WeKnora epistemic report

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

Source-and-claim boundary

Question: what truth-apt content can WeKnora acquire or generate, what checks affect reliance, and what later consumers receive the results? Scope and exclusions are see boundary; evidence identity is see Source register. SRC-1 supplies implementation and doctrine/design, not observed operation or causal experiments. External model weights, provider internals and deployed storage/isolation are outside the inspected implementation. No target execution occurred.

The report directly inspected the additional source passages listed below, at the frozen pin. Supplied runtime routes remain provisional but provide the ordinary turn, RAG, context, document and installer witnesses. Available but uninspected alternatives remain included responsibilities, not access gaps. Coverage therefore supports independent positive findings and bounded limitations, not a system-complete negative about checking or knowledge production.

Entry point or operation	SRC-1 source path	Coverage and records; conclusion prevented
Agent/RAG answers, source enablement, context and tools	internal/application/service/session_agent_qa.go, internal/application/service/session_knowledge_qa.go, internal/application/service/knowledge_process.go, internal/agent/engine.go	Supplied runtime findings: RT-RTE-agent-turn, RT-RTE-rag-answer, RT-RTE-knowledge-admission, RT-RTE-context-rewrite. Exact answer content and model compliance unobserved.
CLI dispatch	cli/cmd/root.go	Direct read of command registrations: search, document/chunk CRUD, chat, model, MCP and skills surfaces corroborate multiple operations, not one agent-only entry. Individual command bodies and generic API passthrough uninspected; prevents exhaustive client-side semantic/admission account.
Registered built-ins	internal/application/service/agent_service.go	Direct read of register switch: search/read/list, graph, conversation/memory, SQL/data analysis, web, wiki read/issues/write/replace/rename/delete. Memory and wiki routes below add omitted maintenance checks. SQL/data-analysis derivation, web acquisition, graph claims and issue resolution bodies uninspected; prevents conclusions about every evaluator or derivation domain.
Model/tool controls and installer	internal/agent/tools/registry.go, internal/agent/tools/mcp_tool.go, internal/application/service/tenant_skill_install.go	Supplied RT-RTE-mcp-call, RT-RTE-sandbox-execution, RT-RTE-skill-install. Schema, approval and dependency checks have different targets from truth checking.
Background memory extraction/write/review	internal/application/service/memory/extract.go, internal/application/service/memory/service.go, internal/router/routes_memory.go	Direct reads; EPI-RTE-memory-extraction, EPI-RTE-memory-write, EPI-RTE-memory-review. Topics/consolidation/vector internals not fully read; prevents complete memory adaptation and lineage account.
Memory later consumption	internal/application/service/memory/service.go, internal/application/service/memory/search.go	Direct reads; EPI-RTE-memory-recall; supplied caller injection in boundary. Ranker internals and all stale-cache failure cases uninspected.
Wiki generation and withdrawal	internal/application/service/wiki_ingest_batch.go, internal/application/service/wiki_ingest.go, internal/application/service/wiki_page.go	Direct bounded reads; EPI-RTE-wiki-rewrite, EPI-RTE-wiki-guard. Map extraction/deduplication, all generation continuation and draft publication branches not fully read; prevents complete generated-content admission account.
Wiki lint and cleanup	internal/application/service/wiki_lint.go, internal/application/service/wiki_ingest.go	Direct reads; EPI-RTE-wiki-maintenance. Link resolution helper and issue-tool internals uninspected; prevents semantic equivalence and complete criticism-loop claims.
Dataset metric hook and consumer	internal/application/service/metric_hook.go, internal/application/service/metric/precision.go, internal/application/service/metric/bleu.go, internal/application/service/evaluation.go	Direct reads; EPI-RTE-dataset-metrics. BLEU tokenization/helpers and other metric formulas not fully read; no measured quality or causal comparison.
Other state-producing surfaces	docreader/main.py, internal/application/service/knowledge_create.go, internal/application/service/knowledgebase_search.go	Boundary survey/runtime coverage only for parser, URL/manual/FAQ ingestion, connectors, OCR/VLM, graph and alternate indexes. End-to-end semantic fidelity and exact evaluator coverage remain unassessed.

Knowledge/warrant claims: RT-CLM-evidence-verification, EPI-CLM-cited-answers, EPI-CLM-automatic-wiki, EPI-CLM-end-evaluation. No inferred stronger promise of infallibility is assessed.

Epistemic-object inventory

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source/local anchor	Gap/limit
RT-OBJ-system-prompt	Verification policy, not itself evidence of verified answers	see record	see record	Instructions do not establish obedience.
RT-OBJ-agent-config	None for grants/limits; configuration is operational policy	see record	see record	Permission is not endorsement.
EPI-OBJ-knowledge-chunks	Imported source assertions over their document scope	see record	see record	Parser fidelity and source truth unestablished.
EPI-OBJ-chat-response	Candidate answer assertions or none, depending on response	see record	see record	Actual propositions and factual entailment not determinable without output instances.
EPI-OBJ-context-checkpoint	Summarized earlier conversation claims	see record	see record	Compression can omit evidence; retained summary is not accepted truth.
EPI-OBJ-memory-item	User/project facts, preferences and tasks; inferred user claims	see record	see record	Truth apt only over named user/environment/time; source pointers and status do not prove truth.
EPI-OBJ-wiki-page	Document synopses, entity/concept assertions and source relationships	see record	see record	Summary refs and chunk refs differ; proposition-level support unproved.
EPI-OBJ-metric-result	Measurements about supplied evaluation pairs, not general truth of answers	see record	see record	Matching can omit unmatched retrieval results; see metric route.

Answers, imported chunks and summaries are separately inventoried here as outputs of the supplied runtime routes, not as duplicate route declarations. Their exact propositions are not determinable without an instance. Model components remain separately typed in RT-CMP-chat-model, RT-CMP-embedding-model, RT-CMP-rerank-model and RT-CMP-vlm-model: weights and exact parameter identities are not inspected. Dense vectors and ranking scores are not equated with localized factual claims.

