Eva-brain epistemic report

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

Source-and-claim boundary

The question is how Eva turns source documents into durable claims, checks or admits them, and makes them available for later reliance. Scope, exclusions and evidence identity are see boundary; the single frozen repository is SRC-1 in the Source register. The desktop, its Rust caller, shipped prompts, MCP and structural evaluator are included. External model/runtime internals, real user sources and installed provider state are not inspected. This is an implementation/doctrine account, not an observed execution or effectiveness assessment.

Knowledge-production intent is EPI-CLM-durable-knowledge; the narrower review claim is RT-CLM-structural-gate. Neither is treated as a blanket promise of truth. Semantic routes include ingestion, query, all seven profile tools and Health, not just the ordinary ingest path.

Entry point or operation	SRC-1 source path	Coverage and supporting records
Tauri command dispatch	apps/desktop/src-tauri/src/main.rs	Direct complete registry read; ingest, query/save/decision, health, profile, bootstrap/settings, import and setup agree with supplied runtime; no extra model-dependent command found in this registry.
Source acquisition, worktree proposal, gate and decision	apps/desktop/src-tauri/src/ingest.rs	Supplied RT-RTE-source-ingest; direct gate and decision reads; content and checks separated below.
Query and all runtime dispatch alternatives	apps/desktop/src-tauri/src/ingest.rs	Supplied RT-RTE-vault-query; direct dispatcher/command reads confirm three adapters; adapter details are supplied findings, not independently reread in full.
Health and profile tools	apps/desktop/src-tauri/src/ingest.rs	Supplied RT-RTE-advisory-tools; direct profile guidance, seven-tool instruction map, shared prompt, validator and dispatch reads. Model criticism remains requested, not observed.
Save formatting, pending review, accept/reject	apps/desktop/src-tauri/src/ingest.rs	Direct formatting/preparation/query decision reads; EPI-RTE-analysis-format and RT-RTE-analysis-save.
Settings, bootstrap, create/import, setup	apps/desktop/src-tauri/src/ingest.rs	Supplied RT-RTE-brain-settings, RT-RTE-runtime-setup; no independent semantic-check inference from setup readiness.
Later desktop reload/display/tool save	apps/desktop/src/main.ts	Supplied runtime findings only: RT-RTE-analysis-save, RT-RTE-source-ingest. Not an independently inspected UI audit.
MCP search, read_page, neighbors, shortest_path	packages/eva-mcp/server.mjs	Direct complete read; EPI-RTE-wiki-navigation; all four registrations covered, none writes.
Wiki loading, parser, link resolution, graph	packages/eva-mcp/vault-io.mjs, packages/wiki-lib/src/parse.ts, packages/wiki-lib/src/vault.ts, packages/wiki-lib/src/graph.ts	Direct complete reads; selection/graph facts are about stored text and resolved links, not truth of page contents.
Structural lint	packages/wiki-lib/src/lint.ts	Direct complete read; EPI-RTE-structural-check.
Shipped operating/format contract	schema/EVA.md, docs/EVA_BRAIN_STANDARD.md	Direct complete reads; epistemic discipline and portable format are doctrine, not evidence of compliance.
Product knowledge claim	README.md	Direct opening/product description read; EPI-CLM-durable-knowledge. Other promotional claims not assessed here.

No independent trace identifies actual generated claims, the evidence a reviewer consumed, or changes in later answers. This prevents attributing exercised acceptance, learning or quality gain. Source truth and extraction fidelity are unknown without real source/claim pairs. External model opacity prevents assessment of internal derivation or parametric change, not of Eva's inspectable admission wiring. Remaining supplied implementation coverage is identified as such; this is not an exhaustive negative about all possible filesystem callers or external hooks.

