Hippocampus epistemic report

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

Source-and-claim boundary

The question is how the personal vault obtains claims, checks them, authorizes reliance and carries results into later sessions. Scope and exclusions are those of boundary; evidence identities are in its Source register. The shipped host contracts and executable utilities are included; external host/model internals and actual private vaults are not. All source anchors below refer to SRC-1 at the reviewed commit. No target execution was performed. Architectural wiring, requested model behavior and observed operation are kept separate.

The consequential warrant claims are EPI-CLM-structural-checks, EPI-CLM-graph-knowledge, and the freshness claim RT-CLM-cache-refresh. These concern inspectability, meaningful links and available recent context, not a formal proof engine. The primary stated objective is transparency and minimal infrastructure rather than superior recall; the source itself acknowledges wording and scale limits (see EPI-CLM-graph-knowledge).

Coverage

Direct reads here supplement, rather than merely repeat, the supplied runtime account. Available but not independently reread material is explicitly marked.

Entry point or operation	Source path (SRC-1)	Coverage / records	Limit and conclusion prevented
Local, outside-file, pasted and URL ingest; batch/re-ingest	.claude/skills/ingest/SKILL.md, _templates/source.md	Direct complete reads; RT-RTE-source-ingest, EPI-OBJ-source, EPI-OBJ-archive	No source-specific extraction trace: semantic preservation and exercised rejection unobserved.
Conversation save	.claude/skills/save/SKILL.md, _templates/note.md	Direct complete reads; RT-RTE-conversation-save	No saved-conversation pair: relation between input and saved propositions remains indeterminate.
Query and cross-project read	CLAUDE.md	Direct complete read; RT-RTE-vault-query	No query/reference-answer set: recall, correctness and transfer unmeasured.
Mechanical lint CLI, both text and JSON output	scripts/vault_lint.py	Direct complete read, including argparse and final return; RT-RTE-vault-lint, EPI-OBJ-findings	Checks structures, not propositions about the world.
Editorial lint, stale claims, deletion/merge/resolution	.claude/skills/lint/SKILL.md, CLAUDE.md	Direct complete reads; RT-RTE-vault-lint	Sampling allowed; human criteria for which conflicting claim is true unspecified.
SessionStart and Stop, both hosts	.claude/settings.json, .codex/hooks.json, scripts/on-stop.sh	Direct complete reads; RT-RTE-lifecycle-hooks	Host execution and interpretation unobserved; timestamp ordering is not summary fidelity.
Host entry/skill adapters	AGENTS.md, .agents/skills/ingest/SKILL.md, .agents/skills/save/SKILL.md, .agents/skills/lint/SKILL.md	Supplied runtime findings; RT-BAP-project-guidance	No independent adapter reread here; runtime establishes redirects, not loading.
Framework update/diff/export	scripts/sync_framework.sh	Supplied runtime findings; RT-RTE-framework-update, RT-RTE-framework-export	Source-level pathname/copy admission only; neither semantics of future revisions nor activation is established.
Public warrant and walkthrough	README.md, docs/why-no-vector-db.md, docs/what-happens-when-you-ingest.md	Direct complete reads; epistemic claims below	Comparisons are design arguments, not causal experiments.
Initial retained surfaces	wiki/hot.md, wiki/index.md, wiki/log.md	Direct complete reads; EPI-OBJ-hot, EPI-OBJ-index	Scaffold only: no provenance-linked production/check/acceptance instance.
Other templates, Obsidian/configuration and non-operative roots	_templates/entity.md, _templates/concept.md, _templates/project.md, .obsidian/, other roots listed by boundary	Supplied boundary coverage, not direct reads here	No independent assessment of other template fields or browsing behavior. Images remain excluded.

There is no separately registered tool server or model executor in the supplied runtime account: tools are host-native and skills are instructions. This report assesses all material route families identified above, but does not promote supplied coverage or uninspected external behavior into a system-complete negative about all possible host-mediated checks. Missing provider internals prevent a conclusion about internal model learning or hidden reasoning; missing observations prevent claims of successful memory activation or benefit. These limits do not make the inspected structural evaluator unknowable.

Epistemic-object inventory

Object fields owned by declarations are referenced rather than duplicated.

