llm-wiki epistemic report

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

Source-and-claim boundary

Can llm-wiki acquire or produce truth-apt content, check it and license later reliance? The assessed boundary is the shipped knowledge-maintenance system, not Claude Code's model or automatic memory loader. See boundary and its Source register. All source anchors below use SRC-1 at that frozen commit. Natural-language maintenance procedures establish doctrine/design, not executable maintenance or observed compliance. The initializer supplies a narrower implementation layer.

The material counterexample to a blanket “no verification” conclusion is prune --l1: it requests per-claim evidence and correction before user admission. Conversely, structural lint, source attribution, recent timestamps and file retention do not by themselves establish factual acceptance. Public/design claims are RT-CLM-cache-precision, EPI-CLM-l1-honesty, EPI-CLM-confidence-lifecycle and EPI-CLM-automatic-repromotion.

Entry point or operation	Source path	Coverage and records	Limit/conclusion prevented
Plugin/copied command dispatch	SKILL.md, .claude-plugin/plugin.json, .claude-plugin/marketplace.json, setup.sh, wiki.md	Direct complete skill/manifest reads; supplied RT-BAP-skill-dispatch and RT-RTE-wiki-setup for remaining packaging	Host permission/loading implementation excluded; no universal mediation claim
Setup and direct initializer	scripts/init-wiki.sh, templates/obsidian/Schema.md	Direct complete reads; EPI-RTE-template-render	Other templates and installer behavior supplied through runtime/boundary, not independently reread
Ingest and bulk import	SKILL.md	Direct complete read; RT-RTE-source-ingest	No actual inputs/outputs; entailment and provenance preservation unknown
Query, fallback, logging, write-back	SKILL.md	Direct; RT-RTE-wiki-query	No answer/retrieval trace; actual synthesized propositions not determinable
L2 prune and re-promotion	SKILL.md, README.md	Direct skill plus README Query/Prune/Lint excerpt; RT-RTE-index-prune	Age/use is not evidence of falsity; README automation conflicts with skill offer
L1 classify, evidence, re-verify, correct, demote, delete, skip	SKILL.md, openspec/specs/prune.md	Direct skill and specification scenarios 9–10; RT-RTE-l1-verification	Spec scenarios are design examples, not executions
Lint, --fix, status, Dashboard	SKILL.md, openspec/specs/schema.md	Direct skill and schema Knowledge/Feedback excerpt; RT-RTE-health-maintenance	No executable independent validator or supplied check result
Other supporting docs, diagrams, examples, specifications	Paths inventoried in boundary	Not independently assessed beyond scoped search of README, architecture and specifications	No exhaustive consistency/historical claim; search hits alone are not corroborating operation
Provider parameters and host L1 consumption	External host	See RT-CMP-claude-code	Weight changes, exact version and endpoint substitution uninspected; no parametric-learning conclusion

No additional target-owned CLI dispatcher, hook or registered executable maintenance tool appears in the supplied tracked-tree account and inspected skill/manifest/initializer. The separate direct-script entry is included, rather than assumed subject to the model's quality gate. This is not a negative about arbitrary host extensions. Missing user vaults/L1 files prevent individuating actual claims or evaluating checks on them; missing traces prevent observed acceptance, activation or benefit; missing comparisons prevent causal retrieval or self-improvement claims. Unread supporting documents prevent a system-complete negative about every design formulation, but do not erase the positively evidenced routes.

Epistemic-object inventory

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
EPI-OBJ-wiki-pages	Source-derived facts, relationships, decisions and current/history states	Retained L2 material for answers	SRC-1, SKILL.md ingest/query	Actual propositions and initial warrant not supplied
EPI-OBJ-hub-index	Page descriptions can assert page subject matter; tags/links primarily select	Routing keys and archive membership	SRC-1, SKILL.md Hub-Index-Routing	Compression fidelity/relevance not checked independently
EPI-OBJ-l1-rules	Current-behavior claims; also non-truth-apt directives/preferences	Session-steering L1 rules and index descriptions	SRC-1, SKILL.md prune --l1	Rule contents absent; claim-class false does not mean every sentence lacks truth value
EPI-OBJ-access-log	Assertion that a page was read on a date, plus selection reason	Evidence for cold/hot selection	SRC-1, SKILL.md Access-Log	Self-reported by maintainer, not independent instrumentation
RT-OBJ-schema	Normative conventions; factual confidence labels	See cited record	SRC-1, templates/obsidian/Schema.md	Schema is not a truth evaluator
RT-OBJ-skill	Procedures plus design claims	See cited record	SRC-1, SKILL.md	Not a learned method or observed execution

