State-Formation Branch Handoff Recovery

Results at a glance 1

Overview

This operational successor lets immutable producer v11 continue after the first branch recovery successfully replaces the retired failed-G0 receipt at its canonical path. It changes no scientific arm, model, data, metric, threshold, or branch rule.

The handoff question is answered and the setup matrix has passed. Continued Stage-B training through this frozen wrapper belongs to the canonical producer and does not make this recovery experiment active.

Research Program

• Program: structured_execution_and_compilers.
• Direct producer: qwen35_4b_state_formation_capacity_adjudication.
• Closest experiment: qwen35_4b_state_formation_branch_recovery, whose exact-prefix seam produced the valid full-rank seed-7411 G0 but whose frozen retirement guard treats any later occupant of that pathname as the retired failure.

Question

Can downstream producer stages continue without weakening failure preservation once a successful recovered G0 legitimately occupies the canonical pathname formerly used by the archived failure?

Hypothesis

The defect is limited to invocation orchestration: the first wrapper asks whether the pathname exists, not whether it still contains the exact retired failure. A new wrapper can reuse its frozen, tested path seam while requiring the archived failure and retirement receipts, rejecting a reappearing mirror or exact failed bytes, and accepting only the exact successful G0 plus its STARTED/COMPLETE lineage.

Setup

• Model-bearing stages remain only Qwen/Qwen3.5-4B revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a, on producer v11's Transformers backend.
• Smoke loads no model and runs no training or evaluation.
• Successful G0: file SHA-256 cdc90cd…c68f, identity e1f1c906…f89dc, status MODEL_SMOKE_PASS, full-rank joint seed 7411.
• Retired failure: SHA-256 47305826…2c71; it must remain in the first recovery archive and must not reappear at either producer failure path.
• First recovery: source contract 55d0a455…56f3, smoke SHA-256 8bf5bb36…6849.
• Allowed stages and argument shapes remain the first recovery's exact registered set. The producer CLI retains canonical path, source snapshot, authorization, setup/control, training, evaluation, lock, and receipt gates.
• Hidden-label boundary: smoke opens no contrast or benchmark payload. The wrapper cannot analyze.

Recovery boundary

The wrapper imports the exact frozen first-recovery source and reuses its safe reads, immutable writes, producer-context checks, argument-shape policy, stage-specific output validation, and src.analysis._canonical_expected_path context manager. It does not call the first wrapper's pathname-only invoke_producer; it owns new immutable STARTED/COMPLETE receipts and directly calls the same exact producer CLI under the same path seam.

Every invocation first rehashes the retirement, archived failure, successful G0, first-recovery G0 STARTED/COMPLETE receipts, authorization, first-recovery smoke/source, producer source/model, and the handoff's own frozen smoke/source. A producer output cannot be adopted unless its handoff STARTED receipt already exists.

Run

Freeze and smoke without loading the model:

.venv/bin/python -B -m unittest discover \
  -s experiments/qwen35_4b_state_formation_branch_handoff_recovery/tests -v
.venv/bin/python -B \
  experiments/qwen35_4b_state_formation_branch_handoff_recovery/scripts/run.py --smoke

Commit and push the successful G0, its first-recovery invocation receipts, and this frozen handoff; require both repository workflows green. Then retry the seed-7411 positive control through this wrapper. Every later branch-authorized model stage uses the same handoff wrapper; producer analysis continues through the already-frozen analysis recovery.

Results

The focused suite passes 8/8. The frozen smoke emitted BRANCH_HANDOFF_RECOVERY_SMOKE_PASS at file SHA-256 0e0409de…6e5b, receipt identity d6fecf0a…de4c, and source contract 4d2ffde3…8acd. All six controls pass: the first wrapper's pathname-only false rejection is reproduced; retirement, failure archive, successful G0, and its invocation lineage are exact; and the retired mirror remains absent. Unit controls additionally reject exact failed bytes at the successful slot, any changed successful-G0 bytes, a reappearing mirror, and unregistered invocation shapes.

Smoke loaded no model, started no training/evaluation, opened no benchmark or contrast path, and changed no scientific interpretation. No positive-control or downstream scientific result is claimed here. At that checkpoint, use remained blocked until publication and both workflows were green; the later paragraphs record that completed progression.

Handoff commit 627254f1…d892 passed both workflows. The exact full-rank seed-7411 positive control then completed once through this wrapper. Producer file SHA-256 is a0d17e2e…7a16, producer receipt identity 6708a4d4…9649, and handoff COMPLETE identity af0dd15c…2262. The producer control passed oracle and intact joint state at 48/48, adaptation-disabled joint state at 0/48, exactly 256 updates, gradient accumulation 16, and 4,096 singleton presentations. Both the 892,272,640-parameter full- rank path and shared-state parameters changed; early stopping and checkpoint selection were false. The receipt authorizes later producer work only through its unchanged barriers and records zero result payload, benchmark, contrast, or scientific-evidence access.

After seed-7411 setup commit c507488c…3c06 passed both workflows, seed 7412 independently passed the same path. Its G0 file SHA-256 / identity are 10bf22fc…1d18 / 62ecb79e…951dc: zero K=1 and checkpoint-roundtrip error, all 62 full-rank deltas with complete finite optimizer state, all joint trainable groups finite/nonzero, finite K=12, and 22.1 GiB free. Its positive-control file SHA-256 / identity are 1cbbd823…8510 / 6575e1d2…4554: oracle/intact 48/48 and disabled 0/48 after exact 256 updates, accumulation 16, and 4,096 presentations. Handoff G0/control COMPLETE identities are d61d6441…9246 and 6334d72f…c14a. Both receipts retain zero result, benchmark, contrast, and scientific-evidence access.

