that path operational through a self-contained offline form. It exposes only salted review identities, exact chapter text, the candidate vocabulary, and frozen evidence rules; requires every row to be explicitly reviewed; supports packet-bound local progress export/import; and emits exactly the existing compiler's JSONL schema. It performs no network request and includes neither the hidden key nor provisional phase labels. Its receipt still records human review, training, and 4B authorization as false. Exact failure and recovery evidence is in docs/SAI_AUTHORED_MODEL_REVIEW_RECOVERY_20260822.md. The final adjudicator now accepts neither arbitrary identity strings nor model-review identities: each side must bind all 127 completed rows to a distinct human identity attestation, exact packet and policy hashes, no model-generated labels, and no hidden-key access before label freeze. Model-model agreement can therefore accelerate review but cannot admit training data.

Live scratchpad — 2026-08-21

• Measured full curriculum/split/stream replay exceeded two CPU wall-hours on Newton before dispatch. Future curriculum-order launchers therefore reserve six CPU hours for the evidence replay; the downstream comparison remains a separate four-hour CPU job, and both 500M-token H100 arms retain their independently measured 18-hour limits. This changes no data, seed, model, or scientific result.

• Exact FLA 0.4.2 Gated DeltaNet and KDA chunk mechanics remain qualified by Newton job 768134. The environment receipt file is SHA-256 778d137224671a44acdcc923270dc7478cded5437780a0ea37e19b764a219f29.

• The benchmark-decontaminated training corpus, mechanically qualified lossless 48K fixed-geometry tokenizer, and exact binary streams are complete. The qualified default tokenizer tree is SHA-256 cf4879ee5b3914b4af187abcc93be5678e41ff942e0b0a14f6eeb1a089f6f76d. It is not an empirical tokenizer-capacity winner: the completed tournament measured 459,376 English-labeled documents but no representative code, math, science, or technical corpus strata. The exact compression tradeoffs and the required capability selection boundary are documented in docs/SAI_TOKENIZER_EVIDENCE_AUDIT.md. The shared 48,828-sequence training stream identity is b50bb94bc4ada3c5949430222d5551b6dc60423378cacd1f80a57641b1546b22; the source-disjoint 1,024-sequence development stream identity is ec533b1faadea0e0974bfce07923f126be5a2dfe3976b5ab3cf10cf0b43c6dd0.

• KDA/MLA (99,594,248 parameters) completed all 191 AdamW updates in job 768546 with zero restarts. It consumed 48,828 sequences and 99,831,130 valid targets. Held-out NLL is 5.64448 per target (282.73 perplexity, 1.20227 NLL/UTF-8 byte). Its immutable result file is SHA-256 7ee0fdc6ae229751976a579187e2d931f9c16a95294612a8bcca1de9e8a7c7e8.

• GDN (100,019,648 parameters) completed the identical work in job 768529 with zero restarts. Held-out NLL is lower at 5.58370 per target (266.05 perplexity, 1.18932 NLL/UTF-8 byte). Its immutable result file is SHA-256 accba7dc5728ffa6317a08bd0d61271778d04f44a89568b7bc8b0cd4d60a601b.

• Gated GQA (100,481,024 parameters) completed the same 191 updates, 48,828 sequences, and 99,831,130 valid targets in job 768523 with zero restarts. It is slowest and has the worst held-out NLL: 5.65844 per target (286.70 perplexity, 1.20524 NLL/UTF-8 byte). Its immutable result file is SHA-256 592bf31ab880b532bb230016e17b77052578fd84da911add5fa86e9a8147afd6. Its eight independent MMLU-Pro shards 768911–768918 and merge 768920 completed without retries: 1,101/12,032 = 9.1506%, also below the exact 11.0877% uniform baseline. The merged result file is SHA-256 0ba400ec78a89f86fcfd002897a45a8ad6529ff3efab32d1c506aad1544f9001. MuSR job 768919 completed at 256/756 = 33.862%, also below the exact 37.099% baseline; its result file is SHA-256 40d5638157f3180f2e06ef61bcbca4a34215138ae498ad942cbdc3541239e8bd.

• On the complete 756-row source-disjoint MuSR development population, KDA scored 257/756 = 33.995% and GDN scored 255/756 = 33.730%. The paired GDN-minus-KDA delta is -0.265 pp, 95% paired normal interval [-2.726, +2.197] pp; both are below the 37.099% uniform-choice baseline. This is evidence of no capability separation, not an architecture win.

