Kompl epistemic report

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

Source-and-claim boundary

This report asks what Kompl acquires, transforms, checks and licenses for later reliance. It adopts the whole-system scope and exclusions in the boundary, and cites its Source register rather than reproducing it. All additional reads below are SRC-1 implementation or doctrine/design at the registered commit; checkout HEAD was checked against that commit. No target execution was performed.

The operative distinction is between installing a page and accepting its claims against evidence. The inspected routes implement the former. Prompts request faithful extraction, attribution and grounding; those requests are not results of evidence-consuming acceptance. Public knowledge/compounding claims are EPI-CLM-compiled-knowledge, EPI-CLM-source-attribution, and EPI-CLM-wiki-grounding. The operational completion claim remains RT-CLM-durable-completion.

Entry point or operation	Source path and inspection basis	Covering records or remaining limit
Staging, finalization, n8n, compile driver; extraction and resolution	Supplied runtime findings at app/src/app/api/compile/run/route.ts, app/src/app/api/compile/extract/route.ts, app/src/app/api/compile/resolve/route.ts; direct extraction prompt/recovery reads at nlp-service/services/llm_client.py	RT-RTE-session-compile, RT-RTE-plan-draft, EPI-RTE-extraction-recovery. Conversion branches and resolution internals not independently audited here; no complete source-fidelity or identity-correctness verdict.
Match and plan	Search of app/src/app/api/compile/match/route.ts; direct planning rules and triage implementation at app/src/app/api/compile/plan/route.ts, nlp-service/services/llm_client.py	EPI-RTE-triage-disposition.
Draft and schema	Supplied app/src/app/api/compile/draft/route.ts, app/src/app/api/compile/schema/route.ts; direct drafting prompts at nlp-service/services/llm_client.py	RT-RTE-plan-draft, RT-RTE-schema-bootstrap. No generated instance or semantic comparison supplied.
Compile crossref; exposed legacy model crossref	Direct complete app handler and injector at app/src/app/api/compile/crossref/route.ts, app/src/lib/wikilink-injector.ts; direct Python dispatch/generation at nlp-service/routers/pipeline.py, nlp-service/services/llm_client.py	EPI-RTE-crossref-transform. Legacy result has no inspected app admission consumer.
Compile commit, explicit approval/rejection	Direct commit gates and complete app/src/lib/approve-plan.ts; other handlers supplied by runtime	RT-RTE-draft-approval.
Chat answer and automatic/manual draft creation	Direct synthesis/feedback and app/src/app/api/chat/save-draft/route.ts; direct insertChatDraft in app/src/lib/db.ts	RT-RTE-chat-feedback, EPI-RTE-chat-save.
Metadata answers, wiki/MCP delivery	Direct getWikiStats in app/src/lib/db.ts, complete registered tool handlers in mcp-server/index.ts; other readers supplied by runtime	RT-RTE-mcp-read, RT-RTE-chat-feedback. External agent dependence remains outside boundary.
Lint and startup hook	Direct complete app/src/app/api/wiki/lint-pass/route.ts, cli/src/startup-tasks.ts, CLI dispatch in cli/src/index.ts; contradiction display in app/src/components/WikiLintResults.tsx	EPI-RTE-wiki-lint. CLI dispatch contains lifecycle, backup and update commands, not another claim-admission workflow.
Contradiction sidebar	Direct app/src/app/api/wiki/[page_id]/contradictions/route.ts and plan activity producer	EPI-RTE-triage-disposition. Advisory challenge, not automatic withdrawal of challenged prose.
Source deletion, redraft, source recompile and import	Supplied maintenance/import implementation; direct source recompile handler at app/src/app/api/sources/[source_id]/recompile/route.ts	RT-RTE-source-maintenance, RT-RTE-archive-import. Recompile queues ordinary session processing, not a new semantic evaluator.
Zombie cleanup and pending-draft dependency cleanup	Direct app/src/app/api/admin/cleanup/zombie-pages/route.ts, findZombiePages and cleanupChatDraftsForDeletedPage in app/src/lib/db.ts	EPI-RTE-zombie-cleanup.
Saved Links regeneration/dismissal	Direct complete app/src/lib/saved-links.ts, app/src/app/api/saved-links/[failure_id]/route.ts; query searched in app/src/lib/db.ts	EPI-RTE-saved-links.
Python registered operations	Direct nlp-service/main.py, pipeline dispatch and nlp-service/routers/llm.py; extraction/resolution/vector/storage routes supplied by runtime	No hidden provider internals assessed; direct storage writes are not evidence of checked claims. Exhaustive metadata/conversion branches unassessed, preventing system-complete negatives.
Product evaluators and provider reports	Direct May 5 runner/report; selected assertions searched in scripts/integration-test.sh; May 22 pending report supplied by boundary	EPI-RTE-corpus-evaluation. Other manual harnesses and unit assertions not read comprehensively; no claim that all shipped tests lack semantic oracles.

Missing evidence is local to its prevented conclusion: absent current-revision traces prevent observed grounding, evaluator activation and acceptance claims; missing exact historical corpus prevents inspecting the historical claims against their original texts; unavailable provider internals prevent inference about hidden reasoning or parameter learning; no controlled intervention/outcome comparison establishes causal improvement from compounding, citations, schema or any individual model. Available uninspected harnesses and conversion branches are inspection limits, not access gaps. No system-complete absence of semantic checking is asserted.