Authority-route ledger

Every row is an architectural assessment, not a claim of activation. Separate rows distinguish checking, disposition and retention even when one record covers their common implementation. Owned route details say see rather than being duplicated.

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-knowledge-admission	content transformation	implemented	EPI-OBJ-knowledge-chunks	truth-apt transformation: acquisition/import	external source → indexed content	Parser/index pipeline; semantic fidelity uninspected	see RT-RTE-knowledge-admission	stored/indexed/enabled	see RT-RTE-knowledge-admission	Source warrant unknown; source storage does not check its truth	Eligible evidence for retrieval	Retrieval services; indexes; permissive; later queries	see RT-RTE-knowledge-admission	none	none	Operation unobserved; see record.
RT-RTE-rag-answer	content transformation	implemented	EPI-OBJ-chat-response	truth-apt transformation: indeterminate	retrieved evidence → answer	Configured chat model; no reference oracle in ordinary QA	see RT-RTE-rag-answer	answer/fallback/error	see RT-RTE-rag-answer	Candidate answer only; entailment and source truth not established	Answer delivered/stored	User and later QA context; answer/history; advisory; current/later turns	see RT-RTE-rag-answer	EPI-CLM-cited-answers	none	Operation unobserved; see record.
RT-RTE-agent-turn	content transformation	implemented	EPI-OBJ-chat-response; tool proposals are a distinct non-truth-apt action channel	truth-apt transformation: indeterminate	open task/context → model content	Configured model; code validates calls, not all factual claims	see RT-RTE-agent-turn	answer, partial content or error	see RT-RTE-agent-turn	Candidate assertions, not an inspected evidence-consuming acceptance decision	Can propose executable actions under distinct eligibility checks	Tools/user/history; calls and text; conditional executable/advisory; turn and later history	see RT-RTE-agent-turn	RT-CLM-evidence-verification	none	Operation unobserved; see record.
RT-RTE-context-rewrite	content transformation	implemented	EPI-OBJ-context-checkpoint; transient tool previews	truth-apt transformation: indeterminate	older context → compressed context	Summary model and code budget rules; token size not semantic equivalence	see RT-RTE-context-rewrite	summary/raw fallback/preview	see RT-RTE-context-rewrite	Semantic warrant unresolved; truncation can degrade evidence visibility	Replacement admitted into later model inputs	Foreground model; summary/checkpoint; advisory; later invocations	see RT-RTE-context-rewrite	none	none	Operation unobserved; see record.
RT-RTE-context-rewrite	retention	implemented	EPI-OBJ-context-checkpoint	no content change	summary → saved checkpoint	Code eligibility, not factual acceptance	see RT-RTE-context-rewrite	checkpoint persisted	see RT-RTE-context-rewrite	Retention is not acceptance	Later history can start from checkpoint	History loader/model; repository/context; permissive/advisory; later turns	see RT-RTE-context-rewrite	none	none	Operation unobserved; see record.
RT-RTE-mcp-call	disposition/acceptance	implemented	tool call arguments	non-truth-apt policy/content update: execution permission	proposed call → permitted/refused	Policy/schema and conditional user approval; effect authorization	see RT-RTE-mcp-call	execute/refuse/suspend	see RT-RTE-mcp-call	No truth license for returned content	Permits or blocks selected external call	Registry/MCP; policy/approval; enforcing when active; selected call	see RT-RTE-mcp-call	none	none	Operation unobserved; see record.
RT-RTE-skill-install	check/evidence production	implemented	skill capability prerequisites	no content change	bundle/install candidate	Tree/script/declared command checks; dependency availability domain	see RT-RTE-skill-install	verification findings	see RT-RTE-skill-install	Bounded prerequisite evidence, not all functional promises	Findings can trigger repair or gate rejection	Installer service; verifier result; enforcing; installation rounds	see RT-RTE-skill-install	none	none	Operation unobserved; see record.
RT-RTE-skill-install	disposition/acceptance	implemented	snapshot candidate	non-truth-apt policy/content update: capability admission	verified install → image pointer	Verifier outcome plus lock/ownership checks	see RT-RTE-skill-install	ready/image switch or failure	see RT-RTE-skill-install	Acceptance for setup use under inspected prerequisite criteria only	New capability environment admitted	Workspace resolver; snapshot pointer; executable/permissive; later sessions	see RT-RTE-skill-install	none	none	Operation unobserved; see record.
EPI-RTE-memory-extraction	content transformation	implemented	EPI-OBJ-memory-item	truth-apt transformation: indeterminate	transcript/existing notes → decisions	Extraction model; JSON/truncation code checks output usability	see EPI-RTE-memory-extraction	add/update/delete/none proposals	see EPI-RTE-memory-extraction	Repeating a statement acquires its claim; exact semantic preservation unproved	Proposes changes but does not itself confirm inferred claims	Write consumer; JSON decisions; advisory; background extraction	see EPI-RTE-memory-extraction	none	none	Prompt example explicitly licenses a non-entailed guess from event questions to organising events; that candidate class is ampliative conjecture. Particular actual outputs unobserved.
EPI-RTE-memory-write	check/evidence production	implemented	EPI-OBJ-memory-item	no content change	proposed memory write	Sanitization, credential policy, tombstones and conflicts; storage-policy domain	see EPI-RTE-memory-write	eligible/redacted/skipped/error	see EPI-RTE-memory-write	No test of source truth or of inferred flag correctness	Determines eligible writes	Memory service; validation results; enforcing; this write	see EPI-RTE-memory-write	none	none	Operation unobserved; see record.
EPI-RTE-memory-write	disposition/acceptance	implemented	EPI-OBJ-memory-item	non-truth-apt policy/content update: memory status	candidate → active/pending	Code uses Inferred and Origin supplied by producer	see EPI-RTE-memory-write	active or pending	see EPI-RTE-memory-write	Active licenses operational recall, not evidence-certified truth	Unflagged extraction can take effect without human confirmation	Repository/recall; status; permissive; later turns	see EPI-RTE-memory-write	none	none	Operation unobserved; see record.
EPI-RTE-memory-write	retention	implemented	EPI-OBJ-memory-item	no content change	eligible candidate → stored row	Code/repository write; capacity and embeddings best effort	see EPI-RTE-memory-write	saved/replaced/archived	see EPI-RTE-memory-write	Source pointers retained; warranted premises not independently checked	Retains claims for later selection	Recall/search; memory row; permissive; cross-session	see EPI-RTE-memory-write	none	none	Operation unobserved; see record.