• The original output-free monolithic MMLU-Pro jobs were stopped after exact workload measurement showed 113,990 independent choice forwards. Eight immutable 1,504-row shards per model completed without retries: KDA jobs 768764–768771 merged through 768772; GDN jobs 768773–768780 merged through 768781. The shared manifest covers all 12,032 identities exactly once and is SHA-256 35d714c6ba8f5f2509be3c71e8fc805b4caa54c2c49812315a05dbbcd2e7ba8b. KDA scored 1,133/12,032 = 9.4166%; GDN scored 1,113/12,032 = 9.2503%. The exact variable-choice uniform baseline is 11.0877%. Paired GDN-minus-KDA is -0.166 pp, 95% interval [-0.533, +0.200] pp. Both recurrent candidates fail the real-capability screen and are not promoted. Their immutable merged result files are SHA-256 06959cdbd5a038870fcc5a9da58e5ce49c8edc63a526f7265b3bd3e30d42b4d6 and 830261b871521297412b589d9ef20371449043bf0fa31292445c13babf3ae04e. GQA is worse on the complete MMLU-Pro board; no family has cleared the capability floor at this token budget. The full comparison file is SHA-256 1558265380edeb701492585477fb16494b280e6c29aa8139d34650123dfef708. The predeclared decision receipt is SHA-256 87b0a0c13eda2a1ab02ef898a48927af51d1dc7346f3d47791cf05ec889b6ed3 and returns no_family_capability_qualified_data_extension_only: no mixer is promoted, and the longer run is strictly a data-starvation diagnostic.

