LLM-WIKI-MCP epistemic report

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

Source-and-claim boundary

The question is how the shipped local wiki acquires and transforms truth-apt content, checks candidates, grants reliance and supplies later consumers. The scope and exclusions are see boundary; evidence identities and layers are see Source register. All added inspection uses SRC-1 at the frozen revision, as implementation except for the explicitly attributed doctrine/design claims below. No target, test, example or provider was executed.

The ledger separates imported statements, model answers, deterministic answer assertions, sidecar suggestions, conversation records and structural/runtime measurements. Import preserves available text and some lineage, not an endorsement of source truth. A blank-answer check, a retrievability warning and a successful write have different targets; none is substituted for factual acceptance. The principal counterexample to treating this as merely retrieval is a deterministic fallback that asserts release readiness even with no page evidence. Conversely, the final context-window answer explicitly declines to guess a provider's native capacity. Both matter; neither licenses a system-wide judgment.

Consequential claims are EPI-CLM-wiki-substrate, EPI-CLM-provenance-refresh, EPI-CLM-answer-fallback, EPI-CLM-note-suggestions, and see RT-CLM-context-budget and RT-CLM-ask-nonmutation.

Coverage

Direct reads below extend rather than merely repeat the provisional runtime. The registered repository is not limited to the boundary's initial path list.

Entry point or operation	SRC-1 source path	Coverage and record	Remaining limit / conclusion prevented
Packaged invocation, interactive ask, Python and MCP	src/llm_wiki_mcp/cli.py; src/llm_wiki_mcp/server.py; pyproject.toml	Supplied dispatch/registration account: RT-RTE-mcp-dispatch. Direct one-shot dispatcher read distinguishes ask, ingest, repair, editor and context commands. Server definition/handler search includes later registry overrides, history, notes, safe self-dispatch and following-page search.	No deployed MCP grants or host continuation; cannot infer host acceptance or planning.
Final agentic ask, second synthesis, blank check and fallback	src/llm_wiki_mcp/context_api.py	Direct final definitions, second-pass builder, public sanitizer and fallback router/composer: RT-RTE-agentic-ask, EPI-RTE-answer-check, EPI-RTE-fallback-compose.	No generated answers or factual evaluation trace.
Plain/store ask and retrieval-only	src/llm_wiki_mcp/context_api.py; src/llm_wiki_mcp/server.py	Supplied RT-RTE-plain-ask, RT-RTE-store-ask; direct unified dispatch and final store-search override read.	Provider semantics uninspected; these prompts do not inherit first-agentic-pack guarantees.
Page/note search, following-page search, full-page expansion, context pack	src/llm_wiki_mcp/context_api.py	Direct final search bodies, context pack and history selection: RT-OBJ-context-pack, EPI-OBJ-page-notes.	Ranking is lexical, not an assessment of source truth; selected metadata need not all survive prompt trimming.
Source extraction and provenance refresh	src/llm_wiki_mcp/context_api.py; src/llm_wiki_mcp/cli.py	Direct extractor/page writer, hash/status database, final ingestion plan/apply and CLI ingest_path: RT-RTE-knowledge-write, EPI-OBJ-wiki-pages.	Optional extractor internals not inspected; one-shot CLI overwrites/imports through a different path, not the provenance planner. No general source-fidelity or latest-source guarantee.
Direct create/update, transcript/decision writes, schema, table of contents	src/llm_wiki_mcp/server.py	Supplied create/update account; direct decision/schema/TOC/dry-run bodies and operation inventory: RT-RTE-knowledge-write, EPI-RTE-structure-check.	Transcript implementation not directly read; represented by supplied local-write account and exposed operation inventory only. No exhaustive semantic assessment of all wrappers.
Notes generation, iteration, candidate links	src/llm_wiki_mcp/context_api.py	Direct template, term extraction and link scorer: EPI-RTE-note-template.	No human/agent promotion trace; suggestion wording is not a gate on retrieval.
Lint, resolve, repair, dry-run write	src/llm_wiki_mcp/server.py	Direct check/repair/resolve bodies: EPI-RTE-structure-check; disposition remains see RT-RTE-knowledge-write.	Structural domain only. No semantic repair oracle.
Rename/delete and later withdrawal	src/llm_wiki_mcp/server.py	Direct bodies: EPI-RTE-page-withdrawal.	No transitive retraction of old answers/history or semantic consequences established.
Ask history, compression, recursive context and recent-history prompt	src/llm_wiki_mcp/context_api.py	Direct recorder/compressor/recursive reader and final pack reader: EPI-RTE-history-reshape, EPI-OBJ-ask-turns.	No deployment history; compression retention and final pack consumption are different paths.
Self-context, context-window answer, doc reconciliation, runtime/metrics tools	src/llm_wiki_mcp/context_api.py	Direct context-window answer, pack and doc reconciliation; supplied config/model account: EPI-RTE-diagnostic-check, RT-BAP-config-control.	Other metric-helper bodies available but not all inspected; prevents a system-complete negative about every diagnostic's accuracy or use.
Docs consolidation/development-history generation	src/llm_wiki_mcp/context_api.py	Supplied consolidation: RT-RTE-docs-consolidate; development-history function inventoried only.	Development-history content transformation unassessed; prevents complete enumeration of generated documentation claims.
Graph, backlinks, related, exports	src/llm_wiki_mcp/server.py; src/llm_wiki_mcp/context_api.py	Direct related/backlinks/TOC and search bodies; supplied export/graph dispatch.	Views carry lexical/structural relationships, not demonstrated explanatory relations. Other view bodies not exhaustively inspected.
Regression/test evaluation	src/llm_wiki_mcp/server.py; tests/ask_history_context_test.py	Direct regression suite and harness result consumer; direct mocked-history/sanitizer/context-window tests: EPI-RTE-regression-check.	Other tests available but uninspected here; no test results or quality/causal comparison supplied.

These available-but-uninspected operations preclude a system-complete negative about all possible acceptance or evaluation routes. The narrower findings below concern the named active answer/write/read/check implementations. Excluded external weights and provider processing prevent inference about internal reasoning, learned truth, training-time warrant or parameter change; see RT-CMP-ollama-model. Missing deployment corpora prevent truth judgments on any actual imported source. Missing provider/run traces prevent observed answer correctness, activation or benefit. Missing interventions/comparisons prevent causal attribution to notes, history, repair or context budgeting.

