Quicky Wiki epistemic report

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

Source-and-claim boundary

The question is how document-derived claims gain reliance, survive criticism and affect later answers. Scope and external exclusions: see boundary; evidence identities and layers: see its Source register. SRC-1 supplies implementation and doctrine/design, not observed operation or causal experiments. The inspected mechanisms establish automatic graph admission and evidence-like delivery, not calibrated belief or demonstrated knowledge improvement.

The runtime is provisional. Direct reads confirmed the compiler, question answering and consensus account and extended it with chunk-prefix filtering, lint/prune and the model-generated slide path. No supplied runtime fact requires supersession; the runtime's renderer coverage is not a complete epistemic account of slide generation. This member adds that material route. Source-tree searches for hooks/registration in src/ and package.json found no additional hook surface; CLI dispatch, library exports, MCP tool names/handlers and dashboard API branches were checked rather than inferred from that search miss.

Entry point or operation Source path (SRC-1) Coverage and records
CLI dispatch and library operations src/cli.ts, src/lib.ts Direct complete reads; RT-RTE-client-access, RT-RTE-runtime-setup. Library exports permit direct store/compiler use.
File, URL, directory, watcher and bookmark ingestion src/compiler/ingest.ts, src/cli/fetch-url.ts, src/cli/ingest.ts, src/cli/watch.ts, src/cli/serve.ts Compiler/URL conversion and bookmark branch directly read; driver variants supplied by RT-RTE-source-ingest.
Comparison, score calculation, resolution, summaries src/compiler/diff.ts, src/compiler/confidence.ts, src/compiler/resolve.ts Direct complete reads; RT-OBJ-knowledge-diff, RT-RTE-source-ingest, EPI-OBJ-page-summaries.
Ordinary query, consensus and save/reingest src/graph/query.ts, src/compiler/consensus.ts, src/cli/save-answer.ts Direct complete reads; RT-RTE-question-answer, RT-RTE-consensus-answer, RT-RTE-answer-save.
MCP read, ingest and metadata tools; dashboard query, edit and deletion src/mcp/server.ts, src/cli/serve.ts Tool/branch inventory searched; MCP ingest/metadata and dashboard mutation branches directly read; other dispatcher details supplied by RT-RTE-client-access. Page DELETE is shadowed, unlike claim DELETE.
Store confidence, challenge, relations, search and removal src/graph/store.ts Direct method reads; broader schema/write context supplied by RT-OBJ-knowledge-store.
Decay, dependency cascade, structural checks src/graph/temporal.ts, src/graph/cascade.ts, src/cli/lint.ts Direct complete reads; RT-RTE-health-review, EPI-RTE-lint-prune.
Health, criticism, resurfacing and discovery src/metabolism/health.ts, src/metabolism/redteam.ts, src/metabolism/resurface.ts, src/cli/metabolism.ts, src/discovery/discover.ts Direct complete reads; RT-RTE-health-review. Suggestions stop at returned/displayed proposals in these callers.
Compile/export and persisted views src/cli/compile.ts, src/render/markdown.ts, src/render/anki.ts, src/render/marp.ts, src/render/obsidian.ts Compile, Markdown, Anki and slides directly read completely; Obsidian write flow and source/index content directly read, remaining page template uninspected. EPI-RTE-slide-output. Graph/timeline and remaining Obsidian templates rely on runtime coverage, not an exhaustive semantic audit.
Setup/model/release machinery src/llm/adapter.ts, src/cli/context.ts, .github/workflows/publish.yml Supplied RT-CMP-llm-adapter, RT-RTE-runtime-setup, RT-RTE-package-release. No graph-driven parameter training or package installation established.

Missing evidence is local to the conclusions it prevents: no source corpus or model input/output pairs means extraction, summary, answer and critique propositions cannot be individuated or classified semantically beyond their implemented request; no run trace means activation, actual acceptance decisions and acted-on criticism are unobserved; no reference-answer/calibration study means scores and consensus accuracy are unevaluated; no intervention/comparison means benefit, causal component effects and successful improvement are unestablished. External model internals and weight identity remain uninspected. The remaining view templates prevent a system-complete negative about every export's semantics; they do not negate the positively traced routes.

Consequential claims: EPI-CLM-living-knowledge and RT-CLM-confidence-citations.