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
EPI-OBJ-archive	Imported source assertions, if present; arbitrary input may contain none	see record	SRC-1 .claude/skills/ingest/SKILL.md	see record
EPI-OBJ-source	Attributed source summary and key claims	see record	SRC-1 _templates/source.md	see record
EPI-OBJ-pages	Concepts, entity facts, saved findings, decisions/rationale	see record	SRC-1 .claude/skills/save/SKILL.md	see record
EPI-OBJ-hot	Recent facts and active-thread descriptions	see record	SRC-1 wiki/hot.md	see record
EPI-OBJ-index	Claims describing what a page covers; journal claims about operations	see record	SRC-1 wiki/index.md, wiki/log.md	see record
EPI-OBJ-findings	Container of the two separately assessed output parts below	see record	SRC-1 scripts/vault_lint.py, .claude/skills/lint/SKILL.md	see record
EPI-OBJ-mechanical-findings	Program-derived predicates about current file structures	see record	SRC-1 scripts/vault_lint.py	see record
EPI-OBJ-editorial-findings	Model allegations of staleness/conflict and proposed repairs	see record	SRC-1 .claude/skills/lint/SKILL.md	see record
RT-OBJ-project-contract	Policies, not newly established factual knowledge	see record	see record	see record
RT-OBJ-hook-config	None in event-dispatch rules	see record	see record	see record
RT-CMP-agent-model	Not determinable at the parameter-content level	see record	see record	see record

Authority-route ledger

Repeated route citations below identify different consequential functions of the same supplied route, not new record identities. For endpoints, owner, state effects, read-back, activation and other generic route fields, see the cited declaration. This overlay adds epistemic classification and distinguishes checking, disposition, retention and consumption. “Doctrine only” describes the semantic route prescribed to a host model; “implemented” describes executable checks/transport, not model compliance.