Epistemic-object inventory

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
EPI-OBJ-extraction-output	Source assertions, identity assignments, relationships and summary	see record	SRC-1 nlp-service/services/llm_client.py	Structured shape and confidence words do not verify source truth or entailment.
EPI-OBJ-draft-content	Proposed wiki statements and source reproduction	see record	SRC-1 nlp-service/services/llm_client.py	Individual content is unobserved; retained raw original-source content is imported, not model-produced knowledge.
RT-OBJ-page-plans	None in the control portion	see record	SRC-1 app/src/app/api/compile/commit/route.ts	A status transition is not a factual acceptance decision.
RT-OBJ-schema	Primarily instructions, not an established truth-apt theory instance	see record	SRC-1 nlp-service/services/llm_client.py	Exact generated rules, premises and assumptions uninspected.
RT-RTE-chat-feedback	Answer prose; selected page text and previous answers	see record	SRC-1 app/src/app/api/chat/route.ts	Counts describe delivered candidates, not actual citation use.
EPI-RTE-triage-disposition	Candidate novelty/contradiction judgment and reason	see record	SRC-1 app/src/app/api/compile/plan/route.ts	Comparison is against summary, not a source-truth oracle.
EPI-RTE-wiki-lint	Database anomaly predicates; model contradiction reports	see record	SRC-1 app/src/app/api/wiki/lint-pass/route.ts	Summary sample is limited; empty result also follows failed scan.
EPI-RTE-saved-links	Import-failure descriptions and count; cleared-state assertion	see record	SRC-1 app/src/lib/saved-links.ts	Dismissal can produce a false imported-all assertion.
EPI-RTE-corpus-evaluation	HTTP/shape/count outcomes about extraction	see record	SRC-1 validation/2026-05-05-multi-provider-corpus/runner.py	No semantic reference answers in that runner; reported operation is historical.
RT-CMP-gemini, RT-CMP-deepseek, RT-CMP-spacy, RT-CMP-key-bert, RT-CMP-mini-lm	Not determinable as individuated propositions in accessible parameter state	see records	SRC-1 nlp-service/services/llm_client.py	Inference/derived-vector writes do not show training; exact weights and hidden provider state unavailable.

Authority-route ledger

The IDs below denote canonical routes or their cited material parts, not new declarations made by table rows. Architectural status refers to inspected wiring; all current-revision operation remains unobserved. Each row separates checking, disposition and retention. “No factual license” means the route supplies no demonstrated factual warrant beyond its named check, not that its content must be false.

