TheKnowledge epistemic report

Type: agentic-system-analyses/types/agentic-system-epistemic-report.md

Source-and-claim boundary

Question: which shipped routes acquire or produce truth-apt content, check it, grant reliance, retain it and affect later behavior? Scope and exclusions are see boundary; evidence identities are see Source register. All additional reads below are SRC-1 implementation or attributed doctrine/design at the frozen revision, not new sources. No target code, tests, examples or providers were executed. No observed-run or causal evidence is supplied.

Assessed families: source/editorial admission; model page and answer production; citation-shaped validation; retained context selection; golden evaluation; candidate policy calibration; contradiction detection/precedence/withdrawal; bibliographic retraction warnings; reversible graph-based cleanup; caller-corrected editorial guidance. Runtime is provisional: direct inspection found additional evaluator, scan and recovery routes. No supplied record needs renaming or supersession; its valid container remains. The deposit identity-copy path was also read directly in SRC-1 src/gateway/ops/deposit.py to test whether source_type was authenticated before conflict precedence. New parts state containment. New mechanisms without a supplied counterpart state their distinction.

Missing run evidence prevents conclusions about exercised acceptance, model compliance, retrieval quality, successful recovery, learning or causal improvement. Provider internals prevent reconstructing NotebookLM synthesis grounding, exact parameter state and any provider-side adaptation. Individual retained domain claims were not assessed: their presence cannot identify which production/check route actually operated. Uninspected included operations below prevent a system-complete negative about truth checking or knowledge production; they are available evidence, not access gaps.

Public warrant claims: EPI-CLM-retrieval-grounding, EPI-CLM-answer-grounding, EPI-CLM-citation-provenance. Gateway-write doctrine is see RT-CLM-gateway-only.

Entry point / operation	Source path	Coverage and records	Remaining boundary
CLI dispatch; full/read MCP; browser operations	SRC-1 `src/gateway/cli.py`, `src/gateway/mcp_server.py`, `src/gateway/web/routes/ops.py`	Direct dispatch searches; supplied runtime account	Every handler was not read in full; no complete enforcement verdict
Ingest/filter/plan/apply/finalize/deposit worker	SRC-1 `src/gateway/ops/ingest.py`, `src/gateway/ops/apply_plan.py`, `src/gateway/ops/finalize.py`, `src/gateway/ops/committer.py`	Supplied RT-RTE-ingest-authorship, RT-RTE-draft-finalization, RT-RTE-deposit-commit; direct citation checker and dedup/gate branches	Acquisition converters individually unassessed; source warrant after extraction unknown
Retrieve/answer; FTS ranking	SRC-1 `src/gateway/ops/answer.py`, `src/gateway/search_index.py`, `src/gateway/ops/retrieve.py`	Direct answer/ranking reads; supplied RT-RTE-retrieve-answer	Dense/multidomain assembly evidence supplied by runtime; no answer-quality run
Research/query/NotebookLM	SRC-1 `src/gateway/research/orchestrator.py`, `src/gateway/ops/query.py`	Supplied RT-RTE-research-query	Remote indexed corpus and inference opaque; adapters individually unassessed
Policy bootstrap/edit/finetune/skill-emit	SRC-1 `src/gateway/ops/bootstrap_domain.py`, `src/gateway/ops/policy_edit.py`, `src/gateway/ops/finetune.py`, `src/gateway/ops/skill_emit.py`	Supplied runtime routes; direct calibration/policy/filter prompt reads	Candidate improvement/promotion unobserved; generated guidance is not checked truth
evaluate; eval-retrieval; calibration readiness/correction	SRC-1 `src/gateway/evaluate/runner.py`, `src/gateway/evaluate/judge.py`, `src/gateway/evaluate/wiki_context.py`, `src/gateway/evaluate/retrieval_eval.py`, `src/gateway/ops/calibration.py`, `src/gateway/ops/filter_calibrator.py`, `src/gateway/ops/filter_correct.py`	Direct reads; EPI-RTE-wiki-evaluation, EPI-RTE-retrieval-measure, EPI-RTE-filter-calibration, EPI-RTE-filter-correction	Embedding bakeoff and merge-map evaluator implementation not independently traced; policy gate coverage is supplied runtime
lint/sweep/drift/structured resolution/revert	SRC-1 `src/gateway/lint/contradictions.py`, `src/gateway/ops/contradiction_sweeper.py`, `src/gateway/ops/contradiction_drift.py`, `src/gateway/ops/contradiction.py`, `src/gateway/commit_gate.py`	Direct reads; EPI-RTE-contradiction-scan, EPI-RTE-conflict-selection, EPI-RTE-conflict-materialization, EPI-RTE-resolution-revert	Automatic reopen detectors and all lint checks unassessed; no complete conflict coverage
pubmed-retractions; retracted-citations	SRC-1 `src/gateway/pollers/pubmed_retractions.py`, `src/gateway/lint/retracted_citations.py`	Direct reads; EPI-RTE-retraction-monitor, EPI-RTE-retraction-warning	arXiv revisions/link-rot and cascade helper not fully inspected; full withdrawal coverage unknown
remediate/restore-depath/reverse-merge	SRC-1 `src/gateway/ops/remediate.py`, `src/gateway/ops/reverse_merge.py`, `src/gateway/commit_gate.py`	Direct selectors/public restoration/apply branches; EPI-RTE-orphan-selection, EPI-RTE-depath-recovery	reverse_merge_plan helper, overlap and crash behavior not fully assessed
cite-add/edit-section/editorial review	SRC-1 `src/gateway/ops/cite_add.py`, `src/gateway/ops/edit_section.py`, `.claude/skills/wiki-editorial-review/SKILL.md`	Direct bounded repair/read-only-review paths; EPI-RTE-citation-check distinguishes citation appearance from support	Underlying cite/cite-suggest/capture, concept-add and other direct authoring helpers not fully traced; no universal semantic admission claim
Hooks/scheduler/watcher/foreground skills	SRC-1 `.claude/settings.json`, `.claude/hooks/precompact-snapshot.sh`, `.claude/hooks/session-start-reanchor.sh`, `src/gateway/scheduler.py`, `src/gateway/watcher.py`	Supplied RT-RTE-session-reanchor and boundary; editorial skill directly read	Host delivery/compliance unobserved; shell schedules are not a truth gate
Other maintenance, agents, migrations, raw/wiki payload	SRC-1 `src/gateway/ops/`, `src/gateway/agents/`, `migrations/`, `wiki/`, `raw/`	Operation inventory only for demand-cluster, discovery/promotion, daily digest/review, orphan discharge, reingest, backfills and publishing	Included but unassessed; prevents exhaustive positive/negative about all content or policy updates