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-source-ingest / acquire	content transformation	doctrine only	EPI-OBJ-archive	truth-apt transformation: acquisition/import	External content enters archive	Host read/fetch/copy under model protocol	On user ingest	Original or fetched extraction retained	Requested source handling	Source warrant unknown; provenance is not endorsement	Source supplies data, not instructions	Foreground model, source text, advisory evidence, current ingest	SRC-1 .claude/skills/ingest/SKILL.md	none	none	URL stores extracted text, not necessarily original bytes; see archive record.
RT-RTE-source-ingest / duplicate check	check/evidence production	doctrine only	EPI-OBJ-source	no content change	Hash equals an existing content_hash	SHA-256 plus model-directed search; byte identity domain	Before extraction	Duplicate candidate reported	Requested check	Content identity only, no truth license	Supplies evidence for stopping	Foreground model, hash/search result, advisory, current request	SRC-1 .claude/skills/ingest/SKILL.md	none	none	Tool algorithm requested, no supplied execution.
RT-RTE-source-ingest / duplicate stop	disposition/acceptance	doctrine only	EPI-OBJ-source	no content change	Whether to process again	Model follows duplicate rule; user can override	After duplicate result	Stop or explicit re-ingest	Requested admission policy	No epistemic acceptance	Blocks repeated processing unless user asks	Model, skill rule, permissive/enforcing by compliance, request	SRC-1 .claude/skills/ingest/SKILL.md	none	none	Not per-claim verification.
RT-RTE-source-ingest / consolidate	content transformation	doctrine only	EPI-OBJ-source, EPI-OBJ-pages	truth-apt transformation: indeterminate	Source to English summaries, selected claims and linked concepts	Model interpretation; likely future reference is page-worthiness criterion	After source/context/template reads	Seed pages, revisions and source links	Requested model transformation	Attributed candidates; confidence/status fields do not establish evidence-consuming acceptance	May write retained claims without universal per-content approval	Foreground model then later readers, Markdown, advisory knowledge, later sessions	SRC-1 .claude/skills/ingest/SKILL.md, _templates/source.md	EPI-CLM-graph-knowledge	unknown	Intended extraction/translation may reshape; no input/output pair establishes preservation, entailment or non-entailment.
RT-RTE-source-ingest / conflict check	check/evidence production	doctrine only	EPI-OBJ-source, EPI-OBJ-pages	no content change	Competing claims in new and existing pages	Model compares propositions; no explicit truth oracle	During ingest	Conflict noticed or missed	Requested semantic check	A tension allegation, not adjudication	Makes conflict available for later resolution	Model, source/pages, advisory, ingest and future maintenance	SRC-1 CLAUDE.md, .claude/skills/ingest/SKILL.md	EPI-CLM-graph-knowledge	none	No guarantee that all competing pages are read.
RT-RTE-source-ingest / flag conflict	retention	doctrine only	EPI-OBJ-pages	truth-apt transformation: indeterminate	Record conflict on both pages without silent overwrite	Model writes callouts; user reserves resolution	On detected conflict	Both candidate assertions remain with warning	Requested preservation protocol	Withholds automatic adjudication, not automatic downstream use	Later readers may see warning; neither side auto-withdrawn	Model/user, warning callouts, advisory, until resolution	SRC-1 CLAUDE.md	EPI-CLM-graph-knowledge	none	No forced query abstention on warned pages.
RT-RTE-conversation-save / synthesize	content transformation	doctrine only	EPI-OBJ-pages	truth-apt transformation: indeterminate	Conversation to durable conclusions and rationale	Model chooses what is worth keeping; ambiguous content prompts user	On save request	Declarative cited note or page update	Requested synthesis	Candidate conclusions; citation does not certify source truth	Can write without general per-content approval	Model, save skill and conversation, advisory content, later sessions	SRC-1 .claude/skills/save/SKILL.md	none	none	Some outputs may be compression, others new inference; no pair individuates these alternatives.
RT-RTE-source-ingest, RT-RTE-conversation-save / file	retention	doctrine only	EPI-OBJ-source, EPI-OBJ-pages, EPI-OBJ-index	no content change	Candidate persistence and graph/catalog connection	Model/host writes; page-worthiness, not warranted truth, selects	On ingest/save completion	Pages, links, index/log entries and archive	Requested retention	Retained candidates, not established accepted ampliative claims	Makes content available to future selection	Future agent and human, files/catalog, advisory/ranking, sessions	SRC-1 .claude/skills/ingest/SKILL.md, .claude/skills/save/SKILL.md	EPI-CLM-graph-knowledge	none	Cross-linking before epistemic acceptance is retention/organization, not lifecycle integration after acceptance.
RT-RTE-vault-lint / scan	check/evidence production	implemented	EPI-OBJ-mechanical-findings	truth-apt transformation: entailed derivation	Parsed fields, names, links, counts and index membership	Python fixed predicates over current snapshot	Explicit CLI; requested after large batch and fixes	JSON/text findings and counts	Reports, does not block writes	Narrow claims about scanner-defined structures	Supplies fix candidates, no automatic admission	Foreground model/user, stdout, advisory, scan/maintenance	SRC-1 scripts/vault_lint.py	EPI-CLM-structural-checks	scope limit	See concrete deduction below; parser and resolver boundaries limit license.