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 source ID and local anchor	Claim IDs or none	Mismatch marker or none	Gap/limit
RT-RTE-session-compile	content transformation	implemented	see record	truth-apt transformation: acquisition/import	External sources to raw markdown	Conversion dependencies; source domain open	see record	see record	Imports selected material	Source warrant unknown; retained identity is not verification	see record	see record	SRC-1 app/src/app/api/compile/run/route.ts	EPI-CLM-compiled-knowledge	none	Conversion fidelity not independently established.
RT-RTE-plan-draft	content transformation	implemented	EPI-OBJ-extraction-output, EPI-OBJ-draft-content	truth-apt transformation: indeterminate	Sources/dossiers to extraction and prose	Models interpret source and prior prose; open subjects	see record	see record	Candidate generation	Requested faithfulness, not checked entailment	see record	see record	SRC-1 nlp-service/services/llm_client.py	EPI-CLM-source-attribution	none	Reshaping versus ampliative content requires actual paired inputs/outputs. Non-entailment not established by synthesis alone.
EPI-RTE-extraction-recovery	check/evidence production	implemented	EPI-OBJ-extraction-output	no content change	Provider output shape	Parsed response/Pydantic schema	After model return	Parsed model or parse error	Gate to return	Shape only	Reject malformed output unless recovery succeeds	Extraction caller; parsed-model channel; enforcing; current call	SRC-1 nlp-service/services/llm_client.py	none	none	Does not test assertions against source.
EPI-RTE-extraction-recovery	content transformation	implemented	EPI-OBJ-extraction-output	truth-apt transformation: indeterminate	Failed response or shortened input	JSON repair; Gemini-specific halved-input fallback	Parse failure	Salvaged extraction or regenerated candidate	Return recovered candidate	No added source-truth license	May continue without complete initial input/output	Extraction caller; return value; permissive; current extraction and later drafts	SRC-1 nlp-service/services/llm_client.py	none	none	Recovery can remove coverage; semantic preservation not established.
EPI-RTE-triage-disposition	check/evidence production	implemented	see record	truth-apt transformation: indeterminate	Novelty/conflict of selected source claims against page summary	Generation model, not expected-answer oracle	Ranked candidates then model call	update/contradiction/skip with reason	Result delivered to planner	Candidate comparison only	see record	see record	SRC-1 nlp-service/services/llm_client.py	EPI-CLM-source-attribution	none	Similarity is candidate selection, not warrant.
EPI-RTE-triage-disposition	disposition/acceptance	implemented	see record	no content change	Result to update/provenance/challenge	Fixed planner branches	After triage	Update plan; provenance-only plan; contradiction activity	Enforcing branches, advisory challenge	No recorded acceptance of source truth	see record	see record	SRC-1 app/src/app/api/compile/plan/route.ts	EPI-CLM-source-attribution	Contradiction does not imply automatic withdrawal	Other rules may still draft related pages.
EPI-RTE-crossref-transform	content transformation	implemented	EPI-OBJ-draft-content	truth-apt transformation: indeterminate	Plain titles/aliases to linked prose	Regex and supplied title map; identity assumptions external to transform	Drafted before commit	Linked prose	Stored crossreffed draft	Prose reshaping conditional on correct title/alias mapping; link association not independently warranted	see record	see record	SRC-1 app/src/lib/wikilink-injector.ts	EPI-CLM-source-attribution	No contradiction scan in app crossref	Legacy exposed model endpoint is a separate alternative within record.
RT-RTE-draft-approval	check/evidence production	implemented	EPI-OBJ-draft-content	no content change	Nonempty prose; compile body length; explicit pending state	Software predicates, not factual oracle	Before write	Pass/thin/failed/not-pending	Enforcing scoped guards	Length/status/nonemptiness only	see record	see record	SRC-1 app/src/app/api/compile/commit/route.ts	none	none	Explicit approval omits compile length gate; original-source exempt.
RT-RTE-draft-approval	disposition/acceptance	implemented	EPI-OBJ-draft-content	no content change	Install, queue or reject proposal	Auto-approve default; otherwise caller approval/rejection	Before installation	see record	Enforcing operational admission	Human evidence review not established; no named factual acceptance criterion recorded	see record	see record	SRC-1 app/src/lib/approve-plan.ts	EPI-CLM-compiled-knowledge	Installation is not epistemic acceptance	Caller identity unresolved.
RT-RTE-draft-approval	retention	implemented	EPI-OBJ-draft-content	no content change	Installed prose/provenance/indexes	SQLite outbox and file flush	Following admission	see record	Persistent availability	Retention of a candidate, not lifecycle integration after factual acceptance	see record	see record	SRC-1 app/src/lib/approve-plan.ts	RT-CLM-durable-completion	none	File-first reads can expose old content after failed flush.
RT-RTE-chat-feedback	operational admission/selection/consumption	implemented	see record	no content change	Wiki evidence delivered for answering	Model selection or hybrid scoring, clipping/fallback	Each question	see record	Ranking and context selection	Selected snippets have no new warrant	see record	see record	SRC-1 app/src/app/api/chat/route.ts	EPI-CLM-wiki-grounding	none	Selected pages may be popular fallback, not sufficient evidence.
RT-RTE-chat-feedback	content transformation	implemented	see record	truth-apt transformation: entailed derivation	Database row counts/timestamps to metadata answer	Fixed query/template; database-defined population	Recognized metadata question before retrieval/synthesis	Counts and latest stored timestamps	Direct answer/history insertion	Counts entailed by database predicates, not evidence that every counted source contributed or every page is true	Bypasses model answering	Chat user; answer text; advisory; current turn and retained history	SRC-1 app/src/lib/db.ts	none	Scope of counts narrower than compiled-knowledge quality	Semantic fidelity of each metadata wording not exhaustively assessed.
RT-RTE-chat-feedback	content transformation	implemented	see record	truth-apt transformation: indeterminate	Selected prose/history to answer	Model; wiki-only request	After retrieval	see record	Returned and retained answer	Requested grounding, no verified entailment or citation correspondence	see record	see record	SRC-1 nlp-service/services/llm_client.py	EPI-CLM-wiki-grounding	none	No actual answer pair supplied.
RT-RTE-chat-feedback, EPI-RTE-chat-save	retention	implemented	see records	truth-apt transformation: non-ampliative reshaping	Answer to pending wiki draft	Retrieved count or explicit content-supplying caller	After answer or save request	see records	Pending candidate, not accepted claim	No new factual license from FAQ wrapper	see records	see records	SRC-1 app/src/app/api/chat/save-draft/route.ts	EPI-CLM-compiled-knowledge	Retrieved count is not verified use	Approval writes page IDs into source-provenance slot.
RT-RTE-schema-bootstrap	behavior/policy adaptation	implemented	RT-OBJ-schema	non-truth-apt policy/content update: generated drafting conventions	Page metadata to standing instructions	Model generator, nonempty write/existence checks	Initial bootstrap; later draft delivery	see record	Requested model guidance	None established for any factual premises of generated conventions	see record	RT-BAP-schema-drafting	SRC-1 nlp-service/services/llm_client.py	none	none	Compliance, criticism and improvement unobserved.
EPI-RTE-wiki-lint	check/evidence production	implemented	see record	truth-apt transformation: indeterminate	Database anomalies and summary conflicts	SQL predicates; model for at most ten same-category pairs	Manual or overdue startup	Findings or empty contradiction list	Return/log/display, advisory	SQL facts within database; model conflicts remain candidate judgments	see record	see record	SRC-1 app/src/app/api/wiki/lint-pass/route.ts	EPI-CLM-compiled-knowledge	Clean scan not certification	Failed model scan collapses to empty list.
EPI-RTE-zombie-cleanup	check/evidence production	implemented	see record	truth-apt transformation: entailed derivation	Existence of joined provenance/source rows	SQL NOT EXISTS, database domain	Dry-run or cleanup	Zombie candidates	List supplied to deletion loop	Exact database dependency predicate, not subject-matter falsity	see record	see record	SRC-1 app/src/lib/db.ts	none	none	Depends on lineage identifiers being correctly typed.
EPI-RTE-zombie-cleanup	disposition/acceptance	implemented	see record	no content change	Dependency failure to deletion	Code loop; caller chooses POST	Following query	Database removal and best-effort file/vector removal	Enforcing withdrawal from normal database readers	Withdraws operational availability, not a refutation	see record	see record	SRC-1 app/src/app/api/admin/cleanup/zombie-pages/route.ts	none	none	Pending chat dependency pruning can leave surviving draft prose unchanged.
RT-RTE-source-maintenance	content transformation	implemented	see record	truth-apt transformation: indeterminate	Surviving sources to replacement prose or provenance note	Cascade rules and drafting model	Source deletion	see record	Direct update/deletion	No checked removal of source's semantic contribution	see record	see record	SRC-1 app/src/lib/recompile.ts	EPI-CLM-compiled-knowledge	Bypasses approval/thin gate	Note fallback retains old prose.
RT-RTE-archive-import	disposition/acceptance	implemented	see record	truth-apt transformation: acquisition/import	Archive to restored state	Manifest/safe-ID/empty-wiki guards	Import request	see record	Direct restore	Imported warrant unknown; structural admission only	see record	see record	SRC-1 app/src/app/api/import/route.ts	none	none	Bypasses per-page approval and semantic checks.
EPI-RTE-saved-links	content transformation	implemented	see record	truth-apt transformation: ampliative conjecture	Empty unresolved set to imported-all statement	Fixed template	Failure resolution/dismissal/lint	Cleared-state wiki page	Direct retained assertion	Not entailed: dismissal is counterexample	see record	see record	SRC-1 app/src/lib/saved-links.ts	none	Unsupported cleared-state assertion	Nonempty rows mostly reshape recorded failures; this row assesses the distinct empty branch.
EPI-RTE-corpus-evaluation	check/evidence production	implemented	see record	no content change	Response key presence/counts and request outcomes	Runner; bounded historical corpus mode	Offline operator invocation	see record	Result file/report; no inspected production gate	Protocol/shape observations only, no source-truth oracle	see record	see record	SRC-1 validation/2026-05-05-multi-provider-corpus/runner.py	none	Provider success not semantic superiority	Quota/timeouts and missing corpus bound comparison.
RT-RTE-mcp-read	operational admission/selection/consumption	implemented	see record	truth-apt transformation: non-ampliative reshaping	Stored page/metadata to tool text	Explicit client requests and formatting	Tool invocation	see record	Advisory/context delivery	Preserves delivered prose, not proof of truth or lineage completeness	see record	see record	SRC-1 mcp-server/index.ts	EPI-CLM-compiled-knowledge	none	External scheduling/action uninspected.

