MindBase epistemic report

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

Source-and-claim boundary

The question is how MindBase acquires or produces truth-apt content, checks it, grants reliance and makes it consequential. The functional inclusions, external host/provider exclusions and frozen SRC-1 identity are those in the boundary Source register. This is a complete repository artifact with a partially traced operational loop, not exhaustive evidence of every shipped route. No target code, tests, examples or providers were executed. All new implementation findings below concern SRC-1 at the reviewed commit; there is no registered observed run or causal experiment.

Public claims assessed are RT-CLM-approved-changes, EPI-CLM-wiki-audit, EPI-CLM-sentence-pruning and EPI-CLM-knowledge-compounds. The session hook's designation of context as curated truth is a reliance label, not an independently checked result: see RT-BAP-session-context.

Coverage distinguishes direct source inspection in this job from provisional supplied findings. Paths in this table are SRC-1 commit-relative anchors, not new sources.

Entry point or operation	Source path	Coverage and records	Limit or conclusion prevented
Ops contribute/build/research/ask	`apps/server/src/ops/runner.ts`, `apps/server/src/ops/recipes/build.ts`, `apps/server/src/ops/recipes/research.ts`, `apps/server/src/ops/gather.ts`, `schema/query.md`	Supplied RT-RTE-contribute-plan, RT-RTE-build-context, RT-RTE-research-page, RT-RTE-ask-answer; direct recipe/gather reads and selected runner ranges	Individual output truth and model compliance unobserved; ask implementation itself used as supplied evidence
Ops lint quote gate	`apps/server/src/ops/runner.ts`, `apps/server/src/ops/gather.ts`	Direct verifyEvidence and runLint reads; EPI-RTE-evidence-quotes, part of RT-RTE-lint-findings	Occurrence does not establish conflict; input caps defeat a whole-wiki guarantee
CLI dispatch, ingest and lint	`apps/cli/src/index.ts`, `apps/cli/src/commands/ingest.ts`, `apps/cli/src/commands/lint.ts`, `packages/core/src/wiki/lint.ts`	Direct dispatch/caller reads; EPI-RTE-cli-admission	CLI query/init/rebuild implementations not traced; compileL1 inspected only through its opening definitions, not its full planning/execution loop
MCP registered tools and health	`apps/mcp/src/tools/index.ts`, `apps/mcp/src/tools/run-wiki-health.ts`, `apps/mcp/src/tools/validate-structure.ts`	Direct registry, health and structure-tool reads; correction rechecked exact-name registration/dispatch and curator grants; EPI-RTE-health-maintenance	Registry enumeration is not implementation inspection of every tool
Server legacy health	`apps/server/src/routes/lint.ts`, `packages/core/src/compile/l2.ts`, `packages/core/src/compile/l2_prompts.ts`, `packages/core/src/tools/executor.ts`, `packages/core/src/graph/crosslinker.ts`, `schema/lint.md`	Direct health/L2/crosslink reads and selected executor methods; EPI-RTE-health-maintenance	Layout/store compatibility and complete behavior of every executor method unobserved
Topic synthesis and contradiction tool	`apps/server/src/lib/synthesis.ts`, `apps/server/src/routes/synthesize.ts`, `apps/mcp/src/lib/synthesis-runner.ts`, `apps/mcp/src/tools/find-contradictions.ts`, `packages/core/src/synthesis/prompts.ts`, `packages/core/src/synthesis/validate.ts`	Direct reads; EPI-RTE-topic-synthesis	Semantic checks weaker than sentence-support promise; other MCP synthesis callers not individually traced
Synthesis retention/recovery	`apps/server/src/lib/synthesis-cache.ts`, `apps/server/src/lib/synthesis-worker.ts`	Direct reads; EPI-RTE-synthesis-cache	Mark-stale producers not exhaustively traced; no universal freshness guarantee
Contradiction probe and consumer	`packages/core/src/compile/contradiction-probe.ts`, `packages/core/src/analysis/cache.ts`, `packages/core/src/analysis/insights.ts`, `packages/core/src/analysis/suggestions.ts`, `apps/server/src/lib/analysis-scheduler.ts`, `apps/server/src/routes/analysis.ts`	Direct reads; EPI-RTE-contradiction-probe, EPI-RTE-probe-consumption	Cue-limited selection and clipped legacy note bodies; no measured accuracy or complete cache-invalidation audit
Plugin commands, agents and session hook	`apps/plugin/hooks/hooks.json`, `apps/plugin/agents/builder.md`, `apps/plugin/agents/curator.md`, `apps/plugin/hooks/session-start`	Direct hook registration and agent-text reads; session-start body via supplied RT-BAP-session-context; RT-RTE-plugin-build	External host obedience and actual context isolation unobserved; curator grant names do not match registered health-tool names, so curator-to-health reachability is unestablished
Other material operations	`apps/server/src/lib/capture-worker.ts`, `apps/server/src/lib/classify-worker.ts`, `packages/core/src/brief/build.ts`, `packages/core/src/srs/extract.ts`, `packages/core/src/compile/l1.ts`, `apps/server/src/routes/answer-file-back.ts`	Unassessed here except limited CLI/L1 opening inspection; no inherited exhaustive finding	Prevent complete enumeration of acquisition, evaluation, acceptance, file-back and policy-adaptation routes
Other exposed paths and evaluators	`apps/mcp/src/tools/ingest-execute.ts`, `apps/mcp/src/prompts/index.ts`, `apps/mcp/src/resources/index.ts`, `apps/server/src/routes/tree/crud.ts`, `apps/server/src/routes/schema.ts`, `docs/testing-walkthrough.md`, `docs/manual-smoke-checklist.md`	Enumerated or supplied as remaining scope; not directly assessed here. Mobile/browser capture, other plugin agents, tests and CI also unassessed	Prevent system-complete negatives, uniform approval or source-warrant assertions, and evaluation-performance conclusions

Requested answer filing (apps/server/src/routes/answer-file-back.ts, packages/core/src/file-back/build-plan.ts), editable schema instructions (apps/server/src/routes/schema.ts) and retained classification guidance (apps/server/src/lib/classify-worker.ts, packages/core/src/classify/samples.ts, packages/core/src/classify/classify-note.ts, packages/core/src/classify/prompt.ts) remain outside this member's declared route inventory. No complete filing/admission, instruction-edit, working-context, retained-authority or policy-adaptation comparison is claimed.

Missing user corpora prevent identifying actual propositions and their original warrant. Missing execution traces prevent identifying which implemented routes produced or checked any real retained page. Missing controlled comparisons prevent attributing improvement, compounding knowledge or model accuracy. Provider weights and host internals are excluded external responsibilities, not evidence of their absence. Supplied runtime findings remain valid scoped accounts; the health pipeline below adds a consequential route omitted there, rather than superseding the ops-lint finding.

