Funes epistemic report

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

Source-and-claim boundary

The question is how Funes turns supplied material into content on which later answers may rely, and what warrants that reliance. The frozen source and exclusions are those in the Source register. Funes includes the instruction-mediated Librarian loop, not just its Python helpers. Semantic work is assigned to an external agent; executable code supplies extraction and mechanical audits. Public claims are RT-CLM-library-guarantees, EPI-CLM-library-abstraction and EPI-CLM-compile-depth.

The following table distinguishes direct inspection for this member from supplied findings. All direct reads use SRC-1 at the reviewed commit. Dispatch was checked against the tracked tree: the four Python main entry points and the CI workflow exhaust the shipped executable operations. No additional registered tools, hooks or model driver were found in that tree; the supplied bounded search is RT-ABS-model-driver.

Entry point or operation	Source path	Coverage and records
Agent entry and local scope	AGENTS.md, starter-library/AGENTS.md	Supplied findings: RT-OBJ-protocol, RT-BAP-librarian-instructions. Host delivery remains unobserved.
Ingest, compilation, registry status, backlinks	protocol.md	Direct complete read: RT-RTE-ingest-compile, EPI-OBJ-raw, EPI-OBJ-wiki.
Wiki Q&A, output retention and filing-back	protocol.md	Direct: RT-RTE-wiki-answer, EPI-OBJ-outputs.
Contradiction, grounding, cleanup and proposed fixes	protocol.md, tools/README.md, tools/compile_lint.py	Direct complete reads: EPI-RTE-fidelity-review, RT-RTE-health-audit.
Link, orphan, frontmatter CLI	tools/check_links.py	Direct complete read, including exclusions and exit: EPI-RTE-mechanical-checks, EPI-RTE-audit-disposition.
Duplicate, echo, depth CLI and strict mode	tools/compile_lint.py	Direct complete read: same two records; EPI-CLM-compile-depth.
CI caller and syntax checking	.github/workflows/funes-lint.yml	Direct complete read: EPI-RTE-audit-disposition. Syntax success concerns Python syntax, not wiki truth.
PDF native/OCR selection, partial failure, force and no-header	tools/pdf_text.py, tools/README.md	Direct complete reads: RT-RTE-pdf-sidecar, EPI-OBJ-pdf-text.
Caption list/fetch, track selection, formatting and output overwrite	tools/youtube_transcript.py, tools/README.md	Direct complete reads: RT-RTE-caption-fetch, EPI-RTE-caption-reshape.
Provisioning and persistent local instructions	library.md	Direct complete read; RT-RTE-library-setup. This creates policy and organization, not verified domain knowledge.
Starter indices, source/concept/topic scaffolds, health/output/changelog folders	starter-library/	Supplied complete coverage: RT-OBJ-library-tree; no populated library instance.
Public design and examples	README.md	Direct complete read: EPI-CLM-library-abstraction. Examples are illustrative doctrine, not observed outcomes.

Excluded host/provider internals prevent a conclusion about the model's internal inference, parameter updates, actual instruction loading or deployment enforcement. External OCR, Poppler and caption-service internals prevent an extraction-fidelity conclusion. No supplied run, populated library or causal experiment establishes actual checking, acceptance, integration, retrieval accuracy or benefit. Linked external examples remain excluded. These are not reasons to erase the implemented wrapper and audit routes. There is no relevant shipped operation left unassessed in the coverage above; semantic content of future arbitrary articles remains not determinable. Evidence: SRC-1, README.md, protocol.md, tools/pdf_text.py, tools/youtube_transcript.py; supplied boundary.