Epistemic-object inventory

Object ID	Candidate truth-apt content or none	Claimed role	Evidence	Gap/limit
EPI-OBJ-source-text	Imported source assertions, bibliographic status	see record	SRC-1 `src/gateway/filter/semantic.py`	Extraction fidelity and original source warrant unknown
EPI-OBJ-page-text	Wiki facts/syntheses/answers/conflict annotations	see record	SRC-1 `src/gateway/ops/answer.py`	Delivered/generated prose not necessarily entailed or accepted
RT-OBJ-plan	Markdown candidate inside symbolic update envelope	see record	SRC-1 `src/gateway/plan.py`	Parser and file checks do not certify every assertion
RT-OBJ-intent	Caller prose or reversal/policy payload inside queued control state	see record	SRC-1 `src/gateway/intent_queue.py`	Receipt is not canonical admission or truth acceptance
EPI-OBJ-eval-results: EvalRunSummary	Wiki/reference-fit judgment and reasoning	see record	SRC-1 `src/gateway/evaluate/runner.py`	Model score is advisory, retained in YAML/trend CSV
EPI-OBJ-eval-results: EvalReport	First expected-slug rank and aggregate hit fraction	see record	SRC-1 `src/gateway/evaluate/retrieval_eval.py`	Reference relevance, not source truth; policy gate consumes metric
EPI-OBJ-eval-results: CalibrationMetrics	Candidate binary-label agreement	see record	SRC-1 `src/gateway/ops/calibration.py`	Optional candidate annotation; no promotion threshold here
EPI-OBJ-eval-results: LintFinding	Candidate conflict/retraction exposure	see record	SRC-1 `src/gateway/lint/contradictions.py`	Advisory findings; scan can persist as draft or snapshot
EPI-OBJ-eval-results: resolution act	Rule-selected winner/loser and recorded rationale	see record	SRC-1 `src/gateway/ops/contradiction.py`	Symbolic precedence feeds annotation/reversal; not a truth verdict
EPI-OBJ-policy-text: editorial criteria	Relevance prescriptions; descriptions may be truth-apt	see record	SRC-1 `src/gateway/filter/policy.py`	Live/candidate YAML have different installation paths
EPI-OBJ-policy-text: corrected examples	Binary editorial labels and rationale	see record	SRC-1 `src/gateway/ops/filter_correct.py`	Prompt selection mediates future influence; caller identity unresolved
EPI-OBJ-policy-text: domain skill	Host instructions plus selected wiki titles	see record	SRC-1 `src/gateway/ops/skill_emit.py`	Persisted text is interpretive; discovery/compliance unobserved
EPI-OBJ-policy-text: session snapshot	Requested assertions about working tree, decisions and next step	see record	SRC-1 `.claude/hooks/session-start-reanchor.sh`	Snapshot generation/checking is requested, not evidenced completion
RT-CMP-qwen3-embedding	not determinable as individuated propositions; numerical retrieval state	see record	SRC-1 `src/gateway/retrieval_encoder.py`	Exact weights/provenance unknown; no truth oracle inferred
RT-CMP-claude-haiku, RT-CMP-claude-opus, RT-CMP-claude-sonnet, RT-CMP-notebooklm	Provider-parametric truth-apt content not determinable from accessible state	see records	SRC-1 `src/gateway/llm/config.py`	Inference outputs are accessible candidates; hidden parameter contents/updates are not

Authority-route ledger

This is an epistemic overlay, not a second declaration of generic route fields. see points to the owning record. All implemented rows have source-only support; activation and observed results remain unobserved. Linked rows separate checking, disposition, retention and consumption. Non-ampliative reshaping is not inferred for opaque model synthesis or citation deletion.

