LLM Wiki epistemic report

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

Source-and-claim boundary

The question is which routes turn incoming research/security material into content licensed for later reliance, and what actually checks that license. Scope and exclusions are those of the boundary; source identity and evidence layers are in its Source register. SRC-1 supplies implementation and doctrine/design, not observed execution or causal experiments. The runtime is a provisional baseline, not an inspection limit.

The principal distinction is between admitting a file and accepting its claims against evidence for an intended use. The former is implemented in several paths. The latter is requested in prompts and governance, but is not established by parsing, nonempty completions, source counts or verification labels. This is a route-specific finding, not a judgment that the imported material is false.

Consequential claims inspected here are EPI-CLM-knowledge-compilation, EPI-CLM-source-integrity, EPI-CLM-retrospective-validation, and RT-CLM-prompt-ratchet. The newsletter adds a machine-written verification assertion in EPI-RTE-newsletter-admission.

Entry point or operation	SRC-1 source paths	Coverage and records	Limit / conclusion prevented
File/URL ingestion, feed acquisition and normalization	tools/ingest.py, tools/arxiv_monitor.py, tools/cve_monitor.py, tools/github_monitor.py, tools/rss_monitor.py, tools/curated_monitor.py, tools/common.py, tools/normalize.py	Supplied RT-RTE-batch-ingestion, RT-RTE-scheduled-drivers; direct normalization and relevance/hash/frontmatter reads	Feed truth, complete provenance and PDF extraction fidelity not independently established
CLI extraction, integration and prompt accessors	tools/extract.py, tools/integrate.py, tools/prompts.py	Direct complete executable reads; RT-RTE-knowledge-admission, EPI-RTE-claim-projection	No model outputs or source-grounded adjudications supplied
Batch, manual watcher, immediate CVE, daemon and unified scheduling	tools/pipeline_all.py, tools/watchdog_monitor.py, tools/cve_monitor.py, tools/daemon.py, tools/unified_daemon.py, tools/run_all_monitors.py	Supplied runtime account plus direct function/entry-point and validation-call searches; RT-RTE-scheduled-drivers	Driver wiring does not show successful execution; different drivers do not inherit each other's checks
Lightweight/deep lint and retrospective CLI	tools/lint.py, tools/retrospective_validator.py	Direct full reads; RT-RTE-research-maintenance, EPI-RTE-reinforcement-selection, EPI-RTE-confidence-update	Semantic evaluator accuracy and human review uptake unobserved; helper check_contradiction is callable but not called by weekly validation
Optimizer, fixture builder and metrics	tools/prompt_optimizer.py, tools/create_eval_set.py, tools/metrics_collector.py, eval_set_index.json, eval_set/	Direct builder and proposal/evaluation/decision reads; supplied RT-RTE-prompt-optimization, RT-OBJ-metrics	No independent expected-answer oracle or installed candidate treatment; fixtures are not certified ground truth
Research execution branches	tools/research_agent.py	Direct execution/validation/synthesis reads plus supplied RT-RTE-research-maintenance	Task labels and generated queries/outlines do not establish completed searches or evidence acceptance
Search, catalog, article edit/read, upload, backlinks and history; browser calls	tools/query.py, tools/index.py, api_server.py, ui/src/App.jsx	Direct query/index and API read/save/backlink reads; API route search and supplied RT-RTE-api-mutation, RT-RTE-wiki-consumption	Git history is not automatic rollback; ranking is not warrant; rendering/build internals not assessed
Newsletter synthesis and verification assignment	tools/newsletter_agent.py	Direct full read; EPI-RTE-newsletter-context, EPI-RTE-newsletter-admission	Assigned verification has no evidence-consuming check at this write site
Ordinary/long paper generation and critique	tools/paper_agent.py, tools/paper_agent_long.py	Direct ordinary analysis/draft/critique/finalization reads; supplied long-form override in RT-RTE-wiki-consumption	No measured novelty, valid experimental results or validated transfer; long form skips critiques
Provider call and model state	tools/common.py	Supplied RT-RTE-model-call, RT-CMP-main-model, RT-CMP-openrouter-model	Weights, provider state, response grounding and training remain inaccessible
Test/configuration/host diagnostic entry points	test_api.py, test_e2e.py, test_ingest.py, test_llm.py, test_phase2.py, test_user_workflow.py, tools/test_daemon.py, tools/test_optimizer.py, tools/test_research_agent.py, tools/check_model_config.py, check_rss_state.py, clear_rss_state.py, configure_lms_max_context.sh, tasks/	Boundary coverage plus direct scoped evaluator/verification searches in Python tools/root tests	Not executed; no benchmark result or host deployment guarantee; detailed diagnostic branches unassessed
Governance and public specification	README.md, AGENTS.md, llm-wiki.md, research_agenda.md	Direct README introduction, AGENTS complete, specification trust section; supplied agenda account	Remaining status/design narratives and individual retained page assertions not audited; no system-complete negative across every claim in the repository

The dispatch search enumerated Python top-level functions/classes, __main__ guards and FastAPI decorators across tools/*.py and api_server.py. A second search across those tools, API and root tests covered validation, acceptance/rejection, withdrawal/retraction, rollback, verification, reinforcement, contradiction helpers, registration, hooks and tool calls. The inspected paths expose CLI functions, HTTP handlers and code-scheduled calls, not a model-native registered-tool loop. This is a bounded search account, not an absence claim about excluded plugin/host behavior. The callable contradiction helper is a counterexample to saying no model contradiction operation exists; the weekly caller does not use it (SRC-1: tools/retrospective_validator.py). Deep lint is another real counterexample to a blanket claim of no semantic checking.