L1 verdicts, query answers and Dashboard metrics are current outputs of the cited route containers, not independently evidenced persisted claim instances. Whether a particular answer adds an ampliative proposition is not determinable without its inputs and output.

Authority-route ledger

Each row separates function even when the same supplied route container covers several phases. Field ownership for the container's progression, owner and read-back remains in the cited record. Architectural status below concerns the specific function, not the host's ability to execute tools. “Policy” force means requested model/human behavior, not a hard interceptor.

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
EPI-RTE-template-render	content transformation	implemented	RT-OBJ-schema	non-truth-apt policy/content update: template parameter substitution	Shipped conventions to configured Schema	Python literal replacements, filesystem existence	Setup, before later commands	Created or skipped files	Executable write/skip	No factual warrant added	Installs conventions available to maintenance	Maintainer, Schema read, advisory policy, later commands	SRC-1, scripts/init-wiki.sh	none	none	Implementation deduction, not execution
RT-RTE-source-ingest	content transformation	doctrine only	EPI-OBJ-wiki-pages	truth-apt transformation: acquisition/import	External URL/file/text to extracted facts or converted notes	Foreground model using source/schema	User ingest/import	Candidate retained facts	Model-interpreted writes	Source warrant unknown; import is not independent production	Creates/updates pages, history and links	Future query agent, page reads, advisory content, later invocations	SRC-1, SKILL.md ingest/import	none	none	Extraction may alter meaning; no actual edge proves preservation or novelty
RT-RTE-source-ingest	check/evidence production	doctrine only	EPI-OBJ-wiki-pages	no content change	Properties, links, routing, credential exclusion, touches	Same model, structural/security-policy domain	After ingest writes; import calls lint	Issues/warnings	Policy report	Checks organization, not source truth	Reports deficiencies; no specified automatic rollback	Maintainer/user, gate report, advisory, current invocation	SRC-1, SKILL.md Phase 4	none	none	Gate name does not establish pre-write fact admission
RT-RTE-source-ingest	retention	doctrine only	EPI-OBJ-wiki-pages	no content change	Current state and superseded wording to Markdown/history/Git	Model decides supersession under schema	Ingest updates	Current headline plus dated old wording	Policy writes/commit	Retention preserves wording, not warrant	Supplies later content; old claims remain inspectable	Query/maintainer, files/history, advisory, across sessions	SRC-1, SKILL.md State + History	none	none	Not lifecycle integration after demonstrated acceptance
RT-RTE-wiki-query	operational admission/selection/consumption	doctrine only	EPI-OBJ-hub-index	no content change	Hubs/descriptions or grep to 3–5 pages	Model relevance judgment; namespace/description domain	Query before synthesis	Selected retained text, optionally L1	Ranking policy	Relevance selection grants no factual endorsement	Chooses evidence/context used in answer	Query agent, Read/search results, ranking, current answer	SRC-1, SKILL.md query Phases 0–1	RT-CLM-cache-precision	none	Fallback can recover archived pages; no guaranteed context eviction
RT-RTE-wiki-query	content transformation	doctrine only	EPI-OBJ-wiki-pages	truth-apt transformation: indeterminate	Selected pages to attributed answer/summary	Model, no reference-answer oracle	After reads	Candidate answer, stale/confidence warnings	Advisory answer	Attribution identifies pages, not established truth	User may rely or request write-back	User, natural-language answer, advisory, current/future actions	SRC-1, SKILL.md query Phases 2–4	none	none	Could reshape, derive or conjecture; inputs/output needed to classify
RT-RTE-wiki-query	disposition/acceptance	doctrine only	EPI-OBJ-wiki-pages	no content change	Useful summary/gap proposal to write-back	User confirms; model proposes	Optional after synthesis	Approval/refusal	Human admission protocol	Confirmation authorizes storage; no named factual acceptance criterion	Permits proposed page write	Maintainer, user response, permissive, proposed write	SRC-1, SKILL.md Optional Write-Back	none	none	No evidence-consuming factual acceptance established
RT-RTE-wiki-query	retention	doctrine only	EPI-OBJ-access-log	truth-apt transformation: acquisition/import	Actual full-page reads to dated log statements	Maintainer self-report	Each full read before synthesis	Append with matched reason	Append policy	Claimed use history, not accuracy or relevance proof	Feeds later prune/status	Prune/status agent, Markdown log, advisory evidence, later invocations	SRC-1, SKILL.md Phase 1b	none	none	Logging not covered by optional content-write confirmation
