Counterexample-Selected Deferred Closure

Status: development and unchanged-weight confirmation pass; typed-source reasoning baseline established.

Hypothesis

DWPC proves that local source facts can be learned and that a single whole-presentation commit makes long composition work. It fails because the commit greedily chooses one completion without using source counterexamples.

Counterexample-Selected Deferred Closure (CSDC) preserves the successful phase separation and changes only the commit:

1. train row-local source evidence without global projection;
2. identify each generator's two least-certain rows;
3. construct the two whole permutation completions for each generator;
4. form at most eight complete multi-generator presentations;
5. execute every source-owned challenge against every presentation;
6. select the one complete presentation with the fewest contradictions; and
7. execute the late query through that selected lineage only.

No query, answer, hidden target, external model, or tool participates in candidate construction or selection. The frozen seed-47 ROW_SOFT compiler is unchanged. This gate isolates whether explicit source-counterexample falsification closes the final ambiguity that DWPC guesses.

The first gate consumes typed challenge fields. It is not yet a free-form language parser claim. The rendered-language compiler has already shown 100% observed-row parsing; a later bridge must recover typed challenge fields from ordinary text. This test focuses on the reasoning operator.

Fixed controls and gate

• CSDC: confidence-derived whole candidates plus challenge selection;
• DWPC: one greedy whole closure with no challenge selection;
• ROW_SOFT: no closure;
• SHUFFLED_CHALLENGE: CSDC with outcomes exchanged across episodes; and
• LINEAGE_SWAP: CSDC with selected complete presentations exchanged before the late query.

Development uses all six existing 1,024-row OOD cohorts. CSDC advances only if it reaches at least 90% macro exact, improves every family by at least 20 points over DWPC, recovers at least 95% of complete presentations and source challenges, and loses at least 20 points under each intervention. A passing development result authorizes one unchanged-weight confirmation on new episode and renderer seeds, requiring at least 85% macro and the same causal directions. Failure closes this candidate constructor without K, confidence, scoring, or duration tuning.

Resource envelope

This is evaluation-only. Development and any confirmation each use one single-H100 job with a 20-minute ceiling and expected use below 0.1 H100-hour. No training or Shohin integration follows from an incomplete causal pass.

Development result

Job 739370 evaluates all six frozen 1,024-row OOD cohorts.

CSDC is 100% exact on both cyclic cohorts, 99.609% / 99.023% on dihedral length 8/12, and 99.316% / 99.512% on random permutation length 8/12. It clears every fixed threshold. Report SHA-256 is 45fd76fcc7702e19a79799c939e318db44a5bb4df84e28861f562ceed17101f4.

Unchanged-weight confirmation

Job 739373 uses new episode seeds (82000+) and renderer seeds (83000+) without changing the checkpoint, candidate constructor, thresholds, or selection rule.

Every confirmation family and depth is at least 99.316% exact. Shuffled challenges remove 47.347 points and lineage swaps remove 85.807 points. The effect therefore replicates and depends on both counterexample outcomes and one coherent selected lineage. Report SHA-256 is c0c565b3f35d7fde306aad955d15891a0e324ada55926cb39db7cb122f798d3a.

Claim boundary and next gate

CSDC is a genuine synthetic causal reasoning result: a learned local compiler constructs whole executable hypotheses, source counterexamples select among them, and unseen long queries execute through the selected lineage. It is not yet general natural-language reasoning. The challenge start, generator word, and outcome enter the selector as typed source fields; only the observation table compiler is learned from rendered records.

The next gate must learn the rendered-language-to-typed-challenge bridge while freezing this selector and executor as the protected oracle ceiling. It must stay within five points of CSDC, preserve both causal intervention effects, and transfer to unseen renderer templates. Failure of the bridge does not invalidate the reasoning operator; it localizes the next problem to semantic compilation.