browzy.ai epistemic report

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

Source-and-claim boundary

The question is what licenses reliance on imported sources, compiled articles, generated answers and interaction-derived material. Scope and exclusions: see boundary; source identity and layers: see Source register. All additional inspection below is SRC-1 at that pin. Code supplies implementation evidence; prompts and README supply doctrine/design. No execution, provider weights, real user corpus or causal comparison was supplied. Consequently actual factual accuracy, successful correction, model compliance, benefit and exercised learning remain unestablished. Accessible service adapters do not reveal changes inside distributed-parametric state; see RT-CMP-llm-provider.

Consequential claims are RT-CLM-wiki-grounding, EPI-CLM-source-compounding, EPI-CLM-compiler-preservation and EPI-CLM-quality-auditor. No broader system-complete negative follows from this assessment.

Entry point or operation	Source path, SRC-1	Coverage and conclusion limit
Executable dispatch	src/cli/entry.tsx	Direct complete read confirms Ink default and registered init/ingest/compile/query/lint/status/search, consistent with runtime. No registered model tool dispatcher here.
Ingest and image description	src/core/ingest/index.ts, src/core/ingest/image.ts	Direct complete reads; EPI-RTE-image-description. Web/PDF/text processor bodies not directly assessed here; supplied source-compilation container covers their dispatch, not extraction fidelity.
Template/model compilation, parser, concepts, backlinks/index	src/core/compile/compiler.ts, src/core/sanitization.ts	Direct complete reads; EPI-RTE-template-article, EPI-RTE-article-admission. Concept suggestions are returned slugs, not newly accepted concept articles.
Answering/cache/history/schema	src/core/query/engine.ts, src/core/query/historyManager.ts, src/core/schema.ts, src/core/prompts.ts	Supplied findings RT-RTE-answer-question, RT-RTE-library-query, RT-RTE-schema-change; prompt claim locations additionally searched. No independent provider compliance observation.
Retrieval and coverage diagnostics	src/core/retrieval/contextBuilder.ts, src/core/discovery/gapResolver.ts	Direct complete reads; EPI-RTE-coverage-diagnostics. Ranking internals, SQLite search implementation and section scorer not independently audited; semantic relevance and search completeness not proven.
Health checks and consumers	src/core/lint/linter.ts, src/cli/commands/lint.ts, src/cli/app.tsx	Linter/CLI complete reads; app health/rebuild range directly read. EPI-RTE-wiki-checks, EPI-RTE-issue-display.
Conversation compaction	src/core/retrieval/compactor.ts, src/cli/app.tsx	Direct complete compactor and app trigger reads; EPI-RTE-conversation-compaction. Shipped caller discards result, despite comment promising future context use.
Digest/crystallizer/session	src/core/query/digest.ts, src/core/query/crystallizer.ts, src/cli/hooks/useSession.ts, src/cli/app.tsx	Supplied RT-RTE-session-digest, RT-RTE-insight-draft. Draft-to-live-wiki promotion remains unestablished, not inferred from filename.
Freshness/refresh/rebuild	src/core/discovery/freshness.ts, src/cli/app.tsx	Freshness complete direct read; supplied refresh and direct rebuild caller findings. HEAD metadata establishes a change signal, not a truth or endorsement check; compiler source-membership predicate can defeat recompilation.
Topic/clipboard/search/demo and static state	src/cli/app.tsx, src/core/demo/seed.ts, src/core/discovery/webSearch.ts	Supplied boundary coverage, app route search only for topic/gap consumers. Not independently traced end to end; no conclusion about source selection quality or demo provenance beyond boundary.
Test evaluators	tests/e2e/retrieval-quality.test.ts	Direct lines 1–210 only: fixture-based search assertions and token heuristics. Expected slug membership is an explicit test oracle for those searches, not factual-answer validation. Tests not run; remaining test bodies/fixtures not independently inspected.
Legacy/exported operations and dependency code	src/cli/index.ts, src/cli/repl.ts, src/index.ts, package-lock.json	Supplied classification and exports only. Full legacy/library operation enumeration and dependency internals not assessed; prevents universal no-evaluator/no-learning claim.