RT-RTE-vault-lint / editorial check	check/evidence production	doctrine only	EPI-OBJ-editorial-findings	truth-apt transformation: indeterminate	Stale contradictory claims, missing links/pages, old seeds	Model search/judgment, possibly sampled; 3+ mentions and 30+ days are editorial triggers	After scan	Proposed repairs, development, merges or deletion	Requested judgment	Allegations/proposals; age does not establish falsity	Can nominate content for revision	Foreground model, maintenance findings, advisory, current maintenance	SRC-1 .claude/skills/lint/SKILL.md	none	none	Sampling prevents full semantic coverage.
RT-RTE-vault-lint / approve fixes	disposition/acceptance	doctrine only	EPI-OBJ-pages, EPI-OBJ-mechanical-findings, EPI-OBJ-editorial-findings	no content change	Which fixes and conflicting claims to retain	Human bulk approval for safe edits, individual approval for merge/delete/contradiction; model proposes	Before model edits	Approve, reject or defer	Model-mediated user veto policy	User decides truth in contradiction cases; criterion/evidence protocol unspecified	Permits approved revisions, not a universal write gate	User → foreground model, approval response, permissive, current maintenance and persisted edits	SRC-1 .claude/skills/lint/SKILL.md	EPI-CLM-graph-knowledge	unknown	A possible human acceptance decision is afforded, not evidenced as a performed, criterion-bound decision.
RT-RTE-vault-lint / apply repair	content transformation	doctrine only	EPI-OBJ-pages	truth-apt transformation: indeterminate	Approved changes replace/remove page content; rerun checks structures	Model/host executes approved edits; user can reject proposals	After approval	Content revision, deletion, merges, updated log/hot	Requested repair	Rescanning licenses structures only, not truth of a resolved claim	Changes later available content; recovery manual	Future readers, edited files, advisory knowledge, sessions	SRC-1 .claude/skills/lint/SKILL.md	none	none	No explicit dependency invalidation propagates withdrawn claims to every derivative.
RT-RTE-lifecycle-hooks / timestamp	check/evidence production	implemented	EPI-OBJ-hot	truth-apt transformation: entailed derivation	Whether any non-hot Markdown file has newer mtime	Shell find and guarded test, filesystem timestamps	Eligible first Stop	Refresh request or no request	Hook stdout; host must honor block	Temporal ordering, not semantic freshness or session causation	Can request another model turn before stopping	Host/model, JSON reason, enforcing only through host, current stopping turn	SRC-1 scripts/on-stop.sh	RT-CLM-cache-refresh	bounded guard	Missing hot, recursive stop, markers and equal/older timestamps bypass.
RT-RTE-lifecycle-hooks, RT-RTE-source-ingest, RT-RTE-conversation-save / refresh	content transformation	doctrine only	EPI-OBJ-hot	truth-apt transformation: indeterminate	Wiki/session content to <=500-word summary	Model instructed to overwrite four sections	After writes or hook request	Replacement summary	Requested rewrite	No established semantic fidelity or claim endorsement	Narrows recent context later injected	Foreground model → later model, hot text, advisory, sessions	SRC-1 scripts/on-stop.sh, CLAUDE.md	RT-CLM-cache-refresh	unknown	Summary may compress or introduce interpretation; no traced pair.
RT-RTE-lifecycle-hooks / persist	retention	implemented	EPI-OBJ-hot, EPI-OBJ-pages, EPI-OBJ-archive	no content change	Working-tree content to Git history	Shell stages content roots and commits existing index	Eligible stop after guard or bypass	Commit or logged failure	Local Git side effects	Versioning is not epistemic acceptance	History can aid manual recovery	Human/Git, local repository, recovery affordance, future sessions	SRC-1 scripts/on-stop.sh	none	none	Failure returns zero; prestaged files also enter commit; see runtime.
RT-RTE-lifecycle-hooks / reload	operational admission/selection/consumption	implemented	EPI-OBJ-hot	no content change	File to model-visible context	Host startup/resume/clear/compact callback runs cat	SessionStart events	Hot text printed if present	Wired supply, host delivery external	Delivered summary without added warrant	Recent facts may influence later answers/actions before new checking	Host/model, hook context, advisory knowledge, invocation	SRC-1 .claude/settings.json, .codex/hooks.json	RT-CLM-cache-refresh	unknown	Delivery/compliance/activation not observed; Codex context limit is external semantics.
RT-RTE-vault-query / read-answer	operational admission/selection/consumption	doctrine only	EPI-OBJ-hot, EPI-OBJ-index, EPI-OBJ-pages	truth-apt transformation: indeterminate	Relevant pages to cited answer or scoped abstention	Model relevance/confidence, search fallback, no reference answer	On question	Cited synthesis, or not-in-vault plus research offer	Requested read/abstain protocol	Source-relative answer; not independently verified world knowledge	Limits training-data substitution; no hard gate enforces it	Foreground/cross-project agent, file reads and citations, advisory/ranking, answer/current task	SRC-1 CLAUDE.md	EPI-CLM-graph-knowledge	unknown	Selection and confidence can miss or misread; not-in-vault is not exhaustive absence.
RT-RTE-framework-update, RT-RTE-framework-export / replace policy	behavior/policy adaptation	implemented	RT-OBJ-project-contract, RT-OBJ-hook-config	non-truth-apt policy/content update: copy framework instructions and executable rules	Path-eligible incoming/local framework differences	see supplied records	see supplied records	see supplied records	see supplied records	No semantic quality or knowledge acceptance established	Future loading can alter guidance, templates and checks	see supplied records and RT-BAP-project-guidance	see supplied records, SRC-1 scripts/sync_framework.sh	RT-CLM-local-edit-protection	see supplied claim	No evidence-responsive improvement criterion or exercised successor evaluation.