Epistemic-object inventory

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
EPI-OBJ-source-text	External/user assertions; actual propositions not determinable without corpus	Inputs	see record; SRC-1 `apps/server/src/ops/gather.ts`	Source identity labels are not source authentication
EPI-OBJ-wiki-pages	Summaries, decisions, learnings, research claims and answers	Maintained knowledge	see record; SRC-1 `apps/server/src/ops/recipes/build.ts`	A truth label or write does not prove acceptance
RT-OBJ-actions	None in command envelope; Markdown is a separately interpreted payload	Execute proposed changes	see record; SRC-1 `apps/server/src/ops/types.ts`	Schema admits command shape, not claim truth
EPI-OBJ-synthesis-prose	Topic summary, threads, headings, contradiction excerpts and gap suggestions	Synthesis and diagnosis	see record; SRC-1 `packages/core/src/synthesis/validate.ts`	Entailment and excerpt occurrence unverified
EPI-OBJ-citation-fields	Symbolic pointers to possible support, not the support relation itself	Lineage/evidence hints	see record; SRC-1 `packages/core/src/synthesis/validate.ts`	Valid line coordinates need not support adjacent prose
EPI-OBJ-probe-reason	Model's stated reason that two pages collide or agree	Factual collision diagnosis	see record; SRC-1 `packages/core/src/compile/contradiction-probe.ts`	A verdict can instead encode missing page or parse/provider error
EPI-OBJ-probe-verdict	Enum encoding a page-pair relation; error fallback is not a substantive judgment	Skip/review selector	see record; SRC-1 `packages/core/src/analysis/cache.ts`	ListConfirmed selects by enum, not reason quality
EPI-OBJ-lint-prose	Claimed contradiction, stale claim, structural issue or gap	Review artifact prose	see record; SRC-1 `apps/server/src/ops/runner.ts`	Diagnosis not established by occurrence
EPI-OBJ-lint-controls	None beyond encoded kinds and occurrence results	Finding admission/display controls	see record; SRC-1 `apps/server/src/ops/runner.ts`	Boolean verified concerns quote occurrence only
EPI-OBJ-compile-plan	Proposed natural-language edits inside symbolic tool calls	CLI proposal and execution input	see record; SRC-1 `apps/cli/src/commands/ingest.ts`	Approval governs calls, not proposition truth; core internals only partially inspected

Authority-route ledger

This ledger splits content production, checks, disposition and retention even where one canonical route implements several functions. Architectural status here concerns inspected capacity, not observed operation. Other members own activation predicates, return/read-back details and authority paths where a cell says see. No row asserts observed acceptance or successful downstream use.