• A separate architecture-independent 30-file FineWeb-Edu prefix is now pinned for the next 300M factor screen. It covers exactly 64,562,434,300 raw upstream bytes at revision 87f09149…b8f9. Attempt 768858 failed before download because ordinary Newton CPU allocations omit SLURM_TMPDIR; no data was written. Corrected job 768891 then scanned exactly 21,855,000 documents and admitted 2,661,644 into a 14,491,695,743-byte source whose SHA-256 is 2f908f5f225de109a21f66fb9fb31baa1f35b4a57f1d6d2a3f60fa95a98ea7e6. Exact decontamination 768892 reached a live source position of at least 6,257,876,992 / 14,491,695,743 bytes (43.18%) with zero restarts, but its maximum RSS had risen to 66,368,664 KiB inside a 64-GiB allocation. Exact source replay proved that implementation stored each SHA-256 shingle as a 64-character Python string; measured RSS growth left only about 740 MiB and projected an unavoidable OOM near halfway. It was therefore cancelled at 20:16 EDT before infrastructure termination, after 12,457 seconds and before publishing an output or receipt. Its sole abandoned partial was explicitly resolved as PID 3211846, measured at 4,150,497,097 bytes, permanently removed, and confirmed absent. No scientific decision changed. A disk-read heuristic initially overstated progress; the process file descriptor is the authoritative scan measurement. Newton rejected a running wall-time extension. Conditional CPU fallback 769225 was originally held on afternotok:768892 with compact, byte-equivalent SHA-256 shingle storage at commit 540d1080ebc5b1cd2b463d8c137e9d2c71567e82. It started on evc1 with zero restarts at 20:16 EDT. The compact representation held resident memory near 48,812,000 KiB, but exact process-file-descriptor replay at 22:51 EDT measured only 3,701,448,704 / 14,491,695,743 source bytes (25.54%) after 2h34m, projecting beyond its immutable eight-hour limit. A deterministic ordered fork implementation was therefore built at commit cc5a7fd4c938e66166c1a9fd7aef91b2f51f2877. A live 20,000-row, one-boundary A/B proved that sequential and eight-worker builds produced the exact same 101,009,890-byte output with SHA-256 d86c43c32f4a0a6649fbd5045515c798f232ea1829d26fce2f85b374d2fc1212. Parallel job 769636 completed in 134 seconds versus 311 seconds for sequential job 769635, an exact measured 2.32x speedup. The slow primary was then cancelled deliberately at 22:55 EDT after 9,516 seconds, before any final output or receipt existed, and parallel recovery 769626 started immediately on evc1 with zero restarts and a 14-hour limit. Its source, twenty ordered boundaries, final output path, and receipt path are identical. The abandoned 2,521,335,906-byte partial from PID 3216825 was proven unreferenced, atomically quarantined, permanently removed, and confirmed absent; the new job owns a distinct staging file. The three stream descendants use an OR dependency on successful completion of 769225 or 769626, so the parallel lineage now exclusively releases shared Sai stream 769226, exact 125M-token Qwen stream 769437, and exact 125M-token SmolLM3 stream 769455. A pre-execution audit found that the first refreshed-population clone 769227 still checked terminal state for cancelled original job 768892 inside its script even though its scheduler dependency had been repaired. That clone and its three bound descendants 769235, 769237, and 769238 were cancelled with zero elapsed time and zero restarts before they could create any output. Exact mutually exclusive replacements now bind both scheduler and in-script custody to the same lineage: population jobs 769355/769627 follow primary/recovery corpus jobs 769225/769626; parent benchmark launchers are 769425/769426; workspace evaluation stages are 769440/769441; and terminal comparisons are 769442/769443. Each pair targets the same collision-safe artifact and only the branch whose corpus lineage completes can run. Workspace launcher 769439 remains staged. Two malformed first launcher clones (769233–769234) were caught while still dependency-held and canceled with zero elapsed time and zero restarts; their corrected replacements bind the exact fallback population and stream job IDs. No recovery job executed concurrently with the primary scan. Recovery-bound parent/workspace/benchmark stagers have been rewired to population job 769627; primary-bound branches are now dependency-dead and cannot race the recovery lineage. Successful decontamination is followed by shared 499,998,720-token stream freeze 769226. The original dependency-held 100M launcher 769232 and benchmark stager 769649 were cancelled with zero elapsed time and zero restarts after end-to-end DeltaMixer qualification found that their older immutable runtime checked only operator-level FLA parity. They submitted no child jobs and created no screen result. Current code now requires an explicit, non-downgrading scope: gqa_only submits exactly one reference GQA baseline over the same 249,999,360-token prefix and labels it non-tournament; three_family fails before output creation or submission unless an exact full-DeltaMixer receipt is hash-pinned and independently revalidated by both launcher and GPU job. Every selected family now first executes an exact B=8, T=2,048, one-update H100 canary; its longer screen is released only by that canary's successful completion, and both job identities are bound downstream. This neither selects a mixer in advance nor authorizes 4B. Once a three-family screen is qualified, the real-benchmark continuation is executable rather than aspirational: after those three checkpoint jobs and the refreshed decontamination population close, a CPU-only stager submits eight independent one-H100 MMLU-Pro shards plus one independent one-H100 MuSR job per family, followed by one deterministic CPU merge per family's MMLU-Pro shards. That is 27 single-H100 evaluation jobs, three CPU merges, and six terminal comparison dependencies across the three families. A final CPU job reopens the six row-complete terminal receipts, verifies all 27 H100 jobs completed cleanly, and emits exact paired family deltas. The benchmark launchers require the population's admitted-source SHA-256 to equal the sole source SHA-256 recorded by the trained checkpoint's token stream. The chain contains no arrays, retries, requeues, implicit substitutions, architecture promotion, or 4B authority. A red-team replay caught the prior six-job draft before execution: its monolithic MMLU-Pro arms repeated a known output-free four-hour failure mode, and its comparison stager could race an already-published fast result. Both are fixed at commit f128faaf507ab35b7782225e8aea273d5b7beea8; the complete tree passes 383 tests. The sealed Newton runtime /lustre/fs1/home/sa305415/sai-initiative-runtime-f128faa-r1 contains 213 files at tree ff7fba7473785368c2e5274e37c29edfdd02c343. Stale pending stager 769619 was cancelled with zero elapsed time and zero restarts. Replacement 769649 was also cancelled dependency-held, with zero elapsed time and zero restarts, when its unqualified upstream launcher was withdrawn.

