NagaYu/dissent-synthetic-clause-ambiguity

Dissent — synthetic financial clauses with seeded formalization ambiguity

Code · Interactive Space

Every clause in this dataset is synthetic. No real contract, client document or third-party text was used, quoted or paraphrased. The language follows standard market forms of drafting so that it is representative of the constructions that cause real formalization disputes.

What this is for

Autoformalization research is crowded at the producing end and empty at the checking end. This dataset supports the checking end: 15 clauses carry a deliberately seeded ambiguity — an axis the text leaves open, annotated with the phrase responsible, the reading the drafter intended, and a fact pattern that separates the readings — and 13 are drafted to be watertight controls, so that false-positive rate is measurable and not just detection rate.

Seeded ambiguity types include: and/or attachment, scope of trailing exclusions and provisos, unstated measurement periods, quarter-end versus at-any-time measurement, calendar versus business day counts and inclusive/exclusive boundaries, what an undefined ratio denominator contains, per-item versus aggregate tests, concentration and aging reference points, within-tier allocation in a payment waterfall, and credit-support (CSA) delivery/return mechanics.

Configurations

config	rows	what it holds
clauses	28	clause text plus the evaluation answer key (seeded axis, gold span, intended reading, distinguishing fact pattern)
formalizations	140	every independent formalization, with stance, executable DSL, decision log and SMT-LIB
disagreements	28	per-clause outcome: status, distinct readings, minimized witness, implicated phrases

Reference results

Produced by Dissent v0.1.0 with N=5 independent formalizations. Try it interactively in the Space:

measure	value
seeded-ambiguity detection rate	100.0% (95% CI 79.6-100.0%, n=15)
false-positive rate on watertight clauses	0.0% (95% CI 0.0-22.8%, n=13)
disputed phrase in the top-1 span	93.3% (95% CI 70.2-98.8%, n=15)
disputed phrase in the top-3 spans	100.0% (95% CI 79.6-100.0%, n=15)
clauses machine-confirmed consistent	13 of 28
mean deviating variables per counterexample	see the disagreements config

How to use it

from datasets import load_dataset

clauses = load_dataset("NagaYu/dissent-synthetic-clause-ambiguity", "clauses", split="train")
ambiguous = clauses.filter(lambda r: r["is_ambiguous"])
print(ambiguous[0]["seeded_axes"][0]["question"])

To score a method of your own: flag clauses, and compare against is_ambiguous for detection and false-positive rates, and against seeded_axes[*].gold_spans for localization. A method that flags everything scores 100% detection and 100% false positives; both numbers have to be reported together.

The shared data dictionary

Formalizations are written against one shared vocabulary of 65 variables (a firm's reporting schema, in effect), so that disagreement between two formalizations is a statement about the clause rather than a naming clash. Machine claims are made under an explicit background theory; its assumptions are:

• Trailing-twelve-month EBITDA is at least the current quarter's EBITDA (equivalently: the prior three quarters do not sum to a loss).
• Trailing-twelve-month Interest Expense is at least the current quarter's.
• EBITDA is positive; clauses are not evaluated on negative-EBITDA quarters, where a ratio covenant is conventionally treated as breached outright.
• Debt Service for the trailing twelve months is positive.
• Debt Service for the fiscal quarter is positive.
• Quarterly scheduled principal is part of the trailing-twelve-month figure.
• Days since the due date never exceed days since the invoice date.
• Stated payment terms are at most 90 days.
• A single Obligor's share of Receivables is at most that of the Obligor together with its Affiliates.
• Business days elapsed never exceed calendar days elapsed.
• Within a two-week window, calendar days exceed business days by at most 4.
• An Event of Default can only be cured if it occurred.
• A Default can only be cured if it occurred.
• The largest single unpaid amount is at most the aggregate unpaid amount.
• Asset Sale costs and taxes do not exceed gross proceeds.
• Amounts reinvested within a later-starting window are a subset of the earlier one.
• Reinvested amounts do not exceed the proceeds received.
• Restricted Payments this year are part of Restricted Payments since closing.
• Some amount is due on the Senior Notes on a Payment Date, so a pro rata allocation is well defined.
• Administrative Fees already paid have never exceeded the cumulative cap.

Limitations

• Synthetic by construction. Detection and false-positive rates measured here describe behaviour on this corpus and are not a guarantee on any real document.
• The corpus is deliberately about half seeded-ambiguous so both rates are measurable. That is not an estimate of ambiguity prevalence in real agreements.
• Non-computational obligations (notice mechanics, governing law, consent rights) are out of scope.
• Clause "pages" describe a notional 312-page agreement and exist to make triage numbers concrete.

Disclaimer

Dissent is a research prototype. It does not give legal advice and it does not certify that any clause has been interpreted correctly. Its only output about a clause is either "independent formalizations of this clause disagree, here is where and here is a fact pattern that separates them" or "independent formalizations of this clause were proved to agree under the assumptions listed in this report". Neither statement is a legal conclusion. All contract text used in development is synthetic.

Citation

@software{dissent_0_1_0,
  title  = {Dissent: finding formalization errors by searching for disagreement},
  year   = {2026},
  note   = {Research prototype. Synthetic contract clauses only. Not legal advice.},
  version = {0.1.0}
}