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-contribute-plan, RT-RTE-research-page	content transformation	implemented	EPI-OBJ-source-text	truth-apt transformation: acquisition/import	Caller contribution or optional web text enters model-visible evidence	Caller/import and optional fetch; no source-authentication check established	see routes	Captured source or fetched/snippet text	Supplies candidate evidence	Source warrant unknown; labels retained but truth/authenticity not established	Can condition proposal generation	see routes	SRC-1 `apps/server/src/ops/runner.ts`; see routes	none	Acquisition is not new knowledge production	Source/capture extraction branches not exhaustively assessed
RT-RTE-contribute-plan	content transformation	implemented	EPI-OBJ-source-text, EPI-OBJ-wiki-pages	truth-apt transformation: indeterminate	Source to candidate takeaways/Markdown	Model; source excerpts and project guidance	see route	Candidate plan	see route	Candidate synthesis only; preservation versus derivation versus addition unresolved	Propose actions	see route	SRC-1 `apps/server/src/ops/recipes/contribute.ts`; see route	RT-CLM-approved-changes	none	No generated instance to classify
RT-RTE-contribute-plan	disposition/acceptance	implemented	RT-OBJ-actions	no content change	Proposed wiki actions	Schema, slug guards and caller indexes; UI offers human veto	see route	Selected or abandoned plan	Enforcing filtering	Operational admission, not a recorded evidence-consuming truth decision	Selected actions can write	see route	SRC-1 `apps/server/src/ops/runner.ts`; see route	RT-CLM-approved-changes	Approval scope limited	Human rationale/criterion unobserved
RT-RTE-build-context, RT-RTE-research-page	content transformation	implemented	EPI-OBJ-wiki-pages	truth-apt transformation: indeterminate	Existing context and selected sources to Markdown	Model instructed to preserve truth/support; no independent semantic verdict	see routes	Rewritten context or research page	Candidate production	Acquired warrant unknown; semantic preservation not demonstrated	Generate executor payloads	see routes	SRC-1 `apps/server/src/ops/recipes/build.ts`, `apps/server/src/ops/recipes/research.ts`	EPI-CLM-knowledge-compounds	Truth not checked	Could reshape, derive or conjecture; actual outputs required
RT-RTE-build-context, RT-RTE-research-page, RT-RTE-plugin-build	retention	implemented	EPI-OBJ-wiki-pages	no content change	Generated content to persisted files	Software shape/path/line constraints; hosted synthesis external	see routes	File writes or partial failures	Enforcing I/O constraints	Retention is not acceptance or post-acceptance integration	Future builds/questions/sessions can consume	see routes and RT-BAP-session-context	SRC-1 `apps/server/src/ops/executors.ts`; see routes	RT-CLM-approved-changes	Direct-write alternatives	MCP writer implemented; hosted build sequence only afforded
RT-RTE-ask-answer	operational admission/selection/consumption	implemented	EPI-OBJ-wiki-pages	no content change	Retained content to answer model	Ranked selection and model file requests	see route	Delivered source blocks and answer	Advisory content	Retrieval licenses availability, not factual endorsement	Model can answer and mark new knowledge for auto-save	see route	SRC-1 `packages/core/src/qa/query.ts`; see route	EPI-CLM-knowledge-compounds	none	Citation instructions are not sentence checks
RT-RTE-ask-answer	content transformation	implemented	EPI-OBJ-wiki-pages	truth-apt transformation: indeterminate	Retrieved sources and question to answer	Model under citation and analysis guidance	see route	Answer, optional new-knowledge marker	Candidate generation	No checked entailment or novelty decision	Supplies user response and potential auto-save input	see route	SRC-1 `schema/query.md`; see route	EPI-CLM-knowledge-compounds	Marker does not prove novelty	No actual answer/source pair supplied
RT-RTE-ask-answer	retention	implemented	EPI-OBJ-wiki-pages	no content change	Answer marker to stored note	autoSave configuration and regex, not evidence comparator	see route	Saved answer or error	Permissive marker then write	No novel-claim acceptance established	Later retrieval can amplify unvalidated answer	see route	SRC-1 `apps/server/src/routes/ask.ts`; see route	RT-CLM-approved-changes	New-knowledge label not warrant	No observed saved instance
EPI-RTE-topic-synthesis	content transformation	implemented	EPI-OBJ-synthesis-prose	truth-apt transformation: indeterminate	Retrieved notes to synthesis	Model receives numbered source lines	see record	Structured candidate prose	Candidate generation	Requested source-bounded synthesis, not proven reshaping or entailment	Supplies validator input	see record	SRC-1 `apps/server/src/lib/synthesis.ts`	EPI-CLM-sentence-pruning	none	No output instance or semantic check
EPI-RTE-topic-synthesis	check/evidence production	implemented	EPI-OBJ-citation-fields	no content change	Citation slug and line range; optional inline marker syntax	Deterministic coordinates and regex; not entailment domain	see record	Valid/invalid coordinates and pruned content	Enforcing limited filter	Validity of pointers only	Feeds retained/displayed result	see record	SRC-1 `packages/core/src/synthesis/validate.ts`	EPI-CLM-sentence-pruning	Conditional pruning	Concrete counterexample retained below
EPI-RTE-topic-synthesis	disposition/acceptance	implemented	EPI-OBJ-synthesis-prose	no content change	Thread admission and contradiction/gap filtering	Code requires nonempty pruned content and at least one valid citation; other fields have weaker gates	see record	Kept/dropped thread; summary survives separately	Enforcing output filtering	Not evidence-consuming acceptance of the prose's truth	Admitted fields returned for caching/UI/tool consumers	see record	SRC-1 `packages/core/src/synthesis/validate.ts`	EPI-CLM-sentence-pruning	Unsupported prose can survive	No accepted ampliative claim established
EPI-RTE-synthesis-cache	retention	implemented	EPI-OBJ-synthesis-prose, EPI-OBJ-citation-fields	no content change	Result to topic cache	Code after generation; no human per-result veto	see record	JSON persists	Enforcing write	Cached candidate, not accepted knowledge	Later endpoint/tool can reuse	see record	SRC-1 `apps/server/src/lib/synthesis-cache.ts`	none	none	Server and MCP cache implementations not presumed identical
EPI-RTE-synthesis-cache	lineage/freshness/recovery	implemented	EPI-OBJ-citation-fields	no content change	Source hashes, reverse index and stale set	Code regeneration; cache read itself does not check hashes	see record	Regenerate or clear failed stale job	Scheduling and reuse	Freshness metadata, not truth validation	Cached result may be served pending refresh	see record	SRC-1 `apps/server/src/lib/synthesis-worker.ts`	none	Refresh not endorsement	Stale producers not fully audited
EPI-RTE-evidence-quotes	check/evidence production	implemented	EPI-OBJ-lint-findings	no content change	Claimed quote occurrence in files	Normalized substring check	see record	Per-item verified booleans	Produces gate inputs	Occurrence only; no logical collision license	Makes findings eligible for retention	see record	SRC-1 `apps/server/src/ops/runner.ts`	EPI-CLM-wiki-audit	Count is not semantic proof	Duplicate evidence not made independent
EPI-RTE-evidence-quotes	disposition/acceptance	implemented	EPI-OBJ-lint-findings	no content change	Contradiction/stale finding retention	Count at least two/one verified items respectively	see record	Kept or dropped finding	Enforcing diagnosis filter	Accepts occurrence threshold for review, not truth of conflict	Parent stores eligible findings; human may dismiss	see record and RT-RTE-lint-findings	SRC-1 `apps/server/src/ops/runner.ts`	EPI-CLM-wiki-audit	Exactness qualified	No checked distinct-page requirement
EPI-RTE-contradiction-probe	check/evidence production	implemented	EPI-OBJ-probe-reason, EPI-OBJ-probe-verdict	truth-apt transformation: indeterminate	Specific factual collision between two page excerpts	Model judge with defined three-way rubric	see record	contradicts/compatible/unrelated, including error fallback	Diagnostic output	Model opinion within excerpt domain; not independent ground truth	Supplies cached verdict	see record	SRC-1 `packages/core/src/compile/contradiction-probe.ts`	EPI-CLM-wiki-audit	Errors collapse to unrelated	No expected-answer oracle or measured calibration
EPI-RTE-contradiction-probe	retention	implemented	EPI-OBJ-probe-verdict, EPI-OBJ-probe-reason	no content change	Verdict to SQL cache	Any returned enum, including fallback	see record	Stored result suppresses matching future judge call	Enforcing cache skip	Caching grants no added warrant	Matching pair/model/prompt reused	see record	SRC-1 `packages/core/src/analysis/cache.ts`	none	No body hash in lookup	Other invalidation paths not exhaustively assessed
EPI-RTE-probe-consumption	operational admission/selection/consumption	implemented	EPI-OBJ-probe-verdict	no content change	Cached enum to high-severity review suggestion	SQL verdict filter and fixed suggestion rule	see record	Contradiction flagged for user review	Ranking/advisory, not automatic factual resolution	UI category is stronger wording than model-only warrant	Prioritizes review suggestions, no compelled correction	see record	SRC-1 `packages/core/src/analysis/insights.ts`	EPI-CLM-wiki-audit	Confirmed is a cache label	Can include other model/prompt rows; no current-key filter in listConfirmed
EPI-RTE-health-maintenance	content transformation	implemented	EPI-OBJ-wiki-pages	truth-apt transformation: indeterminate	Health overview to model-proposed text repairs	L2 model sees summaries/index/structural issues, not every full page	see record	Rewrite, append, summary or other supported action	Direct executor dispatch	Candidate repair, no demonstrated semantic improvement	Changes files subject to handler checks	see record	SRC-1 `packages/core/src/compile/l2.ts`	EPI-CLM-knowledge-compounds	Health API writes; curator mapping unresolved	Proposed content may replace rather than preserve meaning
EPI-RTE-health-maintenance	operational admission/selection/consumption	implemented	EPI-OBJ-wiki-pages	no content change	Parsed L2 actions to handlers	Code dispatch; human_touched veto for inspected rewrite/update handlers	see record	Writes, handler rejection, or return-only flags/merge	Enforcing code checks, no semantic approval	Neither flag success nor queued label establishes resolved knowledge	Later consumers see successful writes, not a validated repair certificate	see record	SRC-1 `packages/core/src/tools/executor.ts`	RT-CLM-approved-changes	Registered API writes; curator grant differs	No pipeline transaction/rollback established
EPI-RTE-health-maintenance	content transformation	implemented	EPI-OBJ-wiki-pages	truth-apt transformation: non-ampliative reshaping	Exact-title mention to wikilink	Fixed title match and score threshold	see record	Existing words linked	Enforcing auto-mode edit	Pointer insertion only; identity of referent remains heuristic	Alters future navigable graph	see record	SRC-1 `packages/core/src/graph/crosslinker.ts`	none	Extracted does not mean semantically verified	Does not warrant title disambiguation
EPI-RTE-cli-admission	disposition/acceptance	implemented	EPI-OBJ-compile-plan	no content change	CLI compile plan to execution	Human all-or-nothing prompt or --yes bypass; distinct from ops envelopes	see record	Aborted or all actions forwarded	Enforcing caller checkpoint	Approval is permission, not evidence of truth	Calls compileL1Execute	see record	SRC-1 `apps/cli/src/commands/ingest.ts`	RT-CLM-approved-changes	--yes alternative	Core planning/execution loop not fully inspected