Missing evidence is localized: source/provider truth prevents certifying imported premises; linked extraction/merge traces prevent classifying actual generated propositions or checking preservation; adjudicated evidence decisions prevent claiming produced accepted new knowledge; deployment traces prevent claiming the routes operated; independent interventions prevent attributing quality gains. Tracked wiki artifacts alone do not identify their producing/checking route. No needed supersession of the supplied runtime was found: its broad containers remain valid; this report adds material parts, including the newsletter's verification assignment and omission of conflict status from its model context.

Epistemic-object inventory

Fields owned by records are cited rather than redefined here. All source anchors below are SRC-1 at the frozen revision.

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
EPI-OBJ-source-text	see record	Normalized input	SRC-1 tools/normalize.py	Source warrant unknown, PDF fidelity untested
EPI-OBJ-extracted-claims	see record	Extraction candidate	SRC-1 tools/extract.py, tools/prompts.py	JSON parsing does not adjudicate claim truth
EPI-OBJ-evidence-fields	see record	Requested evidence/relationships	SRC-1 tools/prompts.py, tools/integrate.py	Not delivered by the integration projection
EPI-OBJ-wiki-body	see record	Retained claims and synthesis	SRC-1 tools/integrate.py, tools/newsletter_agent.py	Import, merge, synthesis and direct-edit lineage differ
EPI-OBJ-wiki-metadata	see record	Reliance labels and selectors	SRC-1 tools/integrate.py, tools/retrospective_validator.py, tools/newsletter_agent.py	Claimed confidence/verification not calibrated warrant
RT-OBJ-prompts	see record	Instructions and registry	SRC-1 tools/prompts.py	Guidance and symbolic selection are separate parts; no truth oracle
RT-OBJ-metrics	see record	Evaluation aggregates	SRC-1 tools/metrics_collector.py, tools/prompt_optimizer.py	Self-confidence/process validity, not adjudicated accuracy
EPI-OBJ-eval-set	see record	Fixed test input	SRC-1 tools/create_eval_set.py	No independently labelled answers in the inspected builder
RT-CMP-main-model, RT-CMP-openrouter-model	not determinable for internal state	Inference	SRC-1 tools/common.py	Unpinned provider-resolved distributed-parametric state; no accessible content-level acceptance account

Catalogs, alerts, research-task notes and paper drafts are covered by their route records below rather than treated as another uniform knowledge store. Paper assertions remain candidate text; queries/outlines can be non-truth-apt instructions as well as carry claims. Their exact generated content is not determinable without outputs.

Authority-route ledger

All architectural statuses concern the route, not observed operation. Every result here is possible under inspected code; none is claimed observed. Repeated route IDs denote distinct consequential functions within a supplied container. The source record owns generic runtime details (see); this ledger supplies the epistemic distinctions. “No license” means the check described does not establish factual warrant, not that no true claim could pass it.