Epistemic-object inventory

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
EPI-OBJ-raw	Imported source assertions; actual propositions depend on user documents	Original evidence	SRC-1 schema/EVA.md; see record	Acquisition is not endorsement; original source warrant unknown.
EPI-OBJ-wiki	Entity/concept claims, summaries and saved syntheses	Durable linked knowledge	SRC-1 schema/EVA.md; see record	Editable prose and links do not prove source entailment.
RT-OBJ-ingest-worktree	see record	Ingest staging	see record	Proposal status differs from merged content.
EPI-OBJ-query-answer	Answer and page/source citations	Answer from wiki	SRC-1 apps/desktop/src-tauri/src/ingest.rs; see record	Nonempty/parseable does not mean supported.
EPI-OBJ-health-report	Alleged contradictions, provenance/staleness problems and questions	Advisory diagnosis	SRC-1 apps/desktop/src-tauri/src/ingest.rs; see record	Questions/next steps need not be truth-apt; diagnoses can be.
EPI-OBJ-profile-result	Evidence maps, decision frames, reflections, threads, study answers	Read-only derived products	SRC-1 apps/desktop/src-tauri/src/ingest.rs; see record	Mixed claims and policy; no product instance examined.
EPI-OBJ-lint-issues	Structural predicates about parsed pages	New-issue comparator input	SRC-1 packages/wiki-lib/src/lint.ts; see record	Does not judge substantive content.
RT-BAP-vault-contract	Operating guidance, not a source-truth oracle	Contract/profile to task behavior	see record	Settings are direct policy inputs; compliance unobserved.
RT-CMP-codex, RT-CMP-claude, RT-CMP-eva-qwen	Internal truth-apt content not determinable	External distributed-parametric reasoning	see records	Weight storage, version fixity and mutation limits remain local to each supplied component.

Authority-route ledger

The entries are function-specific overlays. Repeated citations name one implementation container, not multiple independent observations. Record-owned progression, activation, selection and limits are see the cited record; the fields below assess the epistemic edge and its force.

Route ID: RT-RTE-source-ingest Route function: content transformation Architectural status: implemented Object/candidate ID: EPI-OBJ-raw Content/update relation: truth-apt transformation: acquisition/import Transition or check target: User files to retained raw documents Evaluator/condition and domain: Person chooses sources; filesystem copy, not a truth evaluator Activation and timing: see RT-RTE-source-ingest Possible or observed result: Raw source available before wiki proposal Implemented force: Permissive evidence supply Epistemic authority and scope: Source warrant unknown; preservation of original document is not validation of its assertions Operational authority: Allows source to be read for ingest Behavioral-authority path: External model; raw filesystem; evidence supply; ingest and later source inspection Evidence source ID and local anchor: see RT-RTE-source-ingest Claim IDs: EPI-CLM-durable-knowledge Mismatch marker: none; limits stated locally Gap/limit: No observed execution; implemented request does not prove model compliance.

Route ID: RT-RTE-source-ingest Route function: content transformation Architectural status: implemented Object/candidate ID: EPI-OBJ-wiki Content/update relation: truth-apt transformation: indeterminate Transition or check target: Raw plus existing wiki to proposed page edits Evaluator/condition and domain: External model under source/provenance/profile guidance; semantic fidelity domain is requested, not independently checked Activation and timing: see RT-RTE-source-ingest Possible or observed result: Model may summarize, acquire assertions, synthesize or revise Implemented force: Candidate generation; not yet disposition Epistemic authority and scope: No established warrant beyond whatever source support is actually preserved Operational authority: Proposes durable content and links, cannot pass protected edits through this gate Behavioral-authority path: Worktree/gate; Markdown diff; candidate; pending admission Evidence source ID and local anchor: see RT-RTE-source-ingest Claim IDs: EPI-CLM-durable-knowledge Mismatch marker: none; limits stated locally Gap/limit: No actual input/output semantic pair: import, reshaping, entailment and ampliative conjecture cannot be distinguished at the produced-claim edge. Non-entailment is not inferred from opacity.

Route ID: EPI-RTE-structural-check Route function: check/evidence production Architectural status: implemented Object/candidate ID: EPI-OBJ-lint-issues Content/update relation: truth-apt transformation: entailed derivation Transition or check target: Parsed vault links, incoming edges, metadata and suspicious roots Evaluator/condition and domain: Deterministic lint rules; domain is this parser's structural representation Activation and timing: see EPI-RTE-structural-check Possible or observed result: Issue list; new-string difference computed separately by gate Implemented force: Evidence production for admission Epistemic authority and scope: Only the named structural predicates; no source truth or summary entailment license Operational authority: Supplies issues consumed by gate or displayed reviewer Behavioral-authority path: Rust gate/desktop; issue strings; enforcing hold predicate at gate; one proposal Evidence source ID and local anchor: see EPI-RTE-structural-check Claim IDs: RT-CLM-structural-gate Mismatch marker: none; limits stated locally Gap/limit: No observed execution; implemented request does not prove model compliance.