No post-acceptance lifecycle-integration route is established in these traces: operational writes, links and retrieval occur, but an evidence-consuming acceptance decision for a produced claim was not found in these inspected pipelines. This is a route-bounded finding, not a system-complete negative. No formal derivation route from warranted premises is established here. Failure to establish entailment does not classify all synthesis as ampliative conjecture.

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-approved-changes	see record	SRC-1 `README.md`; doctrine/design	see record	RT-RTE-contribute-plan, RT-RTE-build-context, RT-RTE-research-page, RT-RTE-ask-answer, EPI-RTE-health-maintenance, EPI-RTE-cli-admission	None supplied	No intervention/comparison	Selected UI path affords human selection; several alternate write paths do not require per-content approval	Universal wording unsupported; approval still would not establish warrant
EPI-CLM-wiki-audit	see record	SRC-1 `README.md`; doctrine/design	Recipes and curator request diagnosis	RT-RTE-lint-findings, EPI-RTE-evidence-quotes, EPI-RTE-contradiction-probe, EPI-RTE-probe-consumption	None supplied	Neither excerpt checks nor returned categories establish diagnostic accuracy	Implemented bounded auditing with quote-occurrence gate and separate model-judge review path	Whole-wiki/exact-conflict guarantee exceeds caps, normalization and count-only gate
EPI-CLM-sentence-pruning	see record	SRC-1 `packages/core/src/synthesis/prompts.ts`; doctrine/design	Explicit every-sentence support requirement	EPI-RTE-topic-synthesis	None supplied	Deterministic input trace disproves universal pruning; not a model performance test	Valid reference coordinates can accompany unsupported prose	Citation-free content is preserved if its thread has valid citation metadata
EPI-CLM-knowledge-compounds	see record	SRC-1 `README.md`; doctrine/design	Build requests preservation and contradiction flags; query asks for source-supported analysis	RT-RTE-build-context, RT-RTE-ask-answer, RT-BAP-session-context, EPI-RTE-synthesis-cache	None supplied	No quality baseline, longitudinal corpus or comparison	Persistent material has routes to later models; accumulation and reuse are implemented capacities	Progressively sharper/compounding knowledge remains unestablished, not refuted by mere absence of runs

Bounded conclusion

MindBase has multiple routes from user/source text to retained, behaviorally consequential prose. The inspected system imports sources, selects excerpts, requests synthesis and answers, writes context/research pages and can reload them in builds, questions and sessions. Source warrant is unknown: paths and citations preserve some lineage pointers but do not authenticate content or establish that transformations preserve meaning. See RT-RTE-contribute-plan, RT-RTE-build-context, RT-RTE-research-page, RT-RTE-ask-answer and RT-BAP-session-context.

The implemented checks have different licenses. JSON schemas license command execution; citation coordinates license valid pointers; normalized quote matches license occurrence claims; a model judge supplies a bounded opinion about a pair of excerpts. None establishes a source-backed accepted ampliative proposition merely by retaining or displaying its result. The quote gate can reject fabricated quotations and the synthesis validator can reject invalid line coordinates: these are real checks, not grounds for treating every retained statement as unverified in every respect. Their missing semantic entailment checks remain consequential. See EPI-RTE-topic-synthesis and EPI-RTE-evidence-quotes.

Operational authority is stronger than epistemic warrant in several places. Build/research and marker-triggered answer saving can persist candidates without a content-specific human decision. The registered L2 health API can directly rewrite knowledge; handler checks can veto some edits, while flag/merge success can be return-only. The curator grants and requests mindbase_run_wiki_health, whereas the MCP server registers run_wiki_health and dispatches by exact name. No inspected mapping connects those names, so the curator-to-writer path is unestablished; this mismatch is not a host-safety guarantee. Cached model verdicts influence which issues rank as high-priority review, without forcing a correction or proving a collision. These distinctions follow EPI-RTE-health-maintenance, EPI-RTE-probe-consumption and the supplied runtime routes, not an inference from the word knowledge.

No produced accepted ampliative claim, warranted formal derivation or post-acceptance integration instance is established within these inspected routes. Without actual source/output pairs, model semantic updates remain indeterminate between reshaping, entailed derivation and conjecture. The deterministic crosslink example is bounded non-ampliative reshaping, not new factual knowledge. Other included but unassessed mechanisms prevent a complete system negative.

Persistent updates and advisory delivery afford later behavior changes, not demonstrated learning. No improvement comparison establishes improved future capacity. The code does not update provider parameters through these inference routes; endpoint/version uncertainty remains as stated in RT-CMP-model. Auditing a wiki is not by itself reflection about the system's operating method. Model-owned proposals and software progression establish computational roles independently of correctness; human veto exists in some routes, not all. No exercised self-improvement attribution or autonomous theory-builder verdict follows from these findings. Source-native project rules guide future proposals through RT-BAP-project-rules, but content-directed criticism and iterative revision of those method rules were not established by this bounded trace.

