[abwoo]

Problem and resulting behavior

PR #50 corrects weighted artifact uncertainty, paired covariance, missing HMM observations, spectral endpoints and validation provenance. Its review also exposed a growing mismatch between current behavior, historical diagnostics and execution status. This consolidation keeps the recommendation algorithm unchanged and makes those boundaries explicit.

Current recommendation rule

The current rule was introduced at d908041ff5aa5f8c62cf111e5b75eae397d2f950: retain the proposal unless the feasible baseline is significantly better under the reverse simultaneous bound (baseline - proposal + critical < 0). Indistinguishable evidence is labelled maintain; missing evidence is unassessed. An explicit baseline takes precedence over the average-reference repair comparison.

The earlier b27b435 rule instead retained the baseline when an upgrade was uncertain. Its cached replay numbers are historical diagnostics and do not establish the current rule's accuracy. Current scope and evidence ledger and methods now agree.

Consolidation

• Normalize MATLAB empty/singleton/vector JSON representations only for schema-defined arrays. Keep source values, missing arms, planned coverage and provenance intact.
• Re-aggregate the nine completed artifacts from 40805638604, without EEG fitting. Original algorithm SHA 6b26931 remains distinct from the processing revision.
• Make full shadow searches explicit, so normal PR edits do not launch sixteen long searches or cancel an ongoing shadow run.
• Add one-search checkpoint/slice recovery with code, seed, arm, variance and toolchain identity checks. Preserve completed leaves and ICA work, expose partial completion, and cache completed retries. Checkpointed executions remain separate from old unsliced paired means.
• Preserve eight disconnected prototypes, their tests and four completed one-off audit workflows on archive/pr50-prototypes-20261030, and remove them from the delivery branch.
• Replace appended round-by-round planning notes with maintained methods and an explicit evidence index. Keep docs/RESEARCH_READINESS.md absent.

Scientific limits

The original four-record development comparison at 4116ea1 failed its engineering admission tolerance; the paired difference interval included zero. Its historical accuracy/calibration values do not certify later revisions. The latest original shadow run has eight Standard results and only one complete pair. A one-pair bootstrap interval cannot establish population performance.

Fixed-family reranking and source decomposition do not validate the complete adaptive procedure, stable calibration or cross-record generalization. Independent confirmation remains unopened.

Validation

• 101 affected cases passed in 41052090671; that run also exposed a new fixture failure, corrected before integration.
• The remaining native interrupted/resumed-search proof passed, and the original-fragment MATLAB merge completed, in the successful steps of 41053098369. Its later Python failure is superseded below.
• 41053921374 is successful: 13 Python tests passed, and all original paired values, intervals, verdicts and CSVs match (8 seeds, 10 assessable arm records, one complete pair). No MATLAB or ICA was restarted for this verification.
• Across the scoped runs, 102 MATLAB cases passed. The native proof checks recommendation/score parity, unchanged cached candidate/ICA bytes, cached completion retry and identity mismatch rejection.
• Final tree checks confirm all 42 production implementation blobs are unchanged, the tested helper/checker/test sources match their validation revisions, relative documentation links resolve, and the deleted research-readiness document remains absent.
• Machine-readable provenance records the source, merge and checking commits separately. The native cross-worker workflow is prepared; a full scientific run through it is still pending.

MATLAB validation uses only R2026a + EEGLAB 2026.0.0. The integration uses [skip ci] to preserve completed checks and avoid a full rerun. Linked checks certify their stated source revisions and scope, rather than a new full compatibility run on the final documentation/integration SHA. No merge or release is performed.