Deterministic deduction and attempted counterexamples

For the lint row, the target is link resolution in the scanner's filename map, not the truth of the linked page or whether a heading exists. From inspected implementation, input text [[Plan#Missing]] produces target Plan: link_targets removes the heading portion. With a scanned file whose stem is Plan, the resolver gets a nonempty list and does not emit a dead-link finding for a missing heading. This is a static deduction, not a target run. Conversely an absent stem produces a dead-link finding. It refutes an unqualified reading of “all link correctness” while supporting basename-resolution checks. The scanner's fixed parsing rules warrant the derived predicates only for its actual input snapshot and interpretation, not omitted fields or real-world claims. See EPI-OBJ-mechanical-findings for the implementation passages.

Other stronger interpretations were checked against adverse cases. Ingest's complete-read requirement explicitly allows sectional extraction for very large sources, and URL archival saves extracted content; neither implies lossless raw-byte provenance. The query fallback can find unindexed pages, so “unindexed means invisible” is too strong. Contradiction callouts preserve rather than reject conflicting candidates. Lint approval is not required by ordinary ingest/save or Stop. Stop's recursive/missing-file branches prevent an invariant semantic-refresh claim. These limits follow the supplied routes and the direct source reads above; none establishes that the intended protocol failed in a real run.

No ledger row establishes lifecycle integration after epistemic acceptance: graph connection occurs during candidate filing, and human resolution has no supplied decision trace identifying evidence, criterion and intended scope. This does not deny operational organization or possible human acceptance. Nor does it infer the absence of internal model reasoning from opacity. There is no source-supported input/output pair here that individuates an accepted ampliative proposition; the indeterminate transformations retain non-ampliative reshaping, entailed inference and ampliative conjecture as alternatives pending such evidence.

System-claim versus route comparison

Claim ID	Claimed operation or warrant	Claim source/anchor and layer	Doctrine/design support	Implemented route IDs	Observed-run support	Causal support and design limits	Supported conclusion	Mismatch/unknown
EPI-CLM-structural-checks	see record	see record	see record	RT-RTE-vault-lint	None supplied	No comparison; fixed predicates support only their specified structures	Inspectable deterministic defect reporting, not semantic validation	“Duplicate under slightly different names” exceeds basename equality; missing headings are not checked.
EPI-CLM-graph-knowledge	see record	see record	see record	No executable semantic-link evaluator; RT-RTE-source-ingest, RT-RTE-vault-query are model protocols	None supplied	No component intervention or comparative retrieval benchmark	Links can carry written intent and point to retained claims; explicit conflicts can be preserved for judgment	Link existence does not establish that causal intent is true or that one-hop retrieval beats similarity.
RT-CLM-cache-refresh	see record	see record	see record	RT-RTE-lifecycle-hooks	None supplied	No semantic-fidelity check; guarded timestamps	Wired reload and bounded refresh reminder	see supplied claim; fresh mtime is not endorsement.
RT-CLM-local-edit-protection	see record	see record	see record	RT-RTE-framework-update, RT-RTE-framework-export	None supplied	No exercised successor comparison	Path-based framework change admission only	see supplied claim; not evidence of improved method.

Bounded conclusion

Hippocampus affords acquisition and durable organization of imported source claims and conversation conclusions. The archive/hash/origin protocol can preserve lineage, but source truth is unassessed, fetched extraction may omit original content, and neither a confidence field nor seed/mature status records an evidence-consuming acceptance decision. Source and saved-page transformations remain indeterminate without concrete input/output pairs; novelty or summary labels cannot establish either ampliation or entailment (EPI-OBJ-archive, EPI-OBJ-source, EPI-OBJ-pages).

The inspected program produces narrow, truth-apt structural findings by deterministic derivation from file contents and names. Its findings advise the foreground model/user; they do not prevent writes or certify page truth. Editorial contradiction checks are model prescriptions. Contradiction handling withholds automatic adjudication and preserves warnings; maintenance reserves individual resolution, merging and deletion to user judgment. That is a real decision path, but the source specifies neither a truth criterion nor a performed acceptance with evidence and scope. Rechecking after repairs checks structure, not the truth of the repair (EPI-OBJ-mechanical-findings, EPI-OBJ-editorial-findings, RT-RTE-vault-lint).

Retained pages, catalog descriptions and hot summaries receive operational reliance through model selection, hook injection and cited answers. Their force is advisory knowledge or relevance ranking; policies against training-data substitution and source instructions depend on host/model compliance. No check prevents a still-conflicted page or a faulty summary from influencing a later answer. Versioning is recovery, not validation; recency is not endorsement. The horizon is later invocations/sessions, but neither activation nor benefit is observed (RT-RTE-vault-query, RT-RTE-lifecycle-hooks).

