Distillation

🟡 (05:09) research/oss-governance-coding-protocol-v1.2.md is titled “Performance-blinded OSS governance coding protocol v1.2,” frozen 2026-08-27, with status pre-adjudication evaluator correction; model-informed development protocol.
🟡 (05:09) Protocol V1.2 correction background: the V1.1 run used the intended frozen 40-record manifest and unchanged coders, but its mandatory artifact audit found that the evaluator field “strict edge agreement” compared actor, right, target, direction, and status while omitting edge scope and evidence references; it also failed to route separately coded body/patch-relationship disagreements.
🟡 (05:09) The V1.1 evaluator defects caused 6 records with material differences to be labeled provisional_exact_agreement.
🟡 (05:09) Protocol V1.2 changes, in order: 1. strict edge signatures include scope and complete grounded evidence references; 2. body/patch-relationship disagreement is an explicit route reason; 3. the remaining non-routed state is renamed provisional_core_agreement; 4. full-response agreement is reported separately and is not required because titles and numeric confidence are not canonical research fields.
🟡 (05:09) The V1.2 correction routes 35 records rather than 29, reuses the byte-identical V1.1 coding bundle and both Luna response files from workflow run 33060493337, and repeats only the Terra advisory pass over the corrected route to isolate the evaluator change from model sampling variance.
🟡 (05:09) Protocol V1.2 states that no model response, performance observation, transition date, or desired result influenced the correction; Terra remains advisory, all records remain ineligible for the canonical ledger, and the response schema and frozen envelope remain unchanged.
🟡 (05:16) _edge_signature() in src/institution_lab/governance_coding.py:377-405 returns () for a missing response; for each power change it normalizes actor, right_kind, target, direction, change_status, and scope via str(...).strip(), sorts normalized (source_id, quote) evidence-reference pairs, serializes those references with compact json.dumps(..., ensure_ascii=False, separators=(",", ":")), and returns the sorted tuple of edge signatures.
🟡 (05:16) Assistant said it would confirm each suspected failure through both schema and runtime entry points, rerun only scoped tests, and leave the worktree untouched.
🟡 (05:19) schema/governance-adjudication-v1.schema.json disallows top-level additional properties and requires exactly these named top-level fields: schema_version, adjudication_id, design_status, transition_dates_excluded, outcome_data_used, inputs, reviewers, and records.
🟡 (05:19) The adjudication schema requires the top-level records array to contain exactly 40 records via minItems: 40 and maxItems: 40.
🟡 (05:19) The adjudication schema’s inputs object disallows additional properties and requires 9 fields: workflow_run_id, artifact_id, artifact_sha256, documents_sha256, documents_file_sha256, coding_package_file_sha256, protocol_sha256, coding_schema_sha256, and adjudication_schema_sha256; workflow_run_id and artifact_id must be integers of at least 1.

Distillation

Observations