Route ID: RT-RTE-source-ingest Route function: disposition/acceptance Architectural status: implemented Object/candidate ID: RT-OBJ-ingest-worktree Content/update relation: no content change Transition or check target: Ingest patch admissibility Evaluator/condition and domain: Protected-path verifier; dirty-status condition; no deletions/no new issue strings, else UI accept/reject Activation and timing: see RT-RTE-source-ingest Possible or observed result: Reject, auto-merge, or hold then merge/discard Implemented force: Enforcing at this ingest gate; human overrides structural hold Epistemic authority and scope: Operational admissibility, not evidence-consuming semantic acceptance of each claim Operational authority: Allows merge or blocks pending decision; review resumes queue Behavioral-authority path: Rust merge/worker; Git branch and boolean review; enforcing; current proposal Evidence source ID and local anchor: see RT-RTE-source-ingest Claim IDs: RT-CLM-structural-gate Mismatch marker: none; limits stated locally Gap/limit: Countercases: inherited lint debt does not hold; human may accept new issues; a false but structurally clean claim can pass. These are limits, not demonstrated bad runs.

Route ID: RT-RTE-source-ingest Route function: retention Architectural status: implemented Object/candidate ID: EPI-OBJ-wiki Content/update relation: no content change Transition or check target: Admitted edits to main vault and history Evaluator/condition and domain: Git merge after gate/decision Activation and timing: see RT-RTE-source-ingest Possible or observed result: Persistent Markdown and operational log Implemented force: Durable availability Epistemic authority and scope: Retention does not establish semantic acceptance Operational authority: Later tasks can read merged claims; rejection removes proposal, not existing claims Behavioral-authority path: Desktop and later model tasks; files/search; permissive evidence; subsequent tasks Evidence source ID and local anchor: see RT-RTE-source-ingest Claim IDs: EPI-CLM-durable-knowledge Mismatch marker: none; limits stated locally Gap/limit: No observed execution; implemented request does not prove model compliance.

Route ID: EPI-RTE-wiki-navigation Route function: operational admission/selection/consumption Architectural status: implemented Object/candidate ID: EPI-OBJ-wiki Content/update relation: no content change Transition or check target: Which stored material is returned Evaluator/condition and domain: Lexical top-ten, exact/stem/title read, link graph neighbors and undirected BFS Activation and timing: see EPI-RTE-wiki-navigation Possible or observed result: Excerpts, full body/metadata, neighboring IDs or shortest link path Implemented force: Ranking for search; permissive delivery otherwise Epistemic authority and scope: No endorsement of selected statements; graph connection is not evidential entailment Operational authority: Places selected stored material in agent context Behavioral-authority path: External runtime/model; MCP text result; ranking/permissive; task and later calls Evidence source ID and local anchor: see EPI-RTE-wiki-navigation Claim IDs: none Mismatch marker: none; limits stated locally Gap/limit: Search excerpts can omit qualifications; full read is available. Selection is not an acceptance decision.

Route ID: RT-RTE-vault-query Route function: content transformation Architectural status: implemented Object/candidate ID: EPI-OBJ-query-answer Content/update relation: truth-apt transformation: indeterminate Transition or check target: Wiki plus question to answer Evaluator/condition and domain: External model asked to cite supporting pages; Rust parse/nonempty checks operate on output format Activation and timing: see RT-RTE-vault-query Possible or observed result: Displayed cited answer or error Implemented force: Advisory answer; malformed/empty output rejected Epistemic authority and scope: No independent verification that citations entail answer or match requested working set Operational authority: User receives answer; saving requires another command/decision Behavioral-authority path: UI/person; answer JSON/Markdown; advisory; current query Evidence source ID and local anchor: see RT-RTE-vault-query Claim IDs: EPI-CLM-durable-knowledge Mismatch marker: none; limits stated locally Gap/limit: No semantic pair or execution trace; cannot establish derivation, actual conjecture or scope compliance. Empty citations are allowed rather than proof of an evidence-backed answer.