Epistemic-object inventory

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
EPI-OBJ-raw	Imported assertions, quotations and observations, when present	see record	SRC-1, protocol.md	Preservation of a source is not endorsement of its claims.
EPI-OBJ-pdf-text	Machine transcription of source propositions; page/method metadata	see record	SRC-1, tools/pdf_text.py	Recognition fidelity unknown; original is authoritative.
EPI-OBJ-captions	Imported speech transcription; track and timing metadata depending on format	see record	SRC-1, tools/youtube_transcript.py	Transcription truth and spoken-claim truth are separate unknowns.
EPI-OBJ-wiki	Source summaries, concepts, evidence/caveats and topic relations	see record	SRC-1, protocol.md	No actual populated candidate or semantic trace; relation to source is indeterminate.
EPI-OBJ-outputs	Answers and syntheses may assert propositions; routines/plans can instead prescribe actions	see record	SRC-1, protocol.md	Intended reuse does not establish acceptance.
EPI-OBJ-audit-results	Bounded assertions about paths, text overlap and counts; editorial diagnoses when authored	see record	SRC-1, tools/check_links.py, tools/compile_lint.py, protocol.md	Mechanical and editorial evidence differ.
RT-OBJ-library-tree	Registry and changelog report state/history; navigation mostly selects content	see record	SRC-1, protocol.md	compiled records workflow status, not a truth certificate.
RT-OBJ-protocol	Mostly operating policy, not a candidate domain claim	see record	SRC-1, protocol.md, library.md	Persistent instruction text has requested operational authority.
RT-CMP-librarian-model	Internal truth-apt content not determinable	see record	SRC-1, README.md	Distributed-parametric internals inaccessible; fixity and mutation findings belong to supplied component.

Authority-route ledger

Each row assesses one function, not one script. Multiple rows may cite a valid container record without superseding it. see defers record-owned details; architectural status and the bounded epistemic assessment are this overlay's fields. No row asserts observed operation.