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-batch-ingestion	content transformation	implemented	EPI-OBJ-source-text	truth-apt transformation: acquisition/import	External source to normalized input	File/feed admission, not source-truth judgment	see route	Imported text	File creation	Source warrant unknown	Eligible for extraction	Extractor; file body; permissive; later stage	SRC-1 tools/normalize.py	EPI-CLM-knowledge-compilation	none	PDF and monitor transformations may lose context
RT-RTE-knowledge-admission	content transformation	implemented	EPI-OBJ-extracted-claims	truth-apt transformation: indeterminate	Source body to proposed claims	Model instructed to cite evidence and assess confidence	see route	Candidate JSON text	Advisory model generation	No checked preservation/entailment established	Supplies parser input	Parser; completion string; permissive if parse succeeds; current stage	SRC-1 tools/extract.py, tools/prompts.py	EPI-CLM-knowledge-compilation	unknown semantic relation	Reshaping, derivation and conjecture cannot be separated without content evidence
RT-RTE-knowledge-admission	check/evidence production	implemented	EPI-OBJ-extracted-claims	no content change	Completion syntax as JSON	Python json.loads; JSON domain only	Before sidecar assignment/write	Parse success or exception	Enforcing syntax gate	Parseability only; no schema or factual license	Failed parse exits; parsed values proceed to later field operations	Extractor; parse result; enforcing; current invocation	SRC-1 tools/extract.py	EPI-CLM-source-integrity	syntax is not evidence validation	Parseable non-object values can still fail later
EPI-RTE-claim-projection	content transformation	implemented	EPI-OBJ-extracted-claims, EPI-OBJ-evidence-fields	truth-apt transformation: non-ampliative reshaping	Sidecar to new-claim Markdown	see record	see record	see record	Fixed projection	Claim text retained, evidence lineage degraded at this edge	Reduced claims become direct-write or merge input	Integrator/model; claim body; informational/permissive; current and future pages	SRC-1 tools/integrate.py	EPI-CLM-source-integrity	evidence omitted	Sidecar remains independently available; omission is not source deletion
RT-RTE-knowledge-admission	content transformation	implemented	EPI-OBJ-wiki-body	truth-apt transformation: indeterminate	Old body plus projected new claims to merged body	Model; merge/history/conflict instruction	Existing destination only	Nonempty merge, conflict marker or empty fallback	Guidance; replacement text supplied	No checked semantics; confidence averaging only requested	Supplies write candidate	Writer; completion body; permissive; destination overwrite	SRC-1 tools/integrate.py, tools/prompts.py	EPI-CLM-knowledge-compilation, EPI-CLM-source-integrity	preservation unenforced	Empty fallback appends, rather than rejecting
RT-RTE-knowledge-admission	disposition/acceptance	implemented	EPI-OBJ-wiki-body	no content change	New or merged body admitted to wiki	Existence/nonempty-result rules; marker sets conflict status	see route	Write selected even with conflict	Enforcing software admission, not evidence acceptance	No acceptance criterion naming evidence, use and scope	Permits automatic overwrite	Writer; branch/string result; enforcing; current write	SRC-1 tools/integrate.py	EPI-CLM-source-integrity	admission differs from acceptance	Human escalation does not veto this write
RT-RTE-knowledge-admission	retention	implemented	EPI-OBJ-wiki-body, EPI-OBJ-wiki-metadata	no content change	Selected text to persisted page	see route	see route	Page retained	Durable overwrite	Retention does not add warrant	Future merges/search/synthesis can use it	see RT-BAP-wiki-context	SRC-1 tools/integrate.py	EPI-CLM-knowledge-compilation	none	Not lifecycle integration after evidenced acceptance
RT-RTE-api-mutation	disposition/acceptance	implemented	EPI-OBJ-wiki-body	truth-apt transformation: indeterminate	Caller text replaces article	Path/I/O only; caller controls content	see route	Direct save	Software admission; human browser authorship possible	No source-truth check	Bypasses extraction/merge	API writer then readers; HTTP/file; permissive; subsequent invocations	SRC-1 api_server.py	EPI-CLM-source-integrity	alternate path	Caller text could preserve or revise claims; no output supplied
RT-RTE-research-maintenance	check/evidence production	implemented	EPI-OBJ-wiki-body	truth-apt transformation: indeterminate	Direct factual contradiction among page prefixes	Model sees first 600 body characters with basename identities	lint.py --deep, at least two included pages	JSON contradiction descriptions or parse failure	Produces advisory semantic judgments	Limited to supplied excerpts; neither fact truth nor complete-page consistency certified	Feeds filename-mapped tagging	Lint tagging; parsed list; permissive; current maintenance run	SRC-1 tools/lint.py	EPI-CLM-source-integrity	partial semantic check exists	Basename collisions and truncation limit target resolution/coverage
RT-RTE-research-maintenance	disposition/acceptance	implemented	EPI-OBJ-wiki-metadata	non-truth-apt policy/content update: conflict annotation	Lint result to status/note	JSON list, filename membership; skip already-conflict page	After deep-lint response	Conflict note/status written	Enforcing tagging; advisory review request	Model diagnosis retained, not independently accepted truth	Adds review visibility; does not block query/synthesis	Writer/catalog; file status/note; advisory to human; later reads	SRC-1 tools/lint.py, tools/index.py, tools/query.py	EPI-CLM-source-integrity	no withdrawal gate	No automatic prior-page restoration
RT-RTE-research-maintenance	check/evidence production	implemented	EPI-OBJ-wiki-metadata	no content change	Metadata presence, high-unverified confidence, existing conflict marker	Fixed rules; metadata/string domain	Lightweight lint / weekly validation	Warnings, existing-conflict counts	Console/report advisory; conflict status update	Validates representation predicates only	Exposes issues, not factual rejection	Human/report reader; console/report/status; advisory; maintenance horizon	SRC-1 tools/lint.py, tools/retrospective_validator.py	EPI-CLM-retrospective-validation	no new-source comparison in weekly path	Lightweight broken-link counter stays zero; not evidence links were checked