Epistemic-object inventory

Object ID Candidate truth-apt content or none Claimed role Evidence source ID and local anchor Gap/limit
EPI-OBJ-source-content Document assertions, scope determined by each document Evidence input SRC-1 src/compiler/ingest.ts No corpus supplied; truth and original warrant unknown.
EPI-OBJ-claim-statements Extracted and retained factual statements Atomic, verifiable claims SRC-1 src/compiler/ingest.ts Prompt does not prove fidelity, atomicity or truth.
EPI-OBJ-confidence-state None as policy state; interpreted as degree of support by consumers Confidence, provenance, dependencies, events SRC-1 src/compiler/confidence.ts Numerical derivation is not a calibrated probability or evidential independence test.
EPI-OBJ-page-summaries One-sentence model summaries, potentially asserting facts Page description and retrieval surface SRC-1 src/compiler/resolve.ts Preservation/entailment not checked.
RT-OBJ-knowledge-diff Challenge reasons and gap diagnoses; symbolic action arrays separate from their prose Comparison result SRC-1 src/compiler/diff.ts Evaluates semantic relation, not source truth.
EPI-OBJ-query-results Answer and caveats Returned answer, optional saved source SRC-1 src/graph/query.ts Returned IDs and confidence unvalidated.
EPI-OBJ-consensus-result Response statements and synthesis Consensus categories for CLI SRC-1 src/compiler/consensus.ts Agreement labels have single-response/failure alternatives.
EPI-OBJ-review-proposals Claim critiques Advisory red-team output SRC-1 src/metabolism/redteam.ts No original source evidence or wired acceptance.
EPI-OBJ-resurface-reviews Review questions and suggestions; questions need not assert a proposition Advisory stale/weak review SRC-1 src/metabolism/resurface.ts No verified diagnosis or revision decision.
EPI-OBJ-discovery-proposals Gap, frontier, connection and contradiction diagnoses Advisory discovery SRC-1 src/discovery/discover.ts Model-generated relationships and IDs unverified.
EPI-OBJ-slide-content Presentation assertions and summary Persisted slide deck SRC-1 src/render/marp.ts No post-generation semantic check.
RT-CMP-llm-adapter Not determinable within opaque parametric state External generation/evaluation machinery SRC-1 src/llm/adapter.ts See component for inference-only calls, unpinned weights and mutable endpoints; no claim about inaccessible internal learning.

Authority-route ledger

Fields owned by a declaring record are available through the citation; ledger fields below are this member's epistemic assessment. Each repeated route citation selects a different consequential function of its container, not a duplicate declaration. All implemented rows describe wiring, not observed execution.