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	Imported statements about user-selected topics	see record	SRC-1 src/core/ingest/index.ts	Imported truth warrant unknown; extraction fidelity varies by uninspected processor.
EPI-OBJ-image-description	Model assertions about image contents	see record	SRC-1 src/core/ingest/image.ts	Not the original image's assertions or an independently checked transcription.
EPI-OBJ-wiki-article	Article bodies and summaries about source topics	see record	SRC-1 src/core/compile/compiler.ts	Short copy and model rewrite have different transformation guarantees.
EPI-OBJ-lint-issue	Link/metadata diagnoses and model consistency allegations	see record	SRC-1 src/core/lint/linter.ts	Structural facts and semantic allegations have different evaluators.
RT-OBJ-browzy-schema	Preferences, not necessarily truth-apt	see record	SRC-1 src/core/schema.ts	Instruction supply is not evidence about domain truth.
RT-RTE-answer-question	Answer statements; human and assistant history	see record	SRC-1 src/cli/app.tsx	Route-owned output, not a newly declared duplicate object. Source use list is not an audited citation map.
RT-RTE-session-digest	Claims about prior conversation and its conclusions	see record	SRC-1 src/core/query/digest.ts	Semantic preservation indeterminate; summary is not criticism or acceptance of the conversation's claims.
RT-RTE-insight-draft	Candidate connections extracted from answer	see record	SRC-1 src/core/query/crystallizer.ts	No concrete candidate instance supplied; entailment/non-entailment and actual novelty not determinable.
EPI-RTE-coverage-diagnostics	Lexical match/coverage indicators	see record	SRC-1 src/core/retrieval/contextBuilder.ts	Coverage label is not confidence in factual truth.

Authority-route ledger

Fields owned by cited records are referenced with see; classifications and architectural status below are this overlay's assessments. Each row is one function even where a container implements several. All results are possible implementation results, not observed runs.