Epistemic-object inventory

Fields owned by linked records are see; local object labels do not establish additional accepted knowledge.

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
EPI-OBJ-wiki-pages	Imported or caller-authored statements; source diagnostic placeholders also possible	see	SRC-1 src/llm_wiki_mcp/context_api.py; see record	Actual source warrant unknown; path/hash identifies a file, not truth.
EPI-OBJ-page-notes	Topic assertion and candidate-link/term scores	see	SRC-1 src/llm_wiki_mcp/context_api.py; see record	Template and lexical counts, not model-authored criticism or validated synthesis.
RT-OBJ-context-pack	Selected source statements, operational assertions, questions and prior answers	see	see	Delivery is selected and lossy; not acceptance.
EPI-OBJ-fallback-answer	Release-readiness, inspiration and configuration assertions; extracted lines	see	SRC-1 src/llm_wiki_mcp/context_api.py; see record	Deterministic branches differ from extractive reshaping.
EPI-OBJ-ask-turns	Questions, answers and compressed excerpts; estimates and metadata	see	SRC-1 src/llm_wiki_mcp/context_api.py; see record	Retained answers can be wrong; source list is tool names on final agentic recording path.
EPI-OBJ-check-results	Broken-link/existence/count/blank predicates, warning lists, test booleans	see	SRC-1 src/llm_wiki_mcp/server.py; see record	Domains bounded by executable predicates, not factual-answer validation.
RT-CMP-ollama-model	Specific internal truth-apt content not determinable	see	see	Distributed-parametric provider state inaccessible; model identity not weight-pinned.

Authority-route ledger

Architectural status below is implementation-grounded, not observed activation. Each row separates one function; shared IDs link phases of one source-native route. see delegates owned details to the cited declaration. No row asserts observed epistemic acceptance. All source anchors refer to SRC-1.