EPI-RTE-memory-review	disposition/acceptance	implemented	EPI-OBJ-memory-item	non-truth-apt policy/content update: confirmation or withdrawal	pending claim → confirmed or rejected	API caller decision; service confirms or deletes	see EPI-RTE-memory-review	active or deleted/tombstoned	see EPI-RTE-memory-review	Human/caller endorsement possible; evidence consumed and named truth criterion not recorded	Confirmation permits recall; rejection blocks ordinary read and re-derivation when tombstone exists	Recall/write; status/tombstone; enforcing eligibility; later turns	see EPI-RTE-memory-review	none	none	Generic principal is not evidence of human review. Tombstone failure is logged but deletion continues; absolute non-recurrence is not warranted.
EPI-RTE-memory-recall	operational admission/selection/consumption	implemented	EPI-OBJ-memory-item	no content change	active memories → selected prompt/tool result	Code active queries and lexical/semantic ranking; relevance domain	see EPI-RTE-memory-recall	selected memories	see EPI-RTE-memory-recall	Recall does not add truth warrant	Supplies model context; touch records use rather than truth	Foreground model; prompt/search result; advisory; subsequent model calls	see EPI-RTE-memory-recall	none	none	Subject block cache fallback exists; no complete failure-path withdrawal guarantee established.
EPI-RTE-wiki-rewrite	content transformation	implemented	EPI-OBJ-wiki-page	truth-apt transformation: indeterminate	existing page + sources/additions/retractions → revised page	Chat model guided by chunks, summaries and deleted/remaining sources	see EPI-RTE-wiki-rewrite	new body or generation failure	see EPI-RTE-wiki-rewrite	Source references aid inspection; generated body may be paraphrase, derivation or conjecture	Proposes page content including removal of deleted-source contributions	Wiki write service; generated body; advisory until write; batch/later reads	see EPI-RTE-wiki-rewrite	EPI-CLM-automatic-wiki	none	Operation unobserved; see record.
EPI-RTE-wiki-guard	check/evidence production	implemented	EPI-OBJ-wiki-page	no content change	stored/proposed table row identities	Deterministic identity-loss check; table preservation domain	see EPI-RTE-wiki-guard	loss ratio/refusal	see EPI-RTE-wiki-guard	Checks bounded loss, not cell truth, full evidence preservation or source encoding	Produces enforcing write condition	UpdatePage; loss result; enforcing with exemptions; this rewrite	see EPI-RTE-wiki-guard	none	none	Operation unobserved; see record.
EPI-RTE-wiki-guard	disposition/acceptance	implemented	EPI-OBJ-wiki-page	non-truth-apt policy/content update: page write admission	candidate → saved or deferred	Code source-specific tolerance; intentional shrink exceptions	see EPI-RTE-wiki-guard	write or old page retained/deferred	see EPI-RTE-wiki-guard	No factual acceptance decision for all page propositions	Blocks offending rewrite; ingest retains contributing documents for retry	Ingest/agent writer; explicit error; enforcing; this write/next batch	see EPI-RTE-wiki-guard	none	none	Operation unobserved; see record.
EPI-RTE-wiki-rewrite	retention	implemented	EPI-OBJ-wiki-page	no content change	generated page → persisted content and refs	Wiki service CreatePage/UpdatePage	see EPI-RTE-wiki-rewrite	page/ref update	see EPI-RTE-wiki-rewrite	Retained candidate with lineage, not necessarily accepted knowledge	Available to wiki read/search tools	Wiki clients/model; stored pages; advisory; later requests	see EPI-RTE-wiki-rewrite	none	none	Operation unobserved; see record.
EPI-RTE-wiki-maintenance	check/evidence production	implemented	EPI-OBJ-wiki-page	no content change	wiki link/source availability and short content	RunLint live-slug/source probes and content-size rules; structural domain	see EPI-RTE-wiki-maintenance	lint issues	see EPI-RTE-wiki-maintenance	Existence/freshness evidence only; not endorsement	Selected results feed AutoFix	AutoFix/operator; lint report; advisory/conditional enforcing; maintenance invocation	see EPI-RTE-wiki-maintenance	none	none	Operation unobserved; see record.
EPI-RTE-wiki-maintenance	content transformation	implemented	EPI-OBJ-wiki-page content	truth-apt transformation: non-ampliative reshaping	exact broken-link markup → plain target text	AutoFix literal replacement; text/markup domain	see EPI-RTE-wiki-maintenance	plain text persisted	see EPI-RTE-wiki-maintenance	Preserves the demonstrated label text while withdrawing a navigable reference; not source endorsement	Changes displayed link availability	Wiki readers/model; stored content; advisory; later requests	see EPI-RTE-wiki-maintenance	none	none	Exact [[target]] replacement only; display-text variants remain unmatched in this branch. Semantic equivalence for fuzzy remapping remains unassessed.
EPI-RTE-wiki-maintenance	lineage/freshness/recovery	implemented	EPI-OBJ-wiki-page SourceRefs	non-truth-apt policy/content update: source-reference withdrawal	stale source reference on multi-source page	AutoFix lookup and remaining-source count; existence domain	see EPI-RTE-wiki-maintenance	metadata removed, prose not rewritten by this branch	see EPI-RTE-wiki-maintenance	Removes lineage to unavailable source; does not refute or re-ground the retained prose	Changes source drill-down without necessarily changing factual content	Wiki clients/model; stored refs; enforcing lineage selection; later requests	see EPI-RTE-wiki-maintenance	none	none	Separate model retract route attempts content withdrawal; source lookup failures can be classified as stale.
EPI-RTE-wiki-maintenance	disposition/acceptance	implemented	EPI-OBJ-wiki-page status/existence	non-truth-apt policy/content update: page withdrawal	short non-index page or stale sole-source page	AutoFix size and source-count rules; availability domain	see EPI-RTE-wiki-maintenance	archived or deleted	see EPI-RTE-wiki-maintenance	Withholds page availability, not a truth judgment	Removes page from normal availability according to service status/deletion semantics	Wiki services/readers; repository status/deletion; enforcing; later requests	see EPI-RTE-wiki-maintenance	none	none	Every downstream backend's status filtering and recovery semantics remain uninspected.
EPI-RTE-dataset-metrics	check/evidence production	implemented	EPI-OBJ-metric-result	truth-apt transformation: entailed derivation	supplied QA pairs, mapped result IDs and generated text	HookMetric plus deterministic calculators; dataset/reference-relative domain	see EPI-RTE-dataset-metrics	metric vector	see EPI-RTE-dataset-metrics	Scores follow encoded inputs, not warranted source truth; denominator limited by mapping	Exposed result for caller comparison, no automatic successor admission	Evaluation caller; in-process task metric; advisory; task lifetime	see EPI-RTE-dataset-metrics	EPI-CLM-end-evaluation	none	Unmatched chunks disappear from mapped IDs. Source fidelity, full metric formulas and observed measurements remain unestablished.