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-ingest-compile — raw ingest	content transformation	doctrine only	EPI-OBJ-raw	truth-apt transformation: acquisition/import	Supplied source to raw record	see route	see route	Requested verbatim acquisition	Instruction only	Source warrant unknown; requested preservation of wording and origin	Agent may save raw material before semantic review	see route and RT-BAP-librarian-instructions	SRC-1, protocol.md	RT-CLM-library-guarantees	none	Actual preservation unobserved
RT-RTE-pdf-sidecar — extraction	content transformation	implemented	EPI-OBJ-pdf-text	truth-apt transformation: acquisition/import	PDF selected pages to searchable text	see route	see route	Native/OCR text or marked page failure	File write, see route	Warrant unknown; recognition may degrade content; header and original permit checking	Searchable derivative can exist despite exit 1	see route	SRC-1, tools/pdf_text.py	RT-CLM-library-guarantees	fail-fast is not all-or-nothing	Dependency accuracy uninspected
RT-RTE-pdf-sidecar — text admission	disposition/acceptance	implemented	EPI-OBJ-pdf-text	no content change	Native layer versus OCR; page output versus failure marker	Character threshold, command outcome and nonempty text; see route	Per selected page before final write	Operational admission or fallback/error	Local control flow	No factual acceptance; character count is not accuracy evidence	Native text accepted for output, OCR attempted otherwise; partial output retained	Wrapper branches, enforcing for current command; see route	SRC-1, tools/pdf_text.py	none	none	A long erroneous native layer can pass
RT-RTE-caption-fetch — track acquisition	content transformation	implemented	EPI-OBJ-captions	truth-apt transformation: acquisition/import	Chosen video and track to fetched snippets	see route	see route	Captions or service/dependency error	see route	Source warrant unknown; manual/generated selection is not truth verification	Caller-selected material can be emitted/saved without claim review	see route	SRC-1, tools/youtube_transcript.py	none	none	Listing emits metadata, not speech claims
EPI-RTE-caption-reshape	content transformation	implemented	EPI-OBJ-captions	truth-apt transformation: non-ampliative reshaping	Snippet text to plain-text serialization	see record	see record	Whitespace-normalized lines; optional start timestamps	see record	No new speech-content warrant; timing/provenance reduced in default text	Rendered text available to caller or file consumer	see record	SRC-1, tools/youtube_transcript.py	none	none	Domain excludes interpretation of layout-dependent content
RT-RTE-ingest-compile — source/concept production	content transformation	doctrine only	EPI-OBJ-wiki	truth-apt transformation: indeterminate	Raw sources to summaries, concepts and topic explanations	Librarian interpretation; see route	see route	Requested source-specific abstraction	Instruction only	Neither entailment nor non-entailment established for an actual output	Agent authors/merges wiki without per-claim human approval	see route	SRC-1, protocol.md	EPI-CLM-library-abstraction	none	Need actual inputs/outputs to distinguish reshaping, derivation and conjecture
RT-RTE-ingest-compile — retention/organization	retention	doctrine only	EPI-OBJ-wiki	no content change	Compiled material to files, indices, backlinks and registry	see route	During compile	Requested persistent candidate and compiled status	Instruction only	Filing and linking confer availability, not acceptance	Later sessions can select content through wiki navigation	see route	SRC-1, protocol.md	RT-CLM-library-guarantees	none	Organization before evidenced acceptance is not lifecycle integration
RT-RTE-wiki-answer — reliance	operational admission/selection/consumption	doctrine only	EPI-OBJ-wiki	no content change	Question to wiki-first selection	see route	Later Q&A	Requested retrieved context and explicit noncoverage	Instruction only	Citation supplies traceability, not truth license beyond source support	Wiki can guide answers before any separate acceptance decision	see route	SRC-1, protocol.md	EPI-CLM-library-abstraction	none	Reliance and activation unobserved
RT-RTE-wiki-answer — answer production	content transformation	doctrine only	EPI-OBJ-outputs	truth-apt transformation: indeterminate	Selected wiki to answer/synthesis	see route	At question time	Requested cited outputs	Instruction only	No demonstrated entailed derivation or accepted ampliative claim	Answer/report can reach human and be saved	see route	SRC-1, protocol.md	EPI-CLM-library-abstraction	none	Plans/routines may be non-truth-apt prescriptions instead
RT-RTE-wiki-answer — saving/filing	retention	doctrine only	EPI-OBJ-outputs	no content change	Substantial output to file; durable parts offered for wiki	see route	After answering	Requested persistence and human-facing offer	Instruction only	Durability judgment is not a truth criterion	Human can decline filing offer; saved output already remains available	see route	SRC-1, protocol.md	none	none	No evidence-consuming epistemic acceptance specified for filing
EPI-RTE-mechanical-checks	check/evidence production	implemented	EPI-OBJ-audit-results	truth-apt transformation: entailed derivation	Paths/text predicates within inspected parser domain	see record	On audit invocation	Bounded findings/counts	Produces diagnostics; disposition separate	Conditional warrant for computed predicates, not source truth	No article writes; result made available to caller	see record	SRC-1, tools/check_links.py, tools/compile_lint.py	EPI-CLM-compile-depth	proxy is not semantic extraction proof	File-read fidelity and intended parser interpretation remain premises
EPI-RTE-audit-disposition	disposition/acceptance	implemented	EPI-OBJ-audit-results	no content change	Findings to CLI success/failure and CI step status	see record	After checks	Gate failure or success; advisory display	see record	Operational pass only, not epistemic acceptance of articles	CI step can fail; merge veto not established	see record	SRC-1, tools/compile_lint.py, .github/workflows/funes-lint.yml	EPI-CLM-compile-depth	none	CI omits strict mode
EPI-RTE-fidelity-review — comparison	check/evidence production	doctrine only	EPI-OBJ-wiki, EPI-OBJ-raw	no content change	Compiled article/source fidelity; contradictions and gaps	see record	Editorial review, not a wired pre-write gate	Requested criticism or finding	Instruction/design only	Potential evidence about fidelity; no recorded decision here	Reviewer may recommend correction	see record	SRC-1, tools/compile_lint.py, protocol.md	RT-CLM-library-guarantees	none	This positive editorial route refutes a blanket no-checking verdict
RT-RTE-health-audit — editorial disposition	disposition/acceptance	doctrine only	EPI-OBJ-audit-results, EPI-OBJ-wiki	no content change	Health findings to safe fixes or human proposals	see route	On request/periodically	Requested fixes, proposals and dated report	Instruction only	No universal truth-acceptance criterion or recorded acceptance instance	Librarian decides safe fixes; human can reject proposed remainder	see route	SRC-1, protocol.md	none	none	Rejection/withdrawal of a false claim is possible editorially, not automatically traced to all dependent outputs
RT-RTE-health-audit — editorial revision	content transformation	doctrine only	EPI-OBJ-wiki	truth-apt transformation: indeterminate	Existing claims to revised article	see route	After diagnosis	Requested corrected/merged content	Instruction only	Whether revision preserves, derives or conjectures is undetermined without content	Changed article may guide future Q&A	see route	SRC-1, protocol.md	none	none	No observed rejection, rollback or downstream withdrawal
RT-RTE-library-setup	behavior/policy adaptation	doctrine only	RT-OBJ-protocol	non-truth-apt policy/content update: new library scope and deviations	Human conventions to local instructions and structure	see route	Human-requested setup	Requested retained local AGENTS	Instruction only	No domain-truth acceptance	Later agent may follow new scope/conventions	see route and RT-BAP-librarian-instructions	SRC-1, library.md	none	none	Not an evidenced evaluation-fed method improvement