Route ID: RT-RTE-advisory-tools Route function: check/evidence production Architectural status: implemented Object/candidate ID: EPI-OBJ-health-report Content/update relation: truth-apt transformation: indeterminate Transition or check target: Wiki contradictions, provenance, staleness, gaps Evaluator/condition and domain: External model prompted for page IDs and explained local evidence, not general-world staleness judgments Activation and timing: see RT-RTE-advisory-tools Possible or observed result: Advisory findings/next steps, possibly empty Implemented force: Advisory only; no automatic repair or withdrawal consumer Epistemic authority and scope: A requested evidence-backed diagnosis is a candidate, not established criticism Operational authority: Can inform a person's next action; does not invalidate pages Behavioral-authority path: Person; Health UI; advisory; current report Evidence source ID and local anchor: see RT-RTE-advisory-tools Claim IDs: none Mismatch marker: none; limits stated locally Gap/limit: No finding instance/trace; cannot credit actual content-directed criticism or a closed revision loop.

Route ID: RT-RTE-advisory-tools Route function: content transformation Architectural status: implemented Object/candidate ID: EPI-OBJ-profile-result Content/update relation: truth-apt transformation: indeterminate Transition or check target: Wiki to seven profile-specific products Evaluator/condition and domain: External model uses local record; format/nonempty validation not factual oracle Activation and timing: see RT-RTE-advisory-tools Possible or observed result: Evidence map, reflection, threads map, brief, options review, flashcards or exam/key Implemented force: Advisory/display; optional separate save Epistemic authority and scope: No independent reference-answer acceptance; generated answer key is not an oracle for its own production Operational authority: Supplies user-facing product; may be converted to saved answer Behavioral-authority path: Person and optional save path; JSON/Markdown; advisory; current request or later saved page Evidence source ID and local anchor: see RT-RTE-advisory-tools Claim IDs: none Mismatch marker: none; limits stated locally Gap/limit: Domain differs per tool; shared lack of semantic validator does not establish all outputs false or conjectural.

Route ID: EPI-RTE-analysis-format Route function: content transformation Architectural status: implemented Object/candidate ID: EPI-OBJ-query-answer Content/update relation: truth-apt transformation: non-ampliative reshaping Transition or check target: Supplied answer/citations to analysis Markdown Evaluator/condition and domain: Rust string formatting and reference filtering Activation and timing: see EPI-RTE-analysis-format Possible or observed result: Analysis prose plus formatted evidence and optional sources metadata Implemented force: Candidate retention formatting Epistemic authority and scope: Does not add warrant; reference filtering can degrade explicit lineage Operational authority: Creates reviewable page/index/log, not immediate main-vault admission Behavioral-authority path: Review preparer; Markdown diff; candidate; pending decision Evidence source ID and local anchor: see EPI-RTE-analysis-format Claim IDs: none Mismatch marker: none; limits stated locally Gap/limit: Citation existence/entailment unchecked; unsafe references can disappear while answer prose remains. See concrete deduction in record.

Route ID: RT-RTE-analysis-save Route function: disposition/acceptance Architectural status: implemented Object/candidate ID: EPI-OBJ-query-answer Content/update relation: no content change Transition or check target: Whether saved answer enters main vault Evaluator/condition and domain: UI person requests save, then explicit accept/reject; command callers not authenticated humans Activation and timing: see RT-RTE-analysis-save Possible or observed result: Merge/discard even when lint issues are displayed Implemented force: Enforcing decision requirement Epistemic authority and scope: Acceptance for retention; semantic criterion/evidence-consuming decision not recorded by command Operational authority: Allows or blocks saved content entering later searches Behavioral-authority path: query_decide/Git; review boolean; enforcing; this saved proposal Evidence source ID and local anchor: see RT-RTE-analysis-save Claim IDs: none Mismatch marker: none; limits stated locally Gap/limit: A person could evaluate truth, but no trace or mandatory semantic criterion establishes that here.