The metadata derivation is scoped to the stored database population. Its implementation premise is explicit:

const sourceCount = (db.prepare('SELECT COUNT() AS n FROM sources WHERE status = ?').get('active') as { n: number }).n; const pageCount = (db.prepare('SELECT COUNT() AS n FROM pages').get() as { n: number }).n; --- app/src/lib/db.ts

There is no established lifecycle integration after factual acceptance in these rows. Linking, indexing and later use are real operations, but the inspected admission paths do not supply the required preceding evidence-consuming factual acceptance decision. This is scoped to the named production paths, not to every possible operator practice or uninspected test.

System-claim versus route comparison

Claim ID	Claimed operation or warrant	Claim source ID/anchor and evidence layer	Doctrine/design support	Implemented route IDs	Observed-run support	Causal support and design limits	Supported conclusion	Mismatch/unknown
EPI-CLM-compiled-knowledge	see record	SRC-1 README.md; doctrine/design	Pipeline/planner/update prompts	RT-RTE-plan-draft, RT-RTE-chat-feedback, RT-RTE-schema-bootstrap, EPI-RTE-wiki-lint	No current-revision trace	No intervention/outcome comparison; input caps and manual pending gates restrict coverage	Cross-session accumulation, promotion and update capacity are wired	Seeing patterns across all sources and maintenance-free knowledge improvement not demonstrated; bootstrap is not ongoing criticism.
EPI-CLM-source-attribution	see record	SRC-1 README.md; doctrine/design	Prompts request attribution and conflicts	RT-RTE-plan-draft, EPI-RTE-triage-disposition, EPI-RTE-crossref-transform	No actual claim/citation pair supplied	Three-source relationship threshold is not independent corroboration; extracted mentions may share origin/errors	Provenance retention, comparison planning and advisory challenges implemented	Skip attaches provenance without prose change; conflict can become sidebar activity instead of comparison-page correction. Citation fidelity unverified.
EPI-CLM-wiki-grounding	see record	SRC-1 nlp-service/services/llm_client.py; doctrine/design	Wiki-only and insufficient-information instructions	RT-RTE-chat-feedback	No answer trace	No entailment/citation evaluator inspected in synthesis consumer; clipping can omit needed qualification	Grounding requested over delivered snippets	Provider response schema checks shape, not grounding; returned citations are taken from model output.
RT-CLM-durable-completion	see record	SRC-1 app/src/app/api/compile/run/route.ts; implementation	see record	RT-RTE-draft-approval	see record	see record	File-flush policy, not truth acceptance	Explicit approval and import differ; durable prose can still be an unchecked candidate.

Bounded conclusion

Kompl acquires source texts with identities and retains generated interpretations alongside them. Source warrant remains unknown: source inclusion, extracted confidence labels, mention frequency and provenance counts do not adjudicate subject-matter truth. Generation requests faithful extraction, reproduction, synthesis and citation. Without paired generated artifacts and their actual inputs, these edges remain semantically indeterminate rather than automatically classified as entailed or ampliative. The extraction recovery route can accept repaired structure or a result generated from shorter input; successful parsing does not restore lost source coverage. See EPI-RTE-extraction-recovery and RT-RTE-plan-draft.