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 content transformation implemented EPI-OBJ-source-content truth-apt transformation: acquisition/import File body/converted URL enters extraction File reader, frontmatter parser and URL conversion; byte/text domain On ingest; hash match skips extraction Source body delivered; URL origin metadata retained Permissive input delivery Imported warrant unknown; metadata is not authentication Allows extraction from supplied source Extractor, user-message body, advisory evidence, current call SRC-1 src/compiler/ingest.ts, src/cli/fetch-url.ts EPI-CLM-living-knowledge none HTML stripping can discard context; source is not an oracle.
RT-RTE-source-ingest content transformation implemented EPI-OBJ-claim-statements truth-apt transformation: indeterminate Source prose to extracted statements Chat model instructed to extract factual claims New/changed bytes; chunking above threshold Parsed candidates, or empty array on parsing failure Candidate generation No demonstrated non-ampliative preservation, entailment or acceptance Candidates proceed to comparison Comparator, extracted arrays, permissive, current ingest SRC-1 src/compiler/ingest.ts EPI-CLM-living-knowledge none Need concrete source/output pairs to choose reshaping, derivation or conjecture.
EPI-RTE-chunk-selection operational admission/selection/consumption implemented EPI-OBJ-claim-statements no content change Selection of extracted candidates First occurrence of normalized 80-character prefix Only chunked extraction; before diff Later prefix duplicates removed Enforcing inclusion filter No equivalence license Blocks some candidates from graph comparison Diff, filtered array, enforcing, current ingest SRC-1 src/compiler/ingest.ts none Prefix equality is not proposition equality Distinct negated suffix can be lost; see concrete deduction.
RT-RTE-source-ingest check/evidence production implemented RT-OBJ-knowledge-diff no content change New/existing semantic relationship Model compares statement lists; index handling checks addressability Existing-store comparison Reinforce/challenge/new/gap classifications Supplies resolver inputs Model judgment about relation only Proposes graph changes Resolver, diff arrays, permissive semantic decisions, current ingest SRC-1 src/compiler/diff.ts EPI-CLM-living-knowledge none No source bodies, independent oracle or coverage check; omitted matches can omit candidates.
RT-RTE-source-ingest disposition/acceptance implemented EPI-OBJ-confidence-state non-truth-apt policy/content update: assign confidence and action Admit new claims, reinforce or weaken old ones Code consumes valid indices/actions; quality scaling and fixed formulas Empty store admits all; parse failure in diff admits all as new Stored claim or score/event update Enforcing store mutations Operational admission, not evidence-consuming acceptance for a named factual use/scope Allows future use without per-claim approval; challenge does not remove claim Store writers then readers, rows/scores, enforcing update and advisory later evidence, cross-command SRC-1 src/compiler/diff.ts, src/compiler/resolve.ts, src/compiler/confidence.ts EPI-CLM-living-knowledge Scores do not certify truth Reinforcement can lower confidence; source count is not independence.
RT-RTE-source-ingest retention implemented EPI-OBJ-claim-statements, EPI-OBJ-confidence-state no content change Diff to durable graph, source links and events Store transactions per method During resolution Statements/provenance/scores available later Enforcing retention Lineage links, not endorsement Enables later query/comparison/review Readers, SQLite rows, permissive persistence, cross-command SRC-1 src/graph/store.ts, src/compiler/resolve.ts none none No whole-ingest transaction or immutable raw snapshot.
RT-RTE-source-ingest content transformation implemented EPI-OBJ-page-summaries truth-apt transformation: indeterminate Claims to page summary and page assignment Model; summary nonempty and length greater than five Touched or summary-less pages Stored summary; failures skipped Permissive write No factual-summary acceptance Summary affects page search and views; assignment determines grouping Search/readers, summary/title text, ranking or advisory, cross-command SRC-1 src/compiler/resolve.ts, src/graph/query.ts none none A length check licenses storage, not semantic fidelity; co-citation links are not entailment.
RT-RTE-question-answer operational admission/selection/consumption implemented EPI-OBJ-claim-statements no content change Graph to answer context Lexical prefix-OR search and page fallback Every ordinary query Selected statement/confidence lines Enforcing context selection; advisory evidence No renewed factual warrant Excludes unselected claims; supplies selected claims without a factual veto Answer model, text context, advisory, current call; retained graph crosses calls SRC-1 src/graph/query.ts, src/graph/store.ts RT-CLM-confidence-citations IDs and original evidence omitted No confidence minimum gates ordinary retrieval.
RT-RTE-question-answer content transformation implemented EPI-OBJ-query-results truth-apt transformation: indeterminate Selected claims/question to answer Chat model; permissive JSON parsing After retrieval Answer/confidence/IDs or raw-text fallback Advisory result No checked entailment, citation correctness or calibrated confidence Returned to user/client; optionally saved CLI/dashboard/MCP, result text, advisory, request horizon SRC-1 src/graph/query.ts RT-CLM-confidence-citations Requested citations not verified Concrete answer content not supplied.
RT-RTE-consensus-answer check/evidence production implemented EPI-OBJ-consensus-result truth-apt transformation: indeterminate Responders' claims and convergence Primary model comparing successful parsed answers; model-response domain Two or more successful responses Agreed/disputed/uncertain arrays and synthesis Advisory comparison At most requested agreement judgment, not independent source truth Changes displayed categorization, not graph CLI, consensus result, advisory, current request SRC-1 src/compiler/consensus.ts none Single-response and failed-comparison paths also label claims agreed Shared evidence/errors and independence untested; no reference-answer oracle.
RT-RTE-answer-save retention implemented EPI-OBJ-query-results truth-apt transformation: non-ampliative reshaping Answer to generated raw source User save flag and fixed serializer Standard query only Markdown answer with metadata and resolvable cited statements Permissive retention Explicit generated origin, not independent corroboration Permits later re-extraction/manual or watched ingestion Extractor, raw file, advisory source then automatic admission, cross-command SRC-1 src/cli/save-answer.ts none Compounding can recycle prior output Saving is not factual approval; cited count can include unresolved IDs.
