ai-memex-cli epistemic report

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

Source-and-claim boundary

The question is whether source capture, wiki maintenance and later answering establish warranted content, rather than only retained or operationally usable content. Scope and exclusions are those of the boundary; evidence identities are in its Source register. This is source inspection, not target execution. Public claims are EPI-CLM-memory-compounds, EPI-CLM-synthetic-pages, EPI-CLM-consistent-vault and the supplied RT-CLM-management-use.

Coverage below distinguishes new direct reads from supplied runtime findings. Newly inspected files remain SRC-1 implementation or attributed doctrine/design, not new sources. Dispatch inventory was checked with a scoped search of src/cli.ts, installed command definitions in src/commands/install-hooks.ts, repository slash prompts and installed workflow references. The direct CLI read includes lint, link-check, search, new/log and injection dispatch. No additional provider-native tool registry is established by that inspection; external host internals are outside the target.

Entry point or operation	Source path	Coverage / supporting records
URL/search/crawl capture, delegated capture	src/core/fetcher.ts; src/commands/fetch.ts	Direct helper read plus supplied RT-RTE-capture-source; EPI-RTE-fetch-render adds lineage assessment. Live extraction fidelity remains unknown.
Mechanical and delegated session capture	src/core/distiller.ts; src/commands/distill.ts	Direct parser/renderer and selected command/prompt reads; EPI-RTE-session-render, RT-RTE-session-distill.
Ingest and lint-feedback prompt	src/commands/ingest.ts	Direct prompt read plus supplied RT-RTE-ingest-write.
Mechanical lint, semantic lint	src/commands/lint.ts; src/core/schema.ts; src/core/linker.ts; src/core/wiki-index.ts	Direct complete reads; EPI-RTE-mechanical-check, RT-RTE-semantic-lint.
Watch/loop and after-clean context refresh	src/core/ingest-lint-loop.ts; src/commands/watch.ts	Direct complete loop read, supplied watcher account; EPI-RTE-loop-disposition, RT-RTE-watch-repair.
Standalone link-check, including --fix	src/commands/link-check.ts	Direct complete read and dispatch read; EPI-RTE-link-check. Additional exposed reporting path, not an omitted semantic evaluator.
Query/slash/skill synthesis	templates/skills/ai-memex/references/query-workflow.md; .claude/commands/memex/query.md; src/commands/install-hooks.ts	Direct reads of reference, local query prompt and installed command registry; EPI-OBJ-query-answer, RT-CLM-management-use.
Search/inject/glob/context delivery	src/commands/search.ts; src/commands/inject.ts; src/commands/glob.ts; src/commands/context.ts	Supplied RT-RTE-context-delivery, not a new exhaustive audit of backends or projection helpers.
New page, action log, status	src/commands/new.ts; src/commands/log.ts; src/commands/status.ts; templates/synthesis.md	Direct complete reads. Scaffolding, caller-supplied action entries and counts, not factual acceptance gates; supplied RT-RTE-ingest-write covers the alternate page writer.
Init/onboard/config/migrate/install/update	src/cli.ts; src/commands/install-hooks.ts; src/commands/migrate.ts; src/commands/update.ts	Dispatch inventory plus supplied RT-RTE-install-update, RT-OBJ-agent-profiles and RT-RTE-ingest-write. Onboarding interaction and full helper correctness not independently audited; no system-complete security or evaluator-exhaustiveness claim.
Source, citation and contradiction instructions	templates/AGENTS.md; templates/skills/ai-memex/references/ingest-workflow.md; templates/skills/ai-memex/references/safety-rules.md	Direct complete reads; EPI-CLM-synthetic-pages, RT-BAP-schema-prompt.