No separate post-acceptance lifecycle-integration route is established by the inspected protocol.md, library.md, helpers and CI. Backlinks, index updates and output filing are real prescribed organizational work, but the available evidence places no recorded evidence-consuming truth acceptance before them. This is a boundary on that classification, not a denial of organization or future editorial acceptance. Supplied recovery details, including original preservation, optional force/overwrite, Git history and lack of automatic rollback, remain on RT-RTE-pdf-sidecar, RT-RTE-caption-fetch, RT-RTE-ingest-compile and RT-RTE-health-audit.

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-library-guarantees	see record	SRC-1, protocol.md, README.md; doctrine/design	Raw preservation, relative source links, depth and editorial grounding rules	RT-RTE-pdf-sidecar, EPI-RTE-mechanical-checks, EPI-RTE-audit-disposition provide narrower supports	None supplied	No causal comparison; file capability bypasses protocol; link existence does not check support	A prescribed lineage path, with bounded mechanical detection	Universal immutability and truth grounding are not implemented invariants
EPI-CLM-library-abstraction	see record	SRC-1, README.md; doctrine/design	RT-RTE-ingest-compile, RT-RTE-wiki-answer, RT-RTE-health-audit	Extraction wrappers and audits, not a semantic compiler	None supplied; README examples illustrative	No retrieval/benefit experiment; model interpretation and source quality unresolved	Coherent requested retained-library workflow	No produced accepted new knowledge instance or demonstrated sustained coherence
EPI-CLM-compile-depth	see record	SRC-1, tools/compile_lint.py, tools/README.md; doctrine/design plus attributed reported operation	Explicit gates/advisory separation and editorial fidelity fallback	EPI-RTE-mechanical-checks, EPI-RTE-audit-disposition	Historical failure account only; no inspectable run	Proxies not necessary/sufficient semantic tests; no independent comparison	Can detect selected duplicate/echo patterns and flag thin prose	Distinct fabricated prose can clear duplicate tests without reading a source; editorial fidelity is only doctrine

Bounded conclusion

Funes prescribes source acquisition and retention, abstraction into a linked wiki, wiki-first answering, output saving and editorial repair. Raw retention preserves an evidential route if followed; it does not make imported claims true. PDF and caption helpers implement acquisition, selection and serialization, with explicit but defeasible lineage aids. Default caption text loses timing/track context from the content channel; PDF defaults preserve page/method information but permit partial output and optional header omission. Source warrant is unknown, not upgraded by successful extraction. See RT-RTE-pdf-sidecar, RT-RTE-caption-fetch, EPI-RTE-caption-reshape.

