Query-Valued Evidence Sparse Revision

Status: closed negative after one bounded credit-assignment pilot.

Hypothesis

CGSGR proved that counterexample-conditioned state revision is active, but it lost to fixed evidence coverage because contradiction magnitude is not answer utility. Query-Valued Evidence Sparse Revision (QVESR) keeps the same coherent state, consequence model, sparse write budget, and recurrent revision operator. Its only substantive change is a learned value function over each source-owned evidence item:

u_j = Value(S_t, evidence_j, residual_j, query, visit_count_j)
J_t = HardTopK(u)
S_(t+1) = SparseRevise(S_t, evidence[J_t])

The forward pass is hard top-k and admits exactly two evidence items per round. A straight-through probability path lets final-answer loss train the selector. No target answer, hidden state, tool, teacher, or external verifier is available to the selector at inference. The query is intentionally available during deliberation because the failed hypothesis is precisely that query-agnostic local error predicts answer value.

Prior-art boundary

Learned attention, query-conditioned retrieval, sparse top-k routing, straight-through estimators, iterative refinement, and value functions all have substantial prior art. QVESR is not presented as a novel ingredient. It is the fastest discriminating test of the credit-assignment failure exposed by CGSGR. A broader claim would require a capability win and a more specific separator from those families.

Matched pilot

Three arms instantiate and execute the same modules and parameter count:

• UTILITY: hard top-k learned query-conditioned evidence value;
• FIXED: cyclic evidence coverage;
• RESIDUAL: largest current consequence residual, reproducing CGSGR's rule.

All use seed 29, 1,000 updates, 256 examples/update, four recurrent rounds, two probes/round, two revised state slots/round, and the frozen depth-5/7 noncommuting, repeated-binding, and induction cohorts. The pilot advances only if UTILITY:

1. beats FIXED by at least five absolute macro points;
2. improves the mean of both depths in every family;
3. loses at least three macro points when selected evidence outcomes are shuffled;
4. loses at least three macro points when the selector alone receives a row-shuffled query; and
5. retains exact hard evidence and sparse-write budgets.

A miss closes QVESR and the nearby sparse-revision selector family. It does not authorize another duration, width, seed, temperature, loss, or top-k variant.

Result

The first dispatch (739242--739244) failed before training because the immutable runtime omitted one imported support module. It produced no reports or checkpoints. The corrected fail-fast runtime has SHA256SUMS SHA-256 7c2b11bac068ce8cc47273ecf9719c00d4c10a1f637d5ff70c2ff0a553c9ef3d.

Valid jobs 739246--739248 complete with identical 120,983-parameter models:

UTILITY loses to FIXED by 0.277 points and fails the all-family requirement. Shuffling selected outcomes lowers UTILITY to 10.726%, confirming that the revision still depends on evidence. Shuffling only the selector query lowers it by merely 0.423 points to 17.741%; the learned policy does not acquire material query-specific value assignment. Contradiction reduction is 0.608 / 0.288 / 0.779 for utility/fixed/residual, again showing that better local correction does not imply better final composition.

QVESR and nearby sparse evidence-selector variants are closed. The next architecture must change how structured state transitions and composes, not retune evidence routing into this revision operator.

Report SHA-256 values:

• utility: d50da2a6f7bdee8fb6ae951faeb2e1a1f18672ee0a35293fd04963d6bdc61f82
• fixed: 7b895de5751b3c01486ee734ecd2d86520617cef1eedf3a9a0ffde023630c335
• residual: 5cd99aefc42c1d4fdfc37ad6d49bb3d1121e6f4175b33bf8a1cb33dfc44ec7d3