No real vault, transcript, generated answer, semantic-lint finding or approval record was supplied. Those gaps prevent identifying actual candidate propositions, checking source-to-page fidelity or establishing acceptance and downstream activation. No trace or experiment establishes better answering, factual accuracy, convergence or causal improvement. External inference and qmd internals remain uninspected; model weights, exact versions and training changes are bounded by RT-CMP-coding-agent. Available but unread repository-local prompt variants and design documents prevent an exhaustive doctrine inventory, not the narrower route findings below.

Epistemic-object inventory

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
EPI-OBJ-raw-material	Imported web claims and recorded session utterances; actual propositions not determinable without inputs	see record	SRC-1, src/core/fetcher.ts	A recorded utterance is not evidence that its embedded factual assertion is true. Tool evidence can be omitted.
EPI-OBJ-wiki-content	Source claims, comparisons and proposed insights; actual propositions not determinable	see record	SRC-1, templates/AGENTS.md	Page type and citation fields do not establish warrant.
EPI-OBJ-lint-report	Assertions about indexed links, inbound references and metadata validity	see record	SRC-1, src/commands/lint.ts	Truth scope is the implemented predicates on selected pages, not world facts.
EPI-OBJ-query-answer	Requested source-backed answer, synthesis or inference; actual proposition unavailable	see record	SRC-1, templates/skills/ai-memex/references/query-workflow.md	No answer instance or evidence-consuming acceptance decision.
RT-OBJ-ingest-prompt	Schema directives are policies; embedded index and lint assertions may be truth-apt	see record	SRC-1, src/commands/ingest.ts	Split by consumer below: structural feedback versus live-schema guidance.
RT-CMP-coding-agent	Cannot individuate propositions in distributed-parametric state	see record	SRC-1, src/core/agent-adapter.ts	Invocation is inspectable; internal representations and parameter updates are not.

Authority-route ledger

Fields owned by supplied records are referenced rather than redefined. Linked entries using one route record separate its consequential functions; they do not allocate replacement identities.