Route ID	Route function	Architectural status	Object/candidate ID	Content/update relation	Transition or check target	Evaluator/condition and domain	Activation and timing	Possible or observed result	Implemented force	Epistemic authority and scope	Operational authority	Behavioral-authority path	Evidence	Claim IDs or none	Mismatch marker or none	Gap/limit
RT-RTE-ingest-authorship	content transformation	implemented	EPI-OBJ-source-text	truth-apt transformation: acquisition/import	External source into retained raw text	Converter and source validators; see record	see RT-RTE-ingest-authorship	Canonical source and optional source page	Admission protocol	Source warrant unknown; hash equality licenses body identity, not truth	Persist input for later authorship	Plan/retrieval consumers; filesystem text; permissive; later calls	SRC-1 `src/gateway/ops/ingest.py`	none	none	Converter fidelity/source truth unassessed.
RT-RTE-ingest-authorship	check/evidence production	implemented	EPI-OBJ-source-text	no content change	Domain relevance/quality fit	Model scores capped source against policy/examples; code parses bounded numeric score	see RT-RTE-ingest-authorship	Score/rationale	Advisory score to threshold consumer	Editorial-fit judgment only	Provide source admission input	Ingest code; FilterResult; permissive threshold input; current ingest	SRC-1 `src/gateway/filter/semantic.py`	none	none	Not a source-factuality check; force-include bypasses scoring.
RT-RTE-ingest-authorship	disposition/acceptance	implemented	EPI-OBJ-source-text	no content change	Wiki-source inclusion	Code policy thresholds or caller force; see record	see RT-RTE-ingest-authorship	Include/review/reject/skip/error	Enforcing source-page gate; raw retention separate	Operational editorial admission, not accepted source claims	Admit source page independently of raw and generated-page admission	Ingest code; filter result/override; enforcing; current write	SRC-1 `src/gateway/ops/ingest.py`	none	none	No execution supplied.
RT-RTE-ingest-authorship	content transformation	implemented	RT-OBJ-plan	truth-apt transformation: indeterminate	Model full-page updates	Source/existing-page prompt plus error-fed repair; see record	see RT-RTE-ingest-authorship	Plan candidate	Proposal only before validation	Cannot distinguish faithful reshaping, derivation and conjecture without actual propositions/input comparison	Submit generated changes to validation	apply_plan; parsed Plan; permissive candidate; current call	SRC-1 `src/gateway/plan.py`	none	none	Prompt requests synthesis; novelty/non-entailment and accepted ampliation not evidenced.
EPI-RTE-citation-check	check/evidence production	implemented	EPI-OBJ-page-text	no content change	Claim-shaped sentences citation presence	Deterministic heuristic, not source evidence judge	see EPI-RTE-citation-check	Errors/warnings	Strict-veto input; draft advisory	Citation-shaped coverage only	Permit/reject depending on strict/draft consumer	apply_plan/finalize; ValidationResult; enforcing/advisory; current admission	SRC-1 `src/gateway/validator.py`	EPI-CLM-citation-provenance	citation tokens != semantic warrant	No execution supplied.
RT-RTE-ingest-authorship	disposition/acceptance	implemented	RT-OBJ-plan	no content change	Schema/citation-preservation/update guards	Code validation then write/rollback; see record	see RT-RTE-ingest-authorship	Write or rejection/warning	Enforcing file-admission protocol	Structural acceptance for canonical file use, not truth acceptance	Allow later wiki reuse including drafts	Filesystem/retrieval; accepted write; permissive; subsequent calls	SRC-1 `src/gateway/ops/apply_plan.py`	none	none	No execution supplied.
RT-RTE-ingest-authorship	retention	implemented	EPI-OBJ-page-text	no content change	Validated/draft content persistence	Filesystem write; see record	see RT-RTE-ingest-authorship	Retained pages/backlinks/contradiction log	Durable-by-protocol ledger use	Retention/integration bookkeeping is not evidence-consuming claim acceptance	Later page retrieval and authorship permitted	Models/humans; retained markdown; interpretive/permissive; across calls	SRC-1 `src/gateway/ops/apply_plan.py`	none	none	No lifecycle-integration finding follows without demonstrated claim acceptance; exception rollback is best effort.
RT-RTE-draft-finalization	disposition/acceptance	implemented	EPI-OBJ-page-text	no content change	Strict page validation / draft metadata	Code full validator; see record	see RT-RTE-draft-finalization	Nondraft or failure/escalation; abandon deletes	Enforcing status transition	Nondraft means stricter syntax, not source entailment	Removes draft metadata; draft reading already possible	Retrieval/maintenance; retained status; permissive; later calls	SRC-1 `src/gateway/ops/finalize.py`	none	none	No execution supplied.
RT-RTE-deposit-commit	operational admission/selection/consumption	implemented	RT-OBJ-intent	truth-apt transformation: indeterminate	Caller prose; alias-based entity identity/dedup and Git state	Deterministic same-kind normalized-name criterion; see record	see RT-RTE-deposit-commit	Merge/link/distinct; commit/reject/quarantine	Enforcing queue/Git admission	Co-reference is a policy assumption; names/geometry do not warrant every merged fact	Write/union retained text without apply_plan citation checks	Commit worker/gate; rendered writes; enforcing; current drain and later retrieval	SRC-1 `src/gateway/dedup.py`	RT-CLM-gateway-only	none	Verbatim deposit is import; union with inferred co-reference makes whole semantic relation indeterminate.
RT-RTE-retrieve-answer	operational admission/selection/consumption	implemented	EPI-OBJ-page-text	truth-apt transformation: indeterminate	Retained sections into model context	Rank/quota/budget code; see record	see RT-RTE-retrieve-answer	Ranked/truncated marked blocks	Ranking/permissive delivery	No new truth warrant; source attribution can be lost on truncation	Supply selected wiki prose including drafts to answer generation	Answer model; context blocks; interpretive; current question	SRC-1 `src/gateway/ops/retrieve.py`	EPI-CLM-retrieval-grounding	strict bounded/citation-preserved wording exceeds exceptions	Untruncated selection is non-ampliative reshaping; truncated prose can lose qualifiers/citations, so semantic preservation of the whole assembled result is indeterminate.
RT-RTE-retrieve-answer	content transformation	implemented	EPI-OBJ-page-text	truth-apt transformation: indeterminate	Question-conditioned answer prose	Sonnet/API model prompt; see record	see RT-RTE-retrieve-answer	Generated answer candidate	Prompt-requested context reliance	No output comparison proves entailment or non-entailment	Return candidate to cleanup; optional draft filing	Cleanup/caller; answer string; advisory/permissive; current call	SRC-1 `src/gateway/ops/answer.py`	EPI-CLM-answer-grounding	none	External parametric inference is not a checked derivation.
EPI-RTE-answer-cleanup	content transformation	implemented	EPI-OBJ-page-text	truth-apt transformation: indeterminate	Answer citation membership	Regex against retrieved-ID set	see EPI-RTE-answer-cleanup	Unknown IDs removed, prose survives	Enforcing token deletion only	Membership in supplied block, not semantic support	Return cleaned answer, optionally retain draft	Caller/later wiki readers; surviving prose; interpretive/permissive; current and future calls	SRC-1 `src/gateway/ops/answer.py`	EPI-CLM-answer-grounding	ungrounded citation != unsupported fact	No execution supplied.
RT-RTE-research-query	content transformation	implemented	EPI-OBJ-page-text	truth-apt transformation: acquisition/import	NotebookLM synthesis and source-map rendering	Opaque provider and local citation remapping/application; see record	see RT-RTE-research-query	Cited/draft wiki synthesis	Remote proposal plus local structural admission	Acquired provider prose; semantic entailment and source-map completeness unknown	Retain for later retrieval even if promotion warns	Wiki readers/answer model; synthesis files; interpretive/permissive; future calls	SRC-1 `src/gateway/ops/query.py`	none	none	Provider facts imported, not automatically locally produced knowledge; exact remote corpus opaque.
EPI-RTE-wiki-evaluation	check/evidence production	implemented	EPI-OBJ-eval-results	no content change	Selected domain wiki against golden facts/citations	Sonnet judge + output parser	see EPI-RTE-wiki-evaluation	Per-question and aggregate score	Advisory report, no page gate here	Reference-fit judgment within supplied context only	Record report; no automatic page revision in runner	Operator; YAML/CSV/returned score; advisory; run/history	SRC-1 `src/gateway/evaluate/runner.py`	none	none	Golden validity, omitted source evidence and model score arithmetic unknown.
EPI-RTE-retrieval-measure	check/evidence production	implemented	EPI-OBJ-eval-results	no content change	First expected slug rank	Deterministic curated-slug comparison	see EPI-RTE-retrieval-measure	Hit fraction/MRR	Advisory standalone; floor input in policy-edit	Retrieval of at least one expected slug, not correct answer	Gate may block policy-edit	Policy gate/operator; EvalReport; advisory/enforcing input; current proposal	SRC-1 `src/gateway/evaluate/retrieval_eval.py`	none	none	No execution supplied.
RT-RTE-policy-edit	disposition/acceptance	implemented	EPI-OBJ-policy-text	non-truth-apt policy/content update: proposed policy replacement	Policy write and tested dedup/retrieval domains	Principal and evaluator gates; see record	see RT-RTE-policy-edit	Commit/dead-letter; explicit dev missing-golden skips	Enforcing route-local change control	Tested retrieval floor and proposed dedup golden fit only	Admit full policy despite broader untested fields	Policy-dependent consumers; YAML; enforcing guidance when consumed; later calls	SRC-1 `src/gateway/commit_gate.py`	none	none	No universal policy admission: bootstrap/force bypasses these gates.
EPI-RTE-filter-calibration	check/evidence production	implemented	EPI-OBJ-policy-text	no content change	Candidate binary editorial criteria fit	Model classifier + code against calibration labels	see EPI-RTE-filter-calibration	Precision/recall/F1/accuracy or None	Advisory candidate metrics	Agreement on capped examples, not generalization/source truth	Retain metric beside candidate, no promotion gate here	Reviewer; candidate YAML/result; advisory; later review	SRC-1 `src/gateway/ops/calibration.py`	none	none	No execution supplied.
RT-RTE-policy-distillation	behavior/policy adaptation	implemented	EPI-OBJ-policy-text	non-truth-apt policy/content update: editorial/host guidance	Criteria/skill generation	see RT-RTE-policy-distillation	see RT-RTE-policy-distillation	Retained criterion candidate and optional metrics	Advisory candidate only	No epistemic acceptance inferred from guidance production	Stops before live policy replacement	Reviewer; candidate YAML; advisory; later review	SRC-1 `src/gateway/ops/finetune.py`	none	none	Conditions/learning/reflection/autonomy are see runtime record, independently unresolved where recorded.
RT-RTE-policy-bootstrap	behavior/policy adaptation	implemented	EPI-OBJ-policy-text	non-truth-apt policy/content update: editorial/host guidance	Criteria/skill generation	see RT-RTE-policy-bootstrap	see RT-RTE-policy-bootstrap	Schema-valid live policy, or draft/error	Enforcing live write; criteria interpretive at filter	Schema validity, not criterion truth or quality	Govern future domain filter scoring	Filter model; loaded policy YAML; interpretive; future source requests	SRC-1 `src/gateway/ops/bootstrap_domain.py`	none	none	Conditions/learning/reflection/autonomy are see runtime record, independently unresolved where recorded.
RT-RTE-skill-emission	behavior/policy adaptation	implemented	EPI-OBJ-policy-text	non-truth-apt policy/content update: editorial/host guidance	Criteria/skill generation	see RT-RTE-skill-emission	see RT-RTE-skill-emission	Retained domain SKILL.md	Interpretive host guidance if loaded	Deterministic re-render of selected guidance, not new warrant	Supply domain criteria and selected titles to later host sessions	Foreground agent; generated skill; interpretive; sessions until regeneration	SRC-1 `src/gateway/ops/skill_emit.py`	none	none	Conditions/learning/reflection/autonomy are see runtime record, independently unresolved where recorded.
EPI-RTE-filter-correction	behavior/policy adaptation	implemented	EPI-OBJ-policy-text	non-truth-apt policy/content update: retained corrected label and rationale	Past editorial decision	Caller proposal, code diff check	see EPI-RTE-filter-correction	Pinned example and raw correction	Interpretive guidance through future example selection	Editorial label, not truth acceptance	Can alter future model scoring prompts	Filter model; selected example/rationale; interpretive; future requests	SRC-1 `src/gateway/ops/filter_correct.py`	none	none	No execution supplied.
EPI-RTE-contradiction-scan	check/evidence production	implemented	EPI-OBJ-page-text	no content change	Direct contradiction between extracted sentences	Model pair/rationale proposal; code index checks	see EPI-RTE-contradiction-scan	Potential contradictions; empty/failure results	Advisory diagnostic	Candidate inconsistency judgment, not winner truth	Lint review; no mandatory factual repair	Operator/sweeper; findings; advisory; current scan	SRC-1 `src/gateway/lint/contradictions.py`	none	none	No execution supplied.
EPI-RTE-contradiction-scan	retention	implemented	EPI-OBJ-eval-results	no content change	Scan output as draft/snapshot	Weekly/daily code drivers	see EPI-RTE-contradiction-scan	Draft synthesis and daily diff retained	Direct write and advisory retained diagnosis	Disappeared keys only; resolved label does not license factual resolution	Draft can reach ordinary retrieval; snapshot informs later diff	Wiki readers/operator; markdown/JSON; interpretive/advisory; later calls/days	SRC-1 `src/gateway/ops/contradiction_sweeper.py`	none	none	Same-week skip, batch scope and parse-to-empty can obscure unresolved conflicts.
EPI-RTE-conflict-selection	disposition/acceptance	implemented	EPI-OBJ-page-text	no content change	Unequal extracted claim objects	Code source-type trust then string comparison	see EPI-RTE-conflict-selection	Winner/loser labels and act	Enforcing precedence input	Policy preference only; no evidence-consuming factual acceptance	Permit contested-edge construction while retaining both claims	Gate materializer; act dict; permissive; current deposit	SRC-1 `src/gateway/ops/contradiction.py`	none	old/new comparison is not actual recency; caller-derived new source_type	No execution supplied.
EPI-RTE-conflict-materialization	retention	implemented	EPI-OBJ-page-text	truth-apt transformation: indeterminate	Disputes annotation and loser claim	Deterministic selected-loser render	see EPI-RTE-conflict-materialization	Retained Contested section	Interpretive annotation in canonical prose	Operational dispute claim; factual conflict/winner not independently checked	Later context contains both claims and dispute annotation	Retrieval/answer; markdown; interpretive; subsequent calls	SRC-1 `src/gateway/commit_gate.py`	none	claim-specific loser down-weight not established	No execution supplied.
EPI-RTE-resolution-revert	lineage/freshness/recovery	implemented	EPI-OBJ-page-text	truth-apt transformation: indeterminate	Recorded conflict annotation withdrawal	Caller act request; code source-token scan and section strip	see EPI-RTE-resolution-revert	Committed removal/reverted act or dead-letter	Enforcing annotation deletion	Withdrawal of annotation, not truth of either claim	Later readers no longer receive stripped section	Retrieval/operator; modified markdown/act status; permissive; later calls	SRC-1 `src/gateway/commit_gate.py`	none	none	First matching page and whole sections are coarser than one act.
EPI-RTE-retraction-monitor	content transformation	implemented	EPI-OBJ-source-text	truth-apt transformation: acquisition/import	Bibliographic retraction status	External NCBI XML + broad substring checker	see EPI-RTE-retraction-monitor	Raw retracted flag/cursor	Enforcing metadata write; warning input	Acquired bibliographic signal; source warrant may be degraded, no claim-by-claim invalidation	Permit lint warning without deleting raw claims	Lint; raw YAML flag; advisory downstream; future checks	SRC-1 `src/gateway/pollers/pubmed_retractions.py`	none	none	Parse-false and broad notice matching bound interpretation.
EPI-RTE-retraction-warning	check/evidence production	implemented	EPI-OBJ-page-text	no content change	Direct/transitive cited retracted dependencies	Code link/graph check	see EPI-RTE-retraction-warning	ERROR/WARNING review findings	Advisory, not retrieval exclusion	Dependency exposure only; no claim truth verdict	Request review/replacement; retained prose remains usable	Operator; lint result/report; advisory; later maintenance	SRC-1 `src/gateway/lint/retracted_citations.py`	none	none	No execution supplied.
EPI-RTE-orphan-selection	operational admission/selection/consumption	implemented	EPI-OBJ-page-text	no content change	Orphan de-path eligibility	Code index/provenance reachability	see EPI-RTE-orphan-selection	Queued delete candidate or dry-run list	Permissive intent production	Graph-based disposal, not falsehood rejection	Authorize delayed deletion for selected paths	Worker/gate; typed intent; permissive; later drain	SRC-1 `src/gateway/ops/remediate.py`	none	none	Selection-time safety does not ensure unchanged apply-time graph.
EPI-RTE-depath-recovery	lineage/freshness/recovery	implemented	EPI-OBJ-page-text	no content change	Delete/restoration of recorded text	Code captures content then deletes; public restore reloads provenance	see EPI-RTE-depath-recovery	Removed/recreated file or dead-letter	Enforcing filesystem mutation	Recorded lineage only; restoration does not restore warrant	Remove/reintroduce page to future retrieval	Wiki retrieval; canonical file/provenance; permissive; later calls	SRC-1 `src/gateway/commit_gate.py`	none	none	No execution supplied.
RT-RTE-session-reanchor	other — describe	doctrine only	EPI-OBJ-policy-text	truth-apt transformation: indeterminate	Requested development snapshot/check against tree	Host/agent natural-language instructions; see record	see RT-RTE-session-reanchor	Requested snapshot/reread/verification	Advisory/interpretive host instructions	No observed acceptance of snapshot facts	Agent asked to verify before plan/write and escalate mismatch	Foreground agent; hook stdout/file; interpretive; compaction/later session	SRC-1 `.claude/hooks/session-start-reanchor.sh`	none	none	Actual compression, retained content and compliance opaque; no automatic script-level truth validation.

