OpenKB epistemic report

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

Source-and-claim boundary

The question is what licenses reliance on OpenKB's compiled wiki, answers and portable skills, and what checks actually change their subsequent use. Scope and external exclusions are those of boundary; evidence identity is SRC-1 in its Source register. Source inspection supplies implementation and doctrine/design only. No target execution, tests, provider calls or external research were performed. The frozen checkout HEAD was checked against the reviewed revision.

The supplied runtime is a provisional route map, not an inspection limit. Direct reads added semantic lint, deterministic link cleanup, source-withdrawal internals, skill graders and their CLI consumer. These are finer parts or checks within the runtime's valid containers; no supplied record needs supersession. In particular, RT-RTE-wiki-maintain is not evidence that removing a document removes every proposition learned from it.

Entry point or operation	Source path (SRC-1)	Coverage and supporting records
CLI add, upload, watch and cloud import; commit/log hooks	openkb/cli.py, openkb/api.py, openkb/watch_service.py, openkb/add_coordinator.py	Supplied RT-RTE-document-add; direct CLI hook inspection confirms log hook, not semantic approval. API dispatch searched; watcher/indexer internals rely on supplied findings.
Short compilation; long-summary compilation	openkb/agent/compiler.py	Direct templates, page parser/filter, short/long entry and removal reads; RT-RTE-document-add, EPI-OBJ-summary, EPI-OBJ-concept, EPI-OBJ-entity. PageIndex internals remain excluded.
Query and saved answers	openkb/agent/query.py, openkb/agent/tools.py, openkb/cli.py	Direct query instructions/build opening; tools and save consumers supplied by RT-RTE-wiki-query. Source acquisition differs from answer generation.
Chat resume, installed skills and writes	openkb/agent/chat.py, openkb/agent/chat_session.py, openkb/agent/skills.py	Supplied RT-RTE-session-chat, RT-BAP-skill-instructions; no supplied conversation demonstrates grounding or instruction compliance.
Remove and recompile	openkb/cli.py, openkb/agent/compiler.py	Supplied RT-RTE-wiki-maintain; direct source-removal helpers, EPI-RTE-source-withdrawal.
CLI/chat/REST lint and fix	openkb/agent/linter.py, openkb/lint.py, openkb/cli.py, openkb/api.py	Direct lint agent, CLI and REST report consumer, fix implementation; EPI-RTE-semantic-lint, EPI-RTE-link-cleanup. Structural check inventory partly supplied; no comprehensive semantic negative inferred.
Skill new/validate/rollback and marketplace	openkb/skill/generator.py, openkb/skill/validator.py, openkb/skill/creator.py, openkb/cli.py	Direct validator; creator/generator/admission supplied by RT-RTE-skill-create. Generation does not imply installation.
Skill eval and saved eval-set replay	openkb/skill/evaluator.py, openkb/cli.py	Direct evaluator and entire CLI eval consumer; EPI-RTE-eval-set, EPI-RTE-trigger-grade, EPI-RTE-body-grade.
Deck production and optional critic	openkb/deck/creator.py, openkb/agent/skill_runner.py	Supplied RT-RTE-deck-produce. Full critic styling instructions and validator internals not directly inspected here; prevents a system-complete negative about all possible content checks.
Visualization, browser display and external-agent reading	openkb/visualize.py, frontend/src/pages/ChatSession.tsx, skills/openkb/SKILL.md	Boundary/supplied runtime only; trust policy RT-CLM-wiki-trust. No independent graph-truth evaluator assessment; enclosing host operation excluded.
Samples, model/service/dependency internals	examples/, pyproject.toml	Samples not inspected for outcomes; no actual KB corpus or authenticated execution trace supplied. Provider and dependency implementations outside boundary. Prevents efficacy, complete output lineage and causal attribution.

Consequential claims are RT-CLM-compiled-wiki, RT-CLM-wiki-trust, EPI-CLM-current-synthesis, EPI-CLM-grounded-answer and EPI-CLM-body-alignment. Missing source truth assessments prevent warrant transfer from “ingested” to “true.” Missing candidate/run evidence prevents deciding which generated sentences merely restate premises, follow from them, or add conjectures. Missing independent outcome comparisons prevents attributing retrieval accuracy, learning or improved artifact quality to these routes. Uninspected included helpers are not evidenced absences.

Epistemic-object inventory

The rows separate operative parts, not merely filenames. Candidate truth-apt content is assessed here; source-native form/substrate and producers/consumers are owned by the cited records.