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-fetch-render	content transformation	implemented	EPI-OBJ-raw-material	truth-apt transformation: acquisition/import	see record	see record	see record	see record	see record	URL/date lineage retained; source warrant unknown; full-document lineage degraded by selection	see record	see record	SRC-1, src/core/fetcher.ts	EPI-CLM-memory-compounds	none	Extraction is not verification.
EPI-RTE-session-render	content transformation	implemented	EPI-OBJ-raw-material	truth-apt transformation: non-ampliative reshaping	see record	see record	see record	see record	see record	Retained text records speech, not its correctness; supporting tool-result context can be lost	see record	see record	SRC-1, src/core/distiller.ts	none	lineage-loss	Concrete input deduction in record; not executed.
RT-RTE-session-distill	content transformation	implemented	EPI-OBJ-raw-material	truth-apt transformation: indeterminate	Session to best-practice document	External agent; role/session scope, no oracle	see record	see record	see record	Requested implicit lessons could reshape or add claims; no output supports entailment classification	see record	see record	SRC-1, src/commands/distill.ts	EPI-CLM-memory-compounds	none	Need source/output pair to distinguish compression, derivation and conjecture.
RT-RTE-session-distill	behavior/policy adaptation	implemented	RT-OBJ-ingest-prompt	non-truth-apt policy/content update: requested session-learning notes in live schema	Session experience to future guidance	External agent decides notes; CLI does not validate their use	see record	see record	see record	No increased capacity or factual license established	see record	see RT-BAP-schema-prompt	SRC-1, src/commands/distill.ts	EPI-CLM-memory-compounds	none	Notes can contain truth-apt rationale, but no instance individuates it.
RT-RTE-ingest-write	content transformation	implemented	EPI-OBJ-wiki-content	truth-apt transformation: indeterminate	Sources to source/entity/concept/synthetic pages	Child agent under live schema; collaborative discussion requested, no enforced per-claim gate	see record	see record	see record	Citations, labelled inference and preserved conflicts are requested; no checked derivation or factual acceptance established	see record	see record	SRC-1, src/commands/ingest.ts	EPI-CLM-synthetic-pages	none	Neither synthesis label nor failure to prove entailment establishes ampliation.
RT-RTE-ingest-write	retention	implemented	EPI-OBJ-wiki-content	no content change	Durable writes, index and log	see record	see record	see record	see record	Operational retention precedes any established acceptance	see record	see record	SRC-1, src/commands/ingest.ts	EPI-CLM-memory-compounds	none	Agent writes are requested, not verified by caller. Not lifecycle integration after acceptance.
EPI-RTE-mechanical-check	check/evidence production	implemented	EPI-OBJ-lint-report	truth-apt transformation: entailed derivation	see record	see record	see record	see record	see record	Conditional deduction over parsed index and metadata; not source truth, semantic consistency or useful knowledge	see record	see record	SRC-1, src/core/linker.ts	EPI-CLM-consistent-vault	scope-overclaim	Parser fidelity and selected-page scope bound the result.
EPI-RTE-loop-disposition	disposition/acceptance	implemented	EPI-OBJ-lint-report	no content change	see record	see record	see record	see record	see record	Accepts structural-clean stop condition only, not wiki factual content	see record	see record	SRC-1, src/core/ingest-lint-loop.ts	EPI-CLM-consistent-vault	scope-overclaim	No content rejection/rollback gate; retained edits survive stall/cap.
EPI-RTE-link-check	check/evidence production	implemented	EPI-OBJ-lint-report	truth-apt transformation: entailed derivation	see record	see record	see record	see record	see record	Only exact extracted target membership among basename IDs	see record	see record	SRC-1, src/commands/link-check.ts	none	none	--fix does not fix or withdraw content.
RT-RTE-semantic-lint	check/evidence production	implemented	EPI-OBJ-wiki-content	truth-apt transformation: indeterminate	Cross-page contradictions, stale claims and other schema categories	Child agent scans wiki; no independent answer oracle	see record	see record	see record	Can propose content criticism; no supplied finding establishes its correctness or coverage	see record	see record	SRC-1, src/commands/lint.ts	EPI-CLM-memory-compounds	none	New diagnostic assertions may be reshaping, derived or conjectural; output needed.
RT-RTE-semantic-lint	disposition/acceptance	implemented	EPI-OBJ-wiki-content	no content change	Safe repair versus unresolved review item	Model judges safety/ambiguity; no CLI evidence-consuming factual acceptance record	see record	see record	see record	Direct edit permission, not established acceptance of an ampliative claim	see record	see record	SRC-1, src/commands/lint.ts	EPI-CLM-memory-compounds	none	Ambiguity defers review in prompt; no implementation gate verifies the deferral.
RT-RTE-watch-repair	behavior/policy adaptation	implemented	RT-OBJ-ingest-prompt	non-truth-apt policy/content update: prioritize structural repairs in next invocation	Prior report to repair priorities	Code forwards report; child interprets instructions	see record	see record	see record	Structural defects guide work, not an epistemic oracle for claims	see record	see record	SRC-1, src/commands/ingest.ts	EPI-CLM-consistent-vault	none	This is feedback-driven work selection, not demonstrated theory criticism or improvement.
RT-RTE-context-delivery	operational admission/selection/consumption	implemented	EPI-OBJ-wiki-content	no content change	Select pages/snippets/includes/digest for outer host	see record	see record	see record	see record	Delivery grants availability, no additional truth license	see record	see RT-BAP-host-bootstrap for installed path; other consumers see route	SRC-1, src/commands/context.ts	EPI-CLM-memory-compounds	none	Task/token flags do not enforce relevance/budget; excerpts and old projection files can limit evidence.
RT-RTE-context-delivery	content transformation	doctrine only	EPI-OBJ-query-answer	truth-apt transformation: indeterminate	Wiki evidence to foreground answer	Outer agent instructed to distinguish facts, synthesis, inference and unknowns	see record	see record	Advisory reference text	No answer acceptance or derivation check established	Requested answer in current session; saving is a separate requested workflow	see record	SRC-1, templates/skills/ai-memex/references/query-workflow.md	RT-CLM-management-use	none	CLI implements delivery, not the foreground synthesis algorithm.
RT-RTE-install-update	operational admission/selection/consumption	implemented	RT-OBJ-agent-profiles	non-truth-apt policy/content update: expose workflows or replace distribution	Hook installation or product replacement	see record	see record	see record	see record	Package/build success is not knowledge-quality evaluation	see record	see record	SRC-1, src/commands/update.ts	none	none	Explicit upgrade is not self-directed 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-memory-compounds	see record	SRC-1, README.md, doctrine/design	see record	RT-RTE-ingest-write, RT-RTE-semantic-lint, RT-RTE-context-delivery	None supplied	No comparison or activation evidence; semantic checks are optional and model-dependent	Persistent pages and reuse channels support an accumulation mechanism	Already-flagged contradictions and synthesis reflecting everything read are not established guarantees.
EPI-CLM-synthetic-pages	see record	SRC-1, templates/AGENTS.md, doctrine/design	see record	RT-RTE-ingest-write, RT-RTE-session-distill	None supplied	No concrete input/output propositions or acceptance record	System requests synthetic content and provenance-preserving treatment	Cannot infer actual conjecture, checked entailed derivation or accepted new knowledge from type labels.
EPI-CLM-consistent-vault	see record	SRC-1, src/core/ingest-lint-loop.ts, implementation message/design claim	see record	EPI-RTE-mechanical-check, EPI-RTE-loop-disposition	None supplied	Zero mechanical issues licenses only the selected structural predicates	Deterministic stop and best-effort refresh are wired	“Knowledge base is consistent” exceeds evaluator domain if read semantically.
RT-CLM-management-use	see record	SRC-1, templates/skills/ai-memex/SKILL.md, doctrine/design	see record	RT-RTE-ingest-write, RT-RTE-context-delivery	None supplied	External host controls actual use; schema and query reference differ on saving	Child maintenance and foreground-use guidance are distinct	No uniform read-only query or per-content approval invariant.