RT-RTE-index-prune	behavior/policy adaptation	doctrine only	EPI-OBJ-hub-index	non-truth-apt policy/content update: live-index demotion	Logged access/creation proxy to archive membership	Model threshold/exemptions; user veto	Requested/scheduled externally; confirmation before write	Archived marker and verbatim moved line	Human-gated ranking policy	Coldness does not withdraw factual warrant	Reduces preferred selection; grep/links still available	Query agent, live/archive hub sections, ranking, future queries	SRC-1, SKILL.md prune	RT-CLM-cache-precision, EPI-CLM-automatic-repromotion	README automatic vs skill offer	No deletion/denial; no measured precision benefit
RT-RTE-l1-verification	check/evidence production	doctrine only	EPI-OBJ-l1-rules	no content change	Due individual claims about paths/commands/flags/versions/etc.	Model interprets local files/source or read-only help/version; remote claims inconclusive	Batch, never-verified then oldest	Supports/contradicts/inconclusive; partial flag and evidence basis	Advisory verdict	Only checked claims within local evidence domain; source is not executed behavior	Recommends user actions, does not perform them	User, per-claim verdict/evidence, advisory, current decision	SRC-1, SKILL.md Gather Evidence	EPI-CLM-l1-honesty	none	No omniscient oracle; credentials checked for existence only
RT-RTE-l1-verification	disposition/acceptance	doctrine only	EPI-OBJ-l1-rules	no content change	Verdict/rule to re-verify, correct, demote, delete or skip	User per rule; inconclusive needs user knowledge; contradiction needs corrected text for re-verification	After evidence	Confirmed action or untouched due rule	Human admission protocol	Bounded acceptance procedure for intended continued L1 reliance, not infallibility	Authorizes only selected writes/removal	Maintainer, user response, permissive, per-item action	SRC-1, SKILL.md Act	EPI-CLM-l1-honesty	none	No supplied acceptance episode; user knowledge uninspected
RT-RTE-l1-verification	lifecycle integration	doctrine only	EPI-OBJ-l1-rules	truth-apt transformation: indeterminate	Approved correction to body, description, index and verified date	Maintainer applies approved text in all claim locations	After user admission	Corrected rule connected to refreshed verification metadata	Policy synchronized edits	Only approved claim's local scope; date alone not proof	Changes later rule text delivered by host	Future host agent, L1 body/index descriptions, directive policy, later sessions	SRC-1, SKILL.md Re-verify	EPI-CLM-l1-honesty	none	Conditional post-acceptance integration; semantic relation of unsupplied corrections unknown
RT-RTE-l1-verification	operational admission/selection/consumption	doctrine only	EPI-OBJ-l1-rules	non-truth-apt policy/content update: withdraw session rule	Confirmed demote/delete to L1 removal and reference repair	User chooses; model ensures carry-over precedes deletion	After evidence and confirmation	Rule removed; demotion history retained, credentials never demoted	Withdrawal protocol	No L1 reliance intended after removal; L2 history not renewed endorsement	Reduces future L1 loading; demotion changes channel to on-demand history	Host future loader / L2 query, index/files, selection, later sessions	SRC-1, SKILL.md Demote/Delete	EPI-CLM-l1-honesty	none	No Git undo; skipped/stale rules may still load; see RT-ABS-action-gate
RT-RTE-health-maintenance	check/evidence production	doctrine only	EPI-OBJ-wiki-pages, EPI-OBJ-hub-index	no content change	Page properties/link graph, age, index drift, credentials, L1 due metadata	Foreground model applying listed schema checks	Lint or lightweight status	Counts/severity/fix suggestions/Dashboard metrics	Advisory diagnostics	Structural consistency and age flags only	Informs maintenance; status itself modifies nothing	User/maintainer, report/Dashboard, advisory, current/later maintenance	SRC-1, SKILL.md lint/status	EPI-CLM-confidence-lifecycle	none	No independent regex executable or semantic fact validation
RT-RTE-health-maintenance	behavior/policy adaptation	doctrine only	EPI-OBJ-wiki-pages, EPI-OBJ-hub-index	non-truth-apt policy/content update: confidence downgrade and routing repair	Detected age/drift to confidence/index/stubs/cross-links	Model under caller --fix authorization, no per-fix approval	After lint scan	high to stale, backfilled/moved entries, stubs	Model repair policy	Lower confidence cautions; stub/links add no fact warrant	Changes future warnings and context selection; never writes L1	Query/maintainer, page metadata/index, advisory/ranking, later invocations	SRC-1, SKILL.md Auto-Fix	EPI-CLM-confidence-lifecycle	none	Semantic fidelity of generated descriptions unknown; no automatic multi-file rollback