• Full-model FLA qualification is an evidence-bearing NO-GO, not an assumed pass. Jobs 769650–769654, 769658, and 769659 used independent one-H100, no-requeue requests and executed zero optimizer steps. Across fixed seeds, all structural mappings, row resets, scale=1, and every one of 24 direct packed causal-convolution comparisons passed. The direct recurrence isolation at seed 20260823 produced GDN normalized-RMSE ratios 0.00413/0.00462/0.00457/0.00447 and KDA ratios 0.00213/0.00465/0.00565/0.00434 for lengths 1/63/64/65. Pinned FLA 0.4.2 requires strict forward ratio below 0.005, so KDA length 64 is a real failure and the existing v1 receipt remains unqualified. All direct convolution checks passed; full-layer ratios were about 0.35–1.09%, with sparse maximum-element outliers, so neither threshold relaxation nor a mapping-bug claim is justified. Commit e27a8d723a19cea3c0790568667dee06e5e67b15 preserves the failed evidence and gates hybrid screens on a future qualified receipt. The family-separated v2 primitive gate is now implemented with the upstream FLA forward/backward ratio limits and prospectively frozen seeds 20260824–20260826; calibration seeds cannot be reused, every tensor/case is a veto, and GDN/KDA receive separate statuses. It runs no optimizer and does not authorize training. GQA-only reference training remains independently admissible behind its exact-geometry canary. The three immutable v2 executions 769716, 769718, and 769720 then closed on those exact held-out seeds. Every direct GDN/KDA recurrence forward and backward metric passed, but every seed failed the shared BF16 causal- convolution gate: observed ratios ranged from 0.00137 to 0.00560 against the declared 0.001 limit. Receipt file hashes are respectively 2d022fe508d68327a94f62f79bdf1be0b1eda4957fc14dac8a3822e81b65992e, 005f6c880d44dc23cb018d794d266bf86d7c75ada344a1e5169a247468a1284b, and fa915883f6009fdef430967470626d9b084b365ee49ea45327004e77fed861eb. These receipts remain permanent FAILs. A source audit also established that upstream FLA's Torch-reference 0.001 convolution tests cover FP32/FP16, not BF16, while its relevant varlen check compares two executions of the same FLA operator. Therefore v2 does not establish a Sai mapping failure either: its claimed BF16 threshold grounding is invalid and must be replaced prospectively with fresh seeds, never relaxed or re-signed post hoc.

• The exact pretrained capable-host control is now restored at Qwen/Qwen3.5-0.8B@2fc06364715b967f1860aea9cf38778875588b17 without using a GPU. CPU attempts 769141 and 769144 failed before publication on a Bash digit-separator bug and one transposed README.md Git-blob pin; both output roots remained absent. Corrected job 769148 completed in 33 seconds with zero restarts. Independent replay verifies all 13 upstream members, including the 1,746,942,600-byte weight SHA-256 04b1c301231dd422b8860db31311ab2721511346a32cb1e079c4c4e5f1fe4696. The sealed 15-member local tree contains no links or writable members; its tree identity is 24457a397ecf57057d29636ace857c78f0983cb25647b24f10461bb4943e875d and its receipt is 3c7d25ab4d4bcf4dec81b594f8919636483bd64607ec1ae506c76e6ba815e00b. This is host preparation, not a Sai architecture result.

• First parent-mechanics allocation 769161 ran once on evc37 for 227 seconds, with zero restarts, and failed before model-weight loading, inference, training, or output publication. Its exact stderr SHA-256 is f1a91a4e0d5f7511f8c1f6607f7586995e6e2be7519fc399c7c537c5e88735bf. Read-only replay established that the checkpoint intentionally separates a 248,320-row padded model embedding/logit vocabulary from a tokenizer with base vocabulary 248,044, 33 added tokens, exact length 248,077, and EOS token 248,046. The first mechanics code incorrectly equated tokenizer length with model rows. The repaired implementation now binds all four identities independently in mechanics, stream-freeze, training, and replay validation. Second mechanics allocation 769410 loaded all 320 weight tensors on evc23, then failed in 49 seconds because Transformers 5.15 returns its empty loading-key collections as sets while the validator required the unexpected_keys container itself to be a list. CPU replay 769420 proved exact empty sets for missing, unexpected, and mismatched keys and an empty error list. Independent CPU geometry replay 769435 then proved the exact loaded text host is Qwen3_5ForCausalLM with hidden size 1,024, padded model vocabulary 248,320, 320 parameter tensors containing exactly 752,393,024 parameters, two buffers, and no loading discrepancies. The exact CPU forward replay 769445 also completed in 38 seconds: the frozen ten-token mechanics prompt produced finite logits of shape [1, 1, 248,320], FP32 sum -324265.375, maximum 18.125, and argmax token 279. Thus tokenizer, loading, model-call, and logits contracts are all exercised before the fresh H100 allocation; only CUDA residency and the sealed mechanics receipt remain allocation-specific. The validator now normalizes only list/set/tuple container form, still rejects any non-string or disallowed key, and continues to require no missing, mismatched, or error entries. Full local regression is 354 passed; these are admission-code corrections before any Sai model result, not scientific retries. Exact repaired commit d2cdf5f1fc314fd424c0316df278496a8675c157 is sealed on Newton with zero writable members. Fresh mechanics 769422 is an independent one-H100, no-requeue request with a measured-safe eight-minute limit. Its corrected graph is dependency-staged without occupying GPUs: primary/recovery parent launchers 769425/769426, accelerated Qwen stream 769437, workspace launcher 769439, primary/recovery workspace evaluation stages 769440/769441, and terminal comparisons 769442/769443. Superseded generic-stream jobs 769411, 769427–769429, and 769431–769432 were cancelled with zero elapsed time and zero restarts. Every target was absent at submission; a mechanics failure cancels all scientific descendants before allocation.