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-compilation	content transformation	implemented	EPI-OBJ-source-content	truth-apt transformation: acquisition/import	External material to raw manifest	see route; source warrant unknown	see route	Imported source	see route	Acquisition, not independent confirmation	Raw input becomes compiler-eligible	Compiler/raw manifest/permissive/later compile	SRC-1 src/core/ingest/index.ts	none	none	Processor fidelity unassessed
EPI-RTE-image-description	content transformation	implemented	EPI-OBJ-image-description	truth-apt transformation: indeterminate	Image to model description	see record	see record	Description or pending placeholder	see record	Candidate image interpretation only	Description retained as raw material	see record	SRC-1 src/core/ingest/image.ts	none	none	No instance establishes semantic fidelity or non-entailment
EPI-RTE-template-article	content transformation	implemented	EPI-OBJ-wiki-article	truth-apt transformation: non-ampliative reshaping	Sanitized short source to article/body/summary	see record	see record	Body copied, prefix summary	see record	No new warrant for copied assertions	Eligible for article writes	see record	SRC-1 src/core/compile/compiler.ts	EPI-CLM-compiler-preservation	short-source exception	General Unicode normalization may change source spelling
RT-RTE-source-compilation	content transformation	implemented	EPI-OBJ-wiki-article	truth-apt transformation: indeterminate	Long raw content to model article	see route	see route	Candidate article blocks	see route	Synthesis candidate; no checked entailment	Parser receives generated text	Parser/model response/permissive/current compile	SRC-1 src/core/compile/compiler.ts	EPI-CLM-compiler-preservation	preservation not enforced	Titles only cannot support faithful merging of unseen old bodies
EPI-RTE-article-admission	disposition/acceptance	implemented	EPI-OBJ-wiki-article	no content change	Marker/metadata admission to wiki	see record	see record	Parsed candidate admitted or skipped	see record	No evidence-consuming truth acceptance	Writes permitted without per-article human review	see record	SRC-1 src/core/compile/compiler.ts	EPI-CLM-compiler-preservation	structural admission only	Empty body can pass; citations not validated
RT-RTE-source-compilation	retention	implemented	EPI-OBJ-wiki-article	no content change	Candidate to file and search record	see route	see route	Retained wiki/FTS/index	see route	Retention does not establish acceptance	Later retrieval permitted	Query/Markdown and FTS/permissive/future requests	SRC-1 src/core/compile/compiler.ts	none	none	Link/tag organization is not lifecycle integration after truth acceptance
EPI-RTE-coverage-diagnostics	check/evidence production	implemented	EPI-OBJ-wiki-article	no content change	Query coverage, not claim truth	see record	see record	Confidence/gaps	see record	Heuristic lexical scope only	UI and gap search receive diagnostics	see record	SRC-1 src/core/retrieval/contextBuilder.ts	EPI-CLM-source-compounding	coverage versus correctness	Ranking evidence not a ground-truth answer oracle
RT-RTE-answer-question, RT-RTE-library-query	operational admission/selection/consumption	implemented	EPI-OBJ-wiki-article	truth-apt transformation: non-ampliative reshaping	Wiki subset to model context	see routes	see routes	Selected article blocks, budget omissions	see routes	Source warrant remains unknown; omitted qualifications possible	Selected text presented as primary truth	see RT-BAP-wiki-context	SRC-1 src/core/retrieval/contextBuilder.ts	RT-CLM-wiki-grounding	trust policy not verification	Model obedience unobserved
RT-RTE-answer-question, RT-RTE-library-query	content transformation	implemented	Answer, see routes	truth-apt transformation: indeterminate	Context/question to answer	see routes	see routes	Generated answer	see routes	No post-answer truth license established	Display/cache/session consumption	see routes	SRC-1 src/core/query/engine.ts	RT-CLM-wiki-grounding	grounding requested	Neither entailment nor ampliation classifiable for an absent answer instance
EPI-RTE-wiki-checks	check/evidence production	implemented	EPI-OBJ-lint-issue	no content change	Link structure, fields; summary consistency	see record	see record	Issues or empty result	see record	Structural deductions scoped to loaded files; model allegations unverified	Issues passed to display	see record	SRC-1 src/core/lint/linter.ts	EPI-CLM-quality-auditor	summaries not full claims	Parse failure can look like no semantic issues
EPI-RTE-issue-display	operational admission/selection/consumption	implemented	EPI-OBJ-lint-issue	no content change	Issues to operator	see record	see record	Advisory messages/counts	see record	No acceptance/rejection of article truth	No automatic repair or retrieval veto in these consumers	see record	SRC-1 src/cli/commands/lint.ts	EPI-CLM-quality-auditor	advisory not enforcing	Later user response unobserved
RT-RTE-insight-draft	check/evidence production	implemented	Draft candidate, see route	no content change	Novelty relative to supplied articles	see route	see route	NONE or article proposal	see route	Model comparison only; neither truth nor corpus-wide novelty established	Enables draft parser	Model response/parser/permissive/current answer	SRC-1 src/core/query/crystallizer.ts	none	novelty versus truth	Comparison uses truncated source/answer text
RT-RTE-insight-draft	retention	implemented	Draft candidate, see route	truth-apt transformation: indeterminate	Answer connection to draft	see route	see route	Parsed draft file	see route	Candidate retention, not accepted ampliative knowledge	File write, not live wiki admission	see route	SRC-1 src/core/query/crystallizer.ts	none	none	No automatic later draft consumer established
RT-RTE-session-digest	retention	implemented	Digest, see route	truth-apt transformation: indeterminate	Conversation to summary/banner/wiki attempt	see route	see route	Digest text and optional article	see route	Summary not validation of session assertions	Banner read and attempted wiki insertion	see route	SRC-1 src/cli/app.tsx	none	digest differs from draft gates	Runtime module resolution and retrieval remain unobserved
RT-RTE-schema-change	behavior/policy adaptation	implemented	RT-OBJ-browzy-schema	non-truth-apt policy/content update: owner preferences	Owner file to system instructions	see route	see route	New prompt supply	see route	No truth acceptance implied	Model asked to follow preferences	see RT-BAP-schema-instruction	SRC-1 src/core/schema.ts	none	none	Human-directed customization, not evidence of self-improvement