These ledger splits do not supersede runtime's valid route containers. Acquisition/import is the established entry-edge classification; transformations within extraction/conversion remain semantically untested. No example of warranted-premise entailed derivation or demonstrated non-entailing conjecture is available. The L1 evidence route supplies a real design counterexample to a universal absence of acceptance; the ingest quality gate does not supply a corresponding fact-acceptance route.

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
RT-CLM-cache-precision	Routing/demotion keep L2 precise	SRC-1, SKILL.md; doctrine/design	Query/prune selection and fallback	No executable retrieval evaluator; see RT-RTE-wiki-query, RT-RTE-index-prune for prescribed routes	None	No contrast; age/use need not track correctness/relevance	Preferred-context control is specified	Precision benefit unknown; fallback qualifies “no more grep”
EPI-CLM-l1-honesty	L1 verification keeps rules honest	SRC-1, SKILL.md; doctrine/design	Evidence, verdicts, correction, human choice	None for checks; RT-RTE-l1-verification is prompt doctrine	None	Local/read-only boundary; unchecked/remote claims inconclusive; no action gate	Bounded evidence-responsive checking and acceptance protocol	Reliable truth/currentness or later compliance unestablished
EPI-CLM-confidence-lifecycle	high means verified/reliable/current	SRC-1, openspec/specs/schema.md; doctrine/design	Enum and downgrade semantics	No factual validator; prescribed RT-RTE-health-maintenance	None	Date threshold tests age, not reliability	Intended warning vocabulary and stale repair	No evidence-consuming route for initially assigning high in assessed ingest/query procedures
EPI-CLM-automatic-repromotion	Archived query hits automatically re-promote	SRC-1, README.md; doctrine/design	Main skill says offer re-promotion	None; RT-RTE-wiki-query specifies offer	None	No run chooses between conflicting texts	Re-hit recovery is prescribed as an offer	README overstates automatic disposition