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: behavior permitted, blocked, or changed	Behavioral-authority path: consumer, channel, force, horizon	Evidence source ID and local anchor	Claim IDs or none	Mismatch marker or none	Gap/limit
RT-RTE-knowledge-write	content transformation	implemented	EPI-OBJ-wiki-pages	truth-apt transformation: acquisition/import	Source/caller content to wiki	File extractor; source warrant unknown	Caller import/write	Extracted/wrapped text, or diagnostic placeholder	Produces page text	No endorsement; plain text retained, rich extraction fidelity unestablished	Supplies later indexed content	see route; later retrieval via files/index, permissive until edited	SRC-1 src/llm_wiki_mcp/context_api.py; see object	EPI-CLM-wiki-substrate	none	One-shot CLI and provenance ingestion differ.
RT-RTE-knowledge-write	disposition/acceptance	implemented	EPI-OBJ-wiki-pages	no content change	Write eligibility, not truth	Existing-path/overwrite, update mode, apply flag	Caller request	Admit/reject file operation	Enforcing local write condition	Operational admission only; no evidence-consuming factual acceptance shown here	Writes/reindexes without per-content approval	see	SRC-1 src/llm_wiki_mcp/server.py; see route	none	none	Generic caller authorship unresolved.
RT-RTE-knowledge-write	retention	implemented	EPI-OBJ-wiki-pages	no content change	Admitted text to retained file/index	Write and indexing code	Applied mutation	Persisted page/index	Enforcing storage	Retention, not lifecycle integration after acceptance	Makes content eligible for retrieval	see	SRC-1 src/llm_wiki_mcp/server.py; see route	EPI-CLM-wiki-substrate	none	No populated corpus supplied.
RT-RTE-knowledge-write	lineage/freshness/recovery	implemented	EPI-OBJ-wiki-pages	no content change	Source path/mtime/hash change	Provenance predicate, file identity domain	Explicit re-ingest	new/update/skip; reindex/error	Ranking-independent refresh selection	Byte freshness/lineage only, not endorsement	Changed source selected for replacement	see route; planner to writer, enforcing at apply, future retrieval	SRC-1 src/llm_wiki_mcp/context_api.py; see object	EPI-CLM-provenance-refresh	none	Metadata does not store full prior source bytes; not general rollback.
EPI-RTE-note-template	content transformation	implemented	EPI-OBJ-page-notes	truth-apt transformation: indeterminate	Page/old note to topic assertion and link suggestions	Counts/substrings and template	Explicit notes iteration	Term list, topic wording, scored candidates	Produces/overwrites sidecar	Lexical occurrence facts; topic/semantic relation not checked	Later selection can use suggestions before promotion	see	SRC-1 src/llm_wiki_mcp/context_api.py; see route	EPI-CLM-note-suggestions	suggestion-use	Old-note feedback can influence next term set; semantic preservation unproved.
EPI-RTE-note-template	retention	implemented	EPI-OBJ-page-notes	no content change	Template result to sidecar file	Nonempty page, then file write	Each successful iteration	Sidecar retained	Enforcing write	No promotion/acceptance decision	Search/combined page can consume it	see	SRC-1 src/llm_wiki_mcp/context_api.py; see route	EPI-CLM-note-suggestions	suggestion-use	"To Review" label is not retrieval exclusion.
RT-RTE-agentic-ask, RT-RTE-store-ask, RT-RTE-plain-ask	operational admission/selection/consumption	implemented	EPI-OBJ-wiki-pages, EPI-OBJ-page-notes, RT-OBJ-context-pack	truth-apt transformation: non-ampliative reshaping	Retained evidence to selected context	Lexical scores/top-k, per-page/history/whole-pack limits	Current request	Ranked snippets/page excerpts/history	Ranking then permissive evidence supply	Relevance heuristic, no source truth license	Code decides what provider/caller receives	see routes and RT-BAP-answer-guidance	SRC-1 src/llm_wiki_mcp/context_api.py; see pack	RT-CLM-context-budget	none	Truncation can remove qualification or lineage; semantic equivalence not guaranteed.
RT-RTE-agentic-ask, RT-RTE-store-ask, RT-RTE-plain-ask	content transformation	implemented	RT-OBJ-context-pack, RT-CMP-ollama-model	truth-apt transformation: indeterminate	Prompt to answer text	External provider generation	Nonbasic, nondry request; different prompt families	Requested answer, failure or blank	Candidate generation; advisory output	No inspected entailment or evidence-consuming answer acceptance	Returns answer; unified routes can retain it	see routes; RT-BAP-answer-guidance	SRC-1 src/llm_wiki_mcp/context_api.py; see routes	EPI-CLM-answer-fallback	none	No concrete generated answer to classify as reshaping/derivation/conjecture.
EPI-RTE-answer-check	check/evidence production	implemented	EPI-OBJ-check-results	no content change	Normalized answer availability	Python truthiness, empty-string domain	After synthesis	Blank/nonblank	Predicate feeds continuation	Availability only	Provides branch input	see	SRC-1 src/llm_wiki_mcp/context_api.py; see route	EPI-CLM-answer-fallback	none	A false nonempty answer passes this test.
EPI-RTE-answer-check	operational admission/selection/consumption	implemented	EPI-OBJ-fallback-answer	no content change	Continue, retry or return	Blank predicate, basic-question predicate	In current ask	Extra evidence/second request, deterministic answer or ordinary answer	Enforcing branch	No factual acceptance	Blank triggers recovery; nonblank needs no truth check in this path	see	SRC-1 src/llm_wiki_mcp/context_api.py; see route	EPI-CLM-answer-fallback	none	Top-level ok can retain first-call failure.
EPI-RTE-fallback-compose	content transformation	implemented	EPI-OBJ-fallback-answer	truth-apt transformation: ampliative conjecture	Release-readiness/inspiration branches	Keyword selector and hardcoded prose, no factual evaluator	Recovery asks reaching composer	Specific affirmative conclusion even absent evidence	Advisory answer, then eligible retention	No evidence warrants release readiness from empty input	Can answer and feed future history	see	SRC-1 src/llm_wiki_mcp/context_api.py; see route	EPI-CLM-answer-fallback	unsupported-readiness	Concrete source deduction below; not observed execution.
EPI-RTE-fallback-compose	content transformation	implemented	EPI-OBJ-fallback-answer	truth-apt transformation: non-ampliative reshaping	Generic page-line fallback	Line cleaning/query scoring	Non-special question with read pages	Selected lines with best-effort framing	Advisory answer	Extracted claims remain dependent on source and lost context	Returns text and references	see	SRC-1 src/llm_wiki_mcp/context_api.py; see route	EPI-CLM-answer-fallback	none	Ranking prose does not prove strongest evidential source.
EPI-RTE-history-reshape	retention	implemented	EPI-OBJ-ask-turns	no content change	Completed turn to JSONL	Recorder/config, not factual evaluator	Unified nondry ask or explicit record tool	Stored question/first prompt/final answer	Enforcing append when enabled/succeeds	No accepted-knowledge status	Retains unverified model or fallback assertions	see	SRC-1 src/llm_wiki_mcp/context_api.py; see route	none	none	Not all asks record; first prompt can differ from answer-producing prompt.
EPI-RTE-history-reshape	content transformation	implemented	EPI-OBJ-ask-turns	truth-apt transformation: non-ampliative reshaping	Older turns to bounded Markdown	Token estimate, whitespace flattening, prefix cuts	Budget exceeded and compression enabled	Short answer excerpts; recent JSONL rewrite	Enforcing replacement	No new corroboration; degraded replay/provenance detail	Alters material reachable by later history readers	see	SRC-1 src/llm_wiki_mcp/context_api.py; see route	none	none	Older prompts/full answers removed from live JSONL; not semantic summarization.
EPI-RTE-history-reshape	operational admission/selection/consumption	implemented	EPI-OBJ-ask-turns	no content change	History to later context	Recent-history selector; separate recursive reader	Later request	Bounded recent Q/A in final pack; compressed+recent in recursive API	Advisory model context	History explicitly not immutable truth	Prior answers can guide later answers without acceptance	see	SRC-1 src/llm_wiki_mcp/context_api.py; see route	none	none	No observed dependence or improvement.
EPI-RTE-structure-check	check/evidence production	implemented	EPI-OBJ-check-results	truth-apt transformation: entailed derivation	Link existence, incoming edges, word/paragraph lengths	Graph/file/set/count predicates	Lint/dry-run/repair request	Structural warnings and before/after reports	Advisory; repair consumes before report	Warrant bounded to those measurements of scanned state	Suggests repair actions, no enforced truth gate on create/update	see	SRC-1 src/llm_wiki_mcp/server.py; see route	none	none	Valid link can point to false content.
RT-RTE-knowledge-write	disposition/acceptance	implemented	EPI-OBJ-check-results, EPI-OBJ-wiki-pages	no content change	Repair actions to apply	Caller apply/create-missing flags and resolver	Explicit repair	Rewrite links, link orphans, add scaffold or leave unresolved	Enforcing file edits when enabled	Structural disposition, not factual acceptance/integration	Changes later navigability/retrieval	see; EPI-RTE-structure-check supplies check	SRC-1 src/llm_wiki_mcp/server.py; see route	none	none	Scaffold does not add independently justified domain knowledge.
EPI-RTE-page-withdrawal	lineage/freshness/recovery	implemented	EPI-OBJ-wiki-pages	no content change	Rename/delete active page	Caller instruction and existence tests	Explicit mutation	Renamed/relinked or deleted/tombstoned page; reindex	Enforcing local removal/rename	Operational withdrawal, not recorded refutation	Removes active page, changes future retrieval eligibility	see	SRC-1 src/llm_wiki_mcp/server.py; see route	none	no-cascade	Old history/answers not retracted by these bodies.
EPI-RTE-diagnostic-check	check/evidence production	implemented	EPI-OBJ-check-results	truth-apt transformation: entailed derivation	Expected doc-file existence and command substrings	Filesystem/substring predicates	Reconcile-docs call	Missing/undocumented lists and ok	Advisory returned diagnostics	Only declared file-list/text-presence domain	Caller can inspect; no automatic semantic repair here	see	SRC-1 src/llm_wiki_mcp/context_api.py; see route	none	none	Fixed expected filenames can diverge from consolidated layout.
EPI-RTE-diagnostic-check	content transformation	implemented	EPI-OBJ-check-results	truth-apt transformation: non-ampliative reshaping	Active settings to context-window answer	Code formatting settings/estimates	Basic context-window request	Named settings plus explicit unknown native limit	Advisory answer	Config describes local budgets, not provider capacity	Returns without model call	see	SRC-1 src/llm_wiki_mcp/context_api.py; see route	RT-CLM-context-budget	none	Not an exact tokenizer/runtime probe.
EPI-RTE-regression-check	check/evidence production	implemented	EPI-OBJ-check-results	truth-apt transformation: entailed derivation	Tool surface, fixture mutations/retrieval, history/sanitizer behavior	Fixture assertions and result aggregation	Explicit test/harness invocation	Pass/fail predicates, aggregate exit code	Enforcing harness success/failure, not deployment acceptance	License limited to asserted fixture predicates	Caller receives result/exit; no automatic product promotion shown	see	SRC-1 src/llm_wiki_mcp/server.py; see route	none	none	Tests unexecuted; mock output not real model-quality evidence.
RT-RTE-runtime-config	behavior/policy adaptation	implemented	RT-CMP-ollama-model	non-truth-apt policy/content update: host/model/mode/budget settings	Caller-chosen setting to later controls	Parser/default/minimum rules	Explicit config change	Future dispatch/request selection	Enforcing code selection	No knowledge or provider-capacity acceptance	Can change endpoint, model tag and context policy	see RT-BAP-config-control	SRC-1 src/llm_wiki_mcp/context_api.py; see route	none	none	Not evidence-responsive autonomous improvement or weight update.
RT-RTE-docs-consolidate	content transformation	implemented	Natural-language documentation; see route	truth-apt transformation: non-ampliative reshaping	Fixed mapped documents to concatenated categories	Static mapping/caller archive flag	Explicit consolidation	Overwritten outputs, optional archives	Enforcing write/move	No checked preservation of all claims/qualifications	Later operators may use generated docs	see	SRC-1 src/llm_wiki_mcp/context_api.py; see route	none	none	Not acceptance or demonstrated method improvement.