Bounded conclusion

Retention and reuse are wired; warrant is route-specific. RT-RTE-ingest-write requests durable wiki/index/log writes. RT-RTE-context-delivery makes selected retained content available by search, includes, copies or host digest. Neither route establishes that a user actually relied on it or that reliance improved an answer. Retention is not post-acceptance lifecycle integration.

Acquisition preserves some attribution but not full evidential context. EPI-RTE-fetch-render retains URL/date with extracted Markdown, while full source warrant remains unknown. EPI-RTE-session-render preserves selected utterance text and role metadata but can discard the tool-result evidence behind it. The source-supported countercase to “distill is semantic learning” is its mechanical default; the countercase to “mechanical means lossless transcript” is block-level omission. Both matter for later source citation.

Structural checks have real but narrow force. EPI-RTE-mechanical-check and EPI-RTE-link-check derive diagnostic propositions over parsed pages, not warranted world knowledge. EPI-RTE-loop-disposition consumes zero issues as a criterion for stopping this repair run and attempting registered context refresh. Its intended use is structural maintenance of the selected vault; its horizon is this loop and the subsequently refreshed host projection. Nonzero reports can prioritize another agent invocation, while no-change/cap stops leave edits available. That is operational disposition, not factual acceptance or withdrawal. A missing/unchecked citation cannot be cured merely by zero link issues.

Semantic judgment is delegated, not absent, but its results are unobserved. RT-RTE-semantic-lint requests content-directed diagnosis, direct safe repairs and review pages for ambiguity. It is the source-supported countercase to claiming that the system checks only structure. Schema and reference guidance also request evidence, explicit inference and preservation of contradictions. However, no supplied proposition, checked derivation, evaluator finding or evidence-consuming decision establishes produced accepted ampliative knowledge. Neither novelty nor the synthesis directory settles the transformation class. A later factual acceptance and its lifecycle integration remain unestablished, rather than universally impossible.