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
EPI-OBJ-source-content	Imported assertions, tables and images over the document's own domain	see record	SRC-1, openkb/agent/query.py	Extraction fidelity and original truth unknown; an accessible source is not endorsed.
EPI-OBJ-summary	Claims about a particular document's findings	see record	SRC-1, openkb/agent/compiler.py	Full-text short path differs from PageIndex summary/overview path; no output-grounding check established.
EPI-OBJ-concept	Cross-document topic explanations and synthesis	see record	SRC-1, openkb/agent/compiler.py	New text is not necessarily new knowledge; entailment and non-entailment both unresolved without instances.
EPI-OBJ-entity	Facts about named things accumulated across documents	see record	SRC-1, openkb/agent/compiler.py	Type-enum admission does not assess these facts.
RT-OBJ-agents-md	Mixed schema rules; no distinct empirical claim established here	see record	SRC-1, openkb/schema.py	Instruction authority differs from source-evidence authority.
EPI-OBJ-answer	Answer propositions and saved explorations	see record	SRC-1, openkb/agent/query.py	Wiki-only/citation policy is not evidence of faithful or complete grounding.
EPI-OBJ-lint-report	Claims that particular pages contradict, omit or duplicate content	see record	SRC-1, openkb/agent/linter.py	Model judgments may be wrong; failure text can also be retained as a report.
EPI-OBJ-skill-content	Portable instructions plus reference assertions and activation description	see record	SRC-1, openkb/skill/evaluator.py	Description-only trigger grading differs from body/reference-preview grading.
EPI-OBJ-eval-prompts	Question/expected-label pairs; “this skill should trigger” is scoped to the question	see record	SRC-1, openkb/skill/evaluator.py	Generated or imported labels are not independently established answer references.
RT-CMP-litellm-model, RT-CMP-pageindex-client	Truth-apt internal content not determinable from exposed numerical state	see records	SRC-1, openkb/agent/compiler.py, openkb/indexer.py	Inference interfaces establish no weight-update, provider-memory or inner-evaluator account.
RT-RTE-deck-produce	Deck assertions may be truth-apt; formatting/HTML rules are a separate symbolic target	see record	SRC-1, openkb/deck/creator.py	No artifact supplied to individuate truth claims; critic edits are not necessarily epistemic revisions.

Authority-route ledger

Architectural status describes inspected wiring or the bounded absence of an acceptance route, not observed use. “See” delegates source-native operational fields to their owning record. Relations classify content edges, not entire pipelines. Model-generated summaries, answers and skills remain indeterminate between non-ampliative reshaping, entailed derivation and ampliative conjecture: requesting synthesis alone proves neither entailment nor non-entailment.