RT-RTE-health-review check/evidence production implemented EPI-OBJ-review-proposals truth-apt transformation: indeterminate Weakness of high-confidence claims Model sees up to five strongest claims at least 0.7 and source counts, not original evidence Requested red-team maintenance Critiques and suggested adjustments Advisory only Candidate criticism; not validated criticism or accepted revision Caller displays; no automatic score adjustment Human, report text, advisory, current invocation SRC-1 src/metabolism/redteam.ts, src/cli/metabolism.ts EPI-CLM-living-knowledge Review is not challenge application No observed human action or retained critique loop.
RT-RTE-health-review content transformation implemented EPI-OBJ-resurface-reviews truth-apt transformation: indeterminate Review needs of stale/weak claims Model receives up to five deduplicated claims, weakest first, selected by age over 14 days or confidence at most 0.5 Requested resurfacing Questions/suggestions or fixed fallback Advisory only No factual revision license Displays prompts to verify or update externally Human, console questions, advisory, current invocation SRC-1 src/metabolism/resurface.ts, src/cli/metabolism.ts none none Questions can be non-truth-apt; truth-apt content of actual outputs not supplied.
RT-RTE-health-review check/evidence production implemented EPI-OBJ-discovery-proposals truth-apt transformation: indeterminate Coverage gaps, frontiers, bridges or conflicting statements Model receives page-grouped statements and mode instruction Requested discovery Descriptions, suggested queries, priority, mapped claim IDs Advisory only Candidate diagnosis, not accepted relation or factual finding Returns proposals for display; does not insert contradiction edges Human, discovery result, advisory, current invocation SRC-1 src/discovery/discover.ts EPI-CLM-living-knowledge Claim indices requested but not explicitly numbered in input No independent evidence or verified ID mapping.
RT-RTE-health-review operational admission/selection/consumption implemented EPI-OBJ-confidence-state non-truth-apt policy/content update: time decay and damped dependency score propagation Existing confidence and age/dependency relations Fixed exponential decay/clamp or damped traversal; numeric policy domain Decay request or challenge resolution New scores; dependent traversal only with edges Enforcing update No new evidence about statement truth Changes displayed strength and confidence-selected exports/reviews; claims remain queryable Store then models/views, confidence numbers, ranking/advisory, cross-command SRC-1 src/graph/temporal.ts, src/graph/cascade.ts, src/graph/store.ts EPI-CLM-living-knowledge Dependency extraction not wired to edge insertion Generic confidence update resets last-reinforced, including decay/cascade.
EPI-RTE-lint-prune check/evidence production implemented EPI-OBJ-confidence-state no content change Empty pages, missing source links, low scores, self-dependency and contradiction records Deterministic predicates over graph structure/values User lint request Error/warning report Advisory diagnostic Supports only enumerated structural predicates Does not veto queries or fix claims Human, console, advisory, request horizon SRC-1 src/cli/lint.ts none Self-loop test is not general cycle detection No factual truth check; counts depend on populated edges.
EPI-RTE-lint-prune disposition/acceptance implemented RT-OBJ-knowledge-store no content change Empty-page removal Caller prune request and zero-claim SQL condition CLI lint prune or dashboard prune Empty pages and their links removed Enforcing deletion No factual claim rejected Removes navigation objects, not low-confidence claims Store/readers, deleted rows, enforcing, subsequent reads SRC-1 src/cli/lint.ts, src/graph/store.ts none none Generated files not swept by this operation.
RT-RTE-client-access disposition/acceptance implemented EPI-OBJ-claim-statements truth-apt transformation: acquisition/import Direct statement replacement Caller text; dashboard string check; store update Reachable claim PATCH or library call Replacement statement Enforcing mutation Imported warrant unknown, no factual acceptance criterion New content available under existing claim identity/source links Store/readers, statement row, enforcing mutation then advisory context, cross-command SRC-1 src/cli/serve.ts, src/graph/store.ts none Old provenance can remain attached to changed content Caller identity does not establish human approval; old rendered content can persist.
RT-RTE-client-access disposition/acceptance implemented EPI-OBJ-confidence-state non-truth-apt policy/content update: caller score or metadata assignment Confidence assignment and page metadata merge Dashboard number check or caller metadata object; store methods Claim PATCH, MCP metadata tool or library call New confidence or metadata Enforcing mutation No measurement or factual license from assignment alone Changes confidence-guided use or metadata-filtered selection Store/readers, numeric/JSON rows, enforcing update then ranking/selection, cross-command SRC-1 src/cli/serve.ts, src/mcp/server.ts, src/graph/store.ts none none Metadata may itself assert facts, but no supplied object individuates them; their truth-apt status is not determinable.
RT-RTE-client-access disposition/acceptance implemented EPI-OBJ-claim-statements, EPI-OBJ-confidence-state no content change Claim withdrawal or direct library page removal Caller deletion request and store transaction Reachable claim DELETE; public library deletePage Removed claim and related rows/events Enforcing deletion Caller disposition, not falsification Deleted claims unavailable to live graph consumers Store/readers, deleted rows, enforcing, later reads SRC-1 src/cli/serve.ts, src/graph/store.ts none Dashboard page DELETE shadowed No retained tombstone/criticism or automatic withdrawal of exported artifacts.
EPI-RTE-slide-output content transformation implemented EPI-OBJ-slide-content truth-apt transformation: indeterminate Selected graph claims to slide deck Model prompted to group/summarize, not an independent checker Compile slides; default score threshold, topic override Model text written to slide file Advisory output with permissive persistence No checked semantic preservation or acceptance Presentation artifact available to humans Renderer/file consumer, Markdown, advisory, persisted file horizon SRC-1 src/render/marp.ts, src/cli/compile.ts none none Topic selection bypasses default confidence floor; no factual gate before writing.