Shared records

Operative objects

EPI-OBJ-source-text

Label: Contributor and retrieved source text

SRC-1, apps/server/src/ops/gather.ts, apps/server/src/ops/recipes/research.ts. Natural-language text on project files or supplied search results, temporarily selected into model inputs. Distinct from RT-OBJ-actions: it is input evidence, not a command envelope. Source propositions and warrant are corpus-dependent and not determinable here. Supplied RT-RTE-research-page covers acquisition of optional web results; this is not a new acquisition route declaration.

const MAX_SOURCE_FILES = 30; const MAX_SOURCE_CHARS = 4_000; --- apps/server/src/ops/gather.ts

const fresh = withM.filter((f) => f.mtime > contextMtime).sort((a, b) => b.mtime - a.mtime).slice(0, MAX_SOURCE_FILES); --- apps/server/src/ops/gather.ts

EPI-OBJ-wiki-pages

Label: Context and wiki page prose

SRC-1, apps/server/src/ops/recipes/build.ts, packages/core/src/tools/executor.ts. Natural-language context/research/legacy page bodies on project storage, interpreted by later models or humans. Distinct from RT-OBJ-actions, which identifies transient command envelopes, not the retained corpus; not a part of that envelope. Existing routes cover writes/read-back rather than declaring this corpus object. Actual claim truth is unobserved.

preserve still-true content; flag contradictions between sources explicitly with ⚠️. --- apps/server/src/ops/recipes/build.ts

await this.store.writeText(mdPath, # ${meta.title}\n\n${newContent}\n); --- packages/core/src/tools/executor.ts

EPI-OBJ-synthesis-prose

Label: Synthesis prose fields

SRC-1, packages/core/src/synthesis/validate.ts. Natural-language summary, thread headings/content, contradiction excerpts/explanations and gap suggestions inside transient structured results and cached JSON. Symbolic references have separate identity in EPI-OBJ-citation-fields. No counterpart among supplied runtime objects: these are topic-synthesis outputs, not ops action envelopes. Their validation criteria differ by field; summary is passed through, contradiction excerpt occurrence is not checked.

return { summary: raw.summary ?? '', threads, contradictions, gaps }; --- packages/core/src/synthesis/validate.ts

EPI-OBJ-citation-fields

Label: Synthesis citation coordinates

SRC-1, packages/core/src/synthesis/validate.ts, apps/server/src/lib/synthesis.ts. Symbolic slug/line_range fields, inline marker strings and source_hashes in parsed/cached JSON. Fixed code uses coordinates to filter and hashes as lineage metadata. Distinct from supplied RT-OBJ-actions: citation metadata rather than execution commands. These are not a warrant-bearing interpretation of source text.

const [start, end] = c.line_range; if (!Number.isInteger(start) || !Number.isInteger(end)) return false; if (start < 1 || end > lines.length || start > end) return false; return true;

--- packages/core/src/synthesis/validate.ts

EPI-OBJ-probe-reason

Label: Contradiction probe reason

SRC-1, packages/core/src/compile/contradiction-probe.ts, packages/core/src/analysis/cache.ts. Natural-language model explanation or program-created failure reason, transient then SQL text. Distinct from RT-RTE-lint-findings: a separate legacy-page judge pipeline. The sibling enum has a distinct fixed consumer in EPI-OBJ-probe-verdict.

catch { return { verdict: 'unrelated', reason: 'page A not found' }; } --- packages/core/src/compile/contradiction-probe.ts

EPI-OBJ-probe-verdict

Label: Contradiction probe verdict

SRC-1, packages/core/src/compile/contradiction-probe.ts, packages/core/src/analysis/cache.ts. Symbolic three-way enum in memory and SQLite contradiction_cache; fixed code selects contradicts for review and skips cached pairs. Distinct from ops lint findings and ops command envelopes. Its apparent factual meaning can instead arise from an error fallback.

export type Verdict = 'contradicts' | 'compatible' | 'unrelated'; --- packages/core/src/compile/contradiction-probe.ts

EPI-OBJ-lint-findings

Label: Stored ops lint findings

Part of: RT-RTE-lint-findings

The parent is a route; this object is its retained diagnostic artifact, not a competing route identity. SRC-1, apps/server/src/ops/runner.ts. Natural-language diagnoses and quote text have human/model interpretation; symbolic kinds, evidence verified flags and dismissed state drive filtering/display. Stored as dated JSON. These parts are distinguished by their consumers in EPI-RTE-evidence-quotes and the parent route: verified counts govern retention, dismissal governs later presentation, and prose communicates the proposed issue. No truth decision is encoded merely by verified=true.

const artifact: LintArtifact = { date, findings, dropped }; --- apps/server/src/ops/runner.ts

EPI-OBJ-lint-prose

Label: Ops lint diagnosis and quote prose

Part of: EPI-OBJ-lint-findings

SRC-1, apps/server/src/ops/runner.ts. Natural-language diagnosis text and evidence quotations within the parent's dated JSON artifact. Human/model interpretation supplies their meaning. Distinct part from fixed-rule controls: evidence quote strings are normalized and matched by code, but that match assesses occurrence, not the diagnosis. Supporting passage is the verifyEvidence quotation in EPI-RTE-evidence-quotes, which preserves page/quote fields alongside the computed flag.

EPI-OBJ-lint-controls

Label: Ops lint kinds and verification flags

Part of: EPI-OBJ-lint-findings

SRC-1, apps/server/src/ops/runner.ts. Symbolic kind, per-evidence verified flag, ID and dismissed state in the same dated JSON. Fixed code uses verified-item counts for admission and dismissal for later presentation, unlike the diagnosis prose's interpreted meaning. The parent RT-RTE-lint-findings supplies dismissal/read-back evidence; EPI-RTE-evidence-quotes supplies the exact flag/count passages. This is a part declaration, not a competing finding-retention route.

EPI-OBJ-compile-plan

Label: CLI compile proposal tool calls

SRC-1, apps/cli/src/commands/ingest.ts. Transient plan.proposed entries carry symbolic tool-call names and argument keys plus natural-language names/Markdown edit payloads. Fixed CLI code displays call names and titles and forwards the plan to execution after batch approval. The quoted caller establishes this mixed command object; it does not establish semantic checks on text arguments. Distinct from RT-OBJ-actions: separate core compiler plan used by CLI, not ops' Zod action union or pending-plan map. No persistence of the pending CLI plan is established.

for (const a of plan.proposed) { const args = a.call.arguments as Record; const title = (args['name'] ?? args['concept_name'] ?? args['from'] ?? '?') as string; out.info(${a.call.name} ${title}); }

--- apps/cli/src/commands/ingest.ts

Routes

EPI-RTE-topic-synthesis

Label: Topic synthesis and field admission

SRC-1, apps/server/src/lib/synthesis.ts, apps/mcp/src/lib/synthesis-runner.ts, packages/core/src/synthesis/prompts.ts, packages/core/src/synthesis/validate.ts. Server runSynthesis or MCP runSynthesisMCP selects up to twenty notes, builds a numbered-source prompt, collects model JSON and calls validateSynthesis. Server uses hybrid retrieval; MCP uses keyword retrieval. Both reconstruct legacy wiki/notes paths from hits, skipping unreadable notes; do not infer equivalent v2 coverage. Distinct from RT-RTE-research-page: structured topic-result pipeline, not ops research actions.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: Structured summary, threads, contradictions, gaps, model identity and source hashes; empty output when no notes or unparseable response.
Later read-back: Caller cache writes allow later topic/tool invocations to reuse the result; see the distinct cache route.
Delegated visibility: inapplicable — completion calls, not worker-agent delegation.
Selection predicate: Up to twenty search hits; only successfully loaded legacy note bodies enter the source map. A thread requires at least one valid citation and nonempty content after optional pruning.
Invalidation or expiry: inapplicable — generation/validation itself has no retained state; caller owns cache lifecycle.
Activation or effect: Model proposes prose; fixed code drops invalid coordinates and some threads/sentences; admitted result can be displayed or cached. Behavioral dependence is not observed.
Evidence limits: No executed synthesis or semantic support comparator; malformed shapes beyond the expected interfaces can throw. Source hashes identify bodies, not truth.

