Shohin IDR+AQC Deployable Release

Status: qualified architecture packaged and exercised end to end on 2026-08-09. This is a deployment result, not a new benchmark or causal-mechanism result.

System

The strongest measured Shohin system is now represented by one immutable delta release over pinned Qwen/Qwen3.5-9B revision c202236235762e1c871ad0ccb60c8ee5ba337b9a:

1. draft_adapter.pt produces one complete model-owned internal draft;
2. revision_adapter.pt reads the source and draft and emits a trained revision;
3. draft_adapter.pt produces the matched unchanged continuation from the same source and draft; and
4. commit.pt reads both complete candidates and selects one coherent trajectory.

No external proposal model, verifier, correctness signal, benchmark label, task router, or tool is used at inference. The complete path is implemented by train/hf_idr_aqc_interact.py; immutable packaging is implemented by pipeline/package_idr_aqc_release.py.

Qualified evidence

The release binds the existing protected product evidence:

The learned commit is useful, but the antisymmetric treatment beat its matched independent-score control by only one answer. The release therefore claims coherent learned commitment, not a uniquely causal antisymmetric primitive.

Immutable lineage

• private source commit: 8f0bd8d;
• complete runtime: idr_aqc_8f0bd8d_r2;
• runtime SHA256SUMS SHA-256: 2d4e8fd845bdd023fca5a01c767ed218909ab810fa4e06b25d2888f90f3388c5;
• release: idr_aqc_release_8f0bd8d_r1;
• release manifest SHA-256: 554e841f71edd3a19063411348340e337532db2db05dd5e1e2adc25a3d347e7b;
• release SHA256SUMS SHA-256: 0dad031312dec0859e35bb7e9daea8aef688ef350b9053f587fba5acdc9c58c5;
• interaction prompt SHA-256: 980f025b3da68fc6335ceca283afd5737c3c0577b276102533609fe7c49cf1f3.

The packager verifies the base config, both adapter checkpoints and reports, the trained revision's draft warm start, the qualified commit checkpoint and report, and the protected product report before writing. The runtime verifies every packaged file and the external base config before model load. It refuses unqualified commits, hash drift, incomplete file coverage, output overwrite, candidate truncation, and order-inconsistent commitment.

End-to-end smoke

Job 747423 completed on one H100 in 3m26s. The atomic interaction report is smoke_report.json, SHA-256 47e1bd3ab2bd81132de5f40e038f553eebcd210c436fb01c12acfeb9b2f7171b. The five prompts are original non-benchmark interactions covering arithmetic, modular composition, counterfactual ordering, thermodynamics, and Python.

Commit order consistency is exact and maximum swap error is 0.0. No commit candidate was truncated. One internal draft exhausted its 768-token budget, but it had already derived and verified the correct answer.

Manual transcript inspection finds correct final substance on all four math/logic/science prompts. The code attempt identifies the central predicate-twice defect and proposes a one-pass stable algorithm, but the selected response includes prose and omits the requested standalone function definition, so it fails that interaction contract. The commit chooses the unchanged continuation on all five prompts. This smoke therefore establishes that the packaged architecture executes faithfully; it does not add evidence that commitment improves arbitrary out-of-distribution prompts.

Two preserved pre-result failures preceded the valid run. Job 747421 failed before model load because the first runtime archive was incomplete. Job 747422 executed every model stage but could not write into the intentionally read-only release directory. Both failures changed only packaging/output location; job 747423 used the same model artifacts, prompts, generation budget, seed, and selection policy. Total charged wall time across all three jobs is approximately 0.126 H100-hours.

Boundary and next use

This release makes the strongest existing Shohin reasoning architecture reproducible and directly inspectable. It remains a multi-pass 9B system, not the 125M scratch checkpoint and not a newly proven MoE mechanism. Future work should use this package as the positive dense control. It should not tune the commit policy on these five qualitative prompts or reinterpret this smoke as a benchmark.