Concrete implementation deductions (no target execution):

Given valid ID a, answer input The treatment cures every disease. [[sources/a]] [[sources/b]] becomes The treatment cures every disease. [[sources/a]]: cleanup deletes only the unmatched b link. It does not challenge the universal claim. Evidence is see EPI-RTE-answer-cleanup. Thus link-set enforcement cannot establish factual grounding.
In same-line text This treatment always works for everyone. [[sources/a]], the claim heuristic sees at least five words, terminal period and any source token on the line. It marks cited irrespective of what source a says. Conversely code-fenced/bold-header/exempt text can escape claim detection. SRC-1 src/gateway/citations.py; see EPI-RTE-citation-check. This is a concrete citation-presence check, not source evidence consumption.
For equal source-type trust in gate conflict selection, committed_at="old" is greater lexicographically than "new", so the existing side wins. This contradicts a literal latest-commit interpretation. See EPI-RTE-conflict-selection. No claim truth follows whichever side wins.
Same-kind names ACME, Inc. and acme inc normalize to the same string and trigger merge if selected as candidates; same spelling with different entity_kind never merges. SRC-1 src/gateway/dedup.py. This checks an encoded identity convention, not real-world co-reference or proposition support.

System-claim versus route comparison

Claim ID	Claimed operation/warrant	Source and layer	Doctrine/design support	Implemented route IDs	Observed-run support	Causal support/design limits	Supported conclusion	Mismatch/unknown
EPI-CLM-retrieval-grounding	see record	SRC-1 `README.md`; doctrine/design	Context grounding and preserved citations	RT-RTE-retrieve-answer	none supplied	No retrieval/answer experiment; budget/truncation exceptions	Ranked delivery with controls	Not a strict bounded/provenance-complete invariant
EPI-CLM-answer-grounding	see record	SRC-1 `README.md`; doctrine/design	Context-only answer with ungrounded IDs stripped	RT-RTE-retrieve-answer, EPI-RTE-answer-cleanup	none supplied	Prompt compliance not guaranteed; membership check leaves prose	Only retrieved-ID membership enforced	Semantic grounding remains unestablished
EPI-CLM-citation-provenance	see record	SRC-1 `WIKI.md`; doctrine/design	NLM citation tokens said to prove corpus provenance	EPI-RTE-citation-check, RT-RTE-research-query	none supplied	Opaque token existence not locally checked; corpus-source mapping incomplete cases	Syntactic grounding token accepted	A token plus notebook ID does not prove actual provenance
RT-CLM-gateway-only	see record	SRC-1 `CLAUDE.md`; doctrine/design	Foreground gateway discipline	RT-RTE-ingest-authorship, RT-RTE-deposit-commit	none supplied	No host filesystem isolation; gateway paths have distinct validators	Instruction and route-local checks	Gateway use does not imply one universal factual/citation admission gate

Bounded conclusion

Imported and retained content can influence later answers before any demonstrated factual acceptance. See RT-RTE-ingest-authorship, RT-RTE-retrieve-answer, RT-RTE-research-query. Raw assertions, model-generated pages, caller deposits and drafts are different admission paths. Source identity/hash and citation shape retain some lineage, but original source warrant is unknown, and truncation/opaque provider tokens can weaken attribution. Query imports provider-produced prose; local model authorship and answering are candidate production. No inspected actual proposition/input pair establishes checked entailed derivation or demonstrable non-entailment of an accepted ampliative claim. Local model transformation remains indeterminate, not automatically conjecture because a model generated text; provider synthesis is acquired with unknown source warrant.