The inspected checks license narrower conclusions. SQL can derive database counts and dependency predicates; deterministic title injection can preserve an ordinary proposition's words while adding navigation dependent on a supplied identity map. These have no independent subject-matter warrant beyond their premises. A concrete deterministic counterexample exists in Saved Links: after dismissing the final failure, the empty-set template says all links were imported. The implementation generates a proposition not entailed by its state and demonstrably false for that reachable input history. See EPI-RTE-crossref-transform, EPI-RTE-zombie-cleanup, and EPI-RTE-saved-links.

Operational admission is implemented, but the named routes do not record factual acceptance against evidence for a named intended use. Compile nonemptiness/body-length checks, default auto-approval, explicit caller approval, import structure guards and deletion-redraft rules all permit later wiki/model use without such an acceptance record. A caller may perform meaningful review outside the API; the source does not establish who did so or against which evidence. Candidate retention and use therefore remain distinct from acceptance and subsequent lifecycle integration. Rejection can withhold pending drafts; source/zombie deletion changes availability; neither proves surviving prose has had every removed contribution retracted. See RT-RTE-draft-approval, RT-RTE-archive-import, and RT-RTE-source-maintenance.

Contradiction checks have heterogeneous force. Match triage can change planning, while its contradiction branch creates an advisory sidebar event without automatically withdrawing the page. Lint checks a small summary sample and displays/logs candidate conflicts; a failure can also yield an empty conflict list. Dependency cleanup enforces a database criterion, not semantic refutation. Approval of query-generated drafts also has a lineage caveat: cited/retrieved page IDs occupy a source-ID slot and are then passed to source provenance. Thus stored citation/provenance presence is not a uniform end-to-end warrant path. See EPI-RTE-triage-disposition, EPI-RTE-wiki-lint, EPI-RTE-chat-save, and EPI-RTE-zombie-cleanup.

Retained prose can affect later model context through drafting, answering and pending FAQ feedback, while schema prose directly requests later drafting conventions. These implemented pathways coexist with independently unestablished benefit, learning, content-directed theory criticism and exercised self-improvement. For schema, localized formulation is afforded and delivery is wired; actual decisions depending on its content, criticism and criticism-shaped iteration remain unobserved/uninspected as separated in RT-RTE-schema-bootstrap. There is no basis to infer hidden parameter learning or its absence across opaque providers. The corpus runner supports historical shape/protocol comparisons only; it cannot close a factual-acceptance or causal-improvement gap. No single epistemic verdict follows.

Shared records

Operative objects

EPI-OBJ-extraction-output

Label: Structured source extraction Part of: RT-RTE-plan-draft

This is the material output of the extraction segment of the supplied route, not its encompassing process. It mixes natural-language claims, descriptions and summaries with symbolic JSON fields; app extraction persists it in SQLite and mention tables for later planning/dossiers. Closest supplied object RT-OBJ-page-plans is later control state, not this payload. SRC-1 implementation/doctrine: nlp-service/services/llm_client.py; persistence scope supplied by the parent route. Confidence is a model label about source expression, not a calibrated truth probability.

1. claims — specific factual claims
2. claim: the claim text itself (a string — full sentence)
3. confidence: "stated" (explicit) | "implied" (strongly suggested) | "speculative" --- nlp-service/services/llm_client.py

EPI-OBJ-draft-content

Label: Wiki draft prose Part of: RT-RTE-plan-draft

Natural-language markdown in SQLite page_plans.draft_content, later page files/outbox. This payload is explicitly excluded from the control-state referent of RT-OBJ-page-plans. It is a material output part of the parent process, not a duplicate process. Requested source reproduction and synthesized claims differ from evidence of their production or fidelity. SRC-1 doctrine/design: nlp-service/services/llm_client.py; implementation/persistence supplied by parent and RT-RTE-draft-approval.

Synthesize across sources. Attribute specific claims to their source. --- nlp-service/services/llm_client.py

Routes

EPI-RTE-extraction-recovery

Label: Extraction parse and recovery Part of: RT-RTE-plan-draft

Distinct material parsing/recovery segment of the supplied generation route. The Python function returns parsed extraction, attempts JSON repair, then conditionally repeats Gemini extraction with halved text. SRC-1 implementation: nlp-service/services/llm_client.py. Shape acceptance is not source-truth acceptance.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Parsed/repaired extraction or LLMCompileError.
• Later read-back: Returned output can be persisted by the app and consumed as later dossiers/mentions; no separate recovery memory is claimed.
• Delegated visibility: Initial capped markdown plus NLP hints; fallback sees half the already capped source, with paragraph-boundary trimming.
• Selection predicate: Parsed response first; salvage on parse failure; halved fallback only after unsuccessful salvage with MAX_TOKENS and Gemini model prefix.
• Invalidation or expiry: Errors prevent that return; retry inputs replace the initial candidate for this call, without a semantic completeness test.
• Activation or effect: Return/admission of recovered structure is wired; actual semantic preservation unobserved.
• Evidence limits: Schema and repair validate form, not entailment, and narrowed fallback may lose relevant text.