Route ID: EPI-RTE-conversation-compaction Route function: content transformation Architectural status: implemented Object/candidate ID: Returned conversation summary, see record. Content/update relation: truth-apt transformation: indeterminate Transition or check target: Older messages to summary plus recent tail. Evaluator/condition and domain: see record; no fidelity evaluator. Activation and timing: see record. Possible or observed result: Summary returned; shipped terminal caller ignores it. Implemented force: No downstream force in that caller. Epistemic authority and scope: No checked preservation or truth license. Operational authority: behavior permitted, blocked, or changed: Provider summarization request, no installed future context change. Behavioral-authority path: consumer, channel, force, horizon: No consequential summary consumer in inspected caller; public function can return to an unspecified host whose use is uninspected. Evidence source ID and local anchor: SRC-1 src/core/retrieval/compactor.ts, src/cli/app.tsx. Claim IDs or none: none. Mismatch marker or none: Comment promises future context use but result is discarded. Gap/limit: Extends runtime coverage; no supersession needed for its valid bounded history-trimming account.

Freshness is a separate lineage/change signal, not a truth evaluator: direct SRC-1 src/core/discovery/freshness.ts inspects up to ten oldest eligible web sources using HEAD, compares Last-Modified or stored ETag, and silently skips network errors. Supplied RT-RTE-source-compilation establishes why refresh/rebuild does not guarantee withdrawal or replacement of stale articles. No record here establishes evidence-consuming acceptance against a factual criterion, or lifecycle integration after such acceptance. This is a bounded finding about the inspected writes/checks, not an assertion about inaccessible model cognition or every uninspected operation.

System-claim versus route comparison

Claim ID	Claimed operation or warrant	Claim source ID/anchor and evidence layer	Doctrine/design support	Implemented route IDs	Observed-run support	Causal support and design limits	Supported conclusion	Mismatch/unknown
RT-CLM-wiki-grounding	see record	SRC-1 src/core/prompts.ts, doctrine/design	see record	RT-RTE-answer-question, RT-RTE-library-query	None supplied	No component intervention or answer audit	Context and provenance instructions delivered	Delivery is not enforced grounding; source warrant unknown
EPI-CLM-source-compounding	see record	SRC-1 README.md, doctrine/design	see record	RT-RTE-source-compilation, EPI-RTE-coverage-diagnostics	None supplied	No before/after comparison; finite selection and stale content qualify universal benefit	Sources can expand available content; questions can yield gap suggestions	Every-source/every-answer improvement unsupported
EPI-CLM-compiler-preservation	see record	SRC-1 src/core/prompts.ts, src/core/compile/compiler.ts, doctrine/design	see record	EPI-RTE-template-article, EPI-RTE-article-admission, RT-RTE-source-compilation	None supplied	No merge-preservation or contradiction test output	Model receives preservation request; short route copies sanitized text	Existing bodies not supplied; collision rename and source metadata replacement do not enforce preservation
EPI-CLM-quality-auditor	see record	SRC-1 src/core/prompts.ts, doctrine/design	see record	EPI-RTE-wiki-checks, EPI-RTE-issue-display	None supplied	No sensitivity/specificity measurement; summaries omit body details	Structural checks plus advisory model critique	No automatic correction, truth gate or reliable clean bill of health