EPI-RTE-reinforcement-selection	check/evidence production	implemented	EPI-OBJ-wiki-metadata	no content change	Page deemed reinforced	see record	see record	see record	Threshold classification	Source-count/confidence predicate only; no independence or truth license	Permits confidence-update branch	see record	SRC-1 tools/retrospective_validator.py	EPI-CLM-retrospective-validation	corroboration not established	Integration increments counts per call, not distinct evidence
EPI-RTE-confidence-update	behavior/policy adaptation	implemented	EPI-OBJ-wiki-metadata	non-truth-apt policy/content update: confidence label increment	Reinforced page to stronger reliance label	see record	see record	see record	Enforcing file update	No new evidential license	Changes downstream confidence-driven selection and displayed aggregates	see record	SRC-1 tools/retrospective_validator.py	EPI-CLM-retrospective-validation	confidence without new corroboration	Operational effect capacity, not observed behavioral activation
RT-RTE-prompt-optimization	check/evidence production	implemented	RT-OBJ-metrics, EPI-OBJ-eval-set	no content change	Existing extractor on first ten available fixtures	Process/JSON success, returned confidence; conflict rate fixed zero	see route	Aggregate metrics	Supplies thresholds	Process validity and self-report only	Drives keep/reject label	Optimizer; numeric deltas; enforcing decision; cycle	SRC-1 tools/prompt_optimizer.py	RT-CLM-prompt-ratchet	candidate installation missing	Not independent correctness or a controlled candidate contrast
RT-RTE-prompt-optimization	disposition/acceptance	implemented	RT-OBJ-prompts	non-truth-apt policy/content update: keep/reject experiment bookkeeping	Metric deltas to decision/reset	see route	see route	Keep label or Git reset	Code decision and Git effects	No warrant for proposed prompt effectiveness	Logging/baseline/reset; not candidate activation	see route	SRC-1 tools/prompt_optimizer.py	RT-CLM-prompt-ratchet	keep is not installation	Mixed reset not a content rollback
RT-RTE-research-maintenance	other — describe	implemented	EPI-OBJ-wiki-metadata	non-truth-apt policy/content update: research task selection and result labels	Coverage/confidence/age to tasks	Code thresholds; model query/outline requests	see route	Query, task identified or outline; SUCCESS possible	Task scheduling/logging	No completed corroboration established by labels	Logs tasks, not accepted claims	Research logger; result dictionary; advisory; later metrics/notes	SRC-1 tools/research_agent.py	EPI-CLM-retrospective-validation	none	Claim-bearing generated outlines remain semantically indeterminate
RT-RTE-wiki-consumption	operational admission/selection/consumption	implemented	EPI-OBJ-wiki-body	no content change	Wiki to query/article/paper context	BM25/top five; explicit path; topic matching/truncation	see route	Ranked paths, article or excerpts	Ranking/delivery	Relevance score does not license truth	Conflict/unverified content can be delivered	API/human/model; responses/user message; ranking/informational; request/later calls	SRC-1 tools/query.py, api_server.py, tools/paper_agent.py	EPI-CLM-knowledge-compilation	semantic-ranking wording exceeds token matching	No fact-based consumer exclusion
EPI-RTE-newsletter-context	content transformation	implemented	EPI-OBJ-wiki-body, EPI-OBJ-wiki-metadata	truth-apt transformation: non-ampliative reshaping	Recent pages to newsletter context	see record	see record	Snippets and confidence without status	Fixed selection/projection	Warrant degraded by omitted conflict state/context	Supplies synthesis model	see record	SRC-1 tools/newsletter_agent.py	EPI-CLM-source-integrity	status collected but dropped	Full files remain available outside model context
RT-RTE-wiki-consumption	content transformation	implemented	EPI-OBJ-wiki-body	truth-apt transformation: indeterminate	Wiki insights/snippets to synthesis/draft	Model instructions including novelty/strategic interpretation	see route	Generated newsletter/paper candidates	Generation, not acceptance	Neither entailment nor non-entailment shown for actual outputs	Candidate reaches length gate or draft finalizer	see route	SRC-1 tools/newsletter_agent.py, tools/paper_agent.py	EPI-CLM-knowledge-compilation	requested novelty not demonstrated conjecture	External-source helper supplies placeholder; no measured experiment
EPI-RTE-newsletter-admission	disposition/acceptance	implemented	EPI-OBJ-wiki-body, EPI-OBJ-wiki-metadata	non-truth-apt policy/content update: verification and confidence assignment	Generated body to verified newsletter	see record	see record	Saved with confidence 0.95, verified true	Length admission and durable labels	Verification assertion without evidence acceptance	Enters catalog, search and future synthesis scans	see record	SRC-1 tools/newsletter_agent.py	EPI-CLM-source-integrity	verified label exceeds check	Count includes all recent paths, not independently corroborating sources
RT-RTE-wiki-consumption	check/evidence production	implemented	EPI-OBJ-wiki-body	truth-apt transformation: indeterminate	Ordinary paper draft prefix	Model reviewer asked novelty/rigor/clarity/completeness/impact	Two critique/improve rounds; long form skips	Review and improved draft requested; parser can keep whole result	Model proposal; no independent acceptance veto	Critique request only; no empirical novelty/rigor license	Improved text goes onward and is saved as draft	Drafting model/finalizer; selected completion; informational/permissive; current generation	SRC-1 tools/paper_agent.py	none	review is not validated research	Review need not persist to next round; missing content-directed criticism evidence