Implemented evaluators have distinct licenses and consumers. Citation checking and finalization enforce structural file criteria, not factual entailment; answer cleanup enforces retrieved-ID membership only. Wiki-state evaluation judges selected context against declared golden facts/citations and retains scores without a page gate in its runner. Retrieval measurement checks first expected slug rank and can feed an enforcing policy-edit floor. Calibration compares candidate classifier labels to retained reference labels, but stops at candidate metrics. See EPI-RTE-citation-check, EPI-RTE-answer-cleanup, EPI-RTE-wiki-evaluation, EPI-RTE-retrieval-measure, EPI-RTE-filter-calibration. None of these criteria licenses claims beyond its checked target, reference validity, context and intended use. There is no observed evidence-consuming acceptance decision here naming a generated factual claim, criterion and intended epistemic use; this is an assessed-route limit, not a whole-system absence.

Conflict and withdrawal routes change operational reliance without determining factual truth. See EPI-RTE-contradiction-scan, EPI-RTE-conflict-selection, EPI-RTE-conflict-materialization, EPI-RTE-resolution-revert, EPI-RTE-retraction-monitor, EPI-RTE-retraction-warning, EPI-RTE-orphan-selection, EPI-RTE-depath-recovery. Model contradiction warnings are candidates; weekly driver can write them directly as drafts. Trust/recency policy labels a preferred claim while keeping both; the inspected equal-trust gate does not implement actual recency. A disputes annotation reaches later readers interpretively, not as a line-level ranking/exclusion rule. Public reversals and graph cleanup remove/restore content or annotations, not warrant. Bibliographic retraction flags produce direct/transitive review warnings; they do not automatically repair claims or ban retrieval in the inspected paths. Existing structured-page resolve writes a status/note, not a verified factual adjudication. These counterexamples prevent both a universal acceptance claim and a universal no-withdrawal claim.

Behavior-shaping guidance is separate from knowledge warrant and from demonstrated improvement. See EPI-RTE-filter-correction, RT-RTE-policy-distillation, RT-RTE-policy-bootstrap, RT-RTE-policy-edit, RT-RTE-skill-emission, RT-RTE-session-reanchor. Corrected examples can enter future filter prompts; candidate distillation is not live installation; bootstrap can install schema-valid policy directly; policy-edit gates only selected test domains; host skill/snapshot delivery depends on host activation/compliance. Criteria and rationale are inspectable natural-language guidance in symbolic files; inaccessible model weights remain independently unresolved. Inspected inference does not update parameters, but exact provider weights, endpoint changes and internal learning cannot be assessed. Learning, reflection, theory-builder conditions, autonomy and self-improvement are not inferred from retention or generation; retain the separate runtime findings on those policy routes. No new causal benefit is established by this source-only inspection.

Included but unassessed operations in the coverage table leave broader knowledge-production and lifecycle-integration conclusions open. Ordinary backlinks, retained contradiction acts and corpus maps are bookkeeping/use before demonstrated epistemic acceptance, not evidence that accepted new theory was integrated. The report therefore has no single system-wide epistemic verdict.

Shared records

Operative objects

EPI-OBJ-source-text

Label: Imported source text and metadata

Source-native identity: canonical raw markdown body and mutable bibliographic/frontmatter metadata. Body assertions are natural-language content; source IDs/hash/retraction flags are symbolic inputs to fixed consumers. Substrate: raw//.md filesystem files and current filter/model prompt strings. Separate from RT-OBJ-plan and RT-OBJ-intent: those are operation envelopes, not retained acquired input. SRC-1 src/gateway/filter/semantic.py, src/gateway/pollers/pubmed_retractions.py.

return f"### Frontmatter\n\nyaml\n{front_yaml}\n\n\n### Body (truncated)\n\n{body_head}" --- src/gateway/filter/semantic.py

EPI-OBJ-page-text

Label: Wiki and answer assertion text

Source-native identity: natural-language markdown bodies and generated answer text, with symbolic citation tokens interpreted by link/check code. Substrate: canonical wiki files, current AnswerResult and retrieved context strings. Distinct from RT-OBJ-plan (candidate update envelope) and RT-OBJ-intent (durable queue envelope); this is retained/delivered knowledge content across those mechanisms. It is not a part of exactly one envelope, since direct edits, scans, query and deposits also write it. SRC-1 src/gateway/ops/answer.py, src/gateway/ops/contradiction_sweeper.py.

return AnswerResult(
    answer=cleaned, sections=sections, source_ids=cited,
    stripped=stripped, usage=usage,
)

--- src/gateway/ops/answer.py

EPI-OBJ-eval-results

Label: Evaluator and conflict result records

Source-native identity: EvalResult/EvalRunSummary, EvalReport, CalibrationMetrics, LintFinding and conflict act outputs. Symbolic scores/IDs/status feed fixed program consumers; natural-language reasoning/claims require interpretation. Substrate: current Python values, evaluation YAML/trend CSV, lint JSON/markdown and resolution JSONL. No supplied operative-object counterpart after comparing RT-OBJ-plan and RT-OBJ-intent: these are measurements/diagnoses, not write requests. Material routes below separate their consumers rather than treating all scores as one oracle. SRC-1 src/gateway/evaluate/runner.py, src/gateway/ops/contradiction.py.

    "winner": {"source": winner.get("source"), "claim": winner.get("claim"),
               "trust": _side_trust(winner)},
    "loser": {"source": loser.get("source"), "claim": loser.get("claim"),
              "trust": _side_trust(loser)},

--- src/gateway/ops/contradiction.py

EPI-OBJ-policy-text

Label: Editorial and host guidance text

Source-native identity: inclusion_criteria/exclusion_criteria, example rationale/corrected labels and emitted host/snapshot guidance. Natural-language guidance has interpretive force; symbolic domain/version/threshold fields and labels select/rule program behavior. Substrate: policy/example YAML, generated SKILL.md and requested session-state document. Compared with RT-RTE-policy-distillation, RT-RTE-policy-bootstrap, RT-RTE-skill-emission, RT-RTE-session-reanchor: those are distinct producers/consumers, not the guidance object. Exact generated contents and snapshot existence/use unobserved. SRC-1 src/gateway/filter/policy.py.

    inclusion_criteria=list(data.get("inclusion_criteria", []) or []),
    exclusion_criteria=list(data.get("exclusion_criteria", []) or []),

--- src/gateway/filter/policy.py

Routes

EPI-RTE-citation-check

Label: Citation-shaped claim checking

Part of: RT-RTE-ingest-authorship

This is a material subfunction, not a replacement of its supplied route container.

Plan validators call validate_citation_grounding; code classifies claim-shaped sentences and checks citation presence. Code decides errors/warnings; no human or truth oracle supplies the verdict. Normal strict and draft modes differ. Guidance is the citation heuristic and structural/aggregate exemptions, not source-body entailment. An error can veto strict page admission; draft mode changes it to warning. SRC-1 src/gateway/validator.py; implementation layer.

Immediate return: ValidationResult errors or warnings.
Later read-back: inapplicable — this check itself retains nothing; its downstream page admission belongs to the containing route.
Delegated visibility: inapplicable — deterministic checker, not a delegated model.
Selection predicate: Schema citation_grounded flag and uncited_claims over page body/frontmatter.
Invalidation or expiry: Recomputed per validation; no semantic endorsement persists.
Activation or effect: Strict errors can block the containing apply path; draft warnings cannot.
Evidence limits: SRC-1 src/gateway/citations.py: line-level citation membership, sentence length/punctuation and exemptions leave unsupported prose possible. Finalization invokes this check strictly; deposit renderer is a separate path.
implementation conclusion status: wired
operation conclusion status: uninspected

    if draft:
        result.warnings.append(ve)
    else:
        result.errors.append(ve)

--- src/gateway/validator.py:587-590

EPI-RTE-answer-cleanup

Label: Unknown citation stripping

Part of: RT-RTE-retrieve-answer

This is a material subfunction, not a replacement of its supplied route container.

After model generation, code extracts source IDs from the delivered retrieval block and deletes links to IDs outside that set. Model proposes prose; code decides token deletion without judging facts. Open answering mode has no expected answer oracle. Caller optionally files the surviving answer through draft application; no per-claim approval or rejection is wired in cleanup. SRC-1 src/gateway/ops/answer.py; implementation layer.