• Fresh parent mechanics 769422 completed 0:0 on known-good evc44 in 97 seconds with zero restarts. The 1,845-byte receipt file SHA-256 is e1767a706d3e7aafefb2707cdfcdd8b4af55699efb81c9997f0ea0ee13268ffb; its canonical internal receipt is d94e62c84be0625b3479b671fcd63a264b3358d7713975fb0718c6bfc25ee8a0. Independent replay against the sealed snapshot passes. It binds Transformers 5.15.0.dev0, Torch 2.6.0+cu124, CUDA 12.4, H100 capability 9.0, all 320 parameter tensors and two buffers on CUDA:0, finite [1,1,248320] logits with argmax 279, peak allocation 1,807,927,296 bytes, unchanged model state, zero backward/optimizer calls, and no 4B execution.

• The Qwen factor consumes exactly 61,035 × 2,048 = 124,999,680 ordered training tokens. A dedicated 125M-token freezer now materializes only that exact prefix plus the 256-sequence canary prefix instead of spending CPU, storage, and walltime on the unused 375M-token tail of the generic 500M stream. It preserves tokenizer, source order, boundary masking, and every byte consumed by training; its receipt explicitly rejects any extra prefix or wrong total. Corpus hashing is streaming rather than a 14.49-GB read_bytes() allocation. Full regression after this acceleration is 355 passed. Exact acceleration commit 88275bd9550bd1789f783389c218da188369805b is sealed on Newton with zero writable members.

• The conditional SmolLM3-3B cross-family confirmation consumes the same exact 124,999,680-token prefix. It now has its own 125M-token freezer with the identical no-unused-tail rule, streaming corpus hashing, restored-model manifest/receipt replay, exact Smol tokenizer identity (128,256 vocabulary, EOS 128,012), and no GPU exposure. This removes another 375M tokens of unnecessary preparation if Qwen earns promotion. Full regression is 356 passed. Exact commit a74e1e6e54ad1c57842eafcedb1882348cb6d095 is sealed on Newton with zero writable members. Superseded generic stream 769229 was cancelled with zero elapsed time and zero restarts; accelerated Smol stream 769455 is dependency-held on the same mutually exclusive primary/recovery corpus lineage. Conditional release jobs 769457/769458 are now staged behind the matching primary/recovery Qwen comparison, population, and Smol-stream branches. They target one collision-safe run root and submit the 32-job, independent-one-H100 cross-family graph only if the reopened Qwen receipt passes its predeclared gate; a measured Qwen fail exits before any Smol H100 request.

• The first capable-host Sai factor is now fully executable at commit cc7039d1e5a0653f4581cbe1a7b3ce509fff58e6: a 19,938,304-parameter, 16-slot recurrent workspace attached to the frozen Qwen3.5-0.8B text parent. Its matched reset_average control has identical parameters, initialization, optimizer, data, compiler/reactor/reader calls, and modeled workspace FLOPs; the sole change is whether reactor state carries across the two iterations. Each packed document is passed through the frozen parent exactly once, so no cross-document context leaks into the objective and all eight probe positions reuse the same detached causal hidden states.

• Exact parent H100 mechanics job 769161 and exact Qwen-tokenized 499,998,720- token stream job 769174 are the only unresolved prerequisites. Dependency launcher 769193 will request two independent one-H100 256-sequence canaries, then two independent one-H100 61,035-sequence full arms only if both canaries pass. No 4B model is involved.