The ordinary write paths are model-directed, not universally human-approved. Maintenance decisions and framework synchronization are distinct: one applies a user approval policy to fixes, the other copies behavior-shaping files by pathname/difference rules. Neither establishes autonomous evidence-responsive method improvement. Provider parameter changes remain uninspected, independently of Markdown adaptation. Implemented memory affordances, reflection, learning, autonomy and exercised self-improvement must not be collapsed into one verdict; this evidence establishes no causal benefit or exercised self-improvement claim (RT-CMP-agent-model, RT-RTE-framework-update, RT-RTE-framework-export).

The supported architecture is therefore a source-relative, inspectable retention and retrieval workflow with localized structural evaluation and human-reserved contradiction resolution. It does not, on the supplied evidence, establish produced accepted ampliative knowledge or comparative retrieval superiority. This is a bounded conclusion about those claims, not a negative score for a deliberately personal, low-infrastructure vault.

Shared records

Operative objects

EPI-OBJ-archive

Label: Archived originals

Source-native identity: inbox/_done/ and original input files. Form: natural-language source text when textual; other input/attachment encodings depend on the source and are uninspected in the empty scaffold. Substrate: local files and, conditionally, Git history. Claimed role: original provenance, not a truth oracle. Closest supplied record is RT-RTE-source-ingest, a process rather than this object; no supplied operative-object counterpart. Local files are to be moved/copied without editing; pasted material is saved verbatim; URL material is fetched extraction. Hash/origin identifiers are symbolic lineage metadata, not proof that the attributed proposition is true. No real archive/source pair is supplied. SRC-1 .claude/skills/ingest/SKILL.md.

• A URL — fetch it with the available web tool, save the extracted content to inbox/_done/<slug>-<YYYY-MM-DD>.md with source_url and fetched in frontmatter, then ingest. --- .claude/skills/ingest/SKILL.md

• Decide the archive path now: inbox/_done/<original-name>; if that name is already taken, suffix with date-time (report.md → report-20260711-1030.md). Never overwrite an archived original. --- .claude/skills/ingest/SKILL.md

EPI-OBJ-source

Label: Source pages

Source-native identity: wiki/sources/ pages using _templates/source.md. Natural-language summary, key claims, caveats and entity/concept descriptions coexist with symbolic type/date/hash/path/link fields in local Markdown. Claimed role: attributable source account and provenance bridge. Closest supplied RT-RTE-source-ingest is its producer route, not the retained object; no supplied object counterpart. The confidence: medium default and status: seed do not name an evaluator, criterion, evidence consumption or accepted intended use. A claim may be attributed without being accepted. No populated source pages are supplied. SRC-1 _templates/source.md, .claude/skills/ingest/SKILL.md.

confidence: medium # high | medium | low content_hash: "" origin: "inbox/_done/" --- _templates/source.md

Key Claims

• --- _templates/source.md

EPI-OBJ-pages

Label: Derived and saved pages

Source-native identity: entity/concept/project pages and wiki/notes/ synthesis/decision/session pages. Natural-language propositions, explanations, rationale and warnings; symbolic YAML metadata and wikilinks; local Markdown substrate. Claimed role: reusable durable knowledge, later read by foreground/cross-project agents and humans. Different from EPI-OBJ-source: this is derived or conversation-saved content rather than the source account. Closest supplied RT-RTE-conversation-save and RT-RTE-source-ingest describe production, not an object counterpart. Citation and cross-link requirements provide intended attribution, not guaranteed semantic preservation. Contradiction warnings remain claims awaiting user resolution; not automatic retractions. SRC-1 .claude/skills/save/SKILL.md, .claude/skills/ingest/SKILL.md, CLAUDE.md.

1. Write the page from the matching _templates/ skeleton: declarative present tense, in the language set in the vault CLAUDE.md → Conventions, self-contained (readable cold in 6 months), every mentioned entity/concept wikilinked, claims cited as (Source: [[Page]]). If it answers a specific question, put the question verbatim in the question: field.

--- .claude/skills/save/SKILL.md

1. Check for contradictions with existing pages. Flag both sides with > [!warning] Contradiction with [[Page]] callouts — never silently overwrite.