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-document-add	content transformation	implemented	EPI-OBJ-source-content	truth-apt transformation: acquisition/import	External document to converted text/PageIndex material	see RT-RTE-document-add	see RT-RTE-document-add	Imported content	see RT-RTE-document-add	Source warrant unknown; no endorsement	see RT-RTE-document-add	see RT-RTE-document-add	SRC-1, openkb/converter.py	RT-CLM-compiled-wiki	none	Conversion and indexer supplied findings; fidelity unmeasured.
RT-RTE-document-add	content transformation	implemented	EPI-OBJ-summary	truth-apt transformation: indeterminate	Source/summary and existing page to summary	see RT-RTE-document-add	see RT-RTE-document-add	Candidate authored text	see RT-RTE-document-add	No additional truth license from fluent generation	see RT-RTE-document-add	see RT-RTE-document-add	SRC-1, openkb/agent/compiler.py	EPI-CLM-current-synthesis	completeness unknown	No generated instance establishes semantic preservation or conjectural novelty.
RT-RTE-document-add	content transformation	implemented	EPI-OBJ-concept	truth-apt transformation: indeterminate	Source/summary and existing page to concept	see RT-RTE-document-add	see RT-RTE-document-add	Candidate authored text	see RT-RTE-document-add	No additional truth license from fluent generation	see RT-RTE-document-add	see RT-RTE-document-add	SRC-1, openkb/agent/compiler.py	EPI-CLM-current-synthesis	completeness unknown	No generated instance establishes semantic preservation or conjectural novelty.
RT-RTE-document-add	content transformation	implemented	EPI-OBJ-entity	truth-apt transformation: indeterminate	Source/summary and existing page to entity	see RT-RTE-document-add	see RT-RTE-document-add	Candidate authored text	see RT-RTE-document-add	No additional truth license from fluent generation	see RT-RTE-document-add	see RT-RTE-document-add	SRC-1, openkb/agent/compiler.py	EPI-CLM-current-synthesis	completeness unknown	No generated instance establishes semantic preservation or conjectural novelty.
RT-RTE-document-add	check/evidence production	implemented	RT-OBJ-wiki-pages	no content change	Plan shape, body nonemptiness, entity type and link target set	Python parsers/predicates; structural domain only	Before selected writes	Keep/drop/coerce/error; see RT-RTE-document-add	Constrains writes	Only checked structural properties, not factual accuracy	see RT-RTE-document-add	see RT-RTE-document-add	SRC-1, openkb/agent/compiler.py	none	none	Plain non-JSON prose can be accepted; page failures can be omitted.
RT-RTE-document-add	disposition/acceptance	implemented	RT-OBJ-wiki-pages	no content change	Admit generated pages to persistent wiki	Software; structural conditions and exception handling	During compilation	Persist/skip/fail; see RT-RTE-document-add	Programmatic admission	Operational admission, not evidence-consuming semantic acceptance	see RT-RTE-document-add	see RT-RTE-document-add	SRC-1, openkb/agent/compiler.py	EPI-CLM-current-synthesis	admission is not warrant	No per-page human approval or expected-fact oracle in this path.
RT-RTE-document-add	retention	implemented	RT-OBJ-wiki-pages	no content change	Admitted body to filesystem	see RT-RTE-document-add	see RT-RTE-document-add	Persistent pages	see RT-RTE-document-add	Persistence supplies availability, not truth	see RT-RTE-document-add	see RT-RTE-document-add	SRC-1, openkb/agent/compiler.py	RT-CLM-compiled-wiki	none	Backlinks/indexing organize pre-acceptance candidates; not lifecycle integration after acceptance.
RT-RTE-wiki-query	operational admission/selection/consumption	implemented	RT-OBJ-wiki-pages	no content change	Select evidence for answering	see RT-RTE-wiki-query	see RT-RTE-wiki-query	Selected text/pages/images	see RT-RTE-wiki-query	Access to evidence; correctness not licensed	see RT-RTE-wiki-query	see RT-RTE-wiki-query	SRC-1, openkb/agent/query.py	EPI-CLM-grounded-answer	grounding policy	No independent relevance or source-truth check; whole JSON read alternate remains possible.
RT-RTE-wiki-query	content transformation	implemented	EPI-OBJ-answer	truth-apt transformation: indeterminate	Selected wiki to answer	see RT-RTE-wiki-query	see RT-RTE-wiki-query	Candidate final answer	see RT-RTE-wiki-query	Requested citations do not establish support	see RT-RTE-wiki-query	see RT-RTE-wiki-query	SRC-1, openkb/agent/query.py	EPI-CLM-grounded-answer	outcome unknown	No semantic acceptance gate in inspected answering loop.
RT-RTE-session-chat	retention	implemented	EPI-OBJ-answer	no content change	Answer/history/output for later turns	see RT-RTE-session-chat	see RT-RTE-session-chat	Saved history/explorations	see RT-RTE-session-chat	Prior utterance remains prior utterance, not validated fact	see RT-RTE-session-chat	see RT-RTE-session-chat	SRC-1, openkb/agent/chat_session.py	none	none	Image removal/reload changes context availability, not epistemic endorsement.
EPI-RTE-semantic-lint	check/evidence production	implemented	EPI-OBJ-lint-report	truth-apt transformation: indeterminate	Wiki consistency/coverage/staleness	Model with list/read and schema, not external truth oracle	Explicit lint; see record	Issue report or failure text	Advisory diagnostic	Candidate criticism scoped to read pages	No automatic semantic revision/rejection	see record	SRC-1, openkb/agent/linter.py	EPI-CLM-current-synthesis	flags not repairs	Operational route exists; actual discovery quality unobserved.
RT-RTE-wiki-maintain	retention	implemented	EPI-OBJ-lint-report	no content change	Combined lint report to reports directory	see RT-RTE-wiki-maintain	see RT-RTE-wiki-maintain	Report persisted	Diagnostic availability	Not acceptance of report propositions	see RT-RTE-wiki-maintain	see RT-RTE-wiki-maintain	SRC-1, openkb/cli.py	EPI-CLM-current-synthesis	none	Semantic failure can itself be written as report text.
EPI-RTE-link-cleanup	content transformation	implemented	RT-OBJ-wiki-pages	truth-apt transformation: non-ampliative reshaping	Link syntax against existing target names	Deterministic set/normalization; not proposition truth	Compiler/save/fix consumers; see record	Canonical link or plain text	Enforcing link rewrite	Resolvability only; displayed assertion may survive	see record	see record	SRC-1, openkb/lint.py	none	none	Name normalization may merge intended referents; reshaping classification limited to syntactic edge.
EPI-RTE-source-withdrawal	lineage/freshness/recovery	implemented	RT-OBJ-wiki-pages	truth-apt transformation: non-ampliative reshaping	Remove source metadata/reference; optionally whole page	Exact source-list membership and link text; Python	Explicit remove	Delete page or retain body without that citation	Enforcing metadata/page change	No claim-level retraction or reassessment established	see record	see record	SRC-1, openkb/agent/compiler.py	EPI-CLM-current-synthesis	withdrawal scope	Retained multisource prose can outlive removed evidence.
RT-RTE-skill-create	content transformation	implemented	EPI-OBJ-skill-content	truth-apt transformation: indeterminate	Wiki/intent to portable text	see RT-RTE-skill-create	see RT-RTE-skill-create	Candidate instructions/references	see RT-RTE-skill-create	No semantic truth license from distillation	see RT-RTE-skill-create	see RT-RTE-skill-create	SRC-1, openkb/skill/creator.py	RT-CLM-compiled-wiki	none	Non-truth-apt method directions coexist; no instance to classify each proposition.
RT-RTE-skill-create	check/evidence production	implemented	EPI-OBJ-skill-content	no content change	Skill metadata, sizes, links and optional imports	Deterministic validator, structural domain	After creator and explicit validate	Errors/warnings	Diagnostic; see RT-RTE-skill-create	Checked syntax/portability only	see RT-RTE-skill-create	see RT-RTE-skill-create	SRC-1, openkb/skill/validator.py	none	not universal gate	Generator retains files and regenerates marketplace despite findings.
EPI-RTE-eval-set	content transformation	implemented	EPI-OBJ-eval-prompts	truth-apt transformation: indeterminate	Skill text to question/expected-label set; or import saved labels	Model generator; saved-set loader alternative	Requested eval; see record	Labeled prompts	Grader reference selection	Reference labels express intended trigger scope, not factual answers	see record	see record	SRC-1, openkb/skill/evaluator.py	EPI-CLM-body-alignment	reference provenance	Generated labels may share model biases with graders.
EPI-RTE-trigger-grade	check/evidence production	implemented	EPI-OBJ-skill-content	truth-apt transformation: indeterminate	Description-only activation on labeled questions	Model verdict vs supplied/generated expected label	Requested eval	Trigger/no-trigger; ambiguous defaults no-trigger	Diagnostic comparison	Agreement with labels, not executed task correctness	see record	see record	SRC-1, openkb/skill/evaluator.py	EPI-CLM-body-alignment	proxy evaluator	Malformed verdict can count as a correct negative.
EPI-RTE-body-grade	check/evidence production	implemented	EPI-OBJ-skill-content	truth-apt transformation: indeterminate	Body and reference prefixes sufficient for selected questions	Model coverage verdict, no actual task run	Should-trigger questions only	Supported/unsupported/ambiguous/error	Diagnostic report	Judged textual support, not factual truth or demonstrated capability	see record	see record	SRC-1, openkb/skill/evaluator.py	EPI-CLM-body-alignment	preview and proxy	Excludes ambiguous/error cases from denominator; no automated replacement consumer.
RT-RTE-deck-produce	check/evidence production	implemented	RT-RTE-deck-produce	no content change	Expected artifact existence and optional grammar	see RT-RTE-deck-produce	see RT-RTE-deck-produce	Existence/validation result	see RT-RTE-deck-produce	File/grammar license, not knowledge acceptance	see RT-RTE-deck-produce	see RT-RTE-deck-produce	SRC-1, openkb/agent/skill_runner.py	none	none	Conditional validator; final post-critic validity not established.
RT-RTE-deck-produce	content transformation	implemented	RT-RTE-deck-produce	truth-apt transformation: indeterminate	Optional critic revises artifact	see RT-RTE-deck-produce	see RT-RTE-deck-produce	Possibly patched HTML	see RT-RTE-deck-produce	No truth license from critic role alone	see RT-RTE-deck-produce	see RT-RTE-deck-produce	SRC-1, openkb/deck/creator.py	none	none	Whether any edit changes a truth-apt claim cannot be determined without artifact/trace.