No lifecycle-integration row is inferred from mere retention or operational consumption. The inspected content-write paths do not record an evidence-consuming acceptance decision naming a truth criterion, intended use and scope; this is a bounded finding about these paths, not a system-wide absence. Setup image admission has a distinct operational acceptance criterion. Wiki issue tools and other evaluators remain uninspected as recorded above.

System-claim versus route comparison

Claim ID	Claimed operation or warrant	Claim source/anchor and layer	Doctrine/design support	Implemented route IDs	Observed-run support	Causal support and design limits	Supported conclusion	Mismatch/unknown
RT-CLM-evidence-verification	see record	see record	Prompt asks source verification	RT-RTE-agent-turn, RT-RTE-mcp-call	None registered	None; schema/approval do not check exact quotation truth	Verification is requested; source-reading tools are afforded	Actual verification and entailment unobserved; not an executable universal guarantee
EPI-CLM-cited-answers	see record	see record	Source-cited QA intent	RT-RTE-rag-answer	None registered	None; ordinary QA has no reference oracle	Retrieval and answer supply wired	Whether every displayed citation supports its associated proposition remains unknown
EPI-CLM-automatic-wiki	see record	see record	Generated/editable/wiki rollback intent	EPI-RTE-wiki-rewrite, EPI-RTE-wiki-guard, EPI-RTE-wiki-maintenance	None registered	None; rollback restores content fields, not historical provenance/placement; maintenance can bypass visible version bump	Model-assisted pages, source bookkeeping, bounded write veto and revision revert implemented	Literal “every change” is broader than inspected revert coverage; structural checks do not certify factual claims
EPI-CLM-end-evaluation	see record	see record	Recall/BLEU/ROUGE evaluation intent	RT-RTE-dataset-evaluation, EPI-RTE-dataset-metrics	None registered	None; no intervention/component comparison; fixed evaluation RAG differs from open agent	Reference-relative metrics wired and exposed	Content mapping can hide unmatched retrievals; lexical similarity is not answer truth or mechanism validation

Bounded conclusion

WeKnora acquires documents and uses indexed content as model evidence; indexing/enablement is availability rather than truth acceptance (RT-RTE-knowledge-admission). Source warrant and parsing fidelity remain unknown. RAG and ReAct produce candidate answers, but no actual answer instance here supports an entailment or non-entailment judgment. Model summaries and wiki rewrites likewise remain semantically indeterminate. Fluency and citations alone license no additional truth claim.

Memory is a materially different reliance route. The extraction prompt permits inferences and explicitly illustrates an ampliative user guess. The model-supplied inferred flag can hold that candidate pending; non-inferred extracted content can become active automatically. Confirmation/rejection changes recall eligibility and rejection can persist a tombstone. This defeats a blanket claim that all memory writes take effect silently, but it does not establish evidence-certified acceptance: the service records the caller action, not the evidence or truth criterion. Active memories are selected into later prompt/tool results (EPI-RTE-memory-extraction, EPI-RTE-memory-write, EPI-RTE-memory-review, EPI-RTE-memory-recall). Capacity, conflict, credential and forgotten-content checks constrain retention independently of truth. Flag fidelity, failed tombstones and cached-block fallback bound withdrawal guarantees.

Wiki generation and retraction receive source chunks, summaries, existing content and deleted/remaining source information. References provide navigable lineage but not an independent proposition-level oracle. Table identity preservation can veto a machine rewrite, while deliberate retract, user and revert operations have exceptions. Lint availability/size checks can drive text cleanup, metadata withdrawal, archival or deletion. Those are consequential maintenance checks, not factual acceptance. A multi-source stale-ref AutoFix may leave prose unchanged, unlike the separate model-directed retract route. Revert restores selected revision content while retaining current provenance and placement (EPI-RTE-wiki-rewrite, EPI-RTE-wiki-guard, EPI-RTE-wiki-maintenance). No evidence-consuming lifecycle integration is inferred merely from these edits.