No post-acceptance lifecycle-integration route is established by these writes. That statement is restricted to the admissions above: ordinary retention/linking/indexing is implemented, but an evidence-consuming acceptance decision antecedent to it has not been established. Human/coding-agent review remains doctrine or external behavior under RT-BAP-agent-doctrine, not a hidden Python gate.

System-claim versus route comparison

Claim ID	Claimed operation or warrant	Claim source ID/anchor and evidence layer	Doctrine/design support	Implemented route IDs	Observed-run support	Causal support and design limits	Supported conclusion	Mismatch/unknown
EPI-CLM-knowledge-compilation	see record	SRC-1 README.md; doctrine/design	Extraction/merge instructions	RT-RTE-batch-ingestion, RT-RTE-knowledge-admission, RT-RTE-wiki-consumption	None supplied	No independent variation or benefit measurement	Persistent compilation and reuse are wired; generation is candidate production	Compounding value and semantic correctness not established; BM25 is token-based ranking
EPI-CLM-source-integrity	see record	SRC-1 AGENTS.md, llm-wiki.md, tools/prompts.py; doctrine/design	Attribution/history/trust/human escalation rules	EPI-RTE-claim-projection, RT-RTE-knowledge-admission, RT-RTE-api-mutation, EPI-RTE-newsletter-context, EPI-RTE-newsletter-admission	None supplied	No source-independence or full-history check	Citations/markers and review notifications have routes; not invariant safeguards	Evidence field omitted at merge; direct edit bypass; newsletter verified without cross-validation
EPI-CLM-retrospective-validation	see record	SRC-1 tools/retrospective_validator.py; doctrine/design and implementation contrast	Docstring promises new-evidence validation	EPI-RTE-reinforcement-selection, EPI-RTE-confidence-update, RT-RTE-research-maintenance	None supplied	Metadata repetition is not independent corroboration	Existing conflict detection, threshold reinforcement and confidence writes are implemented	Weekly validation does not invoke new-source contradiction helper; research validation only identifies task
RT-CLM-prompt-ratchet	see record	SRC-1 tools/prompt_optimizer.py; doctrine/design	Metric thresholds and proposal prompt	RT-RTE-prompt-optimization	None supplied	Proposed text not installed; conflict metric constant; no candidate contrast	Working proposal/evaluation/bookkeeping machinery	Keep/revert labels do not establish installed improvement or safe rollback

Bounded conclusion

Acquisition and retention: source text, extracted candidates and wiki content have implemented persistence/reuse paths, with source warrant unknown at acquisition. Deterministic claim projection preserves the supplied claim text but removes the evidence field from the integration input. Later merge semantics are indeterminate, not automatically entailed and not automatically novel. See EPI-RTE-claim-projection and RT-RTE-knowledge-admission.

Checks and authority: JSON parsing can enforce syntax; lint can warn about representation predicates, and deep lint can request semantic contradiction judgments over limited excerpts. Those are real but different checks. Parsed/nonempty/long-enough outputs gain operational authority to be saved and reused, without establishing evidence-consuming acceptance for factual reliance. Conflict tags cause advisory review visibility, not a general block on search or synthesis. The newsletter is a particularly consequential counterpath: it drops status from its model context and assigns verified true/confidence 0.95 after a length gate. See EPI-RTE-newsletter-context, EPI-RTE-newsletter-admission, and RT-RTE-research-maintenance.

Deterministic reinforcement: repeated integration of the same input can raise source_count; the weekly validator then uses that count and existing confidence to strengthen confidence without reading additional source evidence. This supports a metadata-feedback mechanism, not corroborated knowledge growth or observed degradation. See EPI-RTE-reinforcement-selection and EPI-RTE-confidence-update.

Derivation, conjecture and acceptance: no inspected route establishes a produced accepted ampliative claim. Actual model output contents are unavailable, so the extraction/merge/synthesis edges remain indeterminate among reshaping, derivation and conjecture. No checked derivation from independently warranted scientific/security premises is established here. The implemented arithmetic predicates and confidence increments support only their narrow metadata-domain deductions. Draft critique requests and publication-ready wording do not certify novelty, experiments or explanatory transfer. Subsequent indexing is retention/organization before established acceptance, not evidence of post-acceptance lifecycle integration.

Independent properties: automated schedules and software admission exist; they do not establish reliable autonomy in opaque reasoning. Metrics represent aspects of extraction and affect later proposal selection, but the optimizer lacks the proposal-installation link, so its own candidate-change path does not establish exercised self-improvement. Its localized method text and consumption are separately supported; content-directed criticism and criticism-fed iteration remain unestablished, as detailed in RT-RTE-prompt-optimization. Paper critique has a request for criticism and two-round revision wiring, but no supplied output establishing a challenged claim, blame placement and carried criticism; that does not justify a theory-builder finding. Retained wiki context and confidence updates can affect later operations independently of the broken prompt installer. No improved future capacity or causal benefit is demonstrated for either. External model parameters, exact versions and provider-side learning remain uninspected under RT-CMP-main-model and RT-CMP-openrouter-model. These conclusions remain separate; they imply no single epistemic score or system-wide verdict.