The inspected compiler's structural check is explicit (SRC-1, openkb/agent/compiler.py); the supplied RT-RTE-document-add supplies its consequential skip/failure consumers:

def _require_nonempty_content(content, name: str) -> None: """Raise if a concept body is missing or whitespace-only.""" if not isinstance(content, str) or not content.strip(): raise ValueError(f"LLM returned empty content for concept {name!r}") --- openkb/agent/compiler.py

The searched compiler/query/lint/evaluation paths establish operational admission and retention, but no evidence-consuming semantic acceptance decision specifying criterion, intended use and proposition scope. Consequently their index/backlink organization is not established as lifecycle integration after acceptance. This is a bounded finding, not a system-complete negative about every included helper, external host or opaque model process. The positive semantic lint and coverage graders are counterexamples to “OpenKB never checks content”; neither is wired here as a universal truth 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
RT-CLM-compiled-wiki	see record	SRC-1, README.md; doctrine/design	see record	RT-RTE-document-add, RT-RTE-wiki-query	None supplied	No comparison; dependency internals excluded	Persistent structured compilation is wired	Knowledge quality and retrieval accuracy not established
EPI-CLM-current-synthesis	see record	SRC-1, README.md; doctrine/design	see record	RT-RTE-document-add, RT-RTE-wiki-maintain, EPI-RTE-semantic-lint, EPI-RTE-source-withdrawal	None supplied	No corpus-completeness/outcome experiment	Selected old/new pages are synthesized; explicit lint can flag candidate contradictions	“Everything consumed” is stronger than shown coverage; summary bottlenecks, collected failures and partial withdrawal prevent that guarantee
EPI-CLM-grounded-answer	see record	SRC-1, README.md; doctrine/design	Query prompt requests wiki-only cited answers	RT-RTE-wiki-query	None supplied	No groundedness scoring or external truth comparison	Model is given evidence access and grounding instructions	Reliable citation support, completeness and correctness unknown
EPI-CLM-body-alignment	see record	SRC-1, openkb/skill/evaluator.py; doctrine/design	see record	EPI-RTE-eval-set, EPI-RTE-trigger-grade, EPI-RTE-body-grade	None supplied	Shared model, generated labels, truncated references, no task execution	Two distinct diagnostic model judgments and deterministic score aggregation	“Actually answer” is not demonstrated; textual-support judgment is narrower
RT-CLM-wiki-trust	see record	SRC-1, skills/openkb/SKILL.md; doctrine/design	see record	RT-BAP-skill-instructions concerns a separate host/runner path; no deterministic trust filter established	None supplied	External-host compliance outside boundary	Untrusted-data policy is declared	Neither immunity nor host enforcement follows

Bounded conclusion

OpenKB acquires document content with unknown source warrant, then retains model-authored summaries, concepts and entity assertions for later answering and generation. Short compilation receives full converted text; long compilation receives a PageIndex summary and generated overview. The source-preservation request in summary rewrite is real, but it is model guidance rather than a checked semantic invariant. No specific new proposition is established as entailed derivation or ampliative conjecture by the available implementation alone. No warranted premises and checked inference chain establish produced knowledge here.

