Distillation

Date: Sep 9, 2026

• 🟡 (06:18) Search found 38 reviewer-related matches in /home/byk/Code/institutional-transition-lab/src/institution_lab/governance_adjudication.py, including REVIEWER_ROLES = {"source_adjudicator", "evidence_auditor", "synthesizer"}, reviewer validation around lines 320–340, duplicate reviewer-ID checks around lines 542–549 and 694–701, and per-record source_adjudicator requirements around lines 593–599 and 752–758.
• 🟡 (06:22) Scoped test suite result: exactly 78 tests passed in 0.55s, with no failures.
• 🟡 (06:22) Search found 8 source_id matches across schema/governance-adjudication-v1.schema.json and schema/governance-coding-v1.schema.json; both schemas require non-empty string source_id values in evidence references.
• 🟡 (06:22) Investigation conclusion at this stage: the scoped suite’s existing assertions did not expose suspected crash paths. Testing strategy shifted to valid-shape semantic substitutions rather than schema-invalid inputs; a candidate would count as an opening bypass only if it survived both schema and runtime validation.
• 🟡 (06:23) _source_index() in /home/byk/Code/institutional-transition-lab/src/institution_lab/governance_adjudication.py lines 247–260 builds a dictionary from each source’s source_id to its string text, then unconditionally sets sources["record_metadata"] to compact, sorted JSON containing record_id, source_url, published_on, publisher, and source_type.
• 🟡 (06:23) _evidence_errors() in /home/byk/Code/institutional-transition-lab/src/institution_lab/governance_adjudication.py lines 263–287 requires evidence references to contain exactly {"source_id", "quote"} or, for claims, exactly {"source_id", "quote", "field"}; claim fields must be in CLAIM_FIELDS, source IDs must resolve in the frozen source index, and quotes must be non-empty substrings of the selected source text.
• 🟡 (06:23) _response_for() lines 290–295 returns a response only when the named coding result is a dictionary with schema_valid is True and its response is a dictionary.
• 🟡 (06:23) graph_edge_diff() lines 302–313 compares baseline and final power_changes using canonical JSON edge keys from _edge_key(), returning deterministically sorted added and removed arrays.
• 🟡 (06:23) _validate_reviewer() lines 316–341 requires exactly reviewer_id, reviewer_type, model, role, and assistance; IDs and assistance must be non-empty strings, roles/types must be allowed, human reviewers must have model=None, non-human reviewers must name a non-empty model, and any model string containing gemini case-insensitively is rejected because Gemini is restricted to event triage.
• 🟡 (06:23) _validate_record() lines 344–509 requires a matching frozen source_url; a valid decision; a non-empty, unique array of known reviewer_ids; a non-empty rationale; and a source_audit that confirms source-first inspection and exactly matches the frozen document’s ordered source IDs and evidence_bounds.
• 🟡 (06:23) Bounded evidence in _validate_record() means any of: source_text_truncated is True, files_listing_complete is False, patch_selection_truncated is True, or integer patch_unavailable_count > 0; such records require non-empty evidence_limitations.
• 🟡 (06:23) Decision semantics in _validate_record(): accept requires accepted_from in RESPONSE_SOURCES and final_coding exactly equal to that frozen valid response; only accept may set accepted_from; accept, revise, and reject require runtime-valid final_coding; abstain requires both null final_coding and a non-empty unresolved_reason; resolved decisions require null unresolved_reason and may not use final class abstain; reject must resolve to a non-institutional class.
• 🟡 (06:23) Institutional date/status semantics in _validate_record(): effective_institutional_change requires effective_on and permits only edge status effective; announced_institutional_change requires announced_on, forbids effective_on, and permits only announced; proposal_only forbids effective_on and permits proposed, rejected, or unclear.
• 🟡 (06:23) Claim grounding in _validate_record() checks every claim_evidence_refs entry through _evidence_errors() and, for institutional final codings, requires each non-null field in CLAIM_FIELDS to be represented by at least one reference whose field matches; the presence check is field-based after individual reference grounding.