Bounded conclusion

• Imported raw text and model-described images can become retained wiki content through software admission. Original source IDs remain useful lineage labels, but neither selection by a user nor membership in the wiki verifies source truth. The image route introduces a separate unchecked interpretation before compilation; see EPI-RTE-image-description.
• The short-source route copies its sanitized body and computes a prefix summary; this is a demonstrated implementation deduction, not novel knowledge production. Long-source compilation, answering, digesting and draft extraction cannot be classified as entailed derivation or ampliative conjecture without concrete content and checked interpretation. Fluent synthesis and a novelty instruction do not establish either class; see EPI-RTE-template-article and RT-RTE-insight-draft.
• Article markers/title/slug and draft shape/length gates are operational admission criteria. Their intended use is parseable persisted content, not acceptance of factual claims. Wiki admission has future answer-context authority; draft retention has no established automatic answer consumer. Digest insertion is a separate attempted route and defeats a blanket claim that all interaction-derived writes require human approval; see EPI-RTE-article-admission, RT-RTE-session-digest.
• Structural lint results license narrow statements about loaded links/fields. Model consistency checks see summaries rather than full article propositions, and their outputs are advisory. Ranking confidence licenses no claim-truth probability; gap labels are lexical search signals. Neither result enforces revision, rejection, withdrawal or factual acceptance in the traced consumers; see EPI-RTE-wiki-checks, EPI-RTE-issue-display, EPI-RTE-coverage-diagnostics.
• Future model calls receive selected wiki text under a primary-truth policy, and owner schema preferences can change instructions. Those are live authority pathways at the implementation level, not observed compliance. Session/draft writes, schema editing and content accumulation do not alone demonstrate improved capacity, reflective criticism, autonomous theory building or exercised self-improvement. No theory instance or content-directed criticism/revision loop was supplied for those assessments. External model parameters remain opaque independently of these symbolic/textual routes; see RT-BAP-wiki-context, RT-BAP-schema-instruction, RT-CMP-llm-provider.

Shared records

Operative objects

EPI-OBJ-source-content

Label: Raw source content

Part of: RT-RTE-source-compilation

Object part of the route's raw-state output, not a duplicate of that end-to-end route. Source-native identity: source files named by RawSource.path and raw manifest entries. Representational form: natural-language source text plus symbolic source IDs/origin/type records. Storage substrate: local raw files/manifest and source SQLite rows. SRC-1 src/core/ingest/index.ts. Imported statements have unknown warrant; dispatch does not evaluate truth.

fsStore.writeRawManifest(manifest); --- src/core/ingest/index.ts

EPI-OBJ-image-description

Label: Image description

Part of: RT-RTE-source-compilation

Distinct intermediate object within ingestion: natural-language model description appended to a raw Markdown image reference; local raw file, with copied image in raw/images. Its factual claims concern visible image content. Unlike direct text import it is model-authored. SRC-1 src/core/ingest/image.ts; see EPI-RTE-image-description for passage.

EPI-OBJ-wiki-article

Label: WikiArticle

Part of: RT-RTE-source-compilation

Distinct retained output object, not a duplicate route. Source-native identity: WikiArticle.content and frontmatter. Representational form: natural-language body/summary and symbolic slug/source/tag/date fields. Storage substrate: Markdown, SQLite search rows and generated wiki index. Short-source body and model body have distinct producers; their admissions and consumers are traced separately. SRC-1 src/core/compile/compiler.ts; supplied write passage: RT-RTE-source-compilation.

EPI-OBJ-lint-issue

Label: LintIssue

No supplied counterpart after comparison with RT-RTE-source-compilation: diagnosis output, not an article or compilation result. Representational form: natural-language message/suggestion with symbolic severity/article fields. Storage substrate: returned JS array, CLI display and terminal system messages; counts may enter activity log. SRC-1 src/core/lint/linter.ts, src/cli/app.tsx. Structural diagnoses differ from model-generated allegations; see check record.

Routes

EPI-RTE-template-article

Label: Short-source template article

Part of: RT-RTE-source-compilation

Code selects sanitized raw content shorter than 2,000 characters, constructs a WikiArticle without model interpretation, and returns it to the compiler writer. Deterministic witness: sanitized content Rain falls.\nSnow melts. becomes unchanged article body; summary becomes Rain falls. Snow melts. by prefix/newline/trim rules. This establishes non-ampliative reshaping for that concrete input, not preservation for every Unicode normalization or source format. Input source selection is user-initiated; code authors metadata and decides this branch, no per-article veto. Open ingestion mode, no expected-answer oracle; schema/guidance are bypassed. Recovery and subsequent persistence are owned by the parent route.