No lifecycle-integration row is asserted: page assignment, links and summaries organize content after operational admission, but no evidence-consuming acceptance decision for factual reliance was established in the inspected paths. The absence of that decision is not an invented evaluator failure. Source-hash recovery and configuration control remain as described in RT-RTE-source-ingest and RT-RTE-runtime-setup: hashes license skipping, not endorsement; configured entity instructions affect extraction policy, not evidence-responsive policy learning.

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-living-knowledge Living confidence-scored knowledge, contradiction surfacing, claim provenance/dependency and cascade SRC-1 README.md; doctrine/design See claim RT-RTE-source-ingest, RT-RTE-health-review, EPI-RTE-chunk-selection None supplied None; source quality labels and scores do not measure correctness Claims and scores are automatically maintained; semantic conflict detection is requested and score challenge is applied Dependency/contradiction edges are not inserted by ordinary compiler; challenge event does not itself populate contradictedBy. Red-team only advises. Automatic comprehensive surfacing and calibrated uncertainty not established.
RT-CLM-confidence-citations Confidence-aware cited answers SRC-1 src/mcp/server.ts; doctrine/design See claim RT-RTE-question-answer, RT-RTE-consensus-answer None supplied None; model IDs are unvalidated and source evidence omitted Structured answer/confidence feature is implemented Ordinary context lacks claim IDs; MCP reduces returned IDs to count; consensus agreement has non-consensus fallback paths.

Bounded conclusion

Quicky Wiki imports source text of unknown warrant, requests extraction and semantic comparison, and automatically admits their parseable results to a durable graph. Provenance links identify source records, but neither those links nor quality multipliers constitute factual acceptance. Extraction, summaries, answers, reviews and slides remain semantically indeterminate without actual content pairs: the implementation establishes generation requests, not entailed derivation or evidenced non-entailment. No accepted ampliative claim instance is established by this source-only analysis. See RT-RTE-source-ingest, EPI-RTE-slide-output.

Scores have real operational authority. They affect review candidate choice, default slides/Anki inclusion and confidence-bearing model context. They do not enforce factual admissibility in ordinary query retrieval. Decay and damped cascade are deterministic policy updates, not new truth evidence. Reinforcement is a weighted blend and can lower a score: for existing 0.9, new evidence 0.4 and one existing source, weight is 1/1.1 and result about 0.445, within the clamp. This implementation deduction refutes a universal reading that reinforcement necessarily strengthens a claim. No premises about the underlying proposition's truth enter that arithmetic (SRC-1 src/compiler/confidence.ts; EPI-OBJ-confidence-state).

Graph-to-answer delivery has advisory evidential force through model interpretation, not enforced obedience. Saving an answer and reingesting it closes a retained-output loop, but adds no independent evidence merely by making a new source file. Prefix filtering can also remove a distinct proposition before comparison. These are lineage/reliance qualifications, not observed failure rates. See RT-BAP-retained-context, RT-RTE-answer-save, EPI-RTE-chunk-selection.

Checking and disposition differ. Structural lint and health calculations evaluate graph predicates; model red-team/discovery evaluates possible weaknesses and connections from restricted context. Those reports do not automatically revise retained claims. Ingestion comparison does automatically change confidence and admission; direct edits and deletion are separate caller-controlled operations without an implemented factual criterion. Prune removes empty pages, not false statements. Removal from the live store does not establish withdrawal from already exported artifacts. See RT-RTE-health-review, EPI-RTE-lint-prune, RT-RTE-client-access.

Retained content and confidence can shape later behavior without model-weight updates. Actual learning as improved future capacity is unestablished; no improvement outcomes or attribution evidence were supplied. Content-directed comparison is implemented, but reliable criticism, accepted factual revision and retained criticism iteration are not established by the review route. Autonomous host sequencing exists within ingestion; it does not prove autonomous theory building. Health statistics represent selected graph properties, but a criticism-driven revision of the system's own method is not established. Configurable extraction instructions are a policy input, not evidence of exercised self-improvement. These are separate limits, not a single system-wide verdict; provider-internal processing remains opaque (RT-CMP-llm-adapter, RT-RTE-runtime-setup).

Shared records

Components

none declared in this member

Operative objects

EPI-OBJ-source-content

Label: Source document content

Distinct from RT-OBJ-knowledge-store: raw document bodies are read from filesystem paths rather than stored as graph row bodies. Source-native identity is raw/content in ingestion and Markdown returned by fetchUrlToMarkdown. Natural-language assertions occur within text and symbolic frontmatter, on the filesystem and transient strings. Source IDs/hashes identify registered inputs, not immutable evidence snapshots. SRC-1 src/compiler/ingest.ts, src/cli/fetch-url.ts; implementation.

const raw = await readFile(filePath, "utf-8"); const contentHash = createHash("sha256").update(raw).digest("hex"); --- src/compiler/ingest.ts