Route ID: RT-RTE-analysis-save Route function: retention Architectural status: implemented Object/candidate ID: EPI-OBJ-wiki Content/update relation: no content change Transition or check target: Approved saved analysis to later wiki consumers Evaluator/condition and domain: Merge after decision Activation and timing: see RT-RTE-analysis-save Possible or observed result: Persistent analysis/index/log Implemented force: Durable availability Epistemic authority and scope: Ordinary retained candidate claims, not automatically warranted claims Operational authority: Later queries/profile tools can reuse them Behavioral-authority path: Model/person; wiki files/MCP; permissive evidence; later tasks Evidence source ID and local anchor: see RT-RTE-analysis-save Claim IDs: none Mismatch marker: none; limits stated locally Gap/limit: No observed execution; implemented request does not prove model compliance.

Route ID: RT-RTE-brain-settings Route function: behavior/policy adaptation Architectural status: implemented Object/candidate ID: RT-BAP-vault-contract Content/update relation: non-truth-apt policy/content update: human-selected profile, purpose, language, runtime/model settings Transition or check target: Direct configuration edit to future task guidance Evaluator/condition and domain: Software validates settings and clean state; UI person chooses, not a model-evaluated improvement loop Activation and timing: see RT-RTE-brain-settings Possible or observed result: Changed contract/manifest selection Implemented force: Dispatch settings enforcing; natural-language guidance instructional Epistemic authority and scope: No truth endorsement; an extraction lens is not evidence Operational authority: Changes task framing/endpoint; protected from ingest edits Behavioral-authority path: Rust and external model; settings and EVA.md; enforcing dispatch/instructional content; later tasks Evidence source ID and local anchor: see RT-RTE-brain-settings Claim IDs: none Mismatch marker: none; limits stated locally Gap/limit: See record for admission and recovery; neither learning nor model-driven self-improvement follows.

No evidence-consuming semantic acceptance or post-acceptance lifecycle integration is established by these inspected consumers. Updating links/index and merging can be organization or retention before such acceptance; they are not relabelled lifecycle integration. Health may advise rejection or revision, and ingest can replace claims, but no automatic semantic withdrawal consumer links Health findings to those edits. The supplied local-runtime setup route is capability admission by readiness, not a truth-apt knowledge evaluator; see RT-RTE-runtime-setup.

System-claim versus route comparison

Claim ID	Claimed operation or warrant	Source/anchor and layer	Doctrine/design support	Implemented route IDs	Observed-run support	Causal support/design limits	Supported conclusion	Mismatch/unknown
EPI-CLM-durable-knowledge	see record	see record	see record	RT-RTE-source-ingest, RT-RTE-vault-query, RT-RTE-analysis-save, EPI-RTE-wiki-navigation	none supplied	No source/claim accuracy comparison or later-answer intervention	Durable linked intermediary is wired; epistemic discipline is prompted	Knowledge label exceeds what structural validity alone licenses; actual quality unknown.
RT-CLM-structural-gate	see record	see record	see record	EPI-RTE-structural-check, RT-RTE-source-ingest	none supplied	No runtime evidence; gate predicate is inspectable	New structural issues/deletions force review on ingest; not semantic verification	Human override and inherited issue strings limit strength; not a universal write gate.

Bounded conclusion

Eva acquires raw evidence, asks external models to maintain source-linked Markdown, and retains admitted results for later querying. Source warrant is unknown; preserving originals and requesting provenance does not transfer truth to every summary or synthesis (EPI-OBJ-raw, EPI-OBJ-wiki, RT-RTE-source-ingest). Concrete generated content is unavailable, so the semantic edges remain indeterminate rather than being uniformly labelled conjecture.

The deterministic evaluator derives structural facts within its parser/link domain. A syntactically suitable but false summary can satisfy its source-prefix rule. New issues/deletions trigger a review channel; neither a clean gate nor review acceptance records a semantic test criterion, intended epistemic use and evidence-consuming warrant decision (EPI-RTE-structural-check, RT-CLM-structural-gate). Rejection and cleanup prevent proposed content from entering the main vault; they do not retract already merged claims.

Retained claims gain operational availability through later search/read and answers, not an endorsement bit. Search ranking and short excerpts can affect what evidence reaches the model, while full reads remain available. Explicit graph routes derive connectivity, not explanatory or evidential relations (EPI-RTE-wiki-navigation). Query/profile products and Health diagnoses are model candidates. Saving a query/profile product requires a separate review; Health next steps remain advice. Human truth assessment is possible but unobserved (RT-RTE-vault-query, RT-RTE-advisory-tools, RT-RTE-analysis-save).