Shared records

Operative objects

EPI-OBJ-source-text

Label: Normalized source text

Source-native identity: body of raw/normalized/<domain>/<hash>.md; natural-language content stored in filesystem Markdown. Producer is normalization of downloaded/copied text or PDF text; consumers are extraction and fixture builder. Imported propositions can be truth-apt over their source's stated domain; no truth adjudication occurs in normalization. Distinct from RT-OBJ-prompts, which supplies instructions rather than input documents; this input crosses RT-RTE-batch-ingestion, not a separate invocation mechanism. SRC-1: tools/normalize.py.

normalized_content = serialize_frontmatter(metadata, content) --- tools/normalize.py

EPI-OBJ-extracted-claims

Label: Extracted claim text

Source-native identity: claims[*].text in extraction JSON, with model-assigned confidence. Natural-language propositions plus associated numerical labels, stored as a filesystem JSON sidecar. Producer is an external completion; consumers are integration and optimizer analysis. A proposition's scope depends on its actual text, unavailable as a traced run here. This is a material output object of RT-RTE-knowledge-admission, not the prompt referent RT-OBJ-prompts. SRC-1: tools/extract.py, tools/prompts.py.

"claims": [{"text": str, "confidence": float, "evidence": str}], --- tools/prompts.py:8-16

output_path = filepath + ".json" --- tools/extract.py

EPI-OBJ-evidence-fields

Label: Extraction evidence fields

Source-native identity: claims[*].evidence, entities, relationships in the requested JSON. Natural-language evidence/labels with symbolic JSON field structure, persisted if returned in the sidecar. Unlike EPI-OBJ-extracted-claims, these fields do not reach the merge request through the inspected integrator. They are a distinct material object in RT-RTE-knowledge-admission, not another extraction mechanism. Presence in every returned object is not schema-enforced. SRC-1: tools/prompts.py, tools/extract.py, tools/integrate.py.

• Cite specific sentences as evidence for each claim --- tools/prompts.py:18-25

The projection passage in EPI-RTE-claim-projection supports the consumer omission.

EPI-OBJ-wiki-body

Label: Wiki page body

Source-native identity: body of wiki Markdown pages, including synthesis/newsletters. Natural-language claims, historical/conflict notes and links stored in filesystem files. Producers differ: initial claim projection, merge completion, API caller, lint annotation, newsletter completion. Consumers include later integration, query/API delivery, paper/newsletter models. Unlike RT-OBJ-prompts, this is retained subject matter, not shipped instructions; closest supplied route is RT-RTE-wiki-consumption, whose reuse passage supplies evidence. SRC-1: tools/integrate.py, api_server.py, tools/newsletter_agent.py. RT-BAP-wiki-context owns the broad text-to-model path.

EPI-OBJ-wiki-metadata

Label: Wiki frontmatter fields

Source-native identity: source_count, confidence, verified, status and other page frontmatter. Symbolic values under fixed parser/threshold rules, persisted in Markdown YAML; displayed to humans/models as well. Producers are integrator defaults/count increments, retrospective updates, lint, API caller and newsletter constants. Consumers are catalog aggregation, research task selection, validator reinforcement and newsletter context. Verification/confidence can be interpreted as truth-apt assertions about support, but their implemented force is metadata selection/display, not a proof of factual truth. Distinct from RT-OBJ-metrics, which is SQLite/log evaluation state. Closest route containers are RT-RTE-knowledge-admission and RT-RTE-research-maintenance; this object spans both, not a part of exactly one. SRC-1: tools/integrate.py, tools/retrospective_validator.py, tools/newsletter_agent.py, tools/index.py.

'source_count': 1, 'confidence': 0.8, 'verified': False, --- tools/integrate.py

EPI-OBJ-eval-set

Label: Fixed evaluation sources

Source-native identity: eval_set/ Markdown and eval_set_index.json. Natural-language input bodies with symbolic file/domain manifest, stored in filesystem. Builder copies normalized files, balances domains by traversal order and fills from largest domain; optimizer uses first ten manifest items. This differs from RT-OBJ-metrics, which stores evaluation outputs, and from RT-OBJ-prompts, the method being nominally optimized. It belongs to the evaluation context of RT-RTE-prompt-optimization. SRC-1: tools/create_eval_set.py, tools/prompt_optimizer.py. No independently adjudicated expected answers are established by copying inputs.

shutil.copy2(file_info['path'], dest_path) --- tools/create_eval_set.py

Routes

EPI-RTE-claim-projection

Label: Claim body projection Part of: RT-RTE-knowledge-admission

Endpoints: parsed sidecar → new_claims_body → initial page or merge input. Owner is integrate_knowledge; fixed Python formatting consumes claim text/confidence/source_id but not evidence/entities/relationships. It is the deterministic material part of the supplied admission container, not a duplicate complete route. SRC-1: tools/integrate.py.

• Immediate return: Constructed Markdown body within integration; no separate external return.
• Later read-back: Written text is reloaded by subsequent merges/search/synthesis; omitted sidecar fields have no automatic delivery there.
• Delegated visibility: Merge model sees the projected claim body, not the full JSON.
• Selection predicate: Every element of data.get('claims', []); direct keys text/confidence, optional source_id.
• Invalidation or expiry: Reconstructed per integration; old destination overwritten by containing route.
• Activation or effect: Projection wired; semantics beyond exact text formatting not adjudicated.
• Evidence limits: Deduction from implementation, not an executed example; arbitrary field contents may contain delimiters/instructions.
• implementation conclusion status: wired
• operation conclusion status: uninspected

new_claims_body = "## Claims\n" for claim in data.get('claims', []): new_claims_body += f"- {claim['text']} (Confidence: {claim['confidence']}, Source: [[{data.get('source_id')}]])\n" --- tools/integrate.py

Concrete symbolic trace: with source_id = s, one claim text = X is Y, confidence = 0.7, evidence = Sentence E, the constructed body is ## Claims\n- X is Y (Confidence: 0.7, Source: [[s]])\n. Sentence E is not interpolated. This is non-ampliative formatting of the claim, with degraded evidence delivery, not a novel proposition or a check that X is Y. No target execution was needed.