Admission mode is an open topic request or scheduled refresh. The model proposes; software vetoes structurally inadequate thread references, unknown contradiction slugs/confidence values and gap references. There is no per-thread human approval or expected-answer oracle. guidance requests source-bounded synthesis; no persisted criticism-to-theory loop is established. Parse failure returns empty fields; code does not independently decide whether a surviving claim follows from sources.

const validCitations = (t.citations ?? []).filter((c) => citationValid(c, sources)); if (validCitations.length === 0) continue; const prunedContent = pruneUncitedSentences(t.content ?? ''); if (!prunedContent) continue;

--- packages/core/src/synthesis/validate.ts

if (!INLINE_CITE_RE.test(content)) return content.trim(); --- packages/core/src/synthesis/validate.ts

if (!c.with_slug || !sources.has(c.with_slug)) continue; if (c.confidence !== 'low' && c.confidence !== 'medium' && c.confidence !== 'high') continue; --- packages/core/src/synthesis/validate.ts

Concrete implementation deduction, without execution: take source map a → ["The sensor is blue."] and a thread with heading Claim, content The moon is cheese., citations [{slug:"a",line_range:[1,1]}]. citationValid finds source a and integer range 1..1 within its one line. There is no inline marker, so pruneUncitedSentences returns the entire trimmed content. Both thread-admission tests pass and the unrelated sentence survives. The same path also preserves a genuinely supported citation-free sentence; lack of inline markers cannot alone determine falsity. This falsifies universal sentence pruning, not model reliability measured on a corpus. When markers do occur, regex presence rather than semantic support decides which sentence fragments survive.

EPI-RTE-synthesis-cache

Label: Server synthesis cache and refresh

SRC-1, apps/server/src/routes/synthesize.ts, apps/server/src/lib/synthesis-cache.ts, apps/server/src/lib/synthesis-worker.ts. Server owns JSON persistence, reverse source index, stale set, cached SSE replay and serial regeneration. No counterpart in supplied runtime routes: this is not ops research persistence or answer auto-save. MCP find_contradictions similarly calls its cache before generation, but the detailed server cache semantics here are not silently attributed to the separate MCP cache implementation.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: SSE meta, thread, gap and contradiction events, then done, or error. The MCP contradiction tool returns contradiction fields from cached/generated structured synthesis.
Later read-back: Non-forced server topic request directly replays cached fields; MCP find_contradictions returns cached contradictions when present.
Delegated visibility: inapplicable — server worker invokes runSynthesis serially, not an agent.
Selection predicate: Normalized topic key; reverse index maps source slugs to cached topics. Server cache hit unless force; worker selects first stale key.
Invalidation or expiry: markStaleFor adds reverse-index topics to a set; worker polls every five seconds. Cache read does not itself compare source hashes or exclude stale keys. Worker clears stale entry after success and after caught failure.
Activation or effect: Reuses or replaces stored candidate results and influences what a later requester sees; no observed accuracy improvement.
Evidence limits: Mutation callers of markStaleFor not exhaustively traced; no assertion that every source edit marks every affected topic. A cache hit is not a fresh evaluation.

Owner is software; model proposes replacement content during refresh. Rejection uses the generator's parsing/filtering behavior, not a human decision. Failed refresh clears its stale queue entry to avoid retries; no automatic semantic rollback or acceptance gate was found in these files.

const cached = await ctx.synthesisCache.readSynthesis(key); if (cached) { --- apps/server/src/routes/synthesize.ts

const result = await runSynthesis(this.ctx, key); await this.ctx.synthesisCache.writeSynthesis(key, result); await this.ctx.synthesisCache.clearStale(key); --- apps/server/src/lib/synthesis-worker.ts

console.warn([synthesis-worker] regen failed for ${key}:, (e as Error).message); await this.ctx.synthesisCache.clearStale(key);

--- apps/server/src/lib/synthesis-worker.ts

EPI-RTE-evidence-quotes

Label: Lint evidence quote verification

Part of: RT-RTE-lint-findings

SRC-1, apps/server/src/ops/runner.ts, apps/server/src/ops/gather.ts. This is the parent's specific check-and-filter subroute, not a duplicate declaration of its full diagnosis/retention loop. Code resolves model page labels to files, normalizes bodies and quotes, counts matching evidence items and drops insufficient contradiction/stale findings. The check target is quotation occurrence, not factual inconsistency. In particular two items need not refer to different pages or independent passages.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: Kept findings with per-item verified flags and a dropped count.
Later read-back: Parent stores those findings in dated artifacts; later fetch/dismiss is as in the parent route, not a new knowledge-revision consumer.
Delegated visibility: inapplicable — local code after a completion, not agent delegation.
Selection predicate: All proposed evidence entries; contradiction needs at least two verified entries, stale at least one; other model kinds do not have this count gate.
Invalidation or expiry: Parent overwrites same-day artifact; this gate reads current files with a per-call body cache, not a historical source snapshot.
Activation or effect: Can reject findings from retention; it does not change the underlying page, decide which claim is true or require two distinct pages.
Evidence limits: Normalization changes exactness; substring occurrence does not prove that the quote supports the diagnosis or that the model saw all relevant content.

The model proposes semantic findings; software can veto evidence-insufficient items; human dismissal is later and not required for admission. The reference is current page text, not an expected answer to the semantic question. No answer oracle or closed automatic correction follows from this gate.

const normalizeQuote = (s: string): string => s.toLowerCase().replace(/[*_`]/g, '').replace(/\s+/g, ' ').trim();

--- apps/server/src/ops/runner.ts

return { page: e.page, quote: e.quote, verified: haystack !== null && haystack.includes(normalizeQuote(e.quote)) }; --- apps/server/src/ops/runner.ts

const verifiedCount = evidence?.filter((e) => e.verified).length ?? 0; if (need !== undefined && verifiedCount < need) { dropped += 1; continue; } --- apps/server/src/ops/runner.ts

Concrete implementation deduction: for an existing page containing Water is wet., two evidence entries with that same page and quote each receive verified=true. A contradiction finding carrying both reaches count two and is kept. This shows lack of an independent-page or semantic-conflict test; it does not claim any deployed model produced that input.