The dataset evaluator derives numbers from encoded inputs and reference answers. Its operative target is retrieval IDs and lexical response similarity over supplied pairs, not general correctness, source truth or producing mechanism. The content-to-passage map drops unmatched retrieved chunks before precision computation, so an all-relevant mapped subset need not mean all actual returned chunks were relevant. Metrics reach the result caller; no automatic accepted successor is established in this path (EPI-RTE-dataset-metrics).

Execution authority has separate scopes. MCP approvals authorize effects, and skill verification gates a durable environment under dependency/tree/script criteria. Neither certifies every generated claim. Skill image admission can change later capabilities; exercised self-improvement and successful learning require later-operation and improvement evidence not supplied here (RT-RTE-skill-install, RT-BAP-skill-image). Trace-fed memory can alter later model input without proving improved capacity. Theory formulation, content-directed criticism, reflection and autonomy are not inferred from persistence, repair or ranking; issue-tool and other uninspected branches prevent complete negatives. Model weight mutability/version/endpoint limits remain component-local in the runtime records. All implemented routes remain operationally unobserved, and no causal benefit comparison is registered.

Shared records

Operative objects

EPI-OBJ-knowledge-chunks

Label: Indexed document chunks

Source-native identity: document text chunks produced through parsing/indexing. Form: natural-language source content, with symbolic knowledge identifiers and enablement controlling retrieval; numerical vectors are separate retrieval representations. Substrate: source storage and chunk/index repositories. No supplied operative-object counterpart: RT-RTE-knowledge-admission declares the producer/admission route, not this content object. SRC-1, internal/application/service/knowledge_process.go; minimum enablement passage and read-back witness are see that supplied record. Additional direct witness: internal/application/service/wiki_ingest.go reconstructContent filters text chunks and merges them; internal/application/service/wiki_ingest_batch.go resolves cited chunk content for the wiki editor. Chunk content is acquired source material; fidelity and source warrant remain unresolved.

EPI-OBJ-chat-response

Label: ChatResponse candidate content

Source-native identity: types.ChatResponse.Content produced by configured chat inference. Form: natural-language answer or other model content; protocol FinishReason is a separate symbolic usability input. Substrate: transient Go response, with answers subsequently stored in conversation history by application callers. No supplied object counterpart: RT-RTE-rag-answer and RT-RTE-agent-turn own distinct production/consumption routes, not a shared output object declaration. SRC-1, internal/application/service/metric_hook.go, reads chatResponse.Content for generation metrics; internal/application/service/memory/extract.go reads Content and FinishReason for extraction usability. Supplied runtime records establish ordinary answer output and persistence. Exact actual truth-apt claims cannot be individuated from implementation alone.

generatedTexts = h.qaPairMetricList[index].chatResponse.Content --- internal/application/service/metric_hook.go

EPI-OBJ-context-checkpoint

Label: Context summary checkpoint

Source-native identity: eligible compaction summary checkpoint replacing covered conversation context. Form: natural-language summary with symbolic checkpoint coverage/read-back selection. Substrate: message repository through caller-installed checkpoint sink. No supplied object counterpart: RT-RTE-context-rewrite owns transformation, persistence and later read-back, not this output-object declaration. SRC-1, internal/application/service/session_agent_qa.go, internal/agent/compaction/compactor.go; minimum checkpoint-sink quotation and summary-selection findings are see the supplied route. Eligibility to persist and budget reduction do not test semantic preservation. Raw fallback and transient previews are alternate current-context representations, not evidence of a correct accepted summary.

EPI-OBJ-memory-item

Label: MemoryItem claim and status

Source-native identity: types.MemoryItem, including natural-language Content and Topic and separately symbolic Origin, Status, source IDs and expiry. Substrate: memory repository rows; dense embeddings are separate selection material. Distinct from the supplied runtime's prompt/config objects: a durable per-subject claim source, later selected into RT-OBJ-system-prompt, not the prompt container itself. SRC-1, internal/application/service/memory/service.go, internal/application/service/memory/extract.go. Model extraction proposes Content and Inferred; write stores status and source IDs. Source linkage is lineage, not a verified entailment.

Status: statusForWrite(item), SourceSessionID: item.SourceSessionID, SourceMessageID: item.SourceMessageID, --- internal/application/service/memory/service.go

EPI-OBJ-wiki-page

Label: WikiPage content and provenance

Source-native identity: types.WikiPage natural-language Title/Content/Summary; symbolic SourceRefs, ChunkRefs, status, links and version have distinct fixed consumers. Substrate: wiki service repository, with revision snapshots. No supplied object counterpart: the closest RT-RTE-knowledge-admission concerns indexed document state, not generated wiki pages. SRC-1, internal/application/service/wiki_ingest_batch.go, internal/application/service/wiki_page.go. Summary pages have document-level refs but clear chunk refs; entity/concept addition rounds merge chunk refs. Neither establishes that every proposition follows from those refs.

page.ChunkRefs = types.StringArray{} changed = true --- internal/application/service/wiki_ingest_batch.go:2027-2033

EPI-OBJ-metric-result

Label: MetricResult evaluation vector

Source-native identity: types.MetricResult retrieval/generation metrics. Form: symbolic numeric measurements interpreted through fixed calculator rules, not distributed model weights. Substrate: MetricList and evaluation task's in-process storage. No supplied object counterpart; produced inside RT-RTE-dataset-evaluation. SRC-1, internal/application/service/metric_hook.go, internal/application/service/evaluation.go.

params.Metric = metricHook.MetricResult() params.Task.Finished = finished --- internal/application/service/evaluation.go

Routes

EPI-RTE-memory-extraction

Label: Background memory extraction

Distinct from foreground RT-RTE-agent-turn: asynchronous maintenance with a separate model request, JSON decisions and write consumer; no supplied route counterpart. Owner: memory Service. Progression: selected transcript plus existing/forgotten/topic context → chat model → truncation/JSON handling → applyDecisions.