EPI-RTE-reinforcement-selection

Label: Reinforced page selection Part of: RT-RTE-research-maintenance

Endpoints: recently updated page metadata → reinforced list/eligible confidence branch. Owner is find_reinforced_claims; target is the page's recorded count/confidence, not independently verified source agreement. Code decides, no per-page human veto or answer oracle. Mode is open scheduled/CLI maintenance, not a controlled corroboration experiment. SRC-1: tools/retrospective_validator.py.

• Immediate return: List of title/confidence/source_count dictionaries; invokes update branch for eligible pages below cap.
• Later read-back: inapplicable — list is current-run state; durable effects belong to the linked update part.
• Delegated visibility: No source text or model request in this selector; code receives metadata snapshots.
• Selection predicate: source_count >= 3 and confidence >= 0.75; confidence below 0.9 additionally enables increment.
• Invalidation or expiry: Inputs refreshed from mtime-selected files on a later validation invocation; no distinct-source test.
• Activation or effect: Predicate controls update invocation, not acceptance of factual truth.
• Evidence limits: Counts can reflect repeated integrations; no observed invocation or independent evidence comparison.
• implementation conclusion status: wired
• operation conclusion status: uninspected

if page['source_count'] >= 3 and page['confidence'] >= 0.75: --- tools/retrospective_validator.py

metadata['source_count'] = exist_meta.get('source_count', 0) + 1 --- tools/integrate.py

Concrete countercase: integrate the same well-formed claim JSON three times at the same destination, with merge calls returning empty strings. First write sets count 1/confidence 0.8; later calls increment count to 2 then 3, preserve append-fallback content and reset confidence to 0.8. The validator predicate can now succeed without a second independent source. This is a code-path deduction, not observed operation; no conclusion about actual deployment counts follows.

EPI-RTE-confidence-update

Label: Reinforcement confidence update Part of: RT-RTE-research-maintenance

Endpoints: selected page metadata → increment → rewritten frontmatter → later catalog/research/validation consumers. Code owns admission: selected confidence below 0.9 increases by 0.05 capped at 0.9. The intended reinforcement guidance is the selector's interpretation of count/confidence; no new-source evidence or expected-answer oracle is consumed. Proposed change is a reliance label; software decides, exceptions can prevent write, no built-in content rollback or human veto. Metadata and log persist. Mode is open maintenance; trigger is weekly/CLI validation. This part owns the write, separately from the predicate in EPI-RTE-reinforcement-selection. SRC-1: tools/retrospective_validator.py, tools/index.py; downstream research selectors are also described by RT-RTE-research-maintenance.

• Immediate return: Frontmatter update, log entry and adjustment count.
• Later read-back: Later scans reload confidence; catalog averages and research low-confidence selection can depend on it.
• Delegated visibility: File readers receive new confidence with unchanged body; no source text is consulted by this update.
• Selection predicate: Linked reinforcement predicate plus confidence below 0.9.
• Invalidation or expiry: Future integrations can reset confidence; later validators can increment again if still selected; no evidence-expiry check.
• Activation or effect: Durable metadata mutation wired; changed downstream behavior not observed.
• Evidence limits: Numerical increase establishes neither better calibration nor increased factual support.
• implementation conclusion status: wired
• operation conclusion status: uninspected

new_confidence = min(0.9, page['confidence'] + 0.05) self.update_page_confidence(page['path'], new_confidence) --- tools/retrospective_validator.py

meta['confidence'] = new_confidence --- tools/retrospective_validator.py

For the predicate countercase's 0.8 input, the expression yields 0.85, modulo ordinary floating-point representation. It changes a support label, not the underlying claim or its evidence.

EPI-RTE-newsletter-context

Label: Newsletter context projection Part of: RT-RTE-wiki-consumption

Endpoints: recent wiki files → category entries → first five entries per bucket → model context. Owner is newsletter code. Targets include current, unverified or conflict pages because recentness selection does not inspect status. status is collected in entries but omitted from formatting; confidence remains. This distinct part qualifies the broad supplied synthesis container without superseding it. SRC-1: tools/newsletter_agent.py.