Its operational authority is stronger than its demonstrated epistemic license. Parser, enum, nonempty-body and filesystem controls can permit, skip or reject writes; no per-page human acceptance is required. Retained bodies become future query/compiler/generator evidence without an intervening demonstrated truth decision. The consumer, channel, force and horizon are those of RT-RTE-document-add, RT-RTE-wiki-query, RT-RTE-session-chat and RT-RTE-skill-create. These are operative routes, not merely storage affordances; actual compliance and benefit remain unobserved.

The material evaluation criteria differ. Semantic lint requests contradictions, gaps, staleness, redundancy and coverage over read wiki pages, returns an advisory report, and has no implemented semantic repair consumer in the inspected drivers. Structural cleanup instead enforces link resolvability and can erase a citation while leaving the assertion. Source removal deletes single-source pages when configured, but multisource bodies need not lose the removed source's factual contribution. Thus citation withdrawal is not equivalent to proposition withdrawal.

Skill evaluation compares description-only activation to expected labels and asks a different model call whether body/reference excerpts support should-trigger questions. These checks supply advisory diagnostics to the CLI, not an admission veto or an answer-truth oracle. Their intended use is trigger and textual-coverage assessment, not measured skill execution. Skill validation likewise checks structure and portability; it does not gate all generated output. Optional deck criticism has a direct work-product write path, but no outcome evidence establishes factual improvement or final validity.

Schema and installed skill text can change future model guidance through RT-BAP-schema-instructions and RT-BAP-skill-instructions. This is instruction-mediated behavioral authority, distinct from knowledge acceptance. Nothing inspected establishes model-parameter training. Persistence and possible future reuse do not establish improved capacity (learning), self-representation with both causal directions (reflection), or method criticism retained into a next method-revision round (theory building). Independent inference/tool autonomy within bounded requests does not supply those missing properties. No exercised self-improvement or causal quality benefit is established; this does not negate the implemented retention, synthesis, diagnostic and artifact-editing routes.

Shared records

Operative objects

EPI-OBJ-source-content

Label: Imported document content

Identity: converted source Markdown, long-document page JSON and associated images used by compiler/query. Natural-language assertions and numerical/tabular/image content, on filesystem; JSON page selectors are symbolic wrappers, not evidence that the assertions are true. Distinct from RT-OBJ-wiki-pages: input/evidence rather than generated summary/concept/entity bodies. Closest supplied route: RT-RTE-document-add; acquisition/extraction details are supplied there. SRC-1, openkb/agent/query.py (doctrine/design and implemented tool construction).

Summaries may omit details — if you need more, follow the summary's full_text frontmatter field to the source (see step 4). --- openkb/agent/query.py

EPI-OBJ-summary

Label: Summary page bodies Part of: RT-OBJ-wiki-pages

Natural-language Markdown assertions retained in wiki summary files. Short-document producer reads converted text; long-document producer includes indexer output plus overview context. Intended role: describe a document's key findings and route deeper reads. Form/substrate inherited from the container, but generation context differs from concepts/entities. SRC-1, openkb/agent/compiler.py (implementation and template doctrine).

content = source_path.read_text(encoding="utf-8") --- openkb/agent/compiler.py

_LONG_DOC_SUMMARY_USER = """\ This is a PageIndex summary for long document "{doc_name}" (doc_id: {doc_id}):

{content}

Based on this structured summary, write a concise overview that captures \ the key themes and findings. This will be used to generate concept pages. --- openkb/agent/compiler.py

The summary-rewrite template requests preservation, rather than checking it (SRC-1, openkb/agent/compiler.py, doctrine/design):

• Preserve every factual claim, finding, and detail from your draft. Do \ NOT add or remove technical content, examples, or claims. --- openkb/agent/compiler.py

EPI-OBJ-concept

Label: Concept page bodies Part of: RT-OBJ-wiki-pages

Natural-language Markdown topic explanations on filesystem, generated or rewritten using current source/summary context and selected existing page. Intended cross-document synthesis differs from one-document summary and named-entity fact accumulation. SRC-1, openkb/agent/compiler.py (template doctrine; entry implementation supplied in RT-RTE-document-add).

New information from document "{doc_name}" (summarized above) should be \ integrated into this page. Rewrite the full page incorporating the new \ information naturally — do not just append. Preserve the existing structure \ and intent of the page. --- openkb/agent/compiler.py

EPI-OBJ-entity

Label: Entity page bodies Part of: RT-OBJ-wiki-pages

Natural-language Markdown facts about a named entity on filesystem; symbolic configured entity types shape admission separately from factual content. SRC-1, openkb/agent/compiler.py (template doctrine and inspected type filtering). Distinct part from concept abstraction.

Integrate the new facts about this entity from document "{doc_name}" (summarized above). Rewrite the full page — do not just append. Preserve the existing structure and intent. Follow the whitelist rules from the message above for all [[wikilinks]].

--- openkb/agent/compiler.py

EPI-OBJ-answer

Label: Query answer and exploration text

Identity: final model answer, optionally saved exploration, and answer turns retained in chat history. Natural-language propositions, initially SDK/string state then optionally Markdown/session storage. Different referent from RT-OBJ-wiki-pages, which identifies compiled bodies, not answers. Closest supplied RT-RTE-wiki-query and RT-RTE-session-chat establish consumer and persistence paths. SRC-1, openkb/agent/query.py; the instruction supplies intent, not compliance.