• Parent development launcher 769171 and workspace evaluation stage 769194 are wired to the same refreshed, 500M-source-disjoint full MMLU-Pro and MuSR populations. Each workspace arm fans out as eight independent one-H100 MMLU-Pro shards plus one independent one-H100 MuSR job. Comparison stage 769196 will compute exact paired row deltas and 10,000-replicate stratified intervals against both the unchanged parent and matched reset control. Its pass can authorize only another sub-4B confirmation; it cannot authorize the 4B run. These are staged experiments, not positive architecture results.

• The cross-size/cross-family confirmation host is also pinned in advance, but remains unscheduled: HuggingFaceTB/SmolLM3-3B revision a07cc9a04f16550a088caea529712d1d335b0ac1, with 3,075,098,624 text parameters, a 6,167,865,576-byte sealed tree, and tree SHA-256 6badcd593aee3052e3d66afb315b979e2cc62c4a61f9cef31c07203912478a0f. Sai reopens its exact external manifest, receipt, and every weight member; a fresh CPU replay passed all 12 members and all 6,167,865,576 bytes. CPU-only first stream job 769203 was cancelled before allocation (zero elapsed, zero restarts) when preflight found that Hugging Face's base vocab_size excludes Smol's 256 added tokens, including EOS 128012. The packer now binds the full tokenizer length 128256; corrected CPU job 769208 is staged after decontamination to freeze the same 499,998,720-token source under the exact Smol tokenizer. The one-H100 no-training mechanics entry point remains unscheduled. Commit d58937c additionally prepares—but does not launch—the same recurrent-vs-reset factor on this host: a proportional 79,722,496-parameter, 16-slot workspace with identical initialization, optimizer, source prefix, objective, calls, and modeled workspace FLOPs across arms. The sole changed factor remains reactor state carry. This host will be used only if the 0.8B recurrent factor passes; preparing it is not a result and does not consume the terminal 4B boundary. Commits 29c1e1c, dfa4150, and b1980fa now complete the unscheduled Smol execution path: two independent canaries, two matched full training arms, eight independent MMLU-Pro shards plus MuSR per arm, deterministic merges, and a paired terminal comparison. The cross-family comparator must reopen a passing, hash-valid Qwen factor receipt before it can describe a Smol pass as cross-family confirmation. The graph is fail-closed against partial launcher submission, uses one H100 per scientific job, and still records both four_b_training_executed=false and four_b_training_authorized=false. This is executable preparation only; no Smol GPU job has been submitted. Commit 72289ca additionally provides one fail-closed CPU release that can open the passing Qwen comparison and submit the complete Smol mechanics, parent evaluation, matched training, workspace evaluation, and comparison hierarchy. The hierarchy contains 32 eventual independent one-H100 jobs and cannot release on a failed or re-signed Qwen receipt. A clean read-only Newton checkout of this exact commit is sealed at /lustre/fs1/home/sa305415/sai-initiative-72289ca with 226 regular files, zero symlinks, and zero writable members. No release job has been submitted.

• Thirteen obsolete Q36 score jobs (759843, 759860, 760174, 760180, 760185, 760187, 760194, 760201, 760206, 760208, 760215, 760216, and 760217) were terminally cancelled after each was proven held on already-failed August 14–15 dependencies. Every job had zero elapsed time and zero restarts. This released submission slots for Sai's independent benchmark shards without cancelling or changing any live Sai allocation.

• This is a one-seed, approximately 100M-token, iso-data short screen. It is not the frozen three-seed iso-data/iso-FLOP tournament and cannot authorize the 4B run. The user has authorized sub-4B training; actual 4B training remains prohibited pending smaller-scale real-benchmark evidence.

Current target

• Name: Sai
• Size: approximately 4B parameters
• Architecture: not selected; it must win the scale-gated tournament
• Reference model: Qwen/Qwen3.5-4B@851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a
• Final form: approximately 4B total parameters, dense, text-only
• Deployment: one checkpoint, one pass, one H100 or smaller inference tier
• Focus: English, code, math, science, technical reasoning
• Reasoning: direct and deliberate behavior in one model; no mandatory hidden-draft/revision call

The exact Qwen reference metadata is frozen in docs/SAI_PARENT_QWEN35_4B.json. No model weights were downloaded or restored while preparing that receipt. Qwen is a reference and fallback control, not a decision to inherit its full architecture.

