bingreeky/icoder/finalization

Model finalization

Training completion and model completion are different states. This stage freezes measurement, checks cross-stage regressions, proves artifact lineage and loadability, and prepares a model for human-authorized handoff. It does not run a new optimization merely to improve a final table.

Enter only through an active FINAL_EVAL or MODEL_FREEZE context and queue task. If invoked directly without one, load auto-post-training and bootstrap or resume the controller first. Load verification before consuming frozen verifier profiles, denominator policy, unresolved samples, or evaluation integrity evidence.

Read the final model manifest before creating finalization queue items.

Entry requirements

• shortlisted checkpoints have immutable artifact and parent IDs;
• the final evaluation split, sampling protocol, extraction, and verifier

profiles are frozen and contamination-audited;

• required experiments in the registry are evidence-complete or validly waived

where their predeclared policy permits by a scoped human-gate-verdict that names rationale, affected claims, and replacement evidence, or resolve through a unique supersession chain to one of those states; failed-blocking and invalidated experiments do not satisfy entry; and

• known infrastructure defects cannot silently change the denominator.

Finalization queue

1. evaluate every shortlisted checkpoint under the same frozen matrix;
2. compare the selected checkpoint with the base and every applicable SFT/OPSD

handoff, showing both gains and regressions rather than only an aggregate; an optional-stage bypass is represented by its explicit decision, not a phantom checkpoint;

1. audit unresolved samples, truncation, task coverage, and high-scoring outputs;
2. verify checkpoint completeness, tokenizer/config compatibility, conversion,

load, inference, and intended resume state;

1. close code, data, configuration, verifier, reward, and checkpoint lineage;
2. record known defects, limitations, unsupported task properties, and intended

use; and

1. produce an immutable pre-release final model manifest and queue a nonblocking

human release gate.

If the frozen evaluation is invalid, enter VERIFICATION_AUDIT or return to the responsible data state. If a candidate has unacceptable regressions, return to checkpoint selection or the responsible training stage through an explicit decision. Do not change the final evaluation to rescue a preferred checkpoint.

Selection

Select against a predeclared multidimensional regression profile. The highest aggregate is not automatically the model that moves forward. Report the number and depth of regressions, task-family coverage, unresolved fraction, and whether improvement reflects new capability or reliability on already-solvable tasks.

Authority

The Agent may prepare and internally select a checkpoint inside the approved project envelope. Publishing, uploading, or overwriting an existing remote model requires an explicit human gate. Approval moves the controller only to RELEASE_AUTHORIZED. A release task must then execute within that scope and verify the external locator and content digest before the controller may claim RELEASED.

Exit

MODEL_FREEZE closes only when the manifest, frozen results, regression audit, load test, known-defect record, and artifact digests are complete. Frozen-evaluation integrity, checkpoint identity, and load/inference verification are nonwaivable-for-completion; changing the deliverable instead requires a new human-approved project-profile version and a new finalization context. Experiments establishing those same properties use the matching nonwaivable policy; a superseded experiment counts only through its completed successor. COMPLETE_INTERNAL means the internal project deliverable exists. RELEASED is a separate state reached only after both explicit human authorization and a verified external release action; neither is inferred from internal completion.