• Immediate return: Concatenated section-context string for the current completion request.
• Later read-back: inapplicable — this string is current-run context; newsletter retention belongs to the linked admission part.
• Delegated visibility: Model receives titles, paths, confidence and 400-character body prefixes, at most five per bucket, but no status field.
• Selection predicate: Mtime within lookback; omit index/log; bucket by domain/path; traversal order within bucket.
• Invalidation or expiry: Context rebuilt per call; full files remain available on filesystem outside model input.
• Activation or effect: Projection wired; no guarantee snippets contain conflict notes or preserve all qualifications.
• Evidence limits: Code deduction supports lost status channel; no observed false synthesis is asserted.
• implementation conclusion status: wired
• operation conclusion status: uninspected

'status': meta.get('status', 'current'), 'snippet': body.strip()[:400], --- tools/newsletter_agent.py

for e in entries[:max_items]: lines.append(f"TITLE: {e['title']}\nPATH: {e['path']}\nCONFIDENCE: {e['confidence']}\nSNIPPET: {e['snippet']}\n") --- tools/newsletter_agent.py

EPI-RTE-newsletter-admission

Label: Newsletter verification assignment Part of: RT-RTE-wiki-consumption

Endpoints: completion body → length decision → fixed frontmatter → saved wiki newsletter → catalog/search/future scans. Code owns admission; model proposes body under an instruction not to invent information. The proposed product is synthesized prose; software rejects absent/short response, otherwise assigns confidence 0.95, verified true and current status. No source-evidence comparison, human veto or independent answer oracle at that decision. Trigger is CLI generation with recent input pages; mode is open synthesis. Guidance persists as source prompt; body/labels persist as dated Markdown. Re-run may overwrite same-date file; errors are not automatic rollback. This material part is not another complete synthesis route. SRC-1: tools/newsletter_agent.py, tools/query.py, tools/index.py.

• Immediate return: Saved newsletter, console/log output and attempted catalog rebuild.
• Later read-back: Query/index include newsletter; future recent-page synthesis and topic-matched paper selection can include it.
• Delegated visibility: Readers receive generated prose with unconditional support labels; raw source evidence is not inspected by the length gate.
• Selection predicate: Nonempty body with length at least 200 characters; recent_paths nonempty before generation.
• Invalidation or expiry: Same-date path can be overwritten; no verification expiry or independent retraction mechanism here.
• Activation or effect: Durable verified/confidence assignment and read eligibility wired; factual grounding and changed behavior unobserved.
• Evidence limits: Label is not an evidence-consuming acceptance decision; source_count counts paths, not independent corroboration.
• implementation conclusion status: wired
• operation conclusion status: uninspected

if not newsletter_body or len(newsletter_body) < 200: --- tools/newsletter_agent.py

f"source_count: {len(recent_paths)}\n" f"confidence: 0.95\n" f"verified: true\n" --- tools/newsletter_agent.py

"Do not invent information not present in the context." --- tools/newsletter_agent.py

The last passage is a source-supported counterweight: factual restraint is requested. It does not turn the length predicate into verification.

Claims

EPI-CLM-knowledge-compilation

Label: Stateful knowledge compilation

Claimed operation: sources become a persistent cross-linked, compounding wiki with BM25 semantic ranking. SRC-1 README.md, doctrine/design. No counterpart among supplied claims after comparison: RT-CLM-prompt-ratchet concerns prompt improvement, not source-to-wiki compilation. Relevant supplied routes are RT-RTE-batch-ingestion and RT-RTE-wiki-consumption.

The result is a knowledge base that compounds over time — every new source enriches the existing structure rather than existing in isolation. --- README.md

1. Queryable — the wiki is searched with BM25 semantic ranking --- README.md

EPI-CLM-source-integrity

Label: Attribution and trust constraints

Claimed operation: claim attribution/confidence follows source tiers; conflicts preserve history and trigger human escalation; specification requires source tier and verification state per claim and independent corroboration before CVE verification. SRC-1 AGENTS.md, llm-wiki.md, tools/prompts.py, doctrine/design. Distinct from RT-CLM-prompt-ratchet; closest supplied evidence is RT-BAP-agent-doctrine, a delivery/authority path rather than this claim.

• Each extracted claim must include a citation inline, linking back to /raw/normalized/<domain>/<filename>.md.

• Confidence must be rated based on source tiers in Section 7. --- AGENTS.md

• When conflicts arise with a CVSS >= 7.0 or if explicitly flagged by a direct contradiction on a high-confidence topic, pause and raise to the human user by updating the file status to conflict and bringing it to the user's attention appropriately. --- AGENTS.md

- Claims are cross-validated across a minimum of two independent sources before verified: true --- llm-wiki.md

Scope caution: the specification's minimum-two-sources rule explicitly concerns CVE entries, not every newsletter. Newsletter verification nevertheless lacks any evidence check of its own; that finding does not require extending the CVE rule to all products.

EPI-CLM-retrospective-validation

Label: Old claims against new evidence

Claimed operation: retrospective validation checks old claims against new sources, detects contradictions and adjusts confidence. SRC-1 tools/retrospective_validator.py, doctrine/design. Distinct from the supplied prompt-ratchet claim; RT-RTE-research-maintenance supplies the relevant broad implementation container rather than a duplicate claim.

Retrospective validator that checks old claims against new sources. Detects contradictions and adjusts confidence scores. --- tools/retrospective_validator.py

Check for contradictions (simplified - would need actual new sources)

For now, just check if page has [CONFLICT] tags

--- tools/retrospective_validator.py

Evidenced absences

none declared in this member

Behavioral-authority paths

none declared in this member