Immediate return: Cleaned answer and stripped-ID list.
Later read-back: Optional draft answer can reach later retrieval, through the containing route.
Delegated visibility: Answer model sees selected blocks, question and prompt-only source constraint; cleanup sees strings.
Selection predicate: Regex source IDs in retrieval block, not a claim/source support comparison.
Invalidation or expiry: Each answer recomputes valid IDs; no continued check of surviving claims.
Activation or effect: Deletes unknown citation tokens; leaves their prose intact; result returned and optionally filed.
Evidence limits: No factual acceptance or semantic preservation follows from deleting an attribution. Source-supported counterexample: a valid ID can be attached to an unsupported claim.
implementation conclusion status: wired
operation conclusion status: uninspected

    if sid in valid:
        return m.group(0)
    stripped.append(sid)
    return ""

--- src/gateway/ops/answer.py

EPI-RTE-wiki-evaluation

Label: Golden wiki-state judgment

Compared with RT-RTE-ingest-authorship, RT-RTE-deposit-commit and RT-RTE-retrieve-answer: no supplied record declares this specific mechanism; related generation/write containers are not the same referent.

CLI/MCP evaluate dispatches runner → wiki_context loader → Sonnet Judge → parsed EvalResults → run YAML/trend CSV/summary. Model judges the selected wiki against caller-maintained golden must_cite/must_assert/must_not_assert and rubric weights; code parses and aggregates. This is bounded reference evaluation, not generation of a new answer. No page-revision or page-admission consumer is implemented in the inspected runner. Oracle correctness and golden provenance are uninspected. SRC-1 src/gateway/evaluate/runner.py; implementation layer.

Immediate return: EvalRunSummary with mean and per-question results.
Later read-back: Run YAML and trend CSV retain scores for later operator inspection; automatic revision from these files is not established.
Delegated visibility: Judge sees all selected synthesis bodies, first 20 concept lines, first 15 entity lines, and only synthesis-enumerated raw bodies up to 30,000 chars; longer bodies omitted.
Selection predicate: Domain goldens, optional limit, domain-tagged pages and synthesizes source IDs.
Invalidation or expiry: Each run reassembles context; malformed JSON can rescue a numerical score via regex.
Activation or effect: Records and returns an assessment, not a mandatory acceptance gate for wiki prose.
Evidence limits: SRC-1 src/gateway/evaluate/judge.py, src/gateway/evaluate/wiki_context.py: score is trusted from model output, not independently recomputed; no observed evaluation or causal outcome supplied.
implementation conclusion status: wired
operation conclusion status: uninspected

write_run(summary)
append_to_trend(summary)
return summary
--- src/gateway/evaluate/runner.py

EPI-RTE-retrieval-measure

Label: Golden retrieval measurement

CLI eval-retrieval selects FTS authority ranking or grep baseline; evaluator compares first expected slug rank and calculates hit-at-k/MRR. Curated expected slugs are the reference outcomes. Code owns measurement; standalone CLI is advisory, while policy-edit consumes the same FTS metric as a blocking floor. Distinct identity from policy-edit: this is its measurement dependency, not policy replacement. It does not measure the full hybrid section assembly or answer truth. SRC-1 src/gateway/evaluate/retrieval_eval.py; implementation layer.

Immediate return: EvalReport and formatted misses/metrics.
Later read-back: inapplicable — standalone evaluator retains no report itself; policy-edit can retain the gate result through its containing write/provenance path.
Delegated visibility: inapplicable — deterministic reference comparison.
Selection predicate: Golden query/domain and expected slug set, ranked top k.
Invalidation or expiry: Recomputed from current index; expected labels are not source-truth checks.
Activation or effect: Standalone report advises; policy-edit gate can reject under its floor.
Evidence limits: SRC-1 src/gateway/commit_gate.py: policy-edit tests current FTS recall, not every effect of the proposed policy. A hit means at least one expected slug was found, not all expected pages.
implementation conclusion status: wired
operation conclusion status: uninspected

    for i, slug in enumerate(ranked, start=1):
        if slug in g.expect:
            best = i
            break

--- src/gateway/evaluate/retrieval_eval.py

EPI-RTE-filter-calibration

Label: Candidate criterion calibration

Part of: RT-RTE-policy-distillation

This is a material subfunction, not a replacement of its supplied route container.

Distillation optionally invokes score_calibration with proposed inclusion/exclusion criteria. A model classifies a source excerpt; code compares include/exclude against saved calibration labels and computes precision/recall/F1. Candidate metrics are retained with candidate YAML, not used here to install a live policy. Human review/promotion remains declared; readiness routine only recommends distillation. This is bounded label comparison; label accuracy and independence from training examples are not established. SRC-1 src/gateway/ops/calibration.py; implementation layer.

Immediate return: CalibrationMetrics or None for absent/empty set.
Later read-back: Candidate policy versions retain optional metrics for review; future live deployment unobserved.
Delegated visibility: Classifier receives proposed criteria and first 1,500 characters of raw file text, including frontmatter when present; missing source becomes source_id text.
Selection predicate: Calibration entries by domain; include substring decides predicted label.
Invalidation or expiry: Explicit label addition is idempotent by source ID; metrics recomputed.
Activation or effect: Annotates a retained candidate; no measured-quality threshold promotes it.
Evidence limits: SRC-1 src/gateway/ops/filter_calibrator.py recommends a command once a domain crosses count threshold; it does not run distillation. This resolves runtime’s locally uninspected calibration implementation, without establishing results.
implementation conclusion status: wired
operation conclusion status: uninspected

    raw = client.call(prompt).strip().lower()
    predicted = "include" if "include" in raw else "exclude"

--- src/gateway/ops/calibration.py

EPI-RTE-contradiction-scan

Label: Candidate contradiction scan

Lint/weekly sweep/drift invoke a shared model check: extract claim-shaped sentences by domain, truncate to 280 chars, partition into batches of 80, request at most 12 directly conflicting pairs, parse/check pair indices. Model proposes contradiction and rationale; code removes invalid/duplicate pairs. No external evidence oracle adjudicates which claim is true. Weekly sweep writes a draft synthesis directly; drift stores snapshots and labels disappeared pair keys resolved. Separate from queued deposit conflict selection and from explicit structured-page resolution. SRC-1 src/gateway/lint/contradictions.py; implementation layer.

Immediate return: Warnings/pair candidates; failed calls become warnings; malformed JSON can become empty pairs.
Later read-back: Weekly draft synthesis reaches later wiki retrieval; drift JSON reaches next comparison and optional weekly digest.
Delegated visibility: Model receives capped sentence text with page names, not complete source evidence.
Selection predicate: At least four claims per domain, selected page types, contiguous batches; cross-batch contradictions not compared.
Invalidation or expiry: Existing domain/week page skips regeneration; daily snapshot replaces same-date output.
Activation or effect: Reports possible conflicts; weekly driver directly retains them for review without apply_plan; drift does not itself resolve page claims.
Evidence limits: SRC-1 src/gateway/ops/contradiction_sweeper.py, src/gateway/ops/contradiction_drift.py: disappeared detector output is not evidence of successful factual repair; no human acceptance trace supplied.
implementation conclusion status: wired
operation conclusion status: uninspected

    "Task: identify pairs of claims that ASSERT CONTRADICTORY FACTS within this",
    "      domain. A pair must directly conflict on a specific empirical or",
    "      definitional point — not differ in scope, framing, or context.",

--- src/gateway/lint/contradictions.py

EPI-RTE-conflict-selection

Label: Deposit conflict precedence

Part of: RT-RTE-deposit-commit

This is a material subfunction, not a replacement of its supplied route container.

CommitGate extracts first new and existing prefix-matching claim objects for claim_subject; unequal objects invoke auto_resolve. Code proposes the pair and selects winner by source-type trust then lexicographic committed_at. No truth oracle or source evidence assessment decides the conflict. New-side source_type is read from caller-supplied intent.identity; existing side is read/inferred from source page. The rule does not directly consume self-reported numeric trust, but this is weaker than authenticating source category. Normal asynchronous deposit mode needs no per-content human approval. SRC-1 src/gateway/ops/contradiction.py; implementation layer.