Proposer: model then repair/fallback code. Decider/veto: parsed-model checks and exception branches. Oracle: response schema for shape, no expected factual answer. Mode: open source extraction; trigger is parse/truncation failure. Guidance remains extraction prompt; repaired/generated output, rather than a proof, persists through the caller. Recovery can fail and surface both errors; no semantic rollback is established.

if result.parsed is not None:
    return result.parsed  # type: ignore[return-value]

--- nlp-service/services/llm_client.py:752-753

half_markdown = markdown[: len(markdown) // 2]

--- nlp-service/services/llm_client.py

EPI-RTE-triage-disposition

Label: Triage comparison and plan disposition Part of: RT-RTE-plan-draft

This material comparison/disposition segment has separate check and disposition ledger rows. SRC-1 implementation/doctrine: nlp-service/services/llm_client.py, app/src/app/api/compile/plan/route.ts, app/src/app/api/wiki/[page_id]/contradictions/route.ts. Models compare selected claims against a page summary; app rules map the result to a plan or activity event. Comparison-page creation independently requires a supported relationship type, historical count at least three and both entity plans; contradiction triage alone does not guarantee one.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Model decision/reason, then plans and contradiction activity.
• Later read-back: Provenance-only/update plans can affect later pages; persisted challenge events feed contradiction sidebar API.
• Delegated visibility: Selected source claims, existing summary and title; not all underlying sources or page text.
• Selection predicate: Parent route's ranked candidate filter, then update/contradiction/skip; unknown returned decision normalized to skip.
• Invalidation or expiry: A branch can omit a redraft; no automatic challenge resolution or expiry established by inspected plan/reader.
• Activation or effect: Update schedules drafting; skip may attach provenance; contradiction logs a challenge without a plan from this branch.
• Evidence limits: Other planner rules may independently target the page; model conflict does not prove either proposition false.

Proposer/evaluator: triage model; decider: fixed planner; user can cancel/exclude sources but does not necessarily review each result. No reference-answer oracle. Mode: open maintenance. Guidance defines novelty, contradiction and redundancy; reason/activity and plans persist. Operational challenge force is advisory via sidebar API; planner force is enforcing within this branch. No factual acceptance or semantic rollback established.

} else if (match.decision === 'contradiction') {
  // Log a detailed activity row. The getPageContradictions(pageId) helper
  // in db.ts json_extracts from here — keep the field names stable.
  const srcMeta = sessionSourceMetaById.get(match.source_id);
  logActivity('page_contradiction_detected', {

--- app/src/app/api/compile/plan/route.ts

EPI-RTE-crossref-transform

Label: Crossref text transformation Part of: RT-RTE-plan-draft

The parent route's crossref segment is deterministic in the app. SRC-1 implementation: app/src/app/api/compile/crossref/route.ts, app/src/lib/wikilink-injector.ts. Code-derived witness, without execution: input MiniLM encodes text. MiniLM repeats. with target title MiniLM model and alias MiniLM becomes [[MiniLM model|MiniLM]] encodes text. MiniLM repeats. Registration admits the six-character alias, matching finds its first word-boundary occurrence, canonical-used tracking blocks the second, and the alias branch preserves displayed surface text. A plain proposition is reshaped, but the added identity/navigation association relies on the supplied alias mapping. Conflicting aliases are first-writer-wins, so no independent identity proof follows.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: App stores linked drafts and returns counts with contradictions_found zero; exposed Python alternative returns model-edited pages/conflict candidates.
• Later read-back: App crossreffed text enters commit and later wiki consumers; no inspected app read-back of the exposed legacy model result.
• Delegated visibility: App receives draft text and targets/aliases; legacy Python generation sees provided pages, prefix-capped prompt and titles.
• Selection predicate: App nonnull drafted content; excludes own title; four-character surfaces, word boundaries and one link per canonical title; skips frontmatter/headings/code/existing links.
• Invalidation or expiry: Session draft state controls ordinary processing; identity map replacement can change future associations; no retroactive link-identity verification here.
• Activation or effect: App deterministic draft update wired; legacy endpoint model transformation wired but not current compile driver admission.
• Evidence limits: Regex cannot adjudicate the subject of a homonymous mention; legacy model request does not establish faithful reshaping or contradiction accuracy.

Proposer: input drafts and title/alias map; decider: fixed app scanner, or model for exposed Python alternative; caller supplies alternative input. No factual oracle. Mode: open compilation/direct API. Guidance in the legacy prompt requests linking/conflict notes; generated results remain returned candidates absent a demonstrated consumer. Normal rollback is through later plan/content handling, not a semantic check in this transformer.

  if (match.toLowerCase() === canonical.toLowerCase()) {
    return `[[${canonical}]]`;
  }
  return `[[${canonical}|${match}]]`;

--- app/src/lib/wikilink-injector.ts

    pages_updated: pagesUpdated,
    wikilinks_added: wikilinksAdded,
    contradictions_found: 0,

--- app/src/app/api/compile/crossref/route.ts

EPI-RTE-chat-save

Label: Manual chat draft and lineage

No counterpart for the explicit content-supplying endpoint after comparison with RT-RTE-chat-feedback, which describes automatic retrieved-count FAQ creation. Both converge on RT-RTE-draft-approval. SRC-1 implementation: app/src/app/api/chat/save-draft/route.ts, app/src/lib/db.ts, app/src/lib/approve-plan.ts. The endpoint accepts supplied content/citations, adds frontmatter/source display, and inserts a pending plan. It does not retrieve or authenticate a stored answer.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: saved true after draft insertion, or missing-session/content error.
• Later read-back: Pending draft is available for review and potential approval into later wiki context.
• Delegated visibility: API receives content, session and supplied citations; no model is called here.
• Selection predicate: Nonempty session/content; cited page IDs become pages_used then page_plans.source_ids.
• Invalidation or expiry: Pending rejection can block installation; page-deletion cleanup strips cited IDs or deletes empty chat drafts under its chat-session filter.
• Activation or effect: Candidate retention wired; shared approval loops over these same IDs as source provenance, without page-to-original-source expansion in the inspected helper.
• Evidence limits: Human content selection is afforded by UI intent, but arbitrary API caller identity and actual review remain unresolved; source-provenance slot is not typed as page lineage.