Why this repository exists

Always-revise Shohin failed its first broad public test. Across HumanEval+, MBPP+, IFEval, MuSR, and CorrectBench, it scored 42.806% macro versus 54.022% for the original and 49.911% for an equal-compute control. The -33.201 pp MuSR and -20.839 pp CorrectBench regressions close mandatory revision as a route to general intelligence.

Sai starts from those negative results instead of hiding them.

The SAI shift

The earlier plan was too centered on DeepSeek-R1-era post-training. Reasoning distillation and RL with verifiable rewards remain useful later, but they cannot repair a weak base architecture or compensate for lost general capability.

Sai now treats the complete model stack as an empirical tournament. No paper, company, or fashionable mechanism is promoted directly into the 4B model. Every change must first beat declared iso-data and iso-FLOP controls at smaller scales and survive source-disjoint capability and retention tests.

The machine-readable plan is docs/SAI_FRONTIER_ARCHITECTURE_TOURNAMENT.json, and its rationale is documented in docs/SAI_FRONTIER_ARCHITECTURE_TOURNAMENT.md. The first executable CPU oracle and exact 48K scale geometries are documented in docs/SAI_MODEL_GENERATOR_CONTRACT.md and docs/SAI_48K_SCALE_GEOMETRIES.json. The exact no-training 100M comparison planner is specified in docs/SAI_100M_EXPERIMENT_PLAN.md. The tokenizer/data specialization workstream is reconciled with that ladder in docs/SAI_4B_SPECIALIZATION_RESEARCH_PLAN.md. The data-first source, mixture, curriculum, and promotion boundary is specified in docs/SAI_4B_DATA_MIXTURE_PLAN.md and docs/SAI_DATA_CURRICULUM_CONTRACT.md. The exact lossless 64K/48K/32K measurement boundary is frozen in docs/SAI_TOKENIZER_QUALIFICATION_CONTRACT.md. The high-upside conditional-compute thesis and its ordered kill tests are in docs/SAI_ADAPTIVE_COMPUTE_FALSIFICATION_PLAN.md. Its exact Gate-0 workspace mechanics and oracle evidence boundary are in docs/SAI_16_SLOT_WORKSPACE_CONTRACT.md and docs/SAI_ORACLE_SLOW_PATH_CONTRACT.md. Portable checkpoint/run replay and the truthful no-training performance boundary are in docs/SAI_COMPLETED_RUN_LINEAGE_CONTRACT.md and docs/SAI_WORKSPACE_PERFORMANCE_CONTRACT.md.

Sequence-mixer tournament

The core contest is not “Transformer versus one grand invention.” It is:

1. gated GQA as the conventional reference;
2. a Qwen-inspired 3 Gated DeltaNet : 1 gated full-attention hybrid; and
3. a Kimi-inspired 3 KDA : 1 gated MLA hybrid.

KDA, MLA, gated attention, QK normalization, positional treatment, and kernel efficiency are measured separately before they are combined. AttnRes, SiTU-GLU, Engram, and multi-token prediction are second-stage ablations, not assumptions.

Tokenizer and parameter reallocation

The 248,320-token Qwen reference vocabulary contains about 636 million tied embedding/output parameters at width 2,560. A 32K vocabulary would use about 82 million, freeing roughly 554 million parameters; 48K would free roughly 513 million. Those are hypotheses, not free gains.

Sai will compare 64K, 48K, and 32K English/code/math/science/technical tokenizers with byte fallback; 16K remains a stress-test only. The tournament separates two questions:

• does the tokenizer itself improve byte-normalized compression and capability with identical body geometry; and
• does reinvesting the saved parameter budget into depth or FFN capacity improve the fixed-total-parameter system?

Cross-tokenizer comparisons include an iso-data contrast matched by admitted UTF-8 bytes and an iso-FLOP contrast matched by the analytical and measured compute ledger. Multilingual fluency may be deprioritized, but arbitrary Unicode, identifiers, URLs, source code, math, and scientific notation must remain lossless.

Conditional memory and objectives

DeepSeek Engram's deterministic n-gram lookup is a compelling partner for a smaller vocabulary because it could move static phrase memory out of expensive neural computation. It remains an isolated post-tokenizer ablation. NTP plus one or two MTP heads is likewise tested independently; future-summary prediction is exploratory only.