--- .claude/skills/ingest/SKILL.md

EPI-OBJ-hot

Label: Hot cache

Source-native identity: wiki/hot.md. Natural-language recent facts/change/thread summary with symbolic YAML fields, stored as overwriteable Markdown. Claimed role: recent context supplied before later questions. Closest supplied RT-RTE-lifecycle-hooks is its delivery/persistence route, not an object counterpart. Different from the static RT-OBJ-project-contract: accumulated cache is instance state, not the operating contract. Scaffold text establishes an available initial candidate, not execution of summary production. Summary warrant/fidelity remains unknown; source references can be lost in condensation. SRC-1 wiki/hot.md, CLAUDE.md, scripts/on-stop.sh.

Day zero — vault scaffolded, nothing ingested yet. (This cache refreshes automatically after the first session that changes the wiki.) --- wiki/hot.md

EPI-OBJ-index

Label: Catalog and operations journal

Source-native identity: wiki/index.md and wiki/log.md. Natural-language page descriptions and operation/insight statements with symbolic wikilinks/date/action headings; local Markdown. Claimed roles: relevance selection versus operation lineage, respectively. Grouped only as navigation/lineage support here; their consumers differ: queries choose from catalog, humans/maintenance can consult history. No runtime object counterpart; RT-RTE-source-ingest and RT-RTE-conversation-save are their update routes. No full audit invariant follows: journaling is requested in workflows, while hooks commit without checking a log entry. Fallback search counters absolute invisibility of unindexed content. SRC-1 wiki/index.md, wiki/log.md, CLAUDE.md, scripts/on-stop.sh.

Master catalog. One line per page: wikilink + one-sentence description. Newest entries first within each section. --- wiki/index.md

Append-only journal. New entries at the TOP. Never edit past entries. --- wiki/log.md

EPI-OBJ-findings

Label: Mechanical and editorial findings

Source-native identity: the two-pass lint findings presented by the foreground agent, optionally retained as wiki/meta/lint-report-YYYY-MM-DD.md. This is a mixed-output container, not one evaluator or one warrant class. Its symbolic mechanical part is EPI-OBJ-mechanical-findings; its natural-language editorial part is EPI-OBJ-editorial-findings. Substrate: scanner stdout, foreground context, and optional Markdown report. Closest supplied RT-RTE-vault-lint is the producing/consuming workflow rather than an object counterpart. The parts below own form, warrant and consumer findings separately. Retaining this container does not merge their epistemic authority or imply observed operation. SRC-1 scripts/vault_lint.py, .claude/skills/lint/SKILL.md.

Present findings grouped by severity, each with a one-line suggested fix. If the user wants a record (or there are many findings), write it to wiki/meta/lint-report-YYYY-MM-DD.md (type: meta). --- .claude/skills/lint/SKILL.md

EPI-OBJ-mechanical-findings

Label: Mechanical findings

Part of: EPI-OBJ-findings

Source-native identity: scripts/vault_lint.py output findings and summary. Representational form: symbolic severity/check/file/message dictionaries and aggregate counts, rendered as JSON or fixed text messages; substrate: stdout, then foreground context and optionally the combined Markdown report. The Python scanner produces predicates about its decoded file snapshot, parsed fields, filenames and link maps. These warrant scanner-defined structural conclusions, not world-claim truth, heading existence or semantic duplicate detection. The foreground agent is explicitly instructed to trust this output rather than re-derive it; findings then inform proposed fixes presented to the user. Force is advisory evidence for maintenance, not automatic write rejection or epistemic acceptance of page claims. Scope/horizon: the scanned snapshot and maintenance episode; a retained report may outlast that snapshot, with no automatic freshness guarantee. The completed scanner returns zero even with findings. No actual scan output or successful repair is supplied. Unlike EPI-OBJ-editorial-findings, this part is program-derived rather than model editorial judgment. The closest supplied RT-RTE-vault-lint is the containing workflow, not an output counterpart. SRC-1 scripts/vault_lint.py, .claude/skills/lint/SKILL.md; implementation and doctrine/design respectively.

findings.append({ "severity": severity, "check": check, "file": str(path.relative_to(root)) if isinstance(path, Path) else path, "message": message, }) --- scripts/vault_lint.py

print(json.dumps({"summary": summary, "findings": findings}, indent=2, ensure_ascii=False)) --- scripts/vault_lint.py

Trust its output; don't re-derive these by reading pages. --- .claude/skills/lint/SKILL.md

t = m.group(1).split("|")[0].split("#")[0].strip()