• 🟡 (06:23) Graph-revision validation in _validate_record(): abstain requires null graph_edge_revision; every resolved decision requires an object; basis must be a response source or None; an accepted record’s basis must equal accepted_from; and added/removed must exactly equal graph_edge_diff(_response_for(coding_record, basis), final_coding).
• 🟡 (06:23) validate_adjudication() lines 512–656 requires transition_dates_excluded is True, outcome_data_used is False, all required input pins to match verified pins, and the coding package’s documents_sha256 to equal canonical_sha256(documents).
• 🟡 (06:23) validate_adjudication() requires a non-empty reviewer array, validates each reviewer, rejects duplicate reviewer IDs, and requires every adjudication record to list at least one reviewer whose declared role is source_adjudicator.
• 🟡 (06:23) Full adjudication record IDs must match frozen document IDs exactly once and in the same order; coding-package record IDs have the same exact-once/in-order requirement. The function indexes documents and coding rows by record_id before validating matched adjudication records.
• 🟡 (06:23) validate_adjudication() summary includes exact record_count, sorted decision_counts, sorted class_counts, unresolved_record_ids, per-record decisions and reviewer/evidence-limitation data, flattened graph_edge_revisions, and reviewer provenance. canonicalization_gate is "closed" when there are any errors or unresolved abstentions, otherwise "open".
• 🟡 (06:23) validate_adjudication_fragment() lines 659–767 constructs a restricted fragment schema requiring schema_version=1, design_status="performance_blinded_llm_assisted_source_adjudication", and non-empty reviewer and record arrays; it repeats runtime reviewer and record validation.
• 🟡 (06:23) Fragment record IDs must be unique and form a known subset in frozen-document order; coding-package IDs must still exactly match frozen documents in order, and the coding package’s document hash must match canonical_sha256(documents).
• 🟡 (06:23) validate_against_schema() lines 770–775 uses Draft202012Validator with FormatChecker() and emits sorted errors formatted as schema <path>: <message>.
• 🟡 (06:24) validate_response() in /home/byk/Code/institutional-transition-lab/src/institution_lab/governance_coding.py lines 268–302 builds source_map directly from all dictionary sources without requiring string IDs/text during construction; evidence later counts as grounded only if the selected source text is a string and the non-empty quote occurs within it.
• 🟡 (06:28) Search found 35 rejection/graph-revision-related matches across the tests. In tests/test_governance_adjudication.py, graph revision coverage includes fixtures around lines 275–276, 320, 365, 469, 523–524, and 581–582; the accepted-basis mismatch test around lines 452–456; and summary assertion around lines 1020–1021 expecting [{"record_id": "example-1", "basis": "luna_a", "added": [], "removed": []}].
• 🟡 (06:28) Security and integrity tests located in tests/test_governance_adjudication.py include test_input_manifest_rejects_self_consistent_substitute, test_input_manifest_rejects_changed_file_bytes, test_input_manifest_rejects_files_not_from_pinned_artifact, test_cli_rejects_output_aliases_to_frozen_inputs, test_cli_rejects_aliased_outputs, test_cli_rejects_duplicate_manifest_keys, test_input_manifest_rejects_non_object, test_input_manifest_rejects_duplicate_artifact_members, test_input_manifest_rejects_excessive_artifact_member_count, test_regular_file_reader_rejects_fifo_without_blocking, and test_canonical_sha256_rejects_non_finite_numbers.
• 🟡 (06:28) Additional rejection tests found across the repository: tests/test_model_policy.py::test_policy_rejects_scope_drift; tests/test_source_retrieval.py::test_rejects_outcome_derived_protocol_fields and test_rejects_embedded_search_qualifiers; tests/test_robustness.py::test_rejects_a_candidate_inside_the_project_warmup; tests/test_phases.py::test_rejects_nonpositive_values_on_log_scale; tests/test_llm_event_probe.py::test_rejects_hallucinated_evidence, test_rejects_label_leakage_in_model_documents, and test_rejects_missing_model_input_allowlist_audit; tests/test_llm_benchmark_subset.py::test_rejects_unknown_subset_document; and tests/test_llm_benchmark_gate.py::test_rejects_unexpected_configuration.