Mechanical audits derive bounded propositions from inspected file text and path predicates. Their downstream force is local command/CI status, with advisory-only depth results in the shipped caller. Their intended use is repository-health triage; their scope is exactly the parsed graph/sections/counts. They neither certify source truth nor force branch protection. See EPI-RTE-mechanical-checks, EPI-RTE-audit-disposition.

Semantic comparison is not absent: the protocol requests contradiction/grounding review, and the linter design explicitly calls for reading source and article together. Its epistemic target is fidelity and its operational path is editorial criticism followed by safe fixes or proposals. But no supplied comparison result, named-use acceptance decision or ensuing accepted-claim integration is observed. Requested summaries and syntheses cannot yet be classified as entailed derivations or ampliative conjectures; neither entailment nor non-entailment has been shown for an actual produced claim. See EPI-RTE-fidelity-review, RT-RTE-health-audit.

The wiki is licensed by the protocol for operational reliance in later answers without an intervening separately recorded truth-acceptance gate. Cross-library approval controls linking, not the truth of each linked claim. Output filing offers human control over reuse, not an explicit evidence-consuming truth test. Persistent instructions created during setup may alter later behavior without a truth-apt update. See RT-RTE-wiki-answer, RT-RTE-ingest-compile, RT-RTE-library-setup.

No run evidence establishes activation or improved future capacity. Model parameter change remains uninspected, while file-based retention remains a positive independent capacity. Editorial revision is prescribed; content-directed criticism, blame placement and iterative changed reliance for a particular formulated theory are not evidenced here. That limit does not collapse memory, autonomy, reflection or improvement into one system-wide verdict. No exercised evidence-responsive change to Funes's own method and later causally dependent operation is demonstrated. See RT-CMP-librarian-model, RT-BAP-librarian-instructions, EPI-RTE-fidelity-review.

Shared records

Operative objects

EPI-OBJ-raw

Label: Raw source record

Part of: RT-OBJ-library-tree

Identity: verbatim imported source documents under raw/, excluding the mutable registry. Natural-language source content and binary originals live in repository files. Claimed role is immutable evidence for compilation and later checking. Distinct from extracted sidecars: the original remains authoritative. Imported assertions can be truth-apt; factual source warrant is unknown. SRC-1, protocol.md, doctrine/design.

• raw/ — verbatim sources; never edit content. --- protocol.md

EPI-OBJ-pdf-text

Label: PDF extracted text edition

Part of: RT-OBJ-library-tree

Identity: PDF-derived text sidecar described by the protocol and implemented by the helper. Natural-language transcription with symbolic page/method boundaries and provenance text; filesystem text output. Default lineage includes filename, date, settings and versions; header omission is allowed. Different from the raw binary and compiled wiki. Content is a derivative representation, not independently checked testimony. SRC-1, tools/pdf_text.py, implementation; protocol.md, doctrine/design. Closest supplied route: RT-RTE-pdf-sidecar.

"Machine-generated text; verify critical readings against the PDF.", --- tools/pdf_text.py

EPI-OBJ-captions

Label: Fetched captions and track metadata

Identity: result of RT-RTE-caption-fetch, not identical to the entire managed library: output may only go to stdout. Natural-language snippet text, symbolic video/language/generated fields and numerical timestamps live in transient objects or optional output files. JSON preserves a structured provenance envelope; default text does not embed it. Different from PDF transcription, though both are derived source representations. SRC-1, tools/youtube_transcript.py, implementation.

