← Complete research archive
Architecture researchTheory101 lines

Deferred Whole-Presentation Closure

The learned PSPA pilot exposed an optimization conflict. A row-local compiler learns every observed generator row, while end-to-end Sinkhorn projection prevents the same encoder from learning even those source facts. Global consistency is useful at execution but harmful during ea…

docs/research/DEFERRED_WHOLE_PRESENTATION_CLOSURE.mdOpen original Markdown ↗

Deferred Whole-Presentation Closure

Status: strong completion gain but causal gate failed; closed before confirmation.

Hypothesis

The learned PSPA pilot exposed an optimization conflict. A row-local compiler learns every observed generator row, while end-to-end Sinkhorn projection prevents the same encoder from learning even those source facts. Global consistency is useful at execution but harmful during early evidence fitting.

Deferred Whole-Presentation Closure (DWPC) separates those phases:

  1. learn source evidence with independently normalized generator rows;
  2. preserve the resulting local uncertainty through compilation;
  3. at the source/query boundary, project every complete generator to one indivisible permutation; and
  4. execute the late query only through that committed presentation.

No weights, data, source loss, or query supervision change. The only changed factor is when algebraic closure is imposed. The forward commit selects whole generator actions; it never averages fields from incompatible presentations.

This resembles deferred projection, proximal commitment, and neural-symbolic execution. Novelty is not claimed. The potentially useful architectural principle is plastic local evidence first, discrete global closure once, instead of differentiating through the global closure during every update.

Evidence and controls

The frozen source is the failed seed-47 learned-PSPA checkpoint. Its matched row-soft compiler reaches 100% accuracy on observed rows but violates global permutation closure. The jointly Sinkhorn-trained compiler remains near the uniform floor and is the negative control.

The exploratory 128-row-per-cohort read-only diagnostic gives DWPC answer exactness of 52.3/49.2% cyclic, 75.0/75.0% dihedral, and 54.7/43.0% random permutation at word lengths 8/12. This is roughly 58.2% macro, versus 25.8% for unclosed row-soft execution and 9.15% for joint Sinkhorn training.

The complete development evaluation uses all 1,024 rows in each frozen cohort and must report:

  • DWPC, row-soft, and jointly projected answers;
  • source-challenge exactness;
  • complete-table exactness;
  • shuffled-challenge outcomes; and
  • whole-presentation lineage swaps.

The development signal authorizes one unchanged-weight confirmation only if DWPC beats row-soft by at least 10 macro points, gains on every family, and both interventions cost at least five points. Confirmation uses new episode and renderer seeds and requires at least 50% OOD macro, positive gain on every family, and at least ten-point losses under both interventions. Failure closes DWPC without retraining this checkpoint.

Resource envelope

Development and confirmation are evaluation-only. Each is one single-H100 job with a 20-minute ceiling and expected use below 0.1 H100-hour. No training or long pretraining follows unless unchanged weights pass confirmation.

Full development result

Evaluation job 739360 uses the unchanged failed learned-PSPA checkpoint and all 1,024 examples in each of the six frozen OOD cohorts.

Arm or diagnosticSix-cohort OOD macro exact
DWPC58.643%
Unclosed ROW_SOFT25.798%
Jointly projected PRESENTED9.147%
DWPC with shuffled challenge outcomes58.561%
DWPC with whole-presentation lineage swap11.865%
DWPC source-challenge exact71.407%
DWPC complete-table exact40.218%

DWPC gains 32.845 points over its own unclosed checkpoint and improves every family. Cyclic length-8/12 is 55.273% / 49.414%, dihedral is 75.977% / 76.562%, and random permutation is 52.637% / 41.992%. Whole-lineage swapping removes 46.777 points, proving that coherent committed tables cause the answers.

The challenge intervention is decisive in the opposite direction: exchanging all source-challenge outcomes changes macro by only 0.081 points. DWPC is not using counterexamples. It parses observed rows, closes each permutation, and guesses the unresolved completion. The approximately 50% cyclic table score and 13--14% random three-generator table score match that interpretation.

This misses the fixed causal rule, so confirmation remains unopened and DWPC closes without retraining. The reusable result is the phase separation: evidence-first local learning followed by one whole-object commit is far more trainable than differentiating through closure. The successor must add explicit counterexample-conditioned whole-presentation selection at that single commit, not another continuous projection.

Report SHA-256 is 17e4e0da326bf659d89a80804caf3793ac285c064acc7bf671070f3954fe31d2.