The research lens is an important counterexample to a claim that Eva requests only compression: it explicitly asks for competing-explanation revision and evidence maps. It does not establish that a formulated theory was actually criticized and its revision reused. For that requested research pathway, theory-builder condition 1 is afforded by localized Markdown thesis/competing explanations; condition 2 is afforded by instructions to use existing pages, not observed content-dependent reliance; condition 3 is claimed by profile guidance, not evidenced completed blame-assignment/revision; condition 4 is afforded by retaining updates for future ingest, not observed iteration. Addressability is page-level editable natural language; finer internal theory structure depends on actual content. Persistence across tasks is wired. Learning as improved future capacity is uninspected without performance/attribution evidence. Health's subject matter does not establish reflective self-representation of Eva's own operating machinery. End-to-end autonomous theory building is unestablished by opaque model reasoning and human-held admission paths. These are independent limits, not one negative verdict (SRC-1 apps/desktop/src-tauri/src/ingest.rs, profile passages in EPI-OBJ-profile-result).

Human configuration can change future behavior without a truth-apt route. The protected contract/settings path is not an inspected trace-fed self-revision loop; external endpoint changes are likewise not evidence of improved Eva performance (RT-RTE-brain-settings, RT-BAP-vault-contract, supplied model component records). Implementation supports these bounded capacities without establishing exercised semantic acceptance, integrated accepted ampliative knowledge, reliable criticism, or causal benefit.

Shared records

Operative objects

EPI-OBJ-raw

Label: Raw source collection

SRC-1, schema/EVA.md; imported original documents in filesystem raw/, committed by the source-ingest route. Natural-language assertions may occur in text/PDF/etc.; exact input encodings depend on supplied files, not a frozen actual user corpus. Distinct from RT-OBJ-ingest-worktree: original evidence is copied/committed before proposals, not the temporary proposal container. No other supplied object declares the original collection. Import preserves source material, not a truth certificate.

raw/ contains the original source documents. It is immutable: agents may read it but must never edit, move, or delete anything in it. --- schema/EVA.md

EPI-OBJ-wiki

Label: Wiki pages

SRC-1, schema/EVA.md, packages/wiki-lib/src/parse.ts: durable filesystem Markdown. Natural-language body claims reach model/human interpretation; localized metadata and wiki links reach parser/graph rules. This is the retained corpus, distinct from the temporary RT-OBJ-ingest-worktree that can contain proposed edits to it. Metadata is separately tested by EPI-OBJ-lint-issues; it is not authority over body truth.

The wiki is the generated Markdown outside raw/. Agents create and update this layer as new sources and questions add knowledge. --- schema/EVA.md

The profile is not a license to infer facts: preserve uncertainty, distinguish a source's claims from Eva's synthesis, and keep each durable claim close to its evidence. --- schema/EVA.md

EPI-OBJ-query-answer

Label: QueryAnswer

SRC-1, apps/desktop/src-tauri/src/ingest.rs: returned structured answer with natural-language Markdown and symbolic page/source strings; transient response/UI state until separately saved. Closest supplied RT-RTE-vault-query is its producer route, not an object counterpart; RT-RTE-analysis-save can also accept an arbitrary supplied answer. The output label does not attest source support.

1. Return a concise Markdown answer and cite every page that materially supports it. Cite exact brain-relative page ids. For each citation, include raw source paths named by that page when available. If there is no supporting evidence, return an empty citations array. --- apps/desktop/src-tauri/src/ingest.rs

EPI-OBJ-health-report

Label: HealthReport

SRC-1, apps/desktop/src-tauri/src/ingest.rs: transient JSON findings, summary, page IDs and next-step prose; natural-language diagnoses interpreted by a person, symbolic shape parsed by Rust. Distinct product of RT-RTE-advisory-tools, not another declaration of that route. It has no inspected automatic edit consumer.