URL conversion strips navigation and other sections and records source URL/time. Unknown original warrant is further bounded by this lossy conversion; it cannot be treated as authenticated or complete source evidence.

const mainMatch = content.match( /<(?:main|article)[^>]>([\s\S]?)<\/(?:main|article)>/i, ); if (mainMatch) { content = mainMatch[1]; } --- src/cli/fetch-url.ts

EPI-OBJ-claim-statements

Label: Extracted and retained claim statements

Part of: RT-OBJ-knowledge-store

Source-native identity is retained Claim.statement, natural-language text in SQLite claims and transient extracted arrays. This part separates content from its numerical and relational controls. Producer is model extraction or direct caller edit; consumers include comparison, answering, summaries, review and exports. Extraction explicitly requests source-supported factual content but does not check source-to-statement entailment (SRC-1 src/compiler/ingest.ts, src/graph/store.ts; implementation/design).

const systemPrompt = You extract atomic, verifiable claims from source material. Each claim should be a single factual statement. Be precise and specific. Assign an initial confidence based on how well-supported the claim is by the source. ---src/compiler/ingest.ts`

Store persistence/answer delivery passages are supplied by RT-RTE-source-ingest and RT-RTE-question-answer. Nothing in those passages demonstrates a claim's truth.

EPI-OBJ-confidence-state

Label: Claim confidence and relations

Part of: RT-OBJ-knowledge-store

Source-native identities are claim confidence, source quality, source links, dependency/contradiction edges and epistemic event rows. Numerical/symbolic controls live in SQLite and intermediate diffs, with natural-language notes. They are separate from the claim statements whose reliance they affect. Score scaling/clamping and weighted blending are implemented, not calibration measurements (SRC-1 src/compiler/confidence.ts, src/graph/store.ts).

const multiplier = QUALITY_MULTIPLIERS[qualityTier] ?? 0.45; const adjusted = rawConfidence * multiplier; return Math.max(0.01, Math.min(1.0, adjusted)); --- src/compiler/confidence.ts

const weight = 1 / (1 + sourceCount * 0.1); const merged = existing * (1 - weight) + newEvidence * weight; return Math.max(0.01, Math.min(1.0, merged)); --- src/compiler/confidence.ts

The comparison classifies challenges separately from source truth. Its fixed reduction keeps the challenged statement live; the challenge store method records an event, not a new contradiction edge. The complete diff/resolver reads confirm runtime's missing dependency-insertion finding; public store edge methods remain a countercase to any claim that relations cannot exist (RT-RTE-source-ingest; SRC-1 src/compiler/diff.ts, src/compiler/resolve.ts, src/graph/store.ts).

confidenceAfter: Math.max(0.01, existing.confidence * 0.6), --- src/compiler/diff.ts

"UPDATE claims SET confidence = ?, last_reinforced = ? WHERE id = ?", --- src/graph/store.ts:552-558

The latter generic updater resets age even for decay/cascade changes. Confidence is therefore not a pure timestamp-independent evidence measure. Default export/review predicates give it operational consequences, while ordinary question search does not impose a confidence floor.

EPI-OBJ-page-summaries

Label: Page summary text

Part of: RT-OBJ-knowledge-store

Source-native identity is page summary, natural-language text retained in SQLite, used by search and views. It differs from claim statements in its producer/check: a summary model receives page claims and scores, then code checks only minimal nonempty length before writing. Semantic fidelity is indeterminate without inputs/outputs (SRC-1 src/compiler/resolve.ts, src/graph/query.ts; implementation).

const summary = resp.content.trim().replace(/^["']|["']$/g, ""); if (summary && summary.length > 5) { store.updatePageSummary(pageId, summary); } --- src/compiler/resolve.ts

EPI-OBJ-query-results

Label: Ordinary answer result

Part of: RT-RTE-question-answer

This object is the return value of the parent route, not another declaration of that process. Source-native fields are answer, claimIds and confidence: natural-language answer/caveats, symbolic IDs and numeric score in a transient object. CLI/dashboard/MCP consume it; optional saving serializes it for later extraction. Parsing/fallback is not factual acceptance. SRC-1 src/graph/query.ts; implementation.

return { answer: response.content, claimIds: [], confidence: 0.5, }; --- src/graph/query.ts

EPI-OBJ-consensus-result

Label: Consensus result

Part of: RT-RTE-consensus-answer

The object is the parent route's return value: ConsensusResult, a transient symbolic category structure carrying natural-language claims/synthesis and numeric confidence. Model comparison or shortcuts produce it; CLI consumes it, not graph ingestion. SRC-1 src/compiler/consensus.ts; implementation.

agreed: modelResponses[0].claims.map((c) => ({ statement: c.statement, confidence: c.confidence, })), --- src/compiler/consensus.ts

This fallback countercase means agreed is not sufficient evidence that multiple responders converged. The single-response shortcut independently has the same limitation.