Instruction adaptation has a separate path. RT-RTE-session-distill requests appending session learnings to the live schema, and RT-BAP-schema-prompt identifies a later interpreted guidance channel. RT-RTE-watch-repair also feeds structural observations into subsequent priorities. These are inspectable update requests and feedback mechanisms; no parameter-training operation is established by the inspected caller, and external parameter changes remain uninspected. Future-capacity improvement, exercised self-improvement, content-directed theory criticism and autonomous success do not follow from those mechanisms. Their absence is not inferred from model opacity either. No single system-wide epistemic verdict follows.

Shared records

Operative objects

EPI-OBJ-raw-material

Label: Captured raw material

Source-native identity: raw/ web Markdown and session Markdown. Representational form: natural-language assertions/utterances, with symbolic URL/date/role/frontmatter fields. Storage substrate: UTF-8 Markdown files under raw scenes. Content truth and source reliability are not established by capture. Distinct from RT-OBJ-ingest-prompt: persisted evidence inputs, not an invocation prompt. It is an object spanning the outputs of RT-RTE-capture-source and RT-RTE-session-distill, not a replacement or part of exactly one route. Evidence: SRC-1, src/core/fetcher.ts, src/core/distiller.ts, implementation; load-bearing passages appear in the transformation records below.

EPI-OBJ-wiki-content

Label: Wiki page content

Source-native identity: wiki/<scene>/ page bodies and frontmatter. Representational form: natural language consumed by agents and symbolic page IDs, links and metadata consumed by index/lint code. Storage substrate: Markdown in the vault. Distinct from RT-OBJ-ingest-prompt: durable content produced/loaded across invocations, not its prompt carrier. Closest supplied routes are RT-RTE-ingest-write and RT-RTE-context-delivery; object versus writer/reader, not duplicate referents. Evidence: SRC-1, templates/AGENTS.md (doctrine/design), src/core/wiki-index.ts (implementation).

const id = basename(filePath, '.md'); const links = extractWikiLinks(body); --- src/core/wiki-index.ts

EPI-OBJ-lint-report

Label: Mechanical lint report

Source-native identity: LintReport and LintIssue arrays. Representational form: symbolic typed fields interpreted by the loop; issue descriptions/JSON also become natural-language evidence for agents. Storage substrate: current-run objects and output JSON/text; not intrinsically durable accepted knowledge. Distinct from RT-OBJ-ingest-prompt, which can embed it, and RT-RTE-watch-repair, which consumes it. Evidence: SRC-1, src/commands/lint.ts, implementation; report construction and predicate passages below.

EPI-OBJ-query-answer

Label: Foreground query answer

Source-native identity: answer requested by Query Workflow. Representational form: natural-language synthesis/citations, potentially other schema-requested output forms. Storage substrate: outer session response; durable page only via a further writing workflow. Closest supplied RT-RTE-context-delivery describes delivery and foreground use, not a concrete answer object; no supplied object counterpart. Evidence: SRC-1, templates/skills/ai-memex/references/query-workflow.md, doctrine/design.

1. Distinguish:
2. known facts from the wiki
3. source-backed synthesis
4. inference
5. unknown or missing knowledge --- templates/skills/ai-memex/references/query-workflow.md

Routes

EPI-RTE-fetch-render

Label: Direct fetch extraction and rendering Part of: RT-RTE-capture-source

Endpoints/progression: network body → DOM/optional aggressive stripping → Readability selection → Turndown Markdown → URL/date-wrapped raw document; outer fetch command owns persistence. This part assesses direct extraction, not the alternate agent fetch path. Evaluator/condition: HTTP status rejects errors at or above 400; extraction libraries select body text, not factual credibility. No factual oracle. Mode: open source capture. User/query selects sources; code admits converted output, caller may cancel/dry-run in the parent route; no claim-level approval or rollback added here. Evidence: SRC-1, src/core/fetcher.ts, implementation.