1. Be conservative: report a contradiction, stale claim, or provenance weakness only when you can name supporting page ids and explain the evidence. Do not use general-world knowledge to call a claim stale. --- apps/desktop/src-tauri/src/ingest.rs

2. Look for useful findings in: contradiction, provenance, stale-claim, coverage-gap, research-question. A finding must cite exact existing page ids when it relies on them. nextStep is only a suggested action. Return no more than 12 findings. --- apps/desktop/src-tauri/src/ingest.rs

EPI-OBJ-profile-result

Label: ProfileToolResult

SRC-1, apps/desktop/src-tauri/src/ingest.rs: title/content/citations JSON; Markdown is natural-language product, citation/options fields symbolic. Transient until separate analysis-save. Distinct from HealthReport and QueryAnswer by task and consumer wrapper; shares the supplied advisory route container RT-RTE-advisory-tools, not a duplicate route declaration. Personal/threads/decision/options tools mix assertions and suggestions; study tools request answers but supply no independent correctness oracle. Research content may include theories; none is instantiated in the inspected output evidence.

Self::Research => "Separate claims from their evidence, attach provenance, and update competing explanations when the source challenges the current thesis.", --- apps/desktop/src-tauri/src/ingest.rs

Self::ResearchEvidenceMap => "Map the central claims, the evidence for and against each, the confidence or limitations, and the most valuable next question. Keep claims distinct from interpretation.", --- apps/desktop/src-tauri/src/ingest.rs

1. Use only evidence in this brain. If the record is too thin, say exactly what is missing instead of filling gaps from general knowledge. --- apps/desktop/src-tauri/src/ingest.rs

if result.title.trim().is_empty() { return Err("agent returned a tool result without a title".into()); } if result.content.trim().is_empty() { return Err("agent returned an empty tool result".into()); } --- apps/desktop/src-tauri/src/ingest.rs

EPI-OBJ-lint-issues

Label: LintIssue

SRC-1, packages/wiki-lib/src/lint.ts: symbolic in-memory page/rule/message objects, rendered to strings for the supplied RT-RTE-source-ingest comparator. Distinct from that route and from the wiki's substantive assertions; no supplied object counterpart. Predicates concern parsed structure, not agreement with source documents.

export interface LintIssue { /* Page id the issue belongs to. / page: string; rule: LintRule; message: string; } --- packages/wiki-lib/src/lint.ts

Routes

EPI-RTE-structural-check

Label: Structural lint evaluation Part of: RT-RTE-source-ingest

The check is a material function inside the supplied combined ingest route; it is also reused for save-review/desktop diagnostics. Owner: deterministic wiki-lib/lint CLI, not a model. Parsed vault → graph/rules → issue list. Target is frontmatter, link resolution, incoming edges and misplaced path roots. The gate separately compares new strings and decides; no semantic answer oracle. No human judgment is needed to produce this list.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: array of structural issues, rendered by the lint CLI for Rust.
• Later read-back: inapplicable — this check does not itself persist material accumulated through use; gate consumes current/baseline lists.
• Delegated visibility: inapplicable — deterministic in-process computation/CLI, no separate reasoning worker.
• Selection predicate: all loaded wiki pages; summary-specific source checks only on type summary.
• Invalidation or expiry: recomputed from loaded state; baseline/post-state difference belongs to gate.
• Activation or effect: issues reach the ingest gate's hold predicate and review payload, not a factual-correction executor.
• Evidence limits: no target execution; predicate truth is about parsed representation only, not source existence/entailment.

SRC-1, packages/wiki-lib/src/lint.ts:

} else if (!source.startsWith('raw/')) { --- packages/wiki-lib/src/lint.ts

if (!hasIncoming.has(page.id) && page.type !== 'index') { --- packages/wiki-lib/src/lint.ts

Concrete implementation deduction, not a test run: a parsed summary with title, type summary and source: raw/nonexistent.txt does not trigger missing-source or invalid-source, because only presence and prefix are checked. Its prose could contradict the original without affecting those predicates. Likewise two mutually linked non-index pages each have an incoming edge without requiring an index path. These counterexamples constrain evaluator license; other checks may still report unrelated issues.