1. Synthesize a clear, concise, well-cited answer grounded in wiki content. --- openkb/agent/query.py

EPI-OBJ-lint-report

Label: Semantic lint report

Identity: issue claims produced by wiki-linter and combined with structural output under wiki/reports. Natural-language Markdown on filesystem; not the compiled bodies in RT-OBJ-wiki-pages. Closest supplied RT-RTE-wiki-maintain includes this output route but declares no report object. SRC-1, openkb/agent/linter.py, openkb/cli.py (implementation).

return result.final_output or "Knowledge lint completed. No output produced." --- openkb/agent/linter.py

report_content = ( f"# Lint Report — {timestamp}\n\n" f"## Structural\n\n{structural_report}\n\n" f"## Semantic\n\n{knowledge_report}\n" ) report_path.write_text(report_content, encoding="utf-8") append_log(kb_dir / "wiki", "lint", f"report → {report_path.name}") --- openkb/cli.py

EPI-OBJ-skill-content

Label: Skill body, description and references

Identity: SKILL.md activation description and natural-language instructions/body plus Markdown references under generated skill directory. Filesystem storage; YAML metadata is symbolic, description and body are interpreted text. Split consumption: trigger grader receives description only; coverage grader gets body and each reference prefix; installed runner is a further consumer, not the evaluator. Closest supplied RT-RTE-skill-create declares creation route, not this operative object; distinct from compiled wiki input. SRC-1, openkb/skill/evaluator.py.

return f"SKILL.md body:\n{body}\n\nReferences (excerpts):\n{refs}" --- openkb/skill/evaluator.py

EPI-OBJ-eval-prompts

Label: EvalPrompt expected labels

Identity: EvalPrompt question and expected activation label, generated in memory or loaded/saved as JSON under .openkb/eval-sets. Natural-language question plus symbolic trigger/no-trigger label consumed by comparison rules. Distinct from skill content and from supplied RT-RTE-skill-create, whose producer does not establish these evaluation references. SRC-1, openkb/skill/evaluator.py.

@dataclass class EvalPrompt: question: str expected: Literal["trigger", "no-trigger"]

--- openkb/skill/evaluator.py

Routes

EPI-RTE-semantic-lint

Label: Semantic knowledge lint Part of: RT-RTE-wiki-maintain

Endpoints: wiki files/schema to wiki-linter through list/read tools, then final report to CLI/chat/REST lint driver. Model chooses pages and judges candidate issues; host retains report, including exception text on failure. Distinct material check within maintenance container; not semantic disposition or repair. SRC-1, openkb/agent/linter.py, openkb/cli.py (implementation and doctrine/design). Check target: consistency, coverage, possible staleness and redundancy of the pages read. The model proposes findings; code records them, no human approval of each finding or automated factual veto. Mode: explicit open audit, no expected-answer oracle. Report recommendations do not automatically rewrite wiki claims in the inspected consumers.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Markdown issue report or failure/no-output message.
• Later read-back: Combined lint report persists for later inspection; no automatic semantic-fix consumer found in inspected CLI/REST lint drivers.
• Delegated visibility: Lint model sees schema, user audit request and selected list/read tool outputs; no writer registered.
• Selection predicate: Explicit lint invocation; model-selected files; CLI skips report generation when no document hashes.
• Invalidation or expiry: inapplicable — reports are retained diagnostics, not admission tokens with expiry.
• Activation or effect: Report publication is wired; operational use of recommendations unobserved.
• Evidence limits: No trace demonstrates detected contradictions, accuracy, corrective uptake or benefit.

tools=[list_files, read_file],

--- openkb/agent/linter.py

1. Contradictions — Do any pages make conflicting claims about the same fact? --- openkb/agent/linter.py

try: knowledge_report = await run_knowledge_lint(kb_dir, model) except Exception as exc: knowledge_report = f"Knowledge lint failed: {exc}" --- openkb/cli.py

EPI-RTE-link-cleanup

Label: Wikilink cleanup Endpoints: candidate/retained Markdown and valid-target set to strip_ghost_wikilinks; compiler, query-save and lint-fix consumers can persist rewritten text. Python owns target normalization, exact/fuzzy lookup and text substitution; no model judge. Closest supplied RT-RTE-document-add, RT-RTE-wiki-query and RT-RTE-wiki-maintain overlap different callers; this reusable text transformation spans them and is not part of exactly one. SRC-1, openkb/lint.py implementation; caller evidence supplied by those records plus direct fix read. Target/domain: target-name resolvability, not the truth of a link's surrounding sentence. Code deduction, not target execution: with known_targets={concepts/attention}, input “See [[concepts/Missing|a mechanism]].” misses direct and normalized lookup; it becomes “See a mechanism.” and records concepts/Missing as a ghost. With concepts/Attention against that target set, normalized lookup rewrites to concepts/attention. The claim text can remain while its evidential/navigation link is lost. No entailment or semantic endorsement is created.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Cleaned text and ghost-target list; fix wrapper returns changed-file and ghost counts.
• Later read-back: Persisted rewritten pages/explorations are available to later query and compilation.
• Delegated visibility: inapplicable — transformation is deterministic Python, not a delegated model call.
• Selection predicate: Regex wikilinks; supplied known-target set, exact match then normalized match; fix wrapper excludes reports/sources/schema/log by default.
• Invalidation or expiry: Target set is rebuilt from current files for fix; no independent truth-invalidity marker is written.
• Activation or effect: Unresolvable brackets removed or canonical link substituted; fix writes changed files.
• Evidence limits: Deduction from inspected code, not observed execution; normalization is name equivalence, not a referent-identity proof.