• Immediate return: FetchResult, then Markdown string for the command's writer.
• Later read-back: written raw document can enter ingest/search via the parent route.
• Delegated visibility: inapplicable — this extraction part calls no coding agent.
• Selection predicate: supplied URL and optional aggressive DOM stripping; Readability article or body fallback.
• Invalidation or expiry: no extraction TTL; repeated destination writes belong to parent capture route.
• Activation or effect: deterministic extraction and attribution wrapping implemented; factual endorsement not performed.
• Evidence limits: third-party extraction fidelity and live source truth uninspected; no fetched instance supplied.
• implementation conclusion status: wired
• operation conclusion status: uninspected

Source warrant is unknown. URL and fetch date survive, but source completeness is not preserved by article extraction; optional nav/header/footer removal supplies a concrete loss channel. Acquisition is not production of the imported assertions.

const htmlContent = article?.content ?? dom.window.document.body?.innerHTML ?? ''; const content = td.turndown(htmlContent); --- src/core/fetcher.ts

source-url: ${result.url} fetched: ${result.fetchedAt} word-count: ${result.wordCount} --- src/core/fetcher.ts

EPI-RTE-session-render

Label: Mechanical session normalization and rendering Part of: RT-RTE-session-distill

Endpoints/progression: selected file → parseSessionStructured → normalized messages → renderSessionMarkdown → raw Markdown written by mechanical command. Code owns selection/serialization; no model evaluator, oracle or semantic acceptance. Open conversion request; caller selects input/mode. Admission is parser-recognized messages or whole-text fallback when none parse. No content-specific veto or automatic rollback in this part. Evidence: SRC-1, src/core/distiller.ts, src/commands/distill.ts, implementation.

• Immediate return: rendered Markdown and write destination message.
• Later read-back: raw session document is a later ingest source; sources metadata points to original file path.
• Delegated visibility: inapplicable — mechanical branch launches no agent.
• Selection predicate: structured parser recognizes supported message shapes; block flattening selects text and tool names, discards tool-result blocks; malformed/unrecognized records may be skipped.
• Invalidation or expiry: default current-session path uses mtime skip through parent route; explicit mechanical conversion writes selected output, with no general TTL.
• Activation or effect: normalizes and writes a transcript projection; no evidence of learning or correctness checking.
• Evidence limits: input/output deduction below uses inspected code, not a target run; other provider formats remain bounded by parser recognition.
• implementation conclusion status: wired
• operation conclusion status: uninspected

Concrete deduction: take one JSONL object {"message":{"role":"assistant","content":[{"type":"text","text":"The test passed."},{"type":"tool_result","content":"FAIL"}]}}. toStructured takes the Claude-style branch; flattenContent keeps The test passed. and discards the tool_result. Rendering emits an Assistant heading and that text, but no FAIL. This is non-ampliative selection/reshaping of text, not truth validation of the retained assertion. It degrades the retained evidential context. A source path is provenance, not a substitute for omitted counterevidence. Conversely, the structured renderer can preserve a standalone generic role tool message with string content; this is not a claim that all tool evidence is always removed. The legacy mechanicalExtract skips system/tool-use/tool-result lines, but the command calls the structured parser/renderer, not that legacy extractor.

} else if (t === 'tool_result') { / 丢弃工具结果 / --- src/core/distiller.ts

const raw = await readFileUtf8(normalized); const messages = parseSessionStructured(raw); --- src/commands/distill.ts

const bodyText = m.text.trim() || '(空)'; return ## ${label}${ts}\n\n${bodyText}${toolLine}; --- src/core/distiller.ts

EPI-RTE-mechanical-check

Label: Mechanical wiki predicate evaluation

Distinct identity: shared evaluator implementation, rather than the whole controller RT-RTE-watch-repair or delegated RT-RTE-semantic-lint. It serves multiple callers and is not contained in exactly one supplied route. Endpoints/progression: wiki files → parsed page index → optional scene subset → requested predicate families → LintReport. Owner: deterministic code. Target/domain: exact extracted basename-ID membership, inbound links and required/allowed metadata fields on selected indexed pages. Expected factual answers are not supplied; predicates themselves define the structural criterion. No human/model diagnosis in this evaluator. Mode: open lint request. Evidence: SRC-1, src/commands/lint.ts, src/core/linker.ts, src/core/schema.ts, src/core/wiki-index.ts, implementation.

