Shohin ECTR0 Executor-Conditioned Temporal Revision

Status: closed negative on development; no training, holdout, or public output was opened.

Question

CTF1 showed two simultaneous facts on the same 666 source-disjoint GSM8K development identities:

• the untouched Qwen3.5-4B owner produces a useful canonical transaction trajectory whose learned execution solves 419/666 rows causally; and
• forcing every answer through that trajectory loses capability relative to the owner's own direct answer (487/666).

ECTR0 tests the structurally different composition: preserve the complete owner trajectory and learned execution as evidence for an already-qualified later revision owner, then let that owner emit a new complete natural solution. The learned executor is advisory evidence, not the final answer path.

Frozen Owners And Inputs

• Draft/transaction owner: exact immutable CTF1 normal report SHA-256 dc4e939b8186393ad6827f6cecbeacaaf86231d2abf2682ce093d43364b905f0.
• Source board: exact CTE1 development SHA-256 aff466172c74dd7d13a183d117e32ec10d5da2048d10253189e4e3b3a599eb04.
• Revision owner: qualified IDR4 Qwen3.5-4B checkpoint SHA-256 ae3847fe0728b1debcc13049822ea7499f744836b62d6d1c5bcb7c1000d8560b.
• Backbone: Qwen/Qwen3.5-4B@851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a.
• Decoding: greedy, no thinking mode, 512 new tokens, batch four, seed 2026081061, 4,096-token context, no training.
• Public GSM8K test and every other protected board remain sealed.

The revision prompt uses the qualified IDR source-plus-internal-draft format. The internal draft contains the exact CTF1 owner completion. Depending on the arm it then contains either the aligned learned-executor receipt, an unavailable receipt marker, or a deterministic receipt from another row in the same compile-status/register-depth stratum. No gold answer, verifier output, or correctness label is visible.

Arms

1. aligned: exact owner trace plus its own learned-executor receipt.
2. receipt_absent: exact owner trace plus <RECEIPT_UNAVAILABLE>.
3. receipt_shuffled: exact owner trace plus another identity's receipt, deterministically rotated within compile-status/register-depth strata.

All arms use identical source, owner completion, revision weights, generation budget, evaluator, and identity order. Only the executor receipt differs.

Prospective Gates

ECTR0 qualifies only if all conditions hold:

1. aligned revision reaches at least 500/666, thirteen answers above the frozen direct-owner result;
2. aligned exceeds both receipt-absent and receipt-shuffled by at least 13 answers;
3. relative to the direct owner's claimed final answer, aligned semantic repairs minus semantic breaks are at least 13;
4. at least 650 completions contain an explicit final answer;
5. every identity appears exactly once, no prompt is truncated, all protected hashes match, and all three arms complete under the same evaluator.

A miss is evidence that the qualified reviser does not already consume this executor interface. It does not reopen CTE1/CTF1, alter their closed results, or justify a nearby prompt/threshold/checkpoint retry. Receipt-specific training may be considered only as a separately frozen mechanism with a same-source counterfactual objective that makes the receipt identifiable.

The direct-owner reference uses CTF1's exact last #### numeric claim. A trailing number without that marker is not a completed direct answer.

Claim Boundary

A pass would show that model-owned trajectory plus learned execution can causally improve a trained same-family temporal reviser without changing any weights. It would not prove that canonical transactions alone outperform direct generation, nor that the executor is infallible, nor that the result transfers beyond this source-disjoint development board.

Frozen Result

Exact replay jobs 750163--750168 completed all three arms in two shards each. The six valid allocations used 947 H100-seconds (0.2631 H100-hours), including model staging and load. Every row emitted an explicit final answer, no output exhausted 512 tokens, and the largest prompt was 946 tokens under the 4,096 ceiling.

Aligned is eight answers above shuffled, but three below receipt-absent and eleven below the direct owner. It fails the >=500, both +13 causal-margin, and positive repair-minus-break gates; only explicit-final/custody passes. The current qualified reviser therefore does not turn this learned-executor receipt into useful incremental capability. Exact ECTR0 closes without a prompt, checkpoint, decoding, or threshold retry. Aggregate SHA-256 is 869ed7412ec30f7f4095b65fd9370413cdc7633e9ca09c41a94106fc148f0cac.

Read-Only Attribution

A hash-bound paired attribution over the same six reports confirms that the receipt changes behavior without supplying a reliable arbitration rule. Aligned versus receipt-absent helps five rows and hurts eight; aligned versus shuffled helps fourteen and hurts six. Direct and executor predictions differ numerically on 137 rows. On those rows, aligned revision emits the direct prediction 109 times, the executor prediction once, and a third prediction 27 times. Across all rows, aligned emits the direct prediction on 563/666.

The direct/executor oracle is only 498/666, eleven above direct alone. This both localizes the interface failure and bounds the value of a hard selector: the qualified reviser mostly preserves the direct claim, and the few receipt- induced deviations are not reliably beneficial. The attribution cannot rescue the closed gate. Its report SHA-256 is 630830408e051c095a77f216b422dcce2564185d04a7566cec33fbc616fa3cbe; see SHOHIN_ECTR0_ATTRIBUTION_RESULT.json.