Proposer: content-supplying caller; decider: required-fields guard for pending insertion, separate approval caller for installation; veto: rejection. Mode: open manual save. No answer oracle or factual acceptance criterion. Guidance is only wrapper formatting in this endpoint; supplied prose/citation list persists. Recovery/installation follow the shared outbox path, not a new semantic evaluator.

.run(data.plan_id, data.session_id, data.title, JSON.stringify(data.pages_used ?? []), data.draft_content);

--- app/src/lib/db.ts

    for (const sid of sourceIds) {
      insertProvenance({ source_id: sid, page_id, content_hash: contentHash, contribution_type: contribType });
    }

--- app/src/lib/approve-plan.ts

EPI-RTE-wiki-lint

Label: Wiki lint findings

No declared counterpart after comparison with runtime's maintenance account; RT-RTE-source-maintenance changes source/page content, not this evaluator. SRC-1 implementation: app/src/app/api/wiki/lint-pass/route.ts, cli/src/startup-tasks.ts, app/src/components/WikiLintResults.tsx. SQL checks database graph/provenance/time predicates; model checks summary conflict candidates. Results return, persist in activity and display as issue/hint lists. Lint invokes Saved Links regeneration separately; it does not apply conflict-driven prose revisions in this handler.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Findings with duration, or skipped when scheduled lint disabled.
• Later read-back: Activity/results can be displayed later; last-lint timestamp changes next startup scheduling, not factual endorsement.
• Delegated visibility: Model receives flattened summaries from at most ten same-category page pairs, not complete page corpus or source texts.
• Selection predicate: Manual bypass or enabled schedule; overdue startup threshold 36 hours; SQL sample requires nonnull summaries and shared category.
• Invalidation or expiry: New lint replaces displayed latest results; no inspected result-driven automatic withdrawal of challenged prose.
• Activation or effect: Advisory issue lists and scheduling updates wired; model/provider failures can leave contradictions empty.
• Evidence limits: Empty contradictions does not mean a successful complete scan; SQL predicates support database facts, not the factual validity of wiki claims.

Proposer/evaluator: SQL and model; decider: caller/handler for running, result display rather than factual acceptance. No semantic reference oracle. Mode: open maintenance, trigger manual/startup. Guidance requests contradictions, but no truth-directed revision/recovery loop is established here. Consumer: operator; channel issue list; force advisory; horizon until later review/run.

      WHERE a.summary IS NOT NULL AND b.summary IS NOT NULL
      LIMIT 10`,

--- app/src/app/api/wiki/lint-pass/route.ts

} catch { // Contradiction scan is best-effort; never fail the lint for it } --- app/src/app/api/wiki/lint-pass/route.ts

      {result.contradictions.map((c, i) => (

--- app/src/components/WikiLintResults.tsx

EPI-RTE-zombie-cleanup

Label: Provenance dependency withdrawal

No declared counterpart for admin cleanup after comparison with RT-RTE-source-maintenance. Cleanup reuses deletion helpers but independently selects pages from database lineage. SRC-1 implementation: app/src/lib/db.ts, app/src/app/api/admin/cleanup/zombie-pages/route.ts. This is a warranted deduction of a database predicate, not falsification of prose. A page with no provenance joined to an existing source is selected except Saved Links. Chat page IDs in source provenance can therefore satisfy the zombie predicate even when cited pages remain: the query joins sources, not pages.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: GET dry-run list; POST deleted count/list.
• Later read-back: Deleted database pages disappear from ordinary selection; pending chat drafts may have IDs pruned or be deleted when empty.
• Delegated visibility: Vector deletion receives page ID; filesystem helper sees matching page filenames; no model evaluator.
• Selection predicate: NOT EXISTS joined provenance/source row, except page_id saved-links; pending chat cleanup applies only to pending_approval sessions matching chat-%.
• Invalidation or expiry: Database deletion is enforcing; filesystem/vector deletion is fire-and-forget best effort. Surviving pending draft prose is not rewritten by the inspected SQL pruning.
• Activation or effect: Dependency-based withdrawal wired, not observed; cleanup removes lookup state independently of semantic truth.
• Evidence limits: Correct typed lineage is an auxiliary premise; dropping a dependency ID does not remove its semantic contribution from a surviving draft.

Trigger: caller POST; proposer/evaluator: dependency query; decider/executor: fixed loop; veto: caller can choose dry-run/not POST. No subject-matter oracle; bounded database predicate is the criterion for cleanup, intended use removing dangling pages. Mode: open maintenance. No automatic full rollback; backups/archives are distinct recovery possibilities, not truth acceptance.