After seed-7412 setup commit 92473a53…8530 passed both workflows, seed 7413 completed the matrix. Its G0 file SHA-256 / identity are 021a8444…d635 / 4d2316d3…2ff0: zero K=1 and round-trip error, finite/nonzero joint gradients, complete finite optimizer state for all 62 deltas, finite K=12, and 22.1 GiB free. Its control file SHA-256 / identity are 8a4af0d6…fde8 / 976f28ef…14df: oracle/intact 48/48 and disabled 0/48 after exact 256 updates, accumulation 16, and 4,096 presentations, with changed full-rank/shared state. Handoff G0/control COMPLETE identities are fbfb282f…5aef / ccff1520…58c9. All three pairs now pass, remain setup-only, and open no result, benchmark, contrast, or scientific evidence.

Interpretation

The three valid G0/control pairs establish full-rank feasibility and a working state path at every seed, not a capacity result. This recovery safely handed producer authority through setup without changing scientific logic. Capacity interpretation remains unchanged.

Knowledgebase Update

• Program evidence: no scientific evidence from an invocation repair.
• Program backlog: producer Stage B now owns the mandatory LoRA-state-only and full-rank-joint work.
• Claim ledger: unchanged.

Artifacts

• idea_intake.md: nearest-neighbor and novelty decision.
• reports/design_review.md: adversarial review and stop conditions.
• configs/default.yaml: exact producer, recovery, failure, G0, and authorization pins.
• runs/smoke.json: frozen no-model handoff smoke.
• runs/invocations/: handoff-owned immutable producer invocation provenance.
• reports/artifact_manifest.yaml: producer-owned large-artifact boundary.

Report

Rendered from reports/report.md

Status

Finished. The first recovery produced a valid full-rank seed-7411 G0, but its frozen pathname-only retirement guard stopped the following positive control before producer or model work. This handoff then carried all three full-rank G0/control pairs through the unchanged producer barriers. Producer receipts remain scientifically authoritative.

Method

Pin the archived failure, retirement, successful G0, first-recovery G0 invocation lineage, first- recovery source/smoke, producer source/model, and authorization. Reuse the exact frozen path seam while replacing only outer invocation orchestration and emitting handoff-owned STARTED/COMPLETE receipts.

Results

• Focused tests: 8/8.
• Smoke status: BRANCH_HANDOFF_RECOVERY_SMOKE_PASS.
• Smoke file SHA-256 / identity: 0e0409de…6e5b / d6fecf0a…de4c.
• Frozen source contract: 4d2ffde3…8acd.
• Six smoke controls pass; unit controls reject failed bytes, changed G0 bytes, a reappearing mirror, and unregistered invocation shapes.
• Model load, training/evaluation, benchmark/contrast access, and scientific interpretation: zero.
• Full-rank seed-7411 positive control: producer SHA-256 a0d17e2e…7a16, identity 6708a4d4…9649; oracle/intact 48/48, adaptation-disabled 0/48, 256 updates, accumulation 16, 4,096 presentations, and changed full-rank/shared-state parameters.
• Handoff COMPLETE identity: af0dd15c…2262; producer result/benchmark/contrast access and scientific evidence remain zero.
• Full-rank seed-7412 G0: file SHA-256 10bf22fc…1d18, identity 62ecb79e…951dc; zero K=1/ round-trip error, complete finite optimizer state for all 62 deltas, finite K=12, 22.1 GiB free.
• Full-rank seed-7412 control: file SHA-256 1cbbd823…8510, identity 6575e1d2…4554; oracle/intact 48/48, disabled 0/48, 256 updates, accumulation 16, 4,096 presentations, changed full-rank/shared state, and zero result/benchmark/contrast/scientific access.
• Full-rank seed-7413 G0: file SHA-256 021a8444…d635, identity 4d2316d3…2ff0; zero K=1/ round-trip error, complete finite optimizer state for all 62 deltas, finite K=12, 22.1 GiB free.
• Full-rank seed-7413 control: file SHA-256 8a4af0d6…fde8, identity 976f28ef…14df; oracle/intact 48/48, disabled 0/48, 256 updates, accumulation 16, 4,096 presentations, changed full-rank/shared state, and zero result/benchmark/contrast/scientific access.
• Complete matrix: all three independent full-rank G0/control pairs pass; this remains setup-only.

Interpretation

No capacity conclusion changes. A passing recovery licenses only the preregistered producer stage that its inputs independently authorize.

Next action

No recovery action remains. The complete setup matrix was published and the producer's Stage-B cells are tracked in qwen35_4b_state_formation_capacity_adjudication; their scientific barriers remain unchanged.

Experiment log 7

Reproduce

Smoke test

.venv/bin/python -B experiments/qwen35_4b_state_formation_branch_handoff_recovery/scripts/run.py --smoke

Full run

run only an exact branch-authorized producer command documented in the producer GPU runbook through scripts/run.py

Run steps are documented inside the experiment folder (README and scripts).

Browse the experiment folder on GitHub ↗

Related

Related experiments

Nearest by shared name, tags, and programs.

• 2026-07-14 State-Formation Branch Authorization Recovery
• 2026-07-13 → 14 State-Formation Analysis Recovery
• 2026-07-13 State-Formation Capacity Adjudication
• 2026-07-14 Natural-Language State-Table Universal Curriculum
• 2026-07-12 Public-verifier recovery branch tournament
• 2026-06-23 Qwen Prefix-State Process Verifier