Bounded conclusion

• Acquisition and retention: RT-RTE-source-ingest requests extraction/import, current-state replacement and retained history. Source warrant is unknown. It does not establish that a retained fact was independently checked. The structural gate assesses properties, links, routing and credential policy, not the truth of the source's assertions.
• Answers and continued reliance: RT-RTE-wiki-query selects retained text and requests attributed synthesis with inherited confidence/age warnings. Selection and optional user-approved write-back authorize use/storage, not evidence-consuming acceptance of a novel proposition. Actual answer content is unavailable, so non-ampliative reshaping, entailment and ampliative conjecture cannot be separated at that edge.
• Local checking, acceptance and integration: RT-RTE-l1-verification prescribes a distinct chain: named local claim → evidence basis and bounded verdict → user decision → corrected rule/verification metadata and synchronized descriptions/index. Supports recommends continued L1 use; contradiction permits re-verification only with correction; inconclusive requires user knowledge; skip leaves the rule due. This is a design-level acceptance criterion for local behavior claims intended to steer future sessions, not observed acceptance or a guarantee of truth. Demotion/delete withdraw L1 delivery only after user confirmation and reference/carry-over handling. Demoted history remains retrievable but is not endorsed as current. The external host supplies actual later loading, and RT-ABS-action-gate prevents claiming action-time protection.
• Direct adaptation without new factual warrant: RT-RTE-index-prune lets usage evidence change future preferred retrieval through human-approved archive moves. RT-RTE-health-maintenance lets age/drift checks alter confidence and routing under --fix. Neither access frequency nor dates prove factual accuracy. Reports/logs alone are not improvement; no trace closes later behavioral activation or causal benefit.
• Independently bounded properties: Provider parameter changes/version resolution remain uninspected in RT-CMP-claude-code. The L1 theory-property assessment remains the separately qualified one in RT-RTE-l1-verification: localized rule criticism/revision is prescribed, but actual theory contents, iteration and learning benefit are unobserved. Human internal admission decisions prevent an autonomous verification-route claim. There is no supplied evidence of criticism of llm-wiki's own method against operation; no exercised self-improvement or global reflection conclusion follows from file revision alone. These limits coexist with the positive local checking protocol.

Shared records

Operative objects

EPI-OBJ-wiki-pages

Label: Wiki content pages

Natural-language source-derived facts, current-state blocks and dated History stored in Logseq/Obsidian Markdown; properties support model interpretation. Closest supplied record is RT-RTE-source-ingest, the producer route, not this retained object. Distinct also from RT-OBJ-schema, the conventions governing pages. Pages later reach query and maintenance consumers; no actual vault supplied. SRC-1, SKILL.md ingest and State + History.

• Delete nothing: every superseded statement stays readable in the history --- SKILL.md

EPI-OBJ-hub-index

Label: Hub Index and Archive

Natural-language descriptions with link/tag identifiers in retained hub Markdown. The model ranks descriptions; symbolic-looking links do not establish an executable relevance rule. Distinct retained object overlapping consumers/producers RT-RTE-wiki-query, RT-RTE-index-prune and RT-RTE-source-ingest, not a replacement for those routes. SRC-1, SKILL.md Hub-Index-Routing.

A routing line is [[link]] -- description #tags, one per child page. --- SKILL.md

EPI-OBJ-l1-rules

Label: L1 rules and loaded descriptions

Natural-language behavior claims/directives in external user Markdown files, plus frontmatter claim class/verification date and index pointers/descriptions. Actual contents absent. Closest supplied RT-RTE-l1-verification is the checking/change route, not the object; host consumer is separately RT-CMP-claude-code. SRC-1, SKILL.md L1 verification.

• Scan every L1 memory file EXCEPT the index file (e.g. MEMORY.md — an index, not a rule). Its lines are auto-loaded too: keep them in step whenever a rule is corrected, demoted, or deleted --- SKILL.md

EPI-OBJ-access-log

Label: Access-Log entries

Dated page identifiers plus natural-language matched reasons in append-only wiki Markdown. Fixed date/link positions are prescribed for prune/status parsing; reasons are model-interpreted routing transparency. Distinct from the producing RT-RTE-wiki-query and consuming RT-RTE-index-prune. Source is maintainer-reported reads, not external fact verification. SRC-1, SKILL.md Access-Log.