• Immediate return: Single candidate WikiArticle array.
• Later read-back: Future query can receive persisted body through parent route's wiki/search writes.
• Delegated visibility: inapplicable — this branch has no model call.
• Selection predicate: Sanitized rawContent.length < 2000.
• Invalidation or expiry: Parent source-membership skip and file replacement govern future reuse; no separate template expiry.
• Activation or effect: Supplies candidate without semantic evaluation; actual downstream reliance unobserved.
• Evidence limits: Implementation deduction only; no target execution or source truth check.
• implementation conclusion status: wired
• operation conclusion status: uninspected

SRC-1 src/core/compile/compiler.ts, src/core/sanitization.ts.

summary: content.slice(0, 150).replace(/\n/g, ' ').trim(), --- src/core/compile/compiler.ts

content, path: '', --- src/core/compile/compiler.ts

EPI-RTE-article-admission

Label: Parsed article admission

Part of: RT-RTE-source-compilation

Model proposes article blocks; code splits markers, requires slug/title matches, sanitizes slug, and constructs frontmatter with only the current source ID. Compiler then collision-renames and writes/indexes accepted structures. This part owns structural disposition, not a truth check. Starting compile is not per-content human approval; code can skip malformed blocks, with no semantic veto or expected-answer oracle. Open requests. Guidance asks preservation and contradiction handling but does not supply existing article bodies. Rollback is not transactional across file/database writes; see parent recovery limits.

• Immediate return: Parsed article array, potentially empty; accepted entries reach parent writer.
• Later read-back: Persisted article files become later retrieval candidates under parent route.
• Delegated visibility: Model receives truncated raw source and slug/title index; parser receives response text, not an evidence audit.
• Selection predicate: Article marker blocks with slug/title and nonempty sanitized slug; no mandatory nonempty body.
• Invalidation or expiry: Parent source-membership filtering can retain stale represented-source articles; no truth-driven expiry here.
• Activation or effect: Permits wiki/search admission without factual verification or per-item human approval.
• Evidence limits: Does not establish acceptance of source propositions; no run evidence or observed successful persistence.
• implementation conclusion status: wired
• operation conclusion status: uninspected

SRC-1 src/core/compile/compiler.ts.

if (!slugMatch || !titleMatch) continue; --- src/core/compile/compiler.ts

sources: [sourceId], --- src/core/compile/compiler.ts

content: bodyMatch?.[1]?.trim() || '', --- src/core/compile/compiler.ts

EPI-RTE-image-description

Label: Image description generation

Part of: RT-RTE-source-compilation

Input image file → provider multimodal request → natural-language description → raw Markdown. A distinct pre-compilation transformation; caller supplies optional provider, code handles copying/writes, model authors description. File existence/size gates assess admissible input size, not image-description fidelity. No human per-description approval or reference oracle; open ingestion. Guidance is IMAGE_DESCRIPTION_PROMPT. Failure produces pending-description text; no semantic rejection, correction or rollback route is implemented in this function.

• Immediate return: RawSource pointing to persisted description/image reference.
• Later read-back: Parent compile reads the raw file; its resulting wiki article can reach future queries.
• Delegated visibility: Provider receives base64 image, MIME type, filename-derived title and description instructions.
• Selection predicate: Image type dispatch, existing file, size check; model path only with supplied llm.
• Invalidation or expiry: Re-ingest can regenerate raw description; no independent truth-review expiry.
• Activation or effect: Description becomes compiler input without independent fidelity check.
• Evidence limits: No image instance/description pair or run supplied; entailment versus hallucinated addition cannot be determined.
• implementation conclusion status: wired
• operation conclusion status: uninspected

SRC-1 src/core/ingest/image.ts.

description = ![${title}](images/${imgFilename})\n\n## Description\n\n${sanitizeUnicode(response.content)}; --- src/core/ingest/image.ts

EPI-RTE-wiki-checks

Label: Wiki structural and consistency checks

No supplied counterpart after comparison with RT-RTE-source-compilation: separate on-demand evaluator, not compiler admission. WikiLinter.lint loads articles and combines broken-link, orphan, missing-field and model-consistency checks. The structural evaluator's target is the loaded graph/metadata, not subject-matter truth. Example deduction: with only slug a present and body [[missing]], checkBrokenLinks reports missing target. The model evaluator's target is contradiction/duplication/gaps/terminology among supplied titles and summaries, not full-body factual consistency. Human invokes health/lint; code decides structural findings; model proposes semantic allegations; parser accepts issue shape/severity. No reference answer for semantic critique, no automatic human approval, open diagnostic mode. Invalid JSON/shape produces empty semantic findings, not a recorded proof of consistency.