EPI-RTE-contradiction-probe

Label: Cue-selected contradiction judgment and cache

SRC-1, packages/core/src/compile/contradiction-probe.ts, packages/core/src/analysis/cache.ts, apps/server/src/lib/analysis-scheduler.ts, apps/server/src/routes/analysis.ts. Distinct from RT-RTE-lint-findings and its quote subroute: a separate scheduled/manual legacy-page judge with SQL verdict storage. Scheduler starts a first delayed probe and periodic daily probes; manual recompute calls it too. Code selects unique existing-page pairs from contrast cues before wikilinks. Model judges first 4,000 characters from each hardcoded wiki/notes path against a specific factual-collision rubric.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: judged and cached counts; each judge returns verdict and reason.
Later read-back: Matching slug pair/model/prompt cache entry suppresses later judgment; listConfirmed feeds the distinct insights consumer.
Delegated visibility: inapplicable — local completion and cache, not sub-agent.
Selection predicate: Contrast-cue regex, existing target in index, non-self unique pair, default maximum twenty; skip when the pair/model/prompt key already exists.
Invalidation or expiry: Lookup has no body hash or age cutoff. Other deletion/invalidation paths are not exhaustively inspected, so no global permanence claim is made.
Activation or effect: Produces a diagnostic verdict and persists all returned enums, including fallback unrelated results; matching cached result blocks reevaluation in this loop.
Evidence limits: Cue selection misses other conflicts; hardcoded legacy paths can fail for indexed pages elsewhere. Verdict parsing and fallback are not external truth checks; no observed accuracy or calibration.

Diagnosis mode is scheduled bounded candidate evaluation or manual recompute. Software chooses pairs and validates enum; the model proposes/decides semantic classification without per-pair human approval. Missing pages, malformed JSON, adapter exceptions and invalid enums yield unrelated with program-generated reason. No expected-answer oracle. The intended criterion is opposing factual assertions about the same subject, for user review over excerpts; no statement that the system knows which assertion is correct is licensed.

Be strict — only return "contradicts" if there is a specific factual collision you can point at.; ---packages/core/src/compile/contradiction-probe.ts`

const A = bodyA.slice(0, 4000); const B = bodyB.slice(0, 4000); --- packages/core/src/compile/contradiction-probe.ts

if (existing) { cached++; continue; } --- packages/core/src/compile/contradiction-probe.ts

WHERE slug_a = ? AND slug_b = ? AND model_id = ? AND prompt_version = ? --- packages/core/src/analysis/cache.ts

EPI-RTE-probe-consumption

Label: Cached contradictions to review suggestions

SRC-1, packages/core/src/analysis/insights.ts, packages/core/src/analysis/cache.ts, packages/core/src/analysis/suggestions.ts, apps/server/src/routes/analysis.ts. Owner: deterministic analysis code. Endpoints are cached verdicts → listConfirmed → AnalysisInsights/analysis cache → insights API. Distinct from all supplied runtime records and from the judge producer: it consumes rather than evaluates a pair. The consumer does not inspect model reasons to establish factual correctness and does not resolve conflicts.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: Unified insight payload, including contradiction pairs/reasons and ranked suggestions.
Later read-back: Derived suggestions persist in analysis cache; future analysis can reread verdicts. User-visible response is an advisory review cue, not a knowledge rewrite.
Delegated visibility: inapplicable — no model delegation in this consumer.
Selection predicate: SQL verdict equals contradicts, without current model/prompt filter; suggestions label contradictions high severity and cap sorted output at default five.
Invalidation or expiry: Derived insights recompute; source-verdict freshness limitations remain those of the producer/cache. No invalidation of factual claims is implemented by generating a suggestion.
Activation or effect: Promotes selected cached results into high-severity suggestions; future human review is possible, not enforced or observed.
Evidence limits: No downstream accepted correction or observed human action; cached confirmed wording is not a new check.

Behavioral-authority path: insights API/UI caller receives fixed-wording suggestions, ranked ahead of lower-severity issues; advisory horizon lasts through cache reuse/recomputation and human review. Software selects the enum; a human can ignore it. No answer oracle, proposal admission or automatic repair exists in this consumer.

WHERE verdict = 'contradicts' ORDER BY computed_at DESC --- packages/core/src/analysis/cache.ts

message: [[${c.slugA}]] and [[${c.slugB}]] contain contradictory claims — review which is correct., actionable: { slugs: [c.slugA, c.slugB] }, severity: 'high',

--- packages/core/src/analysis/suggestions.ts

EPI-RTE-health-maintenance

Label: Health pipeline crosslink and L2 writes

SRC-1, apps/mcp/src/tools/run-wiki-health.ts, apps/server/src/routes/lint.ts, packages/core/src/compile/l2.ts, packages/core/src/compile/l2_prompts.ts, packages/core/src/graph/crosslinker.ts, packages/core/src/tools/executor.ts, schema/lint.md, apps/plugin/agents/curator.md, apps/mcp/src/tools/index.ts. Distinct from RT-RTE-lint-findings, whose ops endpoint stores findings without rewriting knowledge. This health pipeline writes an insights page, performs automatic crosslinking, then sends a wiki overview to an L2 model and directly dispatches parsed actions to ToolExecutor. It adds an omitted consequential maintenance route; it does not supersede the valid narrower ops finding. The curator is instructed to be read-only and grants mcp__mb__mindbase_run_wiki_health, with process text requesting mindbase_run_wiki_health. The registered API is instead run_wiki_health, and the MCP dispatcher uses exact-name lookup. No inspected mapping establishes curator access to this writer. The API's independent write behavior remains supported; neither curator reachability nor host safety follows from the mismatched names.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: MCP health metrics, crosslink and L2 action counts; server legacy lint SSE progress/results. Individual executor outcomes can include failure even when compileL2 returns ok=true.
Later read-back: Successful writes to wiki bodies, summaries/index and insights can enter future search, compilation or display; actual read-back activation not observed. flag_contradiction and merge handlers inspected here return flags/status without storing a resolution or merge.
Delegated visibility: uninspected — external host controls agent execution and tool-name exposure; the curator's prefixed health grant does not match the registered API name, and no mapping was established. Independently, the registered API performs internal L2 model work and I/O.
Selection predicate: Graph exact-title mentions outside code/frontmatter detection feed crosslinks; exact-title score starts at six and auto mode applies extracted links. L2 gets index, one-liners, metadata and structural issues; actions come from parsed JSON array, not a per-action human checklist.
Invalidation or expiry: Reindex follows MCP health; server reindexes on lint.ok. No pipeline-wide transaction or automatic content rollback is established in inspected orchestration. rewrite/update checks can reject human_touched pages; no equivalence asserted for every handler.
Activation or effect: Auto crosslinks write Markdown. L2 can replace/append prose and change summaries through registered handlers. No expected-answer or semantic truth gate precedes dispatch.
Evidence limits: No successful execution, layout compatibility test or measured improvement; not every executor method inspected. Tool return counts do not establish that all actions succeeded, nor that return-only flags affected later behavior.