ghosts.append(target) if alias: return alias stem = target.rsplit("/", 1)[-1] return stem.replace("-", " ").replace("_", " ")

--- openkb/lint.py

if cleaned != text: atomic_write_text(md, cleaned) files_changed += 1 ghosts_stripped += len(ghosts) --- openkb/lint.py

EPI-RTE-source-withdrawal

Label: Source reference withdrawal Part of: RT-RTE-wiki-maintain

Endpoints: removed document identifier to source-frontmatter/link cleanup in concept/entity files, then delete-or-rewrite and subsequent queries. Python owns exact source membership and filename traversal, not semantic contribution analysis. SRC-1, openkb/agent/compiler.py implementation; caller/retry registry behavior supplied by RT-RTE-wiki-maintain. User initiates/approves remove; code applies changes, with keep_empty controlling the empty-source deletion veto. Intended use is source withdrawal, not criterion-based truth acceptance. Recovery/partial PageIndex failure stays as in the supplied container.

Counterexample to complete factual withdrawal, deduced from inspected implementation: a valid inline sources list with summaries/a.md and summaries/b.md plus body “A says P.” loses only summaries/a.md when a is removed. sources_empty is false, the modified page is written, and this plain prose is unchanged. Thus even correct source metadata cleanup need not retract a's asserted contribution. No target fixture or execution is needed for this deduction.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Modified/deleted slug lists to remove driver.
• Later read-back: Multisource body can remain queryable without the removed citation; deleted pages no longer read by filename.
• Delegated visibility: inapplicable — no semantic model is called by this helper.
• Selection predicate: Concept/entity files referencing the source marker; parsed inline source membership, related-document link and standalone See also patterns.
• Invalidation or expiry: Delete page if last source removed and keep_empty false; otherwise update metadata/references.
• Activation or effect: Filesystem deletion/rewrite changes later available evidence; no proposition-level revalidation is wired here.
• Evidence limits: Malformed/no frontmatter can bypass source-list updates; not a proof of complete cleanup for arbitrary hand-edited pages.

items.remove(source_file) fm_lines[i] = _yaml_list_line("sources", items) return "\n".join(fm_lines) + closing + body, len(items) == 0

--- openkb/agent/compiler.py

if sources_empty and not keep_empty: path.unlink() deleted.append(path.stem) elif new_text != text: atomic_write_text(path, new_text) modified.append(path.stem) --- openkb/agent/compiler.py

EPI-RTE-eval-set

Label: Generate or reload evaluation set Endpoints: skill description/body/reference prefixes to generator model, parsed JSON to expected labels; alternatively saved JSON to load_eval_set. CLI can retain newly generated prompts. Owner: model proposes questions and labels through should_trigger/should_not groups; code parses and records them. SRC-1, openkb/skill/evaluator.py, openkb/cli.py implementation. Closest supplied RT-RTE-skill-create mentions separate diagnostics but does not individuate the reference-generation route; different operation from authoring the skill. Mode: bounded generated/replayed evaluation set, not open artifact generation. Oracle: expected activation labels come from generator or saved-set supplier, not verified task answers. User selects set/count/model; no per-label human approval required. No installed instruction change is admitted by this route.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: List of EvalPrompt questions and expected labels.
• Later read-back: Optional CLI save writes .openkb/eval-sets/.json; --eval-set reloads labels for later diagnostic runs.
• Delegated visibility: Generator gets description, entire extracted body and reference previews; no wiki/tool-executed task result.
• Selection predicate: Fresh generation unless caller supplies eval_set; CLI --save only saves newly generated sets.
• Invalidation or expiry: Explicit file replacement/replay; no automatic invalidation when skill changes established.
• Activation or effect: Expected labels select coverage prompts and reference trigger comparisons; no automatic successor installation.
• Evidence limits: No label accuracy or actual generation trace supplied; defaults use same configured model family for generator and graders.

try: data = json.loads(raw) except json.JSONDecodeError as exc: raise RuntimeError( f"Eval set generator returned non-JSON output: {exc.msg}. " f"Try a more capable model — small models often ignore " f"'output only JSON' instructions. First 200 chars: {raw[:200]!r}" ) from exc prompts: list[EvalPrompt] = [] for q in data.get("should_trigger", []): prompts.append(EvalPrompt(question=q, expected="trigger")) for q in data.get("should_not", []): prompts.append(EvalPrompt(question=q, expected="no-trigger")) return prompts

--- openkb/skill/evaluator.py

if save_flag and eval_set is None: path = save_eval_set(kb_dir, name, result.prompts) --- openkb/cli.py