--- scripts/vault_lint.py

stem = Path(target).stem resolved = by_name.get(stem, []) if not resolved: check = "index_drift" if p.name == "index.md" else "dead_links" add("error", check, p, f"dead link: [[{target}]]")

--- scripts/vault_lint.py

EPI-OBJ-editorial-findings

Label: Editorial findings

Part of: EPI-OBJ-findings

Source-native identity: the lint skill's editorial findings and suggested fixes. Representational form: natural-language allegations and proposals interpreted by the foreground model and user; substrate: foreground report/context, optionally wiki/meta/lint-report-YYYY-MM-DD.md. The model proposes unlinked mentions, missing pages, stale claims and aging seeds through search and judgment; sampling recent pages is permitted for large vaults. These are candidate diagnoses, not deterministic structural evidence or established truths. The user receives grouped findings and suggestions, approves bulk safe fixes or individually decides deletion, merging and contradictions; the model consumes that approval to edit. Force is advisory proposal followed by model-mediated user permission, not a software write gate. Scope/horizon: current maintenance decisions and, if approved edits occur, later page consumers; report retention alone does not establish acceptance or benefit. No performed editorial check or criterion-bound truth decision is supplied. Unlike EPI-OBJ-mechanical-findings, this part is requested model judgment with user-facing proposals. The closest supplied RT-RTE-vault-lint is the workflow, not a duplicate object. SRC-1 .claude/skills/lint/SKILL.md, doctrine/design.

Scale to vault size — sample recent pages if the vault is large: --- .claude/skills/lint/SKILL.md

• One-by-one approval: deleting orphans (isolation may be intentional), merging duplicates, resolving contradictions (needs the user's judgment on which side is true). --- .claude/skills/lint/SKILL.md

• Stale claims: pages whose claims newer pages contradict, without a contradiction callout. --- .claude/skills/lint/SKILL.md

Claims

EPI-CLM-structural-checks

Label: Structural correctness through text checks

Claimed operation/warrant: inspectable text permits deterministic verification of the retrieval substrate and diagnosis of what is broken. Distinct from RT-CLM-cache-refresh, which concerns hook refresh; no supplied claim counterpart. SRC-1 docs/why-no-vector-db.md, docs/what-happens-when-you-ingest.md, doctrine/design. The implementation of scripts/vault_lint.py supports a narrower structural reading: case-insensitive duplicate basenames, not semantic duplicates under arbitrary different names; link targets without heading validation; lenient flat frontmatter parsing. It does not check the truth of page claims or whether an index description is sufficient for recall. Comparisons with vector systems lack experimental support.

Because the entire retrieval substrate is text, a script can verify it: vault_lint.py deterministically checks dead wikilinks, orphan pages, duplicate filenames, alias collisions, frontmatter gaps, index drift and cache size, and tells you exactly what is broken. No tokens, no model judgment. --- docs/why-no-vector-db.md

| duplicates | The same page created twice under slightly different names | --- docs/what-happens-when-you-ingest.md

lower_names[p.stem.lower()].append(p)

--- scripts/vault_lint.py

EPI-CLM-graph-knowledge

Label: Deliberate links carry knowledge

Claimed warrant: ingest-time links encode meaningful intent/causal relationships and make one-hop reading better than similarity-radius expansion. No supplied claim counterpart; distinct from RT-CLM-cache-refresh and EPI-CLM-structural-checks. SRC-1 docs/why-no-vector-db.md, doctrine/design. Source-native objective is visible operation and minimal installation, with explicit scale/wording caveats. RT-RTE-source-ingest requests deliberate linking and conflict flags, while RT-RTE-vault-query requests bounded read/search/citation. Neither establishes link semantics are true or that retrieval comparisons hold. A sources relation can carry attribution; it cannot itself show the cited evidence supports a causal proposition.

So the goal was not better retrieval. It was a system I could always see inside of, that asks you to install as little as possible. --- docs/why-no-vector-db.md

A hand-placed link carries intent that a similarity score cannot reconstruct. "These two pages are 0.83 similar" and "this decision was made because of that constraint" are not the same claim. The first is a measurement; the second is knowledge. Following one hop of real links beats expanding a similarity radius, because the links encode why. --- docs/why-no-vector-db.md

Wording you did not anticipate. If the index line and your question share no vocabulary and no concepts, the read path can miss where semantic search would hit. That is what the grep fallback is for, and grep is a weaker net than embeddings. --- docs/why-no-vector-db.md