• Immediate return: structured report; command can print text/JSON or forward it to semantic pass.
• Later read-back: report can shape a subsequent ingest invocation within the loop; no durable report retention by runLint itself.
• Delegated visibility: semantic/repair agents receive report representation via parent callers, not hidden semantic evidence from this evaluator.
• Selection predicate: configured checks default to orphans, broken-links and missing-frontmatter; optional scene subset is used before link checks.
• Invalidation or expiry: recomputed from files on each call; report superseded by later loop report.
• Activation or effect: diagnostic facts derive from current parsed inputs; parent decides continuation and refresh separately.
• Evidence limits: no check of factual page bodies, citation support or source reliability. Same-basename collisions, parsing and scoped-page selection limit what “link exists” means.
• implementation conclusion status: wired
• operation conclusion status: uninspected

The conditional entailment is concrete: with parsed page IDs {a} and outbound target {b}, membership test entails a broken-link entry for a → b. A page with all required metadata, no outbound links and no inbound links remains an orphan; a false factual body can pass other predicates. sources is optional and its evidential support is not inspected. index.md is outside the wiki index used here; adding only an index entry need not clear an orphan, despite feedback suggesting it. These scope restrictions prevent semantic-consistency inference.

if (!allIds.has(link)) { broken.push({ source: page.id, target: link }); } --- src/core/linker.ts

export const REQUIRED_FIELDS: (keyof Frontmatter)[] = ['name', 'description', 'type', 'scene']; --- src/core/schema.ts

if (options.scene) { pages = pages.filter((p) => p.scene === options.scene); } --- src/commands/lint.ts

EPI-RTE-loop-disposition

Label: Mechanical report disposition Part of: RT-RTE-watch-repair

Endpoints/progression: report → empty/nonempty/repeated predicates → clean return and optional context refresh, another repair iteration or stalled/capped/signal return. Code owns this disposition. Criterion/intended use/scope: zero issues means no reported mechanical defects for the loop's lint invocation, sufficient to stop structural repair and attempt registered-host refresh. It is not factual acceptance of page claims. Proposer: prior ingest may edit pages; decider: loop code; veto/limits: caller stop, cap and repeat guard, with force bypass of repeat guard. Mode: open repair, no answer oracle beyond structural predicates. Recovery: stops do not roll back writes. Evidence: SRC-1, src/core/ingest-lint-loop.ts, implementation.

• Immediate return: iteration count, stop reason and final report.
• Later read-back: nonclean report enters next invocation; clean report can trigger durable host-context refresh, later consumed by host.
• Delegated visibility: next ingest gets report through prompt formatting; no per-claim acceptance record.
• Selection predicate: zero issues; otherwise bounded iteration, signal or sameIssueSet guard. Repeat keys omit frontmatter error detail.
• Invalidation or expiry: report replaced at next iteration; no claim withdrawal, expiry or rollback.
• Activation or effect: enforces loop continuation/stopping and invokes best-effort refresh; does not enforce truth or host compliance.
• Evidence limits: clean can follow failed/unchanged ingestion; neither convergence nor semantic correctness follows.
• implementation conclusion status: wired
• operation conclusion status: uninspected

if (report.issues.length === 0) { logger.success('[loop] lint clean — knowledge base is consistent.'); --- src/core/ingest-lint-loop.ts

if (!opts.skipContextRefresh) { await silentRefreshContexts(opts.vault, opts.cwd); } return { iterations: i, stopped: 'clean', finalReport: report }; --- src/core/ingest-lint-loop.ts