Immediate return: Winner/loser resolution act; appended JSONL act.
Later read-back: Recorded act reaches reversal producer and materialization; both claims remain page content.
Delegated visibility: inapplicable — deterministic rule; no model evaluator.
Selection predicate: Entity/concept identity with claim_subject, one write, existing page, prefix extraction and object inequality.
Invalidation or expiry: Explicit queued reversion can mark act reverted; it is not a new truth check.
Activation or effect: Labels preference and permits contested-edge materialization; does not delete losing claims.
Evidence limits: SRC-1 src/gateway/commit_gate.py: sends committed_at values old and new, so equal-trust comparison favors old ("old" > "new"), contrary to actual-recency wording. Truth of either claim remains undetermined.
implementation conclusion status: wired
operation conclusion status: uninspected

    ca = str(side_a.get("committed_at", ""))
    cb = str(side_b.get("committed_at", ""))
    winner, loser = (side_a, side_b) if ca > cb else (side_b, side_a)

--- src/gateway/ops/contradiction.py

EPI-RTE-conflict-materialization

Label: Contested edge materialization

Part of: RT-RTE-deposit-commit

This is a material subfunction, not a replacement of its supplied route container.

Gate consumes the selected loser and appends a Contested section with a disputes link and loser claim, then commits through normal write machinery. Code admits and retains this annotation; both claims remain available. This is operational annotation, not evidence-consuming acceptance of the winner. Numeric trust affects page ranking through source_type/filter_score, but ordinary authored entity/concept pages default neutral; inspected index does not read the resolution act to separately demote the losing line. SRC-1 src/gateway/commit_gate.py; implementation layer.

Immediate return: Rewritten AuthoredIntent enters commit path.
Later read-back: Contested prose/citation retained in page; act retained separately; later retrieval may supply either claim.
Delegated visibility: inapplicable — deterministic rendering; future answer model interprets page text.
Selection predicate: Conflict act loser fields; normal gate commit success required.
Invalidation or expiry: Queued reversion removes Contested sections and marks act reverted after success.
Activation or effect: Changes later model-visible context through an interpretive disputes annotation, not a line-level enforcing truth gate.
Evidence limits: SRC-1 src/gateway/search_index.py: source-type page trust is a ranking tiebreaker, never eligibility exclusion. Gate’s losing-claim down-weight wording is not established as claim-specific ranking.
implementation conclusion status: wired
operation conclusion status: uninspected

    loser_src = str((act.get("loser") or {}).get("source") or "")
    loser_claim = str((act.get("loser") or {}).get("claim") or "")

--- src/gateway/commit_gate.py

EPI-RTE-resolution-revert

Label: Conflict annotation withdrawal

CLI/MCP revert-resolution queues an act reversal; gate reloads act, scans sorted wiki pages for the first Contested section containing the loser-source disputes token, strips Contested sections, commits and marks act reverted. Caller requests withdrawal; code can veto unknown act/missing matching page. No fact oracle validates the reversal. Matching is by source token, not exact act/page lineage, so several acts citing the same source can be ambiguous. Separate from structured contradiction resolve, which merely writes status/note and can set contested on source pages. SRC-1 src/gateway/commit_gate.py; implementation layer.

Immediate return: Queued receipt, then commit result or dead-letter.
Later read-back: Edited page and reverted-act status survive; both original claims remain.
Delegated visibility: inapplicable — code scans retained pages and acts.
Selection predicate: Named act and first sorted matching loser-source token/Contested page.
Invalidation or expiry: Successful commit marks act reverted; section stripping can remove several Contested sections.
Activation or effect: Withdraws annotation from later context; it does not establish the opposing claim.
Evidence limits: SRC-1 src/gateway/ops/contradiction.py: explicit status resolution neither chooses a factual winner nor rewrites their supporting prose. Other automatic reopen detectors were inventoried, not fully traced.
implementation conclusion status: wired
operation conclusion status: uninspected

    result = self._commit_reversal_writes(
        intent_id, {target_rel: new_content}, [], basis,
        summary=f"reverted contradiction act {act_id}",
    )
    if result.success:
        contradictions_log.mark_act_reverted(act_id, intent_id)

--- src/gateway/commit_gate.py

EPI-RTE-retraction-monitor

Label: PubMed status acquisition

Registered pubmed-retractions poller fetches NCBI XML for retained PubMed sources, checks PublicationType/CommentsCorrections substring retraction, and stamps raw frontmatter while preserving body. Program decides a status annotation; external registry supplies evidence, not a direct truth verdict on every source claim. Open maintenance mode uses a 30-day cursor rule and leaves no reset path when later response is non-retracted. Parse failure returns false, not an explicit unknown result. SRC-1 src/gateway/pollers/pubmed_retractions.py; implementation layer.

Immediate return: Counts/errors and retained retracted metadata.
Later read-back: Raw metadata reaches retracted-citations lint; source bodies stay retained.
Delegated visibility: inapplicable — deterministic XML checker.
Selection predicate: PubMed ID, cursor age, broad case-insensitive retraction substrings.
Invalidation or expiry: No automatic clearing of a retracted flag; cursor controls rechecks.
Activation or effect: Updates a bibliographic warning input; does not remove source or revoke all downstream claims.
Evidence limits: XML containing Retraction of Publication also satisfies substring match; detector is broader than checking only retracted-original status. Provider evidence authenticity and individual claims remain outside the check.
implementation conclusion status: wired
operation conclusion status: uninspected

        if retracted and not front.get("retracted"):
            front["retracted"] = True
            front["retracted_at"] = _now_iso()
            new_text = fm.serialize(front, body)
            write_atomic(src, new_text)

--- src/gateway/pollers/pubmed_retractions.py

EPI-RTE-retraction-warning

Label: Retraction dependency warning

Lint collects flagged raw PubMed/arXiv IDs, scans wiki links for direct cites and invokes cascade_detail for transitive synthesizes dependents. Code returns ERROR direct findings and WARNING transitive findings asking review. No automatic claim replacement, deletion or retrieval exclusion follows in this lint function. Human/operator chooses repair; no mandatory acceptance criterion is recorded here. SRC-1 src/gateway/lint/retracted_citations.py; implementation layer.

Immediate return: LintFinding list with source and depth.
Later read-back: Lint report/operator context can retain findings; inspected function only returns them.
Delegated visibility: inapplicable — deterministic link/dependency warning.
Selection predicate: Raw retracted flags, direct source links and transitive dependency helper.
Invalidation or expiry: Recomputed; flag and linked prose can remain after warnings.
Activation or effect: Advises review without withdrawing retained prose from answer context.
Evidence limits: Transitive graph helper implementation was not fully inspected; complete dependency coverage is not claimed. Source-local search of src/gateway/ops/retrieve.py and src/gateway/search_index.py for retracted/superseded/contested/trust/resolution/disputes found no retracted eligibility test; this is not a negative about every downstream consumer.
implementation conclusion status: wired
operation conclusion status: uninspected

                message=(
                    f"page cites retracted source `{source_id}` — "
                    "review the claim(s) this source supports and remove "
                    "or replace the citation (QUAL-7)"
                ),

--- src/gateway/lint/retracted_citations.py

EPI-RTE-orphan-selection

Label: Corpus de-path candidate selection

CLI/MCP remediate enumerates wiki pages, skips moc/artifact and any indexed inbound count or recursively referenced provenance decision-basis path, and enqueues a typed depath intent otherwise (dry-run only returns candidates). Program chooses cleanup candidates; no semantic truth evaluator or answer oracle participates. Separate from deposit and authored-page admission: this consumer uses graph reachability, not evidence for claims. Selection occurs before later queue draining. SRC-1 src/gateway/ops/remediate.py; implementation layer.