EPI-RTE-trigger-grade

Label: Description trigger grading Endpoints: description/question to trigger-grader model, parsed verdict to deterministic expected-label comparison, aggregate metrics/misses to CLI. Model proposes verdict; code prioritizes NO-TRIGGER then TRIGGER substring and otherwise returns no-trigger. No runtime skill is executed. Distinct from RT-RTE-skill-create's creation/validation and from body support judgment; no identical supplied route after comparison. SRC-1, openkb/skill/evaluator.py, openkb/cli.py implementation. Check target/domain: whether description suggests loading for each labeled question. Expected-label oracle supplied by EPI-RTE-eval-set; not factual answer truth. Mode: bounded evaluation, human chooses invocation/set, software reports, neither vetoes installation. Erroring tasks excluded from scoring. Concrete deduction: grader output “uncertain” becomes no-trigger and counts as passing for a no-trigger reference; this parser fallback is not evidence that the model correctly judged activation.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Trigger/no-trigger verdict; run_eval accumulates misses or errors and CLI prints scores.
• Later read-back: inapplicable — CLI retains optional prompt sets, not automatic diagnostic-driven skill revisions.
• Delegated visibility: Each independent grader receives description only and one question, not body, references or execution trace.
• Selection predicate: All eval-set prompts; expected labels used for mismatch comparison.
• Invalidation or expiry: inapplicable — result is a diagnostic on the current read description, not persistent execution permission.
• Activation or effect: Verdicts change diagnostic counts/miss list; no automatic skill admission or successor selection consumer.
• Evidence limits: Label agreement and substring parsing do not measure achieved task outcomes or factual correctness.

Default: assume no-trigger on ambiguous output

return "no-trigger" --- openkb/skill/evaluator.py

if graded != prompt.expected: result.misses.append(EvalMiss(prompt=prompt, graded=graded)) --- openkb/skill/evaluator.py

EPI-RTE-body-grade

Label: Body support grading Endpoints: current SKILL.md body plus reference prefixes and should-trigger question to coverage-grader model; parsed verdict/reason to aggregates, then CLI diagnostic. Check target/domain: judged availability of substantive textual material for that question, not correctness of an executed answer. SRC-1, openkb/skill/evaluator.py, openkb/cli.py implementation and prompt doctrine. Closest supplied RT-RTE-skill-create includes diagnostic mention but no separately individuated body checker; distinct from trigger checker and creation admission. Model judges, code parses UNSUPPORTED before SUPPORTED; ambiguous and exceptions are separated and excluded from coverage denominator. User chooses invocation; no per-verdict approval or automated skill rewrite/veto. Mode: bounded diagnostic; no independent answer-reference oracle. Each reference contributes at most 4000 Python text characters despite the constant/docstring saying bytes. Excerpts omit later content, so unsupported can reflect a preview limit. A supported verdict remains a model assertion, not proof of truth or successful task execution.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: supported/unsupported/ambiguous plus reason; coverage metrics and gaps printed.
• Later read-back: inapplicable — no automated revised skill or verdict persistence consumer in the inspected CLI evaluator.
• Delegated visibility: Full extracted body, reference prefixes and single question, not trigger grader hidden reasoning or a skill execution.
• Selection predicate: Only EvalPrompt entries expected to trigger; reference Markdown recursively sorted and sliced per-file.
• Invalidation or expiry: inapplicable — current-file diagnostic, not an acceptance token reused by runtime.
• Activation or effect: Coverage verdict changes displayed rate/gap list; diagnostic only.
• Evidence limits: No run or outcome comparison; no facts outside supplied text assessed; denominator omits ambiguous/error judgments.

if len(text) > REFERENCES_PREVIEW_BYTES: text = text[:REFERENCES_PREVIEW_BYTES] + "\n…[truncated]\n"

--- openkb/skill/evaluator.py

"description claims to handle. Decide whether the body has " "substantive material to answer the question.\n\n" --- openkb/skill/evaluator.py

coverage_prompts = [p for p in eval_set if p.expected == "trigger"]

--- openkb/skill/evaluator.py

scored = self.trigger_questions - len(self.coverage_ambiguous) - len(self.coverage_errors) return self.coverage_passed / scored if scored else 0.0 --- openkb/skill/evaluator.py

Claims

EPI-CLM-current-synthesis

Label: Current comprehensive synthesis

Claimed operation: accumulated wiki stays current, flags contradictions and reflects everything consumed. SRC-1, README.md doctrine/design. Distinct stronger coverage/warrant claim than RT-CLM-compiled-wiki's basic compilation product claim. No source statement is treated as proof of quality.

Traditional RAG rediscovers knowledge from scratch on every query. Nothing accumulates. OpenKB compiles knowledge once into a persistent wiki, then keeps it current. Cross-references already exist, contradictions are flagged, and synthesis reflects everything consumed. --- README.md

EPI-CLM-grounded-answer

Label: Grounded cited answers

Claimed operation: produce grounded cited answers, rather than merely emit text after retrieval. SRC-1, README.md doctrine/design. Distinct output-warrant claim from RT-CLM-compiled-wiki; no counterpart claim in supplied runtime after comparison.

| openkb query "question" | A grounded answer with citations (--save to persist to wiki/explorations/) | query & save |

--- README.md

EPI-CLM-body-alignment

Label: Skill body alignment

Claimed operation: body-alignment evaluation tests whether skill can actually answer what it promises. SRC-1, openkb/skill/evaluator.py doctrine/design. Closest RT-RTE-skill-create describes diagnostic admission limits but does not declare this claim; distinct from wiki product/trust claims.

Body alignment — given the full SKILL.md (body + references), can the skill actually answer the should-trigger questions it claims to handle? This catches the failure mode where a well-written description promises capability that the body doesn't deliver — a hollow skill that would trigger but fail in practice. --- openkb/skill/evaluator.py