      AND NOT EXISTS (
        SELECT 1 FROM provenance pr
         JOIN sources s ON s.source_id = pr.source_id
        WHERE pr.page_id = p.page_id
      )`

--- app/src/lib/db.ts

  `UPDATE page_plans
      SET source_ids = (SELECT COALESCE(json_group_array(value), '[]')
                          FROM json_each(source_ids)
                         WHERE value != ?)
    WHERE draft_status = 'pending_approval'
      AND session_id LIKE 'chat-%'

--- app/src/lib/db.ts

EPI-RTE-saved-links

Label: Saved Links status compilation

No counterpart declared after comparison with runtime's lint mention and RT-RTE-source-maintenance. SRC-1 implementation: app/src/lib/saved-links.ts, app/src/app/api/saved-links/[failure_id]/route.ts, app/src/lib/db.ts. Natural-language status markdown is compiled from unresolved failure rows and written directly to page outbox/file/FTS. The empty branch makes a consequential claim available in wiki context, so this deterministic status path is material.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Regeneration promise; dismissal returns deleted true after failure-row deletion.
• Later read-back: Saved Links page/index content enters wiki/search and potentially answering context.
• Delegated visibility: Storage receives generated markdown; metadata descriptions/errors contribute to visible prose.
• Selection predicate: Unresolved-link query; empty list versus nonempty rendering; dismissal or lint can request regeneration.
• Invalidation or expiry: Resolution/dismissal changes queried rows; new generation replaces status content; failed flush retains outbox for startup reconciliation.
• Activation or effect: Direct write without draft approval; empty-list assertion is installed just like nonempty failure listing.
• Evidence limits: Empty unresolved set proves no remaining unresolved rows, not that every saved link was imported.

Concrete implementation deduction: start with one unresolved failed link. DELETE removes its row without ingesting it, then requests regeneration. With no unresolved rows, the template emits “All saved links have been imported into your wiki.” The conclusion is not entailed and is false for this reachable history. This requires no target execution. Proposer/decider: deterministic code, caller can dismiss; no factual oracle. Mode: operational status maintenance. Persistence/outbox recovery does not correct the assertion's semantics.

  'All saved links have been imported into your wiki.',

--- app/src/lib/saved-links.ts

const deleted = deleteIngestFailure(failure_id); --- app/src/app/api/saved-links/[failure_id]/route.ts

void regenerateSavedLinksPage().catch(() => {}); --- app/src/app/api/saved-links/[failure_id]/route.ts

EPI-RTE-corpus-evaluation

Label: Provider corpus shape evaluation

No counterpart after comparison with runtime's production routes and components. SRC-1 implementation/reported operation: validation/2026-05-05-multi-provider-corpus/runner.py, validation/2026-05-05-multi-provider-corpus/REPORT.md. The evaluator's target is provider extraction response shape/counts over a selected file corpus, not factual truth. Historical report success is attributed, not current-revision execution. Selected integration assertions in scripts/integration-test.sh similarly test endpoint fields/counts/progression; exhaustive test semantics remain uninspected.

• implementation conclusion status: wired
• operation conclusion status: claimed
• Immediate return: Per-source summaries/results.json; report attributes DeepSeek seven-of-seven shape success and partial Gemini success.
• Later read-back: Durable report/results are available to developers; no inspected production decision consumer reads them.
• Delegated visibility: Providers receive source markdown with empty NER and no keyphrase/TF-IDF hints via extraction endpoint.
• Selection predicate: Sorted corpus markdown files, both named providers; reported timeouts have a separate rerun.
• Invalidation or expiry: Report is historical; reruns overwrite runner results; exact original corpus unavailable in tracked source.
• Activation or effect: Evaluator wiring supports recording outcomes, not an implemented factual acceptance gate.
• Evidence limits: Key presence/counts are not source fidelity; quota, filtering and timeout differences confound attribution of semantic capability or component effect.

Proposer: provider extractions; evaluator/decider: runner shape summary, developer interprets report; no expected factual answer oracle. Mode: bounded corpus experiment, not production admission. No semantic rollback/rejection downstream consumer inspected. Criteria measure protocol/shape outcomes; inference of knowledge accuracy exceeds their domain.

if "entities" in resp and "concepts" in resp:
    return {
        "ok": True,

--- validation/2026-05-05-multi-provider-corpus/runner.py

Claims

EPI-CLM-compiled-knowledge

Label: Compounding compiled knowledge

No counterpart after comparison with runtime's scoped durability/deployment claims. SRC-1 doctrine/design: README.md. The public claim combines automatic organization/update, cross-source pattern recognition and compiled knowledge; it does not itself establish observed or accepted ampliative knowledge.

Kompl is a compounding LLM-wiki, best fit for a personal second brain that organizes and updates itself as it grows, without maintenance or upkeep on your end. --- README.md

pattern across all ten without you having to ask. --- README.md

EPI-CLM-source-attribution

Label: Conflicts and source attribution

Distinct from RT-CLM-durable-completion: this concerns claim-level attribution and disagreement, not file persistence. SRC-1 doctrine/design: README.md.

promoted into their own pages. When sources disagree, the disagreement surfaces on a comparison page with provenance. Claims cite where they came from. By the time you have a question, the answer is already a page. --- README.md

EPI-CLM-wiki-grounding

Label: Wiki-only answer contract

No counterpart claim after comparison with RT-RTE-chat-feedback, which describes the implementation container. SRC-1 doctrine/design: nlp-service/services/llm_client.py. This is a requested warrant constraint, not an observed grounded answer.

You are a wiki-grounded knowledge assistant. Answer the user's question using ONLY the wiki pages provided. Do not use general knowledge — your answer must be grounded in the provided pages. --- nlp-service/services/llm_client.py