EPI-RTE-link-check

Label: Standalone link reporting

Distinct identity: additional CLI entry point not separately recorded in runtime. It reuses the broken-link evaluator of EPI-RTE-mechanical-check, but reports independently rather than controlling RT-RTE-watch-repair. Endpoints/progression: CLI → vault index → exact link-target membership → terminal diagnostic. Owner/evaluator: deterministic code; target is extracted wiki links, domain is basename IDs in indexed wiki pages. Caller requests scan; no claim proposals or per-content approval. Open diagnostic mode, no semantic oracle. --fix logs an unimplemented message; no revision admission, rejection or rollback. Evidence: SRC-1, src/commands/link-check.ts, src/cli.ts, implementation.

• Immediate return: valid-links message or broken source/target list; optional manual-fix advice.
• Later read-back: inapplicable — this route writes no retained content or report.
• Delegated visibility: inapplicable — no delegated agent call; terminal reader may act manually.
• Selection predicate: all pages returned by buildWikiIndex, then findBrokenLinks.
• Invalidation or expiry: inapplicable — current scan output only, no retained artifact managed here.
• Activation or effect: advisory reporting; --fix performs no automated repair.
• Evidence limits: link existence is not cited-claim support or truth, and output consumption is unobserved.
• implementation conclusion status: wired
• operation conclusion status: uninspected

if (options.fix && broken.length) { logger.info('Auto-fix not implemented in MVP. Agent should resolve these manually.'); } --- src/commands/link-check.ts

Claims

EPI-CLM-memory-compounds

Label: Memory that compounds

Claimed operation: agent-maintained retained knowledge accumulates, cross-references and contradiction flags are already available, and synthesis reflects what has been read. Source: SRC-1, README.md, doctrine/design. Distinct from RT-CLM-management-use: knowledge accumulation/coverage claim rather than work-context separation. No observed-run or causal layer.

Your AI agents, finally with a memory that compounds. --- README.md

Cross-references are already there. Contradictions have already been flagged. The synthesis already reflects everything you've read. --- README.md

EPI-CLM-synthetic-pages

Label: Derived insights and honest contradictions

Claimed operation: synthetic page types carry derived insights and connections; ingestion preserves explicit source conflict rather than silently replacing it. Source: SRC-1, templates/AGENTS.md, with source/inference constraints in templates/skills/ai-memex/references/ingest-workflow.md, doctrine/design. Closest RT-CLM-management-use concerns delegation and saving, not the epistemic role of page content. This claim has no instance-level derivation or acceptance evidence.

• synthesis — derived insights, theses, or non-trivial connections (syntheses/) --- templates/AGENTS.md

• Every new factual claim should trace to a raw file or existing wiki page.

• If a claim is synthesized from multiple sources, cite all important sources.
• If sources conflict, preserve the conflict instead of flattening it.
• If a statement is an inference, label it as an inference. --- templates/skills/ai-memex/references/ingest-workflow.md

Do not choose a winner unless the evidence clearly supports it or the user decides. --- templates/skills/ai-memex/references/safety-rules.md

EPI-CLM-consistent-vault

Label: Clean lint means consistent knowledge base

Claimed operation: a success message describes the knowledge base as consistent after zero mechanical issues. Source: SRC-1, src/core/ingest-lint-loop.ts, implementation text carrying a design/warrant implication, not an observed success. No counterpart among supplied claims after comparison with RT-CLM-management-use. Closest supplied route RT-RTE-watch-repair already limits the predicates; EPI-RTE-loop-disposition supplies the exact passage. Interpreted narrowly as structural cleanliness the message matches; interpreted as factual consistency it exceeds the evaluator's domain.

Evidenced absences

none declared in this member

Behavioral-authority paths

none declared in this member