• Immediate return: LintIssue array combining deterministic diagnoses and parsed model allegations.
• Later read-back: inapplicable — this function does not retain diagnoses for another invocation.
• Delegated visibility: Model sees slug/title/summary list, not article bodies or underlying sources.
• Selection predicate: All loaded articles for structural checks; fewer than two bypasses semantic comparison; valid severity/string fields for model issue admission.
• Invalidation or expiry: Fresh invocation recomputes; no durable issue withdrawal protocol.
• Activation or effect: Results go to separate advisory consumers; no article repair or veto here.
• Evidence limits: Structural license bounded to code's matching rules and loaded state; model allegations lack independent evidence verification; no observed diagnoses.
• implementation conclusion status: wired
• operation conclusion status: uninspected

SRC-1 src/core/lint/linter.ts.

const summaries = articles .map(a => - ${a.slug}: ${a.frontmatter.title} — ${a.frontmatter.summary || '(no summary)'}) .join('\n'); --- src/core/lint/linter.ts

if (!slugs.has(target)) { issues.push({ severity: 'warning', article: article.slug, message: Broken wiki link: [[${target}]], --- src/core/lint/linter.ts

if (!jsonMatch) return []; --- src/core/lint/linter.ts

EPI-RTE-issue-display

Label: Health issue display

No supplied counterpart after comparison with RT-RTE-answer-question: lint result consumption, not answer generation. CLI prints issues and suggestions or healthy status; terminal adds system notices and logs severity counts. Consumer: operator, via console/session messages. Force: advisory, no edit/retrieval veto in these callers. Horizon: present display and retained session notices; no programmatic issue-driven correction loop established. Users may act externally, but invocation is not evidence of such action. Open diagnostic mode; no admission proposal or reference-answer oracle in this display part.

• Immediate return: Console/session messages and severity counts.
• Later read-back: Terminal session notices can be saved, but supplied history/digest routes omit system messages from provider input; no automatic corrective reader established.
• Delegated visibility: inapplicable — display does not invoke a model.
• Selection predicate: Parsed issues displayed; zero issues produces healthy CLI message.
• Invalidation or expiry: Later lint can replace displayed assessment; no durable rejection/withdrawal semantics.
• Activation or effect: Changes operator-visible advice only; no automatic article correction or blocking.
• Evidence limits: User reliance and subsequent edits unobserved; healthy message can follow absent/failed semantic parsing.
• implementation conclusion status: wired
• operation conclusion status: uninspected

SRC-1 src/cli/commands/lint.ts, src/cli/app.tsx.

success('Wiki is healthy — no issues found'); --- src/cli/commands/lint.ts

console.log(${icon} [${issue.article}] ${issue.message}); --- src/cli/commands/lint.ts

EPI-RTE-coverage-diagnostics

Label: Query coverage diagnostics

Part of: RT-RTE-answer-question

Material check part of context preparation, also reused by RT-RTE-library-query. Code turns rank scores/count into high/medium/low coverage; per-term FTS with punctuation/stop-word/length filters yields gaps. Target is lexical corpus coverage, not proposition truth. Human supplies question, software decides scores/gaps, no expected-answer oracle or per-diagnostic veto. Open query mode. Consumer is result/UI and terminal's first-gap suggestion path; code can request search through gapResolver with blacklist/length filtering. Force: advisory, not factual answer admission. Gap suggestions do not validate fetched sources.