SRC-1 implementation, internal/application/service/memory/extract.go.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Parsed decisions or extraction error.
• Later read-back: Decision content can enter memory through EPI-RTE-memory-write; temporary model context itself is not retained.
• Delegated visibility: Model gets segment, existing notes, forgotten notes/topics and workspace extraction instructions; foreground does not receive this whole temporary frame.
• Selection predicate: Configured extraction model; isTruncated retries once; malformed output rejected.
• Invalidation or expiry: Error prevents usable extraction result; per-item invalid target/action can be skipped. Full scheduling/watermark recovery not inspected.
• Activation or effect: Implemented proposal of memory addition, replacement, supersession; no observed output.
• Evidence limits: Semantic equivalence and correct source indexing not guaranteed by valid JSON; model processing opaque.

Open transcript distillation, not an experiment. Model proposes and can mark Inferred; code rejects unusable output, without an expected-answer oracle. Guidance names durable user notes, prohibits credentials and requests source-line IDs. The prompt example gives explicit non-entailment evidence: asking about ceremonies/venues does not entail organising events. For that candidate class the relation is ampliative conjecture; unobserved actual decisions remain indeterminate. Temporary summarizer-like work frame returns decisions by value; stored notes are later selected separately, not live references to the discarded frame.

• Set "inferred" to true when you are deducing something about the user rather than repeating what they said — for example concluding from questions about award ceremonies and venue clearing that they organise events. Such entries are shown to the user for confirmation instead of taking effect silently, so a reasonable guess is welcome; a confident assertion is not. --- internal/application/service/memory/extract.go

EPI-RTE-memory-write

Label: Memory write admission

No supplied route counterpart; follows EPI-RTE-memory-extraction and also handles explicit/manual inputs. Owner: memory Service. Progression: content/policy checks → conflict/duplicate selection → SaveItem → capacity/block/embedding updates.

SRC-1 implementation, internal/application/service/memory/service.go.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Stored/unchanged item, rejected sensitive/forgotten/conflicting candidate, or error.
• Later read-back: Repository rows and rendered block can reach EPI-RTE-memory-recall.
• Delegated visibility: Embedding subsystem receives stored claim; foreground later sees selected rendered text, not write checks.
• Selection predicate: Nonempty content; redaction and tombstone rules; inferred non-explicit/non-manual claims pending, other claims active.
• Invalidation or expiry: Save replaces target where selected; capacity can archive; expiry stored. Exact repository transaction semantics uninspected.
• Activation or effect: Active/pending retention and block rebuild wired; best-effort embedding does not gate save.
• Evidence limits: Model flag can misclassify an inference; policy checks do not evaluate truth.

Trigger: explicit/manual input or extracted decision. Input supplier proposes; code decides status and can reject; not per-content human approval on extraction. No expected-answer oracle. Admission criteria are policy/usability, intended use later recall over the user subject. Pending claims are retained but withheld from active-query consumers until confirmation. Checks do not prove source truth. Failed extraction writes can skip selected policy errors and preserve other decisions.

func statusForWrite(item types.MemoryItem) string { if item.Inferred && item.Origin != types.MemoryOriginExplicit && item.Origin != types.MemoryOriginManual { return types.MemoryStatusPending } return types.MemoryStatusActive } --- internal/application/service/memory/service.go

if forgotten { logger.Infof(ctx, "memory: skipped a statement the user previously deleted") return nil, ErrPreviouslyForgotten } --- internal/application/service/memory/service.go

EPI-RTE-memory-review

Label: Memory confirmation and rejection

Distinct subsequent caller decision, not extraction or RT-RTE-mcp-call effect approval. Owner: memory Service behind POST confirm/reject endpoints in SRC-1, internal/router/routes_memory.go. Progression: scoped lookup → confirm pending/rebuild or DeleteItem/tombstone → later recall/write eligibility.

SRC-1 implementation, internal/application/service/memory/service.go.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Confirmed item or deletion/error.
• Later read-back: Confirmation activates stored claim; rejection removes row and records forgotten fingerprint/source for future write checks.
• Delegated visibility: inapplicable — no model is called for confirmation/rejection; later recall is separate.
• Selection predicate: Resolved caller scope and existing item; confirmation calls repository ConfirmPendingItem.
• Invalidation or expiry: Delete adds tombstone before removal; tombstone failure logs warning but deletion proceeds.
• Activation or effect: Status change/removal shapes later eligibility, not observed human deliberation or changed answer.
• Evidence limits: Principal may be software; service does not retain evidence consumed or named truth criterion. Repository pending-state preconditions not fully inspected.

Caller chooses confirm/reject, code performs it. Open memory-manager use, no expected-answer oracle. Rejection has operative force, but deletion with failed tombstone prevents an absolute no-reappearance guarantee. Human endorsement is possible, not established by endpoint invocation.

if err := s.repo.ConfirmPendingItem(ctx, scope, id); err != nil { return nil, err } --- internal/application/service/memory/service.go

if err := s.repo.AddTombstone( ctx, scope, existing.Topic, types.MemoryFingerprint(existing.Content), existing.SourceMessageID, ); err != nil { logger.Warnf(ctx, "memory: record tombstone failed: %v", err) } --- internal/application/service/memory/service.go

EPI-RTE-memory-recall

Label: Memory prompt and search consumption

Distinct stored-claim selection route supplying RT-RTE-agent-turn and RT-OBJ-system-prompt. Owner: memory Service. Progression: enabled scope → active resident/situational queries → rank/budget/render → prompt or search result → foreground model; supplied boundary establishes caller SetMemoryPrompt.