EPI-OBJ-review-proposals

Label: Red-team critiques

Part of: RT-RTE-health-review

This object is a returned output part of the parent route, not its score mutations. redteamClaims returns a transient array of natural-language critique with symbolic claim ID and numeric suggested adjustment, based on strongest selected claims and source counts. The CLI/human consumes it; no caller acceptance or persistence is wired. SRC-1 src/metabolism/redteam.ts, src/cli/metabolism.ts; implementation. The supplied parent route supplies return-only evidence. Red-team proposals name claim indices but do not consume original sources; adjustments are only printed.

console.log( chalk.dim( Suggested adjustment: ${c.suggestedConfidenceAdjustment > 0 ? "+" : ""}${(c.suggestedConfidenceAdjustment * 100).toFixed(0)}%, ), ); --- src/cli/metabolism.ts

EPI-OBJ-resurface-reviews

Label: Resurfacing review prompts

Part of: RT-RTE-health-review

Source-native identity is resurface returned reviews: natural-language question/suggestion plus symbolic reference to the selected claim, in a transient array. This is the parent route's review output part. A model generates it from stale/weak statements, with fixed fallback questions on parsing failure; the CLI displays it to a human, without acceptance or persistence. Unlike red-team, the target is a review action rather than a strongest-claim critique. SRC-1 src/metabolism/resurface.ts; implementation/design.

question: Is this still accurate: "${c.statement}"?, suggestion: "Review and reinforce or remove this claim.", --- src/metabolism/resurface.ts

EPI-OBJ-discovery-proposals

Label: Discovery proposals

Part of: RT-RTE-health-review

Source-native identity is Discovery[]: natural-language description/title/suggested queries plus symbolic priority, mode and related claim IDs, in a transient array returned for display. This is the parent route's discovery output part. The model receives page-grouped claims and a selected mode instruction. It does not receive explicitly numbered claims although it is asked for claim indices; code maps those indices onto the independent claims array, without verifying semantic correspondence. No contradiction-edge insertion or acceptance follows. SRC-1 src/discovery/discover.ts; implementation/design.

relatedClaims: (d.relatedClaimIndices ?? []) .map((i: number) => claims[i]?.id) .filter(Boolean), --- src/discovery/discover.ts

EPI-OBJ-slide-content

Label: Slide deck content

Source-native identity is renderSlides response text and _slides.md. Natural-language claims/summary plus symbolic Marp formatting live in a transient string and then a filesystem artifact. Distinct from RT-OBJ-knowledge-store and the supplied answer route: a presentation caller generates it and writes a file, rather than returning an ordinary question answer or storing graph claims. SRC-1 src/render/marp.ts, src/cli/compile.ts; implementation. Evidence passages are in EPI-RTE-slide-output.

Routes

EPI-RTE-chunk-selection

Label: Chunk prefix selection

Part of: RT-RTE-source-ingest

The ingestion code owns a deterministic filter between sequential chunk extraction and diff consumption. This part is distinct from model semantic comparison: lowercased alphanumeric prefix equality determines candidate retention (SRC-1 src/compiler/ingest.ts; implementation).

  • Immediate return: First candidate for each normalized prefix, as an extracted array.
  • Later read-back: inapplicable — filter has no retained state; surviving claims can persist through the parent ingestion route.
  • Delegated visibility: Comparator receives survivors, not removed candidates or duplicate reasons.
  • Selection predicate: Chunked extraction when body length exceeds 12000; normalized statement prefix of 80 characters; first occurrence wins.
  • Invalidation or expiry: Seen set is fresh for each chunked extraction.
  • Activation or effect: Candidate omission is wired before diff; actual omissions unobserved.
  • Evidence limits: Static text deduction establishes non-equivalence risk, not frequency in a real corpus.
  • implementation conclusion status: wired
  • operation conclusion status: uninspected

Code decides; neither model nor human vetoes individual filter decisions. No answer oracle; mode is open ingestion. Guidance is a fixed deduplication heuristic; it persists as code, not revised learned policy. No rollback within the filter.

const key = c.statement .toLowerCase() .replace(/[^a-z0-9]/g, "") .slice(0, 80); if (seen.has(key)) return false; seen.add(key); return true; --- src/compiler/ingest.ts

Concrete implementation deduction, without target execution: let the shared prefix be the concatenation of four copies of The recorded system configuration, followed by is enabled. for statement A and is not enabled. for B. Together the repeated clauses normalize to more than 80 letters, so the first 80 normalized characters are identical and occur before either differing suffix. If A precedes B, the filter retains A and drops B. The opposite factual polarity is never compared by the diff. This is candidate selection, not a checked non-ampliative semantic deduplication.