Immediate return: Candidate paths and skipped reachable paths; live mode queues intents.
Later read-back: Queue intents reach commit-worker; retained provenance protects pages during later scans.
Delegated visibility: inapplicable — deterministic graph/index selectors.
Selection predicate: Zero inbound and not provenance-referenced; exempt page types.
Invalidation or expiry: Content-addressed enqueue; next scan recomputes eligibility; gate does not recheck reachability in inspected depath helper.
Activation or effect: Grants delayed removal authority to selected intents, not epistemic rejection.
Evidence limits: Concurrent new references after selection are not protected by an apply-time eligibility recheck in the inspected helper; no concurrency execution supplied.
implementation conclusion status: wired
operation conclusion status: uninspected

    if counts.get(rel, 0) > 0:
        # Has inbound links — safe, skip.
        continue
    # Zero inbound links. Check provenance reachability.
    if rel in prov_reachable:
        skipped_reachable.append(rel)
        continue

--- src/gateway/ops/remediate.py

EPI-RTE-depath-recovery

Label: De-path and recorded restoration

Worker/gate depath captures page text, then calls reversal commit machinery to delete; restore-depath producer retrieves latest recorded content from provenance and queues recreation. Missing target/content vetoes. Caller requests recovery; deterministic program applies it, without evaluating source truth. This record is the paired deletion/restoration mechanism, not orphan selection. Reverse-merge is a different exposed reversal: it strips recorded contributed aliases/sections/exact claim lines from the canonical and deletes the tombstone; its plan helper was not fully inspected. SRC-1 src/gateway/commit_gate.py; implementation layer.

Immediate return: Queued receipt then commit/dead-letter result.
Later read-back: Provenance carries deleted content; restore producer reloads it; recreated page can reenter retrieval.
Delegated visibility: inapplicable — deterministic persistence/recovery.
Selection predicate: Typed depath/restore payload with target; restore content supplied from recorded provenance by public producer.
Invalidation or expiry: Deletion removes canonical file; later restore recreates previous text, not a newly checked proposition.
Activation or effect: Changes retrievability and retained lineage; not acceptance or factual rehabilitation.
Evidence limits: SRC-1 src/gateway/ops/reverse_merge.py: public restoration is provenance-fed. Low-level typed payload handling does not independently authenticate restored truth. Full reverse-merge lineage/overlap safety remains unassessed.
implementation conclusion status: wired
operation conclusion status: uninspected

    return self._commit_reversal_writes(
        intent_id, {}, [target_rel], basis,
        summary=f"de-pathed {target_rel}",
    )

--- src/gateway/commit_gate.py

EPI-RTE-filter-correction

Label: Caller label correction

CLI/MCP/browser filter-correct takes include/exclude and rationale; validates mutable raw-frontmatter diff, stamps user_correction and pins a user-correction example. Caller proposes the label; code vetoes malformed input, unknown source/domain and diff failure. A user-correction name does not authenticate a human caller. Next filter prompt can use selected corrected examples; this is direct editorial decision adaptation, not source-claim acceptance. SRC-1 src/gateway/ops/filter_correct.py; implementation layer.

Immediate return: Corrected result and pinned example.
Later read-back: Raw correction block/example bank survive and can enter later filter prompts.
Delegated visibility: Later filter model receives selected examples/rationales, not necessarily this example on every call.
Selection predicate: Known raw source/domain; explicit binary decision; example selection remains separate.
Invalidation or expiry: Repeated correction overwrites correction block; raw body unchanged.
Activation or effect: Changes future editorial guidance through natural-language rationale and binary label; no improved capacity shown.
Evidence limits: SRC-1 src/gateway/filter/semantic.py: model consumes policy plus examples, not a certified truth oracle; correction need not immediately create/delete the existing wiki source page.
implementation conclusion status: wired
operation conclusion status: uninspected

    new_filter["user_correction"] = correction_block
    front["filter"] = new_filter

--- src/gateway/ops/filter_correct.py

Additional implementation passages for the linked records:

ident = dict(identity)
ident.setdefault("page_type", payload["page_type"])
iid = compute_intent_id(payload, ident, semantics="deposit")

--- src/gateway/ops/deposit.py

        "committed_at": "old",

--- src/gateway/commit_gate.py

        "committed_at": "new",

--- src/gateway/commit_gate.py

        "source_type": str(ident_d.get("source_type", "")),

--- src/gateway/commit_gate.py

        has_citation = (
            bool(_CITATION_LINE_RE.search(raw_line))
            or bool(_NLM_CITATION_LINE_RE.search(raw_line))
            or _line_cites_via_footnote(raw_line, footnote_map)
        )

--- src/gateway/citations.py

if src_type:
    trust_val = _trust.server_trust_tier(src_type, filter_score)
else:
    trust_val = _trust.NEUTRAL_TRUST

--- src/gateway/search_index.py

key += _W_TRUST * (h.trust - 0.5)
return key

--- src/gateway/search_index.py

page = _build_page(domain, week, all_findings)
out_path.parent.mkdir(parents=True, exist_ok=True)
with file_lock(_LOCK_NAME):
    write_atomic(out_path, page)

--- src/gateway/ops/contradiction_sweeper.py

new = [x for x in current if _finding_key(x) not in prev_keys]
resolved = [x for x in previous if _finding_key(x) not in curr_keys]
return new, resolved

--- src/gateway/ops/contradiction_drift.py

                if len(body) <= max_body_chars:
                    bodies[sid] = body

--- src/gateway/evaluate/wiki_context.py

    # RULE 2 (authority): same-kind + alias/canonical exact-or-normalized match → merge.
    if id_names & cand_names:
        return Verdict("merge", c.slug, "alias-canonical-exact-match", basis)

--- src/gateway/dedup.py

# Skip any prose preamble before the first `{`. Sonnet sometimes emits
# "Here is the evaluation:\n\n{...}" despite the strict-output instruction.

--- src/gateway/evaluate/judge.py

        if predicted == "include":
            fp += 1
        else:
            fn += 1

--- src/gateway/ops/calibration.py

Claims

EPI-CLM-retrieval-grounding

Label: Bounded cited retrieval promise

Claimed operation: ranked bounded context with preserved citations grounds later answering. SRC-1 README.md; doctrine/design. Distinct from RT-CLM-gateway-only: this claim concerns read delivery, not write discipline.

wiki retrieve "<question>" [--domain X] — the default grounding call: a ranked, bounded context block of the most relevant sections, citations preserved. Prefer over reading index.md or grepping. --- README.md

EPI-CLM-answer-grounding

Label: Grounded local synthesis promise

Claimed operation: answer performs grounded synthesis and strips ungrounded citations. SRC-1 README.md; doctrine/design. Distinct from RT-CLM-gateway-only and EPI-CLM-retrieval-grounding: this concerns newly returned answer prose after model inference.

| wiki answer "<question>" [--domain X] [--file] | Retrieve + one grounded Claude call; cites only retrieved sources (ungrounded citations stripped). NotebookLM-independent. --file drafts a synthesis page | --- README.md

EPI-CLM-citation-provenance

Label: Opaque corpus-token provenance claim

Claimed operation: NLM corpus tokens count as valid grounding and prove provenance to the page’s corpus, despite no local UUID existence verification. SRC-1 WIKI.md; doctrine/design. Distinct from RT-CLM-gateway-only and EPI-CLM-answer-grounding: this claims token/corpus provenance rather than local answer ID membership.

NotebookLM-internal corpus citations: [[nlm:<uuid>]]. NLM-authored synthesis pages (those carrying nlm_notebook_id: in frontmatter) frequently cite the corpus via chunk-level UUIDs that NotebookLM emits inline (e.g., [[nlm:a17f0b4f-ba96-46b0-b4e1-6be759219e89]]). These are accepted as valid grounding by the validator alongside [[sources/<id>]] — the citation maps into the page's NLM corpus, not a wiki source page. The validator does not verify UUID existence (that would require live NLM API access; out of scope for the in-page validator). Treat NLM citations as opaque grounding tokens — they prove provenance to the corpus listed in nlm_notebook_id, but do not by themselves create wiki backlinks. --- WIKI.md

Evidenced absences

none declared in this member

Behavioral-authority paths

none declared in this member; route-specific consumer/channel/force/horizon assessments are in the ledger, and supplied prompt/host paths remain see RT-BAP-authorship-prompt, RT-BAP-domain-skill, RT-BAP-session-state.

Keys	Action
`?`	Open this help
`n`	Next page
`p`	Previous page
`s`	Search