SRC-1 implementation, internal/application/service/memory/service.go.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Recall.Prompt/Items; SearchMemory result in SRC-1, internal/application/service/memory/search.go.
• Later read-back: Later invocations select retained active notes; touchAsync changes usage metadata.
• Delegated visibility: Foreground receives rendered selected notes, not complete memory repository, tombstones or pending list.
• Selection predicate: Resident interests query-selected; active fact/task candidates and lexical/semantic selection with item/rune budgets.
• Invalidation or expiry: Superseded/archived items excluded from inspected active queries; block cache fallback exists on empty rendering. Full failure-path cache invalidation uninspected.
• Activation or effect: Delivered memory has advisory prompt/tool-result force; actual obedience/benefit unobserved.
• Evidence limits: Ranking checks relevance not truth; vector internals and every expiry/cache branch uninspected.

Open queries with no reference answer. Code chooses delivery; model may request SearchMemory. Normal active selection withholds pending claims; cache fallback bounds any universal withdrawal claim. Usage updates establish access bookkeeping, not an evaluation reward for correct beliefs.

situational, err := s.repo.ListActiveByKinds(recallCtx, scope, []string{types.MemoryKindFact, types.MemoryKindTask}, lexicalPoolSize(cfg)) --- internal/application/service/memory/service.go

prompt := types.WrapMemoryForPrompt(block, types.RenderMemoryRecall(matched)) --- internal/application/service/memory/service.go

EPI-RTE-wiki-rewrite

Label: Wiki page generation and retraction

Distinct from RT-RTE-knowledge-admission indexed-document enablement and RT-RTE-agent-turn foreground loop. Owner: wikiIngestService/wiki page service. Progression: additions/retracts plus existing page → cited chunks/summary and source bookkeeping → template model call → decode/body split → guarded CreatePage/UpdatePage.

SRC-1 implementation, internal/application/service/wiki_ingest_batch.go.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Updated page/change flags, deferred write or generation failure.
• Later read-back: Generated body and document/chunk refs persist; registered wiki read/search tools consume pages later (SRC-1, internal/application/service/agent_service.go).
• Delegated visibility: Editor gets existing page, source contexts, new/deleted/remaining source content and available link handles; summary context is not equivalent to every raw source passage.
• Selection predicate: Entity/concept additions or retract/retractStale; summary branch stores document synopsis separately.
• Invalidation or expiry: Retraction removes deleted-source refs and requests content removal; generation errors can leave current content. Guard can defer.
• Activation or effect: Body/ref changes are implemented, not observed correct source-preserving synthesis.
• Evidence limits: Actual generated propositions unobserved; partial source context and merged refs cannot prove each sentence grounded.

Worker/model proposes change; code can reject generation or guarded write. Guidance: WikiPageModifyUserPrompt, workspace ContentInstructions scoped to wiki_content and source blocks. No expected page/answer oracle. Open source-driven maintenance. Retention comes after structural admission, not epistemic acceptance. Intentional retractions pass WithWikiShrinkAllowed. Recovery/retry budget details beyond deferred result remain supplied/uninspected.

updatedContent, err = s.generateWithTemplate(ctx, chatModel, agent.WikiPageModifyUserPrompt, map[string]string{ "HasAdditions": hasAdditionsStr, "HasRetractions": hasRetractionsStr, --- internal/application/service/wiki_ingest_batch.go

page.SourceRefs = newRefs --- internal/application/service/wiki_ingest_batch.go

page.ChunkRefs = mergeChunkRefs(page.ChunkRefs, additions) --- internal/application/service/wiki_ingest_batch.go

EPI-RTE-wiki-guard

Label: Wiki rewrite preservation gate

Distinct check/disposition at page write, compared with RT-RTE-skill-install: both can veto changes but this checks stored table row identity loss, not dependency availability. Owner: wikiPageService.UpdatePage → explicit error consumed by ingest and agent writer.

SRC-1 implementation, internal/application/service/wiki_page.go.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Write refusal with ErrWikiWriteDroppedTableRows, or normal update.
• Later read-back: Refusal preserves prior body; ingest deferred result retains documents for later batch. Revisions support later content revert.
• Delegated visibility: Caller/model can receive refused-write explanation; no separate truth evaluator is invoked.
• Selection predicate: Machine edit source without deliberate shrink flag; missing identities and source-specific loss tolerance.
• Invalidation or expiry: User/revert/intentional-shrink exemptions; identity preservation does not constrain all cell changes.
• Activation or effect: Gate blocks a bounded set of rewrites and permits others; operation unobserved.
• Evidence limits: Internal row identity helper/tolerance values not fully inspected; no guarantee of full table semantics.

Worker/agent supplies candidate, code can veto. Existing content is the preservation reference, not an independently warranted answer oracle. Ingest classifyWikiPageWrite consumes the sentinel as deferred, not successful version change. Revert restores Title/Content/Summary/PageType/Status/Aliases, retaining current placement/provenance. This is bounded recovery, not every-change rollback.

source := types.WikiEditSourceFromContext(ctx) if source != types.WikiEditSourceUser && source != types.WikiEditSourceRevert && !types.WikiShrinkAllowedFromContext(ctx) { --- internal/application/service/wiki_page.go

return nil, fmt.Errorf( "%w: slug %s lost %d of %d table row identities (e.g. %s); the stored page is unchanged", ErrWikiWriteDroppedTableRows, page.Slug, len(missing), storedIdentities, examples) --- internal/application/service/wiki_page.go

if errors.Is(updateErr, ErrWikiWriteDroppedTableRows) { return true, nil } --- internal/application/service/wiki_ingest_batch.go

target.Content = rev.Content --- internal/application/service/wiki_page.go

EPI-RTE-wiki-maintenance

Label: Wiki structural lint and cleanup

No supplied counterpart; separate structural maintenance alongside EPI-RTE-wiki-rewrite. Owner: WikiLintService. Progression: live-slug/source/content probes → issues → AutoFix → page edits/archive/delete → RebuildLinks.