• Immediate return: Coverage confidence and gaps bundled with context.
• Later read-back: Library cached query result can retain diagnostics; no diagnosis-driven method revision established.
• Delegated visibility: Model receives assembled context; diagnostic fields alone do not establish an independent model check.
• Selection predicate: Ranked candidate scores/count; search terms longer than four characters and not stop words; one FTS hit counts as coverage.
• Invalidation or expiry: Recompute on new context build; parent cache TTL/invalidation qualifies reuse.
• Activation or effect: Result can trigger source suggestions and UI labels; no semantic truth veto.
• Evidence limits: Ranking/search internals not independently verified here; threshold is not measured epistemic calibration; suggestions/benefit unobserved.
• implementation conclusion status: wired
• operation conclusion status: uninspected

SRC-1 src/core/retrieval/contextBuilder.ts, src/core/discovery/gapResolver.ts; supplied terminal gap consumer: RT-RTE-answer-question.

if (topScore > 15 && ranked.length >= 3 && avgScore > 5) return 'high'; --- src/core/retrieval/contextBuilder.ts

const results = this.db.search(term, 1); if (results.length === 0) { gaps.push(term); --- src/core/retrieval/contextBuilder.ts

EPI-RTE-conversation-compaction

Label: Conversation compaction

Distinct from RT-RTE-answer-question's ordinary history selection: a separate model summary call whose result is discarded by the shipped terminal caller. SRC-1 src/core/retrieval/compactor.ts, src/cli/app.tsx. After more than twenty session messages the terminal requests compaction retaining six recent messages. Code slices older messages, truncates each to 3,000 characters, asks a model for topic/findings/corrections/pending-thread summary and returns a system-summary message plus recent tail. The app awaits the call without assigning its result. This is candidate generation without a subsequent answer consumer in this path, despite the nearby comment. No factual fidelity check, reference oracle or human per-summary approval; open conversation mode. On repeated provider failure original messages are returned.

• Immediate return: Summary and keptMessages, or original messages after retry failure.
• Later read-back: inapplicable — inspected terminal call discards return; no persistence or future prompt installation occurs there.
• Delegated visibility: Summarizer receives older role-tagged messages with per-message truncation, not retained recent tail.
• Selection predicate: Terminal message count greater than twenty and provider available; function summarizes all but requested recent count.
• Invalidation or expiry: Returned candidate has no retained lifetime in this caller; original displayed/session messages remain.
• Activation or effect: Adds a provider call, not a demonstrated or wired replacement of later terminal answer context.
• Evidence limits: No run; unspecified library host could consume return, but that behavior is not established.
• implementation conclusion status: wired
• operation conclusion status: uninspected

await compactConversation( session.messages.map(m => ({ role: m.role, content: m.content })), currentLlm, 6, ); --- src/cli/app.tsx

// Compaction result is used internally for LLM context in future queries. --- src/cli/app.tsx

const summaryMatch = response.content.match(/

([\s\S]*?)<\/summary>/); --- src/core/retrieval/compactor.ts

Claims

EPI-CLM-source-compounding

Label: Source addition compounds answer quality

No supplied counterpart after comparison with RT-CLM-wiki-grounding: improvement claim rather than grounding policy. SRC-1 README.md, doctrine/design; universal future-answer benefit is not observed or causally supported.

The loop that compounds: every source you add makes every future answer better. Every question you ask can surface gaps — browzy suggests sources to fill them. --- README.md

EPI-CLM-compiler-preservation

Label: Compiler source preservation

Distinct from answer provenance claim RT-CLM-wiki-grounding: compiler instructions request preservation, merging and contradiction handling. SRC-1 src/core/prompts.ts, src/core/compile/compiler.ts, doctrine/design. Code input limitations prevent interpreting request as a guarantee.

1. If this source's content fits into an existing article, output an UPDATED version of that article with the new information merged in. Preserve all existing content and sources — add to them, don't replace. --- src/core/compile/compiler.ts

EPI-CLM-quality-auditor

Label: Wiki quality auditor

Distinct from RT-CLM-wiki-grounding: diagnostic mandate rather than answer policy. SRC-1 src/core/prompts.ts, doctrine/design. Mandate does not establish reliable detection or corrective force.

export const LINTER_SYSTEM_PROMPT = You are browzy's browzy quality auditor. Your job is to find real problems in the knowledge base — not style preferences, not nitpicks, but issues that would cause a researcher to get wrong answers, miss connections, or waste time. ---src/core/prompts.ts`