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-wiki-substrate	see	SRC-1 README.md; doctrine/design	see	RT-RTE-knowledge-write, RT-RTE-store-ask	None supplied	No causal comparison with raw chunks	Markdown is authoritative storage for page retrieval	Storage "source of truth" does not establish factual truth.
EPI-CLM-provenance-refresh	see	SRC-1 METHODOLOGY.md; doctrine/design	see	RT-RTE-knowledge-write	None supplied	Source identity/change predicates, not automatic verification	Explicit provenance re-ingestion is wired	Not all import paths use provenance; lineage can be redacted/truncated downstream.
EPI-CLM-answer-fallback	see	SRC-1 README.md; doctrine/design	see	EPI-RTE-answer-check, EPI-RTE-fallback-compose	None supplied	Blank recovery is implemented; no semantic adequacy check or benefit contrast	Can return answer-first text and evidence lists	Readiness branch overstates evidential basis; generic "states uncertainty" does not make each assertion warranted.
EPI-CLM-note-suggestions	see	SRC-1 src/llm_wiki_mcp/context_api.py; emitted design guidance	see	EPI-RTE-note-template, RT-RTE-agentic-ask	None supplied	No promotion trace or retrieval-quality experiment	Candidate wording distinguishes suggestions from canonical edits	Notes already affect ranking/snippets before promotion.
RT-CLM-context-budget	see	see; doctrine/design and implementation distinguished there	see	RT-RTE-agentic-ask	None supplied	Character-based estimate, alternate prompts	First-pack budgeting/inspection is implemented	Second prompt may produce answer while first prompt is recorded; native capacity unknown.
RT-CLM-ask-nonmutation	see	see	see	RT-RTE-agentic-ask, RT-RTE-mcp-dispatch	None supplied	Scope is ask's code-selected evidence actions	Prompt policy and local ask path differ from general capability surface	History/config writes remain possible; mutation tools are separately callable.

Bounded conclusion

Acquisition and use are implemented; epistemic acceptance is not established for the named answer/write paths. Page import acquires content rather than producing its warrant. Local metadata permits file-change tracking, and indexed files supply later readers. Plain-text content is copied with wrappers; optional rich extraction, placeholder text, public redaction, snippet selection and truncation limit semantic and lineage preservation. Source warrant remains unknown. See RT-RTE-knowledge-write, EPI-OBJ-wiki-pages and RT-OBJ-context-pack.

The answer producers differ. Model generation is requested, but no concrete model answer or entailment check is available. Its content relation remains indeterminate. Generic deterministic fallback selects and reframes page lines; a separate keyword branch produces an affirmative release-readiness conjecture without requiring supporting pages. The final runtime context-window answer, in contrast, returns configured estimates and explicitly withholds a native-capacity claim. These findings coexist; see EPI-RTE-fallback-compose and EPI-RTE-diagnostic-check.

Checks have narrow, consequential domains. Blankness controls retries/fallback, not truth. Lint and dry-run checks derive structural facts; repair consumes them to alter navigation and scaffolding after caller application. Test code contains fixture oracles and an exit-status consumer, but no passing execution is supplied. No inspected active answer path records an evidence-consuming acceptance decision with a factual criterion, intended use and scope. Accordingly, retention, indexing, relinking and prompt delivery here are not lifecycle integration after such acceptance. This is a scoped finding, not an absence theorem over uninspected diagnostics/tests/document generators. See EPI-RTE-answer-check, EPI-RTE-structure-check and EPI-RTE-regression-check.

Unaccepted content has operational influence. Notes can shape ranking and snippets despite their suggestion label. Unified ask can retain unverified answers; later agentic prompts include bounded recent Q/A as advisory follow-up context. Compression keeps shortened answer excerpts and removes older exact turns from live JSONL. The final pack does not itself read the compressed-history file, whereas the separately exposed recursive-context helper does. Page deletion/reindexing is local withdrawal, not a cascade retracting historic answers. See EPI-RTE-note-template, EPI-RTE-history-reshape, EPI-RTE-page-withdrawal and RT-BAP-answer-guidance.

Behavioral control and learning are separate. Explicit settings control later model/host/mode/budget choices; history provides an implemented experience-to-later-context path. Neither supplies evidence of improved capacity, content-directed criticism of a formulated theory, iterative theory revision or exercised self-improvement. No new theory-route claim is made. Provider processing/parameter changes remain uninspected; no negative about all internal learning follows. No run or intervention establishes model compliance, activation of a particular retained claim or benefit attributable to notes/history/context budgets. See RT-CMP-ollama-model, RT-RTE-runtime-config and EPI-RTE-history-reshape.