SRC-1 implementation, internal/application/service/wiki_lint.go.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Lint report or fix count/error.
• Later read-back: Persistent fixes shape wiki later availability and references; no retained acceptance rationale for factual claims established.
• Delegated visibility: Operator gets report; AutoFix program consumes fixable issue type and target.
• Selection predicate: Broken-link, short-content and stale-source conditions; not every lint issue is automatically fixable.
• Invalidation or expiry: Stale source can delete sole-source page or remove metadata when other sources remain; short pages archived except index.
• Activation or effect: Code-generated structural results have direct maintenance consumers; factual evaluation not implied.
• Evidence limits: Lint “health” is structural; source lookup errors can classify refs stale; partial stale-ref cleanup need not remove unsupported prose.

Caller triggers maintenance; code proposes/decides fixable updates and may skip failed pages. No reference factual answer. Concrete implementation deduction without target execution: with TargetSlug ghost and Content "See [[ghost]].", ReplaceAll produces "See ghost." and UpdateAutoLinkedContent persists on success. With "See [[ghost|label]].", that exact replacement leaves text unchanged, though a successful write can increment fixed. Separate stripDeadWikiLinks branch in SRC-1, internal/application/service/wiki_ingest.go, handles display labels and fuzzy resolution, whose helper equivalence is uninspected. Cleanup preserves prose in the exact plain-text branch, but withdrawing stale references changes lineage. RebuildLinks follows positive fix count; no universal semantic repair follows.

target := issue.TargetSlug page.Content = strings.ReplaceAll(page.Content, "[["+target+"]]", target) if err := s.wikiService.UpdateAutoLinkedContent(ctx, page); err == nil { fixed++ } --- internal/application/service/wiki_lint.go

if len(remaining) == 0 { if err := s.wikiService.DeletePage(ctx, kbID, page.Slug); err == nil { fixed++ } } else if len(remaining) != len(page.SourceRefs) { page.SourceRefs = remaining --- internal/application/service/wiki_lint.go

EPI-RTE-dataset-metrics

Label: Dataset reference metric calculation

Material metric-computation part of supplied evaluation route, not a duplicate of its whole request/KB/pipeline lifecycle. Owner: HookMetric/MetricList; progression: QA pair + selected retrieved content + response → passage-ID mapping → calculators → task Metric result.

SRC-1 implementation, internal/application/service/metric_hook.go.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Per-pair scores and mean metric vector.
• Later read-back: In-process result lookup described in parent route; durable cross-restart retention unestablished.
• Delegated visibility: Calculators receive mapped IDs, reference IDs, generated text and reference answer; no new judging model call here.
• Selection predicate: Prefer rerank result when nonempty, otherwise search; skip empty chunks/passages; first substring containment match maps PID and duplicate IDs collapse.
• Invalidation or expiry: uninspected — parent in-process task lifetime/cleanup not fully traced; immutable references do not establish their truth.
• Activation or effect: Computed numbers exposed through evaluation.go task update; no automatic successor admission in inspected consumer.
• Evidence limits: Unmatched result disappears before metric denominator. Remaining calculator formulas/reference fidelity uninspected; no runs or causal experiments.

Part of: RT-RTE-dataset-evaluation

Dataset provider supplies oracle QA pairs/PIDs/answer; code selects and computes, caller interprets. Bounded dataset mode. Precision domain is mapped PIDs, not all original retrievals. Implementation deduction: given one exact relevant chunk plus one unrelated unmatched chunk, mapping retains only the relevant PID; PrecisionMetric divides one hit by one mapped ID, yielding 1, not precision over both returned chunks. This follows inspected mapping and denominator, not an executed test. BLEU compares normalized text n-grams; semantic truth is beyond its licensed domain. No evidence-consuming acceptance/automatic optimization trigger in exposed task-metric consumer.

if strings.Contains(passage, r.Content) || strings.Contains(r.Content, passage) { pid := qaPair.PIDs[i] --- internal/application/service/metric_hook.go

totalPrecision += float64(hits) / float64(len(ids)) --- internal/application/service/metric/precision.go

candidate := splitIntoWords(splitSentences(metricInput.GeneratedTexts)) references := []Sentence{splitIntoWords(splitSentences(metricInput.GeneratedGT))} --- internal/application/service/metric/bleu.go

params.Metric = metricResult params.Task.Finished = finished --- internal/application/service/evaluation.go

Claims

EPI-CLM-cited-answers

Label: Source-cited quick QA

Claimed operation: README claims quick QA answers from knowledge base with cited sources. Distinct from RT-CLM-evidence-verification exact-evidence instruction; this is a public answer-consumption claim. SRC-1 doctrine/design, README.md. Conclusion status: claimed.

Two ways to ask. Quick Q&A answers from the knowledge base with RAG and cites the sources it used. In smart reasoning the agent plans multi-step work, searching, reading documents and calling tools and skills, and shows each step in the conversation. Docs → --- README.md

EPI-CLM-automatic-wiki

Label: Automatic wiki and rollback

Claimed operation: README claims source-cited generated pages, editable with every change reversible. No supplied claim counterpart; distinct from foreground exact-evidence policy. SRC-1 doctrine/design, README.md. Conclusion status: claimed.

Documents organized into a browsable wiki. With Wiki enabled, WeKnora extracts people, products and concepts from knowledge-base documents into pages with source citations, organized by directory. The knowledge graph shows how pages relate; pages can be edited directly and every change can be rolled back. Docs → --- README.md

EPI-CLM-end-evaluation

Label: End-to-end retrieval evaluation

Claimed operation: README claims end-to-end evaluation with recall and BLEU/ROUGE. No supplied claim counterpart; RT-RTE-dataset-evaluation is implementation, not the public claim. SRC-1 doctrine/design, README.md. Conclusion status: claimed.

| Retrieval | Keyword + vector hybrid search, rerank, parent-child chunking and GraphRAG (Neo4j); end-to-end evaluation with recall and BLEU / ROUGE | --- README.md

Evidenced absences

none declared in this member

Behavioral-authority paths

none declared in this member; route records name their consumers/channels/force/horizons and the supplied runtime owns RT-BAP-prompt-instructions and RT-BAP-skill-image.