Date: 2026-10-05. Three new independently owned research candidates; none automatically promoted.

The locked protocol compares KL preservation, supervised replay proposal projection, and projection after native QA preparation. All arms share initialization and source-component fit/selection/calibration partitions. The61licensed QA examples retain native labels; no yes/no/maybe-to-stance conversion occurs.

Ownership and scientific limits

Each artifact has 682,008new adapter parameters plus2,307native QA-head parameters. The unused QA head in no-preparation arms is stored but not trained. The preparation arm trains both owners during QA preparation, then freezes the QA head during stance fitting. Named stage-specific optimizer scopes, primary trace/data hashes and immutable parent hashes are in every adjacent manifest.

Projection enforces one first-order supervised replay half-space on the actual Adam proposal; it does not guarantee finite-step loss preservation or unseen retention. The small science pilot does not reproduce MultiVerS domain preparation. New-task improvement without retention/confidence is rejected. Two stages of one adapter do not establish successive distinct agent acquisitions.

Matched candidate union-oracle balanced accuracy: 76.43%; all three miss 39of183 cases. This is a diagnostic upper bound, not a deployable policy. Paired gained/lost case IDs and balanced differences are preserved in matched-errors.json; one seed does not establish statistical superiority.

Development was already opened and is diagnostic. The187held-out pairs remain unscored. Temperature is fitted only on the 46 training-calibration records; raw confidence gates are unchanged. Native QA post-stance scores concern training examples, not unseen QA ability.

Protected-artifact integrity: {"allowed_new": ["[retained internal evidence] "[retained internal evidence] "[retained internal evidence] "passed": true, "protected_artifacts": 196, "unchanged": 196}. Fresh-process restoration: 3arms, all passed; maximum probability differences recorded in restored.json. Focused tests must pass before this report is written. Original training logs and bookkeeping failures are preserved; the post-fit metadata closure never refits a model.

If an arm manifest marks evaluation_recovered_without_refitting, its elapsed time covers recovery evaluation only; the interrupted original fitting duration is unavailable. The original checkpoint and logged selection history were reused. The missing science-preparation arm was fitted once. Original interruption and failed closure logs are retained; no cause is asserted without evidence.

Individual artifacts

• no-preparation-KL-control-seed-5301: [checksum retained in the private evidence record]; selected epoch3; elapsed214.9s; peakGPU1518020608 bytes. Gates: {"balanced_admitted_80": false, "balanced_all_80": false, "coverage": true, "gain_memory_2": true, "gain_parent_5": true, "high_confidence_wrong_2": false, "retention_95": false}; calibration: {"calibrated": {"brier": 0.41194379483264815, "ece_10_bins": 0.05650031210537013, "nll": 0.7181681371828096, "temperature": 2.0}, "raw": {"brier": 0.4323308847443796, "ece_10_bins": 0.16628327120747455, "nll": 0.7742045982190746, "temperature": 1}, "temperature_fit_on_training_calibration_only": true}.
• no-preparation-seed-5301: [checksum retained in the private evidence record]; selected epoch5; elapsed16.3s; peakGPU673976832 bytes. Gates: {"balanced_admitted_80": false, "balanced_all_80": false, "coverage": true, "gain_memory_2": true, "gain_parent_5": true, "high_confidence_wrong_2": false, "retention_95": false}; calibration: {"calibrated": {"brier": 0.39266361777461667, "ece_10_bins": 0.07584109673009712, "nll": 0.6788366633023492, "temperature": 1.5}, "raw": {"brier": 0.42662028667622, "ece_10_bins": 0.16188170195109342, "nll": 0.7543099239650212, "temperature": 1}, "temperature_fit_on_training_calibration_only": true}.
• science-preparation-seed-5301: [checksum retained in the private evidence record]; selected epoch4; elapsed16.6s; peakGPU673976832 bytes. Gates: {"balanced_admitted_80": false, "balanced_all_80": false, "coverage": true, "gain_memory_2": true, "gain_parent_5": true, "high_confidence_wrong_2": false, "retention_95": false}; calibration: {"calibrated": {"brier": 0.399292647888532, "ece_10_bins": 0.1004878933033632, "nll": 0.6823848773061649, "temperature": 1.5}, "raw": {"brier": 0.43617596395739955, "ece_10_bins": 0.15677003489304425, "nll": 0.7830540590193817, "temperature": 1}, "temperature_fit_on_training_calibration_only": true}.

Research and next decision

Locked runbook, source admission, MultiVerS official training, GEM official implementation.

If these gates fail, stop this preparation/projection mechanism. Do not extend declining fits, run a second seed, score the sealed set or relax confidence gates. Use the matched contrasts to decide whether source scale, scientific representation or finite-step retention needs changing. A development gate pass requires separately locked unseen/seed/runtime qualification before any promotion.

The active goal remains open until semantic learning and distinct persistent acquisitions are qualified. The standardized private backup closure follows this report.