Shared records

Components

none declared in this member

Operative objects

EPI-OBJ-wiki-pages

Label: Wiki pages

Source-native Markdown files under the vault's wiki/, interpreted as natural-language content by humans/models and parsed symbolically for frontmatter/links/index fields. Storage is local files plus derived SQLite index. Imported bodies, caller edits, decision-log entries, TOC text and repair scaffolds can all enter this substrate. This is not a duplicate of RT-OBJ-context-pack: files persist outside the selected invocation prompt. Closest write account is RT-RTE-knowledge-write; object identity is the resulting page corpus, not that route.

SRC-1 implementation: src/llm_wiki_mcp/context_api.py, src/llm_wiki_mcp/server.py, src/llm_wiki_mcp/cli.py. Provenance stores source path/hash/timestamps and page associations, not a factual endorsement or full prior file snapshot. Text extraction can produce diagnostic content instead of document text.

body=release_extract_text_from_source_file(p) content=f"# {title}\n\nSource file: {p}\nSource type: {p.suffix.lower() or 'file'}\n\n---\n\n{body}\n" wiki_path.write_text(content, encoding="utf-8") release_record_ingest_result(vault_path, p, wiki_title=title, wiki_path=str(wiki_path), action="updated_or_new") --- src/llm_wiki_mcp/context_api.py

if int(row["mtime_ns"])==rec["mtime_ns"] and row["current_sha256"]==rec["sha256"]: return {"action":"skip","reason":"same datestamp and hash","record":rec,"source_file_id":row["id"],"wiki_title":row["wiki_title"]} return {"action":"update","reason":"file changed","record":rec,"source_file_id":row["id"],"wiki_title":row["wiki_title"],"previous_sha256":row["current_sha256"]} --- src/llm_wiki_mcp/context_api.py

return f"[No extractable DOCX paragraph text found in {p.name}.]" --- src/llm_wiki_mcp/context_api.py

The public-object sanitizer can remove the local source locator before delivery; this limits downstream auditability even when retained local provenance is available separately.

Keep local raw provenance out of answer prompts. Filename/wiki title remain.

if str(k).lower() in {"path", "full_path", "wiki_path", "source_path", "db", "vault"}: cleaned[k] = "[local path redacted]" --- src/llm_wiki_mcp/context_api.py

One-shot CLI import instead constructs a Source - {stem} title, copies available raw bytes and writes through create/overwrite, then reindexes. This is a counterexample to attributing the provenance planner to all ingestion.

page = store.create_page(title, body, tags=["source", "ingested", path.suffix.lower().lstrip('.') or "file"], source=str(path), overwrite=overwrite) --- src/llm_wiki_mcp/cli.py

EPI-OBJ-page-notes

Label: AI sidecar notes

Source-native generated Markdown sidecars under notes/. Natural-language topic/suggestion text is interpreted by people/models; extracted terms/scores are symbolic algorithm outputs. Files persist and can be overwritten on iteration. Distinct from the transient RT-OBJ-context-pack; closest supplied route RT-RTE-knowledge-write includes their generation but is not an object counterpart. Mechanism/evidence is see EPI-RTE-note-template, SRC-1 src/llm_wiki_mcp/context_api.py.

EPI-OBJ-fallback-answer

Label: Direct fallback answer

Part of: RT-RTE-agentic-ask

This cross-kind containment identifies a material output part of agentic ask, not the entire execution route. Source-native answer string from release_evidence_extract_answer/composer. Natural-language assertions exist in process memory and can be returned, displayed and recorded. It is not RT-OBJ-context-pack, which supplies input context; it is not an inspected model response. Evidence is see EPI-RTE-fallback-compose, SRC-1 src/llm_wiki_mcp/context_api.py.

EPI-OBJ-ask-turns

Label: Ask history

Source-native ask_history.jsonl and ask_history_compressed.md records. JSON keys/numeric estimates are symbolic; question/prompt/answer text and compressed Markdown are natural-language material. Local filesystem storage is distinct from RT-OBJ-context-pack: selected text enters the pack, but the retained record also contains omitted first-prompt/metadata fields. Overlaps the output described by RT-RTE-agentic-ask and RT-RTE-plain-ask, and explicit recorder tools can populate it without either ask. Evidence is see EPI-RTE-history-reshape, SRC-1 src/llm_wiki_mcp/context_api.py.

EPI-OBJ-check-results

Label: Check results

Source-native booleans, warning/action lists and counts returned by answer branches, lint/dry-run/repair, doc reconciliation and regression harness. Symbolic dictionaries/booleans in process memory, serialized for callers; no common persistent acceptance register is established. Parts are separated by evaluator/consumer in the ledger. No operative-object counterpart in supplied runtime; RT-RTE-knowledge-write and RT-RTE-agentic-ask describe some consuming routes. Evidence is see EPI-RTE-answer-check, EPI-RTE-structure-check, EPI-RTE-diagnostic-check and EPI-RTE-regression-check.

Routes

EPI-RTE-answer-check

Label: Answer availability check

Part of: RT-RTE-agentic-ask

The final local ask code owns answer normalization, truthiness check, second synthesis and final fallback selection. State is current answer/evidence/provider status; action effects are additional retrieval/request or deterministic output, not factual candidate acceptance. Check and continuation remain separate ledger functions.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Normalized answer or recovery answer in the enclosing return dictionary; availability predicate is consumed internally.
• Later read-back: Returned answer may enter automatic history and later prompts; the predicate itself is not a retained warrant.
• Delegated visibility: Provider receives a new second-pass prompt; caller receives answer/evidence/status.
• Selection predicate: Basic-question classification bypasses generation; otherwise empty normalized first answer triggers second synthesis and empty continuation triggers fallback.
• Invalidation or expiry: Recomputed for each candidate in the current invocation; no enduring approval.
• Activation or effect: Code branches on blankness; a nonempty false assertion is not rejected by this check.
• Evidence limits: Implementation only; first-call top-level status is not a factual correctness result, nor proof of which prompt produced the returned answer.