EPI-RTE-lint-prune

Label: Structural lint and empty-page prune

Part of: RT-RTE-health-review

CLI code reads pages/claims/sources and emits structural warnings; optional pruning delegates to store deletion. This part separates graph predicates and deletion from model review (SRC-1 src/cli/lint.ts, src/graph/store.ts; implementation).

  • Immediate return: Console issues/counts; optional deleted-empty-page count.
  • Later read-back: Pruned page/link rows are absent from subsequent store reads; warnings themselves are not persisted by the caller.
  • Delegated visibility: Human sees errors/warnings; store receives prune request, not a factual diagnosis.
  • Selection predicate: No claims, no source links, confidence below 0.1, self-ID dependency, or stored contradiction relations; prune selects pages with zero claims.
  • Invalidation or expiry: Recomputed per call; independent store writes can change results; generated files are not deleted here.
  • Activation or effect: Reporting is advisory; prune deletes navigation rows on explicit request. Operation unobserved.
  • Evidence limits: Self-dependency check does not establish general graph acyclicity; missing edges cannot establish lack of semantic contradiction.
  • implementation conclusion status: wired
  • operation conclusion status: uninspected

User initiates lint/prune; code judges only the named structural predicates. No expected factual answer or improvement experiment; errors do not veto answering. Prune admits deletion by empty-page criterion, with no factual rejection and no backup/undo in this caller. Dashboard exposes the same prune method independently.

if (claim.dependsOn.includes(claim.id)) { --- src/cli/lint.ts

const pruned = store.deleteEmptyPages(); --- src/cli/lint.ts

"SELECT p.id FROM pages p LEFT JOIN claims c ON c.page_id = p.id GROUP BY p.id HAVING COUNT(c.id) = 0", --- src/graph/store.ts

EPI-RTE-slide-output

Label: Claim-to-slide output

No exact route counterpart after comparison with RT-RTE-source-ingest and RT-RTE-client-access: this is a shipped compiler output caller with separate model generation and durable file consumption. CLI owns progression: select claims, request deck, write response. The input is retained natural-language claims plus scores; the output is Marp text (SRC-1 src/render/marp.ts, src/cli/compile.ts; implementation/design).

  • Immediate return: renderSlides returns model text or a fixed no-claims deck; CLI writes _slides.md.
  • Later read-back: Persisted deck is available to external presentation/human consumers; automated graph reimport is not wired by this caller.
  • Delegated visibility: Model gets selected statements, percentages and topic, not source bodies or evidence histories.
  • Selection predicate: Default claims have confidence at least 0.5; an existing named topic replaces that selection with all claims on the page.
  • Invalidation or expiry: Subsequent compilation overwrites the same file; no automatic factual revalidation or deletion on graph change.
  • Activation or effect: Model output is written without factual gate; actual output and downstream use are unobserved.
  • Evidence limits: Prompt requests grouping and summary, but no deck instance or semantic comparison establishes faithful reshaping rather than additional propositions.
  • implementation conclusion status: wired
  • operation conclusion status: uninspected

User selects open compile request; code selects inputs and accepts the response for file writing, model proposes output prose. User may decline use externally. No answer oracle, named factual acceptance criterion, improvement trigger or rollback; fixed presentation instructions persist in code.

if (topic) { const page = store.getPageByTitle(topic); if (page) { claims = store.getClaimsByPage(page.id); pageTitle = page.title; } } --- src/render/marp.ts

return response.content; --- src/render/marp.ts

const slides = await renderSlides(store, llm, opts.topic); const outPath = join(config.paths.wiki, "_slides.md"); await writeFile(outPath, slides, "utf-8"); --- src/cli/compile.ts

Claims

EPI-CLM-living-knowledge

Label: Living confidence-scored knowledge

Distinct from RT-CLM-confidence-citations: this public design claim concerns source extraction, temporal changes, contradiction surfacing, provenance/dependencies and cascade rather than query result shape. SRC-1 README.md; doctrine/design only. It does not explicitly promise calibrated probability or infallible truth, so those are not attributed promises. The contrast is between its source-native operations and their actual implemented force.

Quicky Wiki extracts claims from your sources, tracks how confident each claim is, watches for contradictions, and gives you a visual dashboard to explore everything. Think of it as a personal Wikipedia that actually tells you what it's unsure about. --- README.md

  • Claim — An atomic, verifiable statement extracted from a source. Has a confidence score, provenance, and dependency chain. --- README.md

  • Cascade — When a foundational claim is challenged, confidence changes propagate through dependent claims. --- README.md

Evidenced absences

none declared in this member

Behavioral-authority paths

none declared in this member; retained context authority is supplied by RT-BAP-retained-context. Output-specific consumers and horizons are recorded in the ledger and slide route.