• prune/status parse the date + [[page]] from fixed positions (split on --); the matched: suffix is irrelevant to LRU aggregation and does not affect parsing --- SKILL.md

Routes

EPI-RTE-template-render

Label: Schema template rendering Part of: RT-RTE-wiki-setup

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Created/skipped path messages; generated Schema file if absent.
• Later read-back: Setup installs static shipped conventions, so use-derived read-back is inapplicable; later maintenance reads the generated Schema as guidance.
• Delegated visibility: inapplicable — Python receives flags/templates, not a delegated model task.
• Selection predicate: Tool flag chooses template directory; namespace flags supply replacement text; existing target is skipped.
• Invalidation or expiry: No automatic expiry or overwrite; existing Schema survives rerun.
• Activation or effect: Python writes configured conventions; their later influence depends on host reads/model interpretation and is unobserved.
• Evidence limits: Source inspection and deterministic deduction only; no initializer execution or vault inspection.

This is the template-rendering part of the setup container, not another setup entry. Bash supplies tool/root/namespaces/date; embedded Python reads the selected Schema, replaces namespace/date tokens and calls write_file. For concrete symbolic input tool=obsidian, namespaces Tech Reference, date 2026-10-17, the inspected code maps {{NAMESPACES}} to Wiki/Tech, Wiki/Reference and {{DATE}} to 2026-10-17, targeting Wiki/Schema.md. If that file exists it is skipped; otherwise rendered text is written. This deduction establishes non-truth-apt convention instantiation, not a newly warranted proposition. User flags guide the admitted configuration; unsupported tool/invalid namespace fails before writes. No factual oracle or improvement trigger applies. Recovery is rerun/skip, not transaction rollback. SRC-1, scripts/init-wiki.sh, templates/obsidian/Schema.md.

schema = read_template("Schema.md") schema = schema.replace("{{NAMESPACES}}", ", ".join(f"Wiki/{ns}" for ns in namespaces)) write_file(page("Schema"), schema.replace("{{DATE}}", today)) --- scripts/init-wiki.sh

if os.path.exists(path):
    print(f"Skipped (already exists): {label}")
    return

--- scripts/init-wiki.sh

Claims

EPI-CLM-l1-honesty

Label: L1 honesty objective

SRC-1, SKILL.md claims L1 verification keeps rules honest. Distinct from RT-CLM-cache-precision, which concerns L2 selection; supported design mechanism is RT-RTE-l1-verification, not observed correctness.

One mechanism keeps L1 honest: --- SKILL.md

• Verdict per CLAIM (a rule usually makes several): supports / contradicts / inconclusive, with what was checked and what was found --- SKILL.md

• Overall verdict: contradicts if any claim contradicts; supports if all checked claims support; else inconclusive. Mark it partial when claims stayed unchecked, and name them --- SKILL.md

The source-supported countercase to overbroad endorsement is an unchecked or remote claim: partial support is not whole-rule truth. The design scenario below also individuates intended contradiction handling, not an observed episode:

THEN the system SHALL report verdict contradicts, with both paths as evidence AND recommend demote or delete AND offer re-verify only together with a corrected rule text naming the new path AND write nothing until the user chooses --- openspec/specs/prune.md

EPI-CLM-confidence-lifecycle

Label: Confidence lifecycle semantics

SRC-1, openspec/specs/schema.md assigns high confidence a verified/reliable/current meaning. No supplied claim counterpart after comparison with RT-CLM-cache-precision; RT-OBJ-schema is the governing artifact, not this warrant claim. Ingest requires confidence but its quality gate does not establish factual verification; lint's age downgrade is narrower than these semantics.

• REQ-533: The confidence property follows this lifecycle:
• high → verified, reliable, up-to-date
• medium → probably correct, verify on next ingest
• low → uncertain, incomplete, or limited sources

• stale → was high/medium but updated is 90+ days old --- openspec/specs/schema.md

• Downgrade stale confidence from high to stale --- SKILL.md

EPI-CLM-automatic-repromotion

Label: Automatic archived-page re-promotion

SRC-1, README.md Prune calls re-promotion automatic. Distinct claim from RT-CLM-cache-precision; closest RT-RTE-wiki-query already correctly describes an offer. This flags a source-document mismatch, not a defect or supersession of runtime.

the page is still found by the L3 grep fallback (and re-promoted automatically if queried again). --- README.md

• If the L3 fallback hits an archived page (archived:: set) — a re-hit on an evicted page: offer to re-promote it — move its routing line back into the hub ### Index, drop the archived:: property --- SKILL.md