Proposer is model or deterministic router, decider is code, caller can select dry-run/synthesis suppression; no human per-answer approval. Open user requests, no reference-answer oracle. Recovery seeks nonempty text, with no inspected content-based rollback/acceptance gate. SRC-1 implementation, src/llm_wiki_mcp/context_api.py:

if not answer and "release_try_second_pass_synthesis" in globals(): second = release_try_second_pass_synthesis(vault_path, original, trace, evidence) answer = second.get("answer") or "" llm = {**llm, "second_pass": second.get("llm")} evidence = evidence + release_sanitize_public_obj(second.get("extra_evidence", [])) --- src/llm_wiki_mcp/context_api.py:4438-4442

EPI-RTE-fallback-compose

Label: Evidence fallback composition

Part of: RT-RTE-agentic-ask

Code owns the router → evidence extractor → direct-answer composer → supporting-evidence rendering path. It uses current question and available page/search/capability results. Distinct material producer from the containing ask's external model generation; the combined record remains valid and needs no supersession.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Natural-language direct answer plus optional evidence/support/tool sections.
• Later read-back: Enclosing unified ask can record it and later select its answer as follow-up context.
• Delegated visibility: Returned to caller; not a provider critique or validation request.
• Selection predicate: Ordered keyword branches, then best-scored cleaned page lines, then weak-evidence messages.
• Invalidation or expiry: Recomputed per fallback; retained prior answers are not automatically invalidated by later evidence.
• Activation or effect: Recovery can supply a nonempty answer without any model generation or truth check.
• Evidence limits: Concrete deduction below is implementation evidence only; it does not assert that a user observed this answer.

Concrete counterexample, without executing target code: pass question Is this ready?, read_pages=[], search_hits=[], caps=None to the composer. low contains ready; titles and docs are empty; the branch chooses the available release documentation and returns Yes — ... broadly ready for a GitHub release. The readiness assertion is not entailed by an empty evidence set (a not-ready project satisfies those same inputs). It is an ampliative conjecture, not an extractive summary or acceptance decision. The final nonbasic fallback router reaches the evidence extractor; empty provider output is sufficient to reach recovery. The extractor calls the composer before deciding whether it has pages to list. This route challenges the broader supplied characterization of all deterministic fallback as merely "evidence fallback" without making the runtime container defective.

SRC-1 implementation, src/llm_wiki_mcp/context_api.py:

if any(word in low for word in ["ready", "release", "github"]): docs=[t for t in titles if any(k in t.lower() for k in ["readme", "install", "cli", "mcp", "architecture", "product", "books", "function"])] doc_phrase="the available release documentation" if not docs else ", ".join(f"{t}" for t in docs[:5]) return ( "Yes — based on the wiki evidence, this looks broadly ready for a GitHub release, with the normal caveat that a final clean checkout, " --- src/llm_wiki_mcp/context_api.py

lines=["## Direct answer", "", release_compose_best_effort_direct_answer(q, read_pages, search_hits, caps), ""] --- src/llm_wiki_mcp/context_api.py

The inspiration branch likewise returns source-native lineage prose by keyword without inspecting page content. Generic fallback instead selects lines and adds source-ranking/best-effort framing; it can lose qualifiers through line selection, cleaning and cutting. No evidence-consuming factual acceptance criterion, intended-use license, oracle or promotion decision is implemented in this composer. The caller may reject the displayed product externally, but that decision is not inspected.

EPI-RTE-note-template

Label: Sidecar note iteration

Part of: RT-RTE-knowledge-write

Code reads a page and its prior sidecar, counts terms, scores other pages by substring overlap/backlink-title bonus, renders a fixed template and overwrites the sidecar. Natural-language "appears to discuss" and proposed semantic links are distinguished from measured occurrence. This is a part, not a competing identity for the combined write route.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Note path, terms and candidate-link dictionaries, or missing-page error.
• Later read-back: Subsequent iteration reads old note; page/note search scores note terms and may select its snippet; combined-page context concatenates page and note.
• Delegated visibility: Calling CLI/MCP/Python consumer receives result; model can receive selected note snippets in later evidence.
• Selection predicate: Existing page; frequency terms from page plus old note; other-page title/first 3000 characters plus backlink-title bonus for candidate links.
• Invalidation or expiry: Rerun replaces note; no semantic invalidation or evidence-based promotion gate shown.
• Activation or effect: Sidecar becomes available to ranking and context independently of canonical-page promotion.
• Evidence limits: No retrieval improvement or accepted semantic relationship is demonstrated; old-note boilerplate can affect extracted terms.

Caller triggers iteration, code proposes and writes without per-content human approval; missing page can block it, caller can withhold invocation. Guidance persists inside generated note; later editors can promote links, but this implementation does not perform an evidence-consuming promotion. No expected-answer oracle or experiment is wired. Topic output relation is indeterminate: occurrence-driven wording may be loose description or unsupported semantic inference; source content/semantic criterion is needed to decide for a concrete page.

SRC-1 implementation and emitted doctrine/design, src/llm_wiki_mcp/context_api.py:

terms = release_extract_terms(page_text + "\n" + old_note, 25) links = release_find_candidate_links(vault_path, title, max_links) --- src/llm_wiki_mcp/context_api.py

This page appears to discuss: {', '.join(terms[:12]) if terms else 'no clear extracted terms yet'}. --- src/llm_wiki_mcp/context_api.py

• Candidate links are suggestions until promoted into the source page or accepted by a human/agent. --- src/llm_wiki_mcp/context_api.py

excerpt_source = note if note and any(t in note.lower() for t in terms) else source win = release_centered_match_excerpt(excerpt_source, matched or terms, context_chars=context_chars) --- src/llm_wiki_mcp/context_api.py

The final search loop also adds a score bonus when the query term occurs in a note. This is operational ranking authority, not factual or canonical status.