Behavior-preserving skill learning

After a base architecture wins, Sai post-training may train a narrow adapter on verified, benchmark-decontaminated math, code, logic, science, technical, and instruction data. Every optimizer window also replays broad selected-base behavior. The candidate minimizes task loss plus frozen-base token KL. The equal-compute control executes identical forwards with KL weight zero.

Reasoning without compulsory verbosity

After the base architecture wins, post-training may use verified multi-teacher distillation and RL with verifiable rewards for math, code, formal logic, and tool use. Direct-response examples remain in the same mixture. Only an SFT checkpoint that survives the public gate may enter bounded outcome-based RL.

Long reasoning is a selectable inference mode, not a ritual imposed on every prompt. Fixed-direct and fixed-deliberate controls must show that adaptive compute actually helps.

Scale ladder

The generator must instantiate comparable models at approximately 100M, 300M, 1B, and 4B parameters. After an official training order, the sequence is:

• 100M: mechanics, stability, kernel, memory, and throughput qualification;
• 300M: three-seed factor screens on frozen development data;
• 1B: confirmation of only the surviving factors and interaction checks; and
• 4B: one selected stack, followed by the complete public gate.

The 4B run is prohibited until the smaller-scale evidence exists. This is the lesson from Shohin made executable: benchmark evidence chooses the architecture.

First public gate

The complete official HumanEval+, MBPP+, IFEval, MuSR, and CorrectBench boards are conjunctive. A candidate must:

• beat original and equal-compute macros by at least 1.0 point;
• remain within 1.0 point of both comparators on every benchmark;
• beat each comparator on at least four of five benchmarks; and
• be nonnegative against both comparators on MuSR and CorrectBench.

One serious regression vetoes a favorable average.

Build status

• retire always-revise compute and free every GPU request;
• encode the five-board benchmark gate and historical falsification;
• prototype frozen-parent replay KL with a matched zero-weight control;
• begin a lossless tokenizer-capacity auditor;
• implement deterministic, benchmark-disjoint freezing for skill, direct, deliberate, replay, and RL-prompt banks;
• replace the R1-centered plan with a verified 2026 architecture tournament;
• freeze the 100M → 300M → 1B → 4B promotion ladder and factor isolation;
• implement causal CPU reference mixers and exact parameter ledgers;
• freeze matched 48K geometries for all three families at every scale;
• prove packed-document isolation across attention, RoPE, convolution, and recurrent state in the CPU architecture oracle;
• implement an exact three-family, three-seed iso-data/iso-FLOP planner;
• implement deterministic binary token packing with exact UTF-8 prefix and cross-document boundary receipts;
• implement the exact 64K/48K/32K tokenizer qualifier and protected Unicode suite without building or selecting a candidate;
• define an oracle-first falsification ladder for latent workspace, fixed-point recurrence, regret gating, and sparse semantic memory;
• implement exact 16-slot workspace accounting, a bitwise fast bypass, and a row-level equal-FLOP oracle analyzer without training;
• make oracle evidence reopen portable checkpoint/run lineage and add a mutation-free CPU workspace performance receipt without claiming H100 speed;
• implement deterministic 64K/48K/32K tokenizer construction, exact stream loading, masked AdamW training, held-out NLL evaluation, and atomic resume;
• qualify GDN/KDA FLA chunk forward/backward mechanics and complete stable full-model optimizer updates for both frozen 100M delta-family geometries;
• implement replayable word/code benchmark decontamination and pin the FineWeb-Edu mechanics source prefix;
• run that freezer on the exact admitted source populations;
• qualify 64K/48K/32K tokenizer candidates on the admitted corpora;
• freeze matched data/FLOP/seed manifests for the 100M tournament;
• run the 100M mechanics tournament, then evidence-gated 300M/1B stages;
• package exactly one winning 4B architecture and matched controls;
• run all five complete public boards;
• promote only if every gate conjunct passes.

Repository layout

• src/sai/gates/ — real-benchmark promotion decisions
• src/sai/model/ — scalable configurations, parameter ledgers, and CPU oracles
• src/sai/data/ — verified role populations and contamination filtering
• src/sai/training/ — behavior preservation and reasoning training
• src/sai/tokenizer/ — vocabulary capacity measurement and surgery
• tests/ — fail-closed regression coverage
• docs/ — frozen contracts and experimental evidence

Historical Shohin evidence remains in GodlyDonuts/shohin-ettr. Sai-specific implementation and results live here.