EPI-RTE-wiki-navigation

Label: MCP wiki navigation

Distinct from supplied RT-RTE-vault-query: reusable read service used by agents, not the query orchestration. No supplied standalone MCP route counterpart. Owner: Node MCP server and wiki-lib. Files → filtered vault → lexical search/page read/graph → JSON tool result. Search ranks lexical matches; neighbors and shortest_path derive explicit link connectivity under fixed resolution and an undirected traversal, not semantic support. The model decides when to request each tool.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: ranked excerpts, full page, adjacency IDs or shortest undirected path/error.
• Later read-back: each call reloads retained wiki, including accepted earlier edits and accessible current worktree edits.
• Delegated visibility: tool caller receives selected JSON text; raw and infrastructure excluded by loader, though native filesystem tools are separate channels.
• Selection predicate: search scores title/alias/stem hit plus body term hits and returns ten; read accepts ID/stem/title; graph follows resolved explicit links.
• Invalidation or expiry: no persistent result cache in inspected service; next call rebuilds from files.
• Activation or effect: delivered text can supply model evidence; no proof it controls a later answer.
• Evidence limits: snippets may lack context; graph truth is bounded by parser/resolver, not relationship validity in the world.

SRC-1, packages/eva-mcp/server.mjs:

const excerpt = at >= 0 ? page.body.slice(Math.max(0, at - 60), at + 120).replace(/\s+/g, ' ') : ''; --- packages/eva-mcp/server.mjs

return asText({ query, matches: results.slice(0, 10) }); --- packages/eva-mcp/server.mjs

SRC-1, packages/eva-mcp/vault-io.mjs:

.filter((p) => !p.startsWith('raw/')) --- packages/eva-mcp/vault-io.mjs

Concrete selection deduction: for body Alpha is uncertain. and query alpha, body match position is zero and the returned excerpt is the original-cased whole short body, not lowercased output or a new factual conclusion. Longer bodies can be cut at the fixed window. Full read returns body/frontmatter without that excerpt truncation. For graph operations, results concern resolved link paths only; there is no evidence-strength edge type.

EPI-RTE-analysis-format

Label: Analysis Markdown formatting Part of: RT-RTE-analysis-save

Distinct material transformation inside the supplied save route, not another admission mechanism. Owner: Rust analysis_markdown. Supplied answer/citations → trimmed answer prose, safe evidence links, selected source metadata, heading structure. Candidate file/index/log is prepared before review; admission/rejection remains see RT-RTE-analysis-save. This transformation does not independently authenticate caller provenance or inspect cited source truth.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: formatted Markdown string, subsequently written to proposal worktree.
• Later read-back: after separate merge, page is loaded as wiki and can be used in future tasks.
• Delegated visibility: inapplicable — Rust formatting, not a model invocation; UI reviewer receives diff.
• Selection predicate: prose trimmed; citation references rejected if empty/control/bracket/pipe-containing; sources metadata selects raw paths, with summary-page fallback.
• Invalidation or expiry: rejected worktree removed by separate decision route; accepted file has no automatic semantic expiry.
• Activation or effect: preserves answer text while constructing evidence display and optional provenance metadata.
• Evidence limits: filtering may remove references but never establishes entailment; no observed saved instance.

SRC-1, apps/desktop/src-tauri/src/ingest.rs:

let value = value.trim(); if value.is_empty() || value.chars().any(char::is_control) || value.contains(['[', ']', '|'])

--- apps/desktop/src-tauri/src/ingest.rs

.filter_map(|source| safe_reference(source)) .filter(|source| source.starts_with("raw/")) --- apps/desktop/src-tauri/src/ingest.rs

answer.answer.trim(), --- apps/desktop/src-tauri/src/ingest.rs

Concrete implementation deduction: input answer Claim A. with citation page concepts/a and source raw/a.txt retains prose Claim A., emits a wiki evidence link and sources metadata. If the source string instead contains a pipe, it is excluded by safe_reference while the answer prose survives; formatting is non-ampliative, but explicit lineage can be degraded. Neither example is executed or a claim that a real answer had those values.

Claims

EPI-CLM-durable-knowledge

Label: Durable linked knowledge

SRC-1, README.md, schema/EVA.md: product/design claim of a durable Markdown intermediary and source-derived knowledge, not an explicit infallibility claim. Distinct from RT-CLM-structural-gate, which concerns a narrower review predicate. Prompted provenance/uncertainty gives discipline; structural checks do not validate that discipline semantically.

Eva helps an agent maintain a durable, linked Markdown brain between the sources and the user. --- README.md

The purpose of the vault is to turn immutable source material into a durable, interlinked body of knowledge — not to act as a chat transcript or a temporary retrieval index. --- schema/EVA.md

Evidenced absences

none declared in this member

Behavioral-authority paths

none declared in this member