EPI-RTE-history-reshape

Label: Ask history retention

The recorder appends a turn, compresses when configured estimates exceed a threshold, and exposes recent/recursive readers. Code owns each phase. This spans the recorder used by RT-RTE-agentic-ask, unified RT-RTE-plain-ask and explicit record/compression tools in RT-RTE-mcp-dispatch; not wholly a part of exactly one supplied route. Distinct identity is the history write/reshape/read chain, not a duplicate ask execution.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Record/path, compression metadata or reconstructed context; enclosing ask can return metrics/error even if persistence fails.
• Later read-back: Final pack uses up to eight recent Q/A with per-answer cut and history budget; recursive-context API separately includes compressed Markdown and recent memory.
• Delegated visibility: Callers can inspect history/context; selected history enters subsequent provider prompt. Stored prompts are not copied verbatim by final recent-Q/A renderer.
• Selection predicate: History enabled for append; compression enabled and estimate over max-history threshold for compaction; keep-recent-turns setting splits older/recent.
• Invalidation or expiry: Compaction rewrites live JSONL to recent turns, appends compressed excerpts to existing Markdown; no factual rejection/invalidation based on later answers shown.
• Activation or effect: Retained answers become eligible for later advisory context, not automatically accepted beliefs or revised policies.
• Evidence limits: No observed later dependence or benefit; first prompt can be retained when second synthesis supplies answer; source list on final agentic route is tool names, not passage-level warrant.

Proposer is automatic current-turn recorder or explicit caller-supplied record. Code/config decides append/compaction; caller can disable history/compression or omit explicit invocation. No per-content human approval or factual oracle. Compression guidance is shipped truncation rules, not a model critique; there is no semantic rollback protocol. SRC-1 implementation, src/llm_wiki_mcp/context_api.py:

a = " ".join(str(row.get("answer", "")).split())[:1000] src = row.get("sources") or [] summary_parts.append(f"## Turn: {q[:120]}") --- src/llm_wiki_mcp/context_api.py

with live.open("w", encoding="utf-8") as f: for row in recent: f.write(_release_json.dumps(row, ensure_ascii=False) + "\n") --- src/llm_wiki_mcp/context_api.py

lines = ["# Recent Ask/Reply History", "", "These are recent local CLI ask turns. Use them for follow-up context, not as immutable truth.", ""] --- src/llm_wiki_mcp/context_api.py

if ask.get("recursive_enabled", True): if compressed: parts.append("# Compressed Long-Term Ask Memory\n\n" + compressed) if recent: parts.append("# Recent Ask Memory\n\n" + recent) --- src/llm_wiki_mcp/context_api.py

Working-context control differs from a separate-agent frame: code copies selected history by value, truncates answers and whole sections, and can reload external history through exposed readers. No sub-agent isolation or frame return is established. Current-call pruning does not remove the page corpus; compaction, separately, changes retained exact-turn availability. Delivery does not establish that the provider treated the historical answer as true.

EPI-RTE-structure-check

Label: Structural wiki checks

Code computes graph/file existence and count predicates in lint, word/paragraph/link predicates in dry-run-write, and before/after reports for repair. Closest supplied RT-RTE-knowledge-write includes repair admission, not the entire independently callable lint/preview evaluator family. This evaluator record does not replace that write route.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Broken links, orphan/short-page lists, preview warnings/ok or before/after repair report.
• Later read-back: Repair edits can persist and be retrieved; reports have no independent durable acceptance read-back established.
• Delegated visibility: CLI/MCP/Python caller gets report; no model factual evaluation occurs here.
• Selection predicate: Current wiki graph/files, supported preview mode, lexical word counts and paragraph lengths; repair invokes lint before and after apply.
• Invalidation or expiry: Results describe scanned state and are recomputed on later call; no immutable certificate.
• Activation or effect: Repair consumes structural findings to plan changes; callers receive warnings. Direct create/update need not call preview.
• Evidence limits: Domain is retrievability/navigation predicates, not source correctness or semantic adequacy. File scan is not guaranteed atomic against concurrent edits.

Open caller requests; code proposes actions, apply and create-missing flags control admission, caller can veto. No expected factual answer/reference outcome; fixtures are separate test oracles. Recovery is reindex/recheck/edit, not automatic old-content restoration. Checked counts are entailed by scanned strings under these functions; interpretation outside that formal domain is not licensed. SRC-1 implementation, src/llm_wiki_mcp/server.py:

if len(plain_words(body)) < 20: short.append(p.stem) return {'broken_links': broken, 'orphans': sorted(orphans), 'very_short_pages': sorted(short), 'page_count': len(existing), 'ok': not broken} --- src/llm_wiki_mcp/server.py

if wc<50: warnings.append('doc-too-short: consider adding enough context for retrieval') if wc>1800: warnings.append('doc-too-long: consider splitting into linked sub-pages') --- src/llm_wiki_mcp/server.py

'Repair improves navigation; it does not replace human review of source meaning.' --- src/llm_wiki_mcp/server.py

Concrete deduction: a preview containing fewer than 50 plain_words triggers the short-document warning; it does not imply that any proposition in the preview is false. Conversely, arbitrary false prose with enough words and no broken links need not trigger a warning. Thus a structural result can guide maintenance but cannot stand in for factual acceptance.

EPI-RTE-page-withdrawal

Label: Page rename deletion

Caller → rename_page/delete_page → file mutation and reindex → subsequent active-page readers. Code owns changes with process filesystem authority. Closest supplied RT-RTE-knowledge-write describes create/update/ingest/notes/repair; rename/delete are distinct exposed operations, not a correction to that account.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Rename metadata/changed pages or deleted/tombstone/backlinks-before-deletion result.
• Later read-back: Later page readers/index consumers see changed active files; tombstone copy can remain in deleted directory. Existing ask-history records are not rewritten by these bodies.
• Delegated visibility: Invoking caller gets result; other consumers depend on subsequent read/cache behavior.
• Selection predicate: Caller-named page, source/target existence, update-links/tombstone flags; rename uses selected literal wiki-link replacements.
• Invalidation or expiry: Reindex called after mutation; no transitive historic-answer/sidecar/provenance retraction shown in these methods, and cached search can retain prior results until its TTL.
• Activation or effect: Removes active page or changes its identity/link targets; old answer assertions can remain retained.
• Evidence limits: Local operational withdrawal is not evidence of epistemic refutation, complete cascade invalidation, atomic rollback or observed fresh results.