Admission mode is a user/host health request. Fixed code proposes exact-title links; L2 proposes repairs under schema/lint guidance. Software selects/dispatches, with argument/path/state checks in handlers; no human per-action approval in these wrappers. Human_touched metadata supplies a software veto in the inspected rewrite/update handlers, not a general human truth decision. Failed operations can leave earlier effects; no rollback implementation was inspected in this path. What persists is changed wiki content/metadata/index and insights, not an evidenced accepted theory or its criticism history.

Concrete crosslink deduction: with indexed target title Alpha, a page body Alpha is useful. and no existing edge, titlePattern matches; score begins at six, so confidence becomes extracted. In auto mode replaceFirstOutsideCode yields [[Alpha]] is useful. and the changed page is written. This changes navigation without establishing a new factual proposition. It does not establish that every occurrence of a title refers to that indexed entity.

const xlink = await crossLink(ctx.store, { mode: 'auto' }); const adapter = ctx.getAdapter(); const lint = await compileL2({ adapter, store: ctx.store, model: ctx.config.model }); --- apps/mcp/src/tools/run-wiki-health.ts

const call = actionToToolCall(action, i); const result = await executor.execute(call); toolResults.push({ call, result }); --- packages/core/src/compile/l2.ts

if (meta.edit_state === 'human_touched') { return { ok: false, error: concept ${slug} is human_touched, skipping rewrite }; } --- packages/core/src/tools/executor.ts

result: { slug_a: args.slug_a, slug_b: args.slug_b, flagged: true }, --- packages/core/src/tools/executor.ts

result: { keep: args.keep, absorb: args.absorb, status: 'queued_for_review' }, --- packages/core/src/tools/executor.ts

let score = 6; // exact title match (+6 per plan, guarantees EXTRACTED) --- packages/core/src/graph/crosslinker.ts

await store.writeText(pageNode.path, modifiedBody); --- packages/core/src/graph/crosslinker.ts

const handler = handlers.get(req.params.name); --- apps/mcp/src/tools/index.ts

name: 'run_wiki_health', --- apps/mcp/src/tools/run-wiki-health.ts

Run wiki health via mindbase_run_wiki_health({ projectId }). This calls find_orphans + find_contradictions + find_gaps + suggest_links and returns a composite report. --- apps/plugin/agents/curator.md

description: Sub-agent dispatched by /mb:lint. Read-only — has Read + lint MCP tools but no Edit/Write/Bash. Surfaces orphans, contradictions, stale claims, gaps. Suggests fixes; doesn't apply them. --- apps/plugin/agents/curator.md

EPI-RTE-cli-admission

Label: CLI ingest all-action approval

SRC-1, apps/cli/src/index.ts, apps/cli/src/commands/ingest.ts. Distinct from RT-RTE-contribute-plan: separate CLI caller of core compileL1Plan/Execute, not the ops pending-plan route. Local file ingestion stores raw content before planning; URL mode passes the URL string to ingestPaste rather than fetching its page in this caller. CLI uses an empty hybridSearch callback. Human sees action kinds/titles and an all-or-nothing confirmation unless --yes. The core compile implementation was not fully traced, so this record owns caller admission, not every executor guarantee.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: Proposed action list, aborted message or execution results printed to terminal.
Later read-back: uninspected — caller forwards to compileL1Execute; full persistence/read-back semantics of this core path not traced here.
Delegated visibility: inapplicable — no sub-agent in this caller.
Selection predicate: User confirms entire proposal, or --yes bypasses interactive confirmation; execution receives empty approvals object documented locally as all approved.
Invalidation or expiry: inapplicable — current invocation plan, no TTL or persistent proposal cache in this caller.
Activation or effect: Confirmation/flag gates the call to compileL1Execute; no evidence-consuming truth criterion is enforced in the CLI caller.
Evidence limits: No runtime or human deliberation evidence; downstream core guarantees not inferred from caller intent.

Open ingest request; model proposes via core planner, human or --yes flag authorizes the batch. Aborting vetoes execution but does not reverse prior source ingestion. No expected-answer oracle. No rollback assertion beyond this caller's inspected code.

.option('--yes', 'auto-approve all proposed actions') --- apps/cli/src/commands/ingest.ts

const approvals = {}; // empty = all approved --- apps/cli/src/commands/ingest.ts

Claims

EPI-CLM-wiki-audit

Label: Whole-wiki exact-conflict audit

SRC-1, README.md; doctrine/design conclusion status: claimed. Distinct from RT-CLM-approved-changes: audit coverage and quote quality rather than permission to write.

It audits its own knowledge. One command re-reads the whole wiki and reports contradictions with the exact conflicting sentences quoted. Notion and NotebookLM structurally cannot do this. --- README.md

The competitor comparison is attributed marketing, not independently assessed evidence. The concrete MindBase audit claim is bounded by forty research excerpts of 1,200 characters in ops gather, model judgment and normalized occurrence counts; probe uses selected legacy-page pairs. The implementation supports diagnosis capacity, not exhaustive or sound contradiction detection.

const MAX_LINT_PAGES = 40; const MAX_EXCERPT_CHARS = 1_200; --- apps/server/src/ops/gather.ts

EPI-CLM-sentence-pruning

Label: Every synthesis sentence backed and pruned

SRC-1, packages/core/src/synthesis/prompts.ts; doctrine/design conclusion status: claimed. No supplied claim counterpart: this concerns the structured topic-synthesis prompt, not ops action approval. It explicitly claims a postprocessing behavior as well as requesting citation support. The validator's no-inline-marker branch is a countercase; see EPI-RTE-topic-synthesis.

Every sentence in "content" MUST be backed by at least one citation. Post-processing drops uncited sentences. --- packages/core/src/synthesis/prompts.ts

EPI-CLM-knowledge-compounds

Label: Contributions sharpen retained knowledge

SRC-1, README.md; doctrine/design conclusion status: claimed. Distinct from supplied approval claim: a longitudinal quality/benefit assertion. Persistence and later consumers are implemented; improving truth/coverage or future action capacity requires evidence not supplied in this run.

Knowledge compounds. Your context.md gets sharper every time you contribute. --- README.md

Evidenced absences

none declared in this member

Behavioral-authority paths

none declared in this member; supplied runtime paths and new route declarations above give the assessed consumers, channels, forces and horizons without duplicating their identities.

Keys	Action
`?`	Open this help
`n`	Next page
`p`	Previous page
`s`	Search