Caller proposes and can withhold operation; code rejects missing source/existing rename target. No human per-content acceptance or factual oracle is inserted. Optional tombstone preserves page bytes on delete, not all dependent artifacts; rename writes new file then removes old, without a demonstrated full transaction. SRC-1 implementation, src/llm_wiki_mcp/server.py; cache selection corroborated by final search in src/llm_wiki_mcp/context_api.py.

if tombstone: shutil.copyfile(path, deleted_dir/f'{title}.{int(time.time())}.md') path.unlink(); self.reindex() return {'title': title, 'deleted': True, 'tombstone': tombstone, 'incoming_links_before_delete': backlinks['incoming']} --- src/llm_wiki_mcp/server.py

new_path.write_text(make_front_matter(meta)+body, encoding='utf-8') old_path.unlink() changed=[] --- src/llm_wiki_mcp/server.py

EPI-RTE-diagnostic-check

Label: Runtime document diagnostics

Two separated ledger functions share this diagnostic family: doc reconciliation returns fixed expected-file/command-presence measurements; the final context-window answer renders active settings and uncertainty. No counterpart in supplied runtime beyond adjacent RT-RTE-runtime-config and RT-CLM-context-budget: these diagnostics read/report rather than set budgets or build the pack.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Missing/undocumented-doc lists and ok, or context-window answer text naming estimates and unknown native model capacity.
• Later read-back: inapplicable — these diagnostic bodies do not independently retain updated content; enclosing ask may separately record returned text.
• Delegated visibility: Caller receives result; basic context-window ask bypasses model generation.
• Selection predicate: Fixed expected-document names and command substrings for reconciliation; basic intent plus active settings/history for context answer.
• Invalidation or expiry: Recomputed from current filesystem/settings; no continuing certificate of documentation completeness/provider capacity.
• Activation or effect: Diagnostic output advises caller; context answer returns directly. Reconciliation result has no automatic repair/disposition consumer in inspected body.
• Evidence limits: Presence checks do not establish semantic documentation accuracy; settings estimates do not probe native provider token capacity.

Open requests, program predicates not model judgment; caller can act later, no automatic acceptance gate/oracle or retained machinery update in these bodies. SRC-1 implementation, src/llm_wiki_mcp/context_api.py:

missing_docs = [d for d in expected_docs if not (docs / d).exists()] commands = list(release_command_tool_map().keys()) fn_text = function_list.read_text(encoding="utf-8", errors="replace") if function_list.exists() else "" undocumented = [c for c in commands if c not in fn_text] --- src/llm_wiki_mcp/context_api.py

"I do not know the exact native context-window length of the configured Ollama model from the wiki alone. The ask context is now configurable and less conservative than the earlier release.", --- src/llm_wiki_mcp/context_api.py

EPI-RTE-regression-check

Label: Regression fixture checks

Explicit developer/test invocation → suite using tool client or mocked generation → assertions/results aggregation → printed report and harness exit status. Program owns comparison against fixture expectations. No route counterpart in supplied runtime; RT-RTE-mcp-dispatch is exercised by one harness but is not this evaluator.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Regression result booleans/aggregate or assertion failures; async harness returns 0 only if all suite predicates pass.
• Later read-back: inapplicable — inspected harness does not retain a candidate acceptance artifact for later production use.
• Delegated visibility: Human/CI caller can receive printed JSON or test-script output; no provider quality judgment follows.
• Selection predicate: Explicit shipped test cases; expected tool-name set, seeded counts, fixture retrieval/mutation predicates, mocked history/sanitizer/context-window assertions.
• Invalidation or expiry: Test results would apply to that invocation/fixture only; none supplied here.
• Activation or effect: Aggregate controls harness exit code; no automatic deployment/successor-selection consumer shown.
• Evidence limits: Test source is not passing results; mocked answer establishes no actual model quality. Other tests remain uninspected in this member.

Mode is bounded fixture checks, not ordinary open asks. Test authors provide reference expectations; program decides pass/fail, developer chooses invocation and any later response. This constitutes narrow fixture oracles, not an answer oracle for arbitrary user questions. No production update proposal, rollback or promoted successor is implemented by result aggregation. SRC-1 implementation, src/llm_wiki_mcp/server.py, tests/ask_history_context_test.py:

results.append({'name':'tools_list_expected_surface','pass':names==expected,'details':{'count':len(names),'names':sorted(names)}}) --- src/llm_wiki_mcp/server.py

return 0 if suite['all_passed'] else 1 --- src/llm_wiki_mcp/server.py:1744-1744

return {"ok": True, "answer": "The first answer mentions a working prototype.", "elapsed_seconds": 0.01} --- tests/ask_history_context_test.py

Claims

EPI-CLM-wiki-substrate

Label: Markdown source of truth

SRC-1 doctrine/design, README.md. Claims retained Markdown authority for retrieval and model context. Distinct from RT-CLM-context-budget, which concerns sizing/inspection, not the storage authority claim. No counterpart among supplied runtime claims.

The core idea is simple: Markdown remains the human-readable source of truth, SQLite provides a fast local index, and an Ollama-compatible local model can answer questions using retrieved wiki context. --- README.md

EPI-CLM-provenance-refresh

Label: Provenance source refresh

SRC-1 doctrine/design, METHODOLOGY.md. Claims preserved provenance enables refresh; no supplied claim counterpart. Closest implementation account is RT-RTE-knowledge-write, not a separate claim identity.

1. Preserve provenance so pages can be refreshed when source files change. --- METHODOLOGY.md

EPI-CLM-answer-fallback

Label: Answer first fallback

SRC-1 doctrine/design, README.md. Claims a blank-response fallback gives direct best-effort answer, uncertainty and references. Distinct from the sizing and nonmutation claims RT-CLM-context-budget and RT-CLM-ask-nonmutation.

/ask uses the configured Ollama model when available. If model synthesis returns a blank response, the CLI now falls back to an answer-first evidence summary: it gives a direct best-effort answer, states uncertainty, and then lists the wiki pages/tools used. This avoids returning only references when the local model is overloaded or returns an empty response. --- README.md

EPI-CLM-note-suggestions

Label: Candidate link suggestions

SRC-1 emitted doctrine/design in src/llm_wiki_mcp/context_api.py. Generated notes distinguish suggested links from canonical or accepted links while requiring their inclusion in search. No supplied claim counterpart; closest route RT-RTE-knowledge-write is their producer. Passage is see EPI-RTE-note-template.

Evidenced absences

none declared in this member

Behavioral-authority paths

none declared in this member; paths are assessed per ledger row and supplied RT-BAP-answer-guidance, RT-BAP-config-control.