"video_id": transcript.video_id, "language": transcript.language, "language_code": transcript.language_code, "generated": bool(transcript.is_generated), "snippets": transcript.to_raw_data(), --- tools/youtube_transcript.py

EPI-OBJ-wiki

Label: Compiled source notes concepts and topics

Part of: RT-OBJ-library-tree

Identity: wiki/sources/, wiki/concepts/, wiki/topics/ content, distinct from source originals and generated outputs. Natural-language summaries/explanations/maps plus symbolic links and frontmatter, stored as Markdown. Semantic content enters through the external Librarian, not a shipped compiler. Claimed role: later question-answering context. SRC-1, protocol.md, doctrine/design. Minimum production passage is supplied in RT-RTE-ingest-compile; consumption passage in RT-RTE-wiki-answer.

EPI-OBJ-outputs

Label: Generated outputs

Part of: RT-OBJ-library-tree

Identity: substantial generated answers/reports/plans/routines saved to outputs/; durable portions may be offered for wiki filing. Natural-language prose and prescriptions plus symbolic citation links in Markdown files. Different consumer/admission horizon from wiki articles: initially human-facing, with optional later reuse. Not every prescription is truth-apt. SRC-1, protocol.md, doctrine/design; passage and closest route: RT-RTE-wiki-answer.

EPI-OBJ-audit-results

Label: Mechanical diagnostics and editorial health reports

Identity: audit CLI stdout/status and requested dated meta/health/ reports. This crosses RT-RTE-health-audit and RT-OBJ-library-tree, rather than being wholly part of the file tree: CLI status need not persist there. Symbolic findings/counts/status have fixed code consumers; prose editorial reports require model/human interpretation. Storage is stdout/CI logs for mechanical results and requested Markdown for editorial reports. Claimed role is triage, not fact certification. SRC-1, tools/check_links.py, tools/compile_lint.py, implementation; protocol.md, doctrine/design. Consumer details and passages are on the routes below.

Routes

EPI-RTE-mechanical-checks

Label: Mechanical health evidence production

Part of: RT-RTE-health-audit

Identity: checking phase only, distinct from exit-based disposition. Python reads the selected root/library Markdown and computes findings. The link target/domain is existence of parsed relative destinations, not actual support for cited propositions. Remote URLs and pure anchors are accepted without checking; fences, template files and documented missing-binary cases are excluded as coded. Orphan target is absence of a basename occurrence elsewhere in the library corpus; frontmatter target is opening ---, not YAML validity.

Compile targets are normalized n-gram overlap in selected source summaries/concept leads, linked takeaway echoes, lead word counts and a selected raw/wiki word ratio. No source-fact oracle is consulted. Code decides matches under caller-selectable numeric thresholds; user or CI selects invocation, not individual findings. Ordinary maintenance mode; triggered by request or push/PR, not a controlled experiment. No product edit is proposed or admitted by this checking phase. SRC-1, tools/check_links.py, tools/compile_lint.py, implementation.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: diagnostic findings and counts, subsequently consumed by the disposition phase.
• Later read-back: inapplicable — the check phase retains no knowledge or policy update for later agent invocation.
• Delegated visibility: inapplicable — local Python computation; no delegated agent or model call.
• Selection predicate: root/library discovery and parser-specific file/section filters, with optional orphan/frontmatter flags.
• Invalidation or expiry: fresh computation on invocation; no persistent certificate or incremental invalidation.
• Activation or effect: diagnostics produced without modifying content; status consumer is EPI-RTE-audit-disposition.
• Evidence limits: no execution; derived findings are conditional on the file contents and coded parser domain, not semantic truth.

if dest.startswith(("http://", "https://", "mailto:", "#")): return True # external or pure in-page anchor — not our concern --- tools/check_links.py

if not open(p, encoding="utf-8").read().startswith("---"): missing.append(p) --- tools/check_links.py

if overlap(items[i][1], items[j][1]) >= pair_threshold: pair_findings.append(sorted([items[i][0], items[j][0]])) --- tools/compile_lint.py

A deterministic counterexample to semantic certification follows directly from these predicates, without target execution. Two summaries A cactus is a planet. and Whales are made of copper. tokenize into disjoint short token sequences. ngrams returns one joined token string per summary when fewer than eight tokens; overlap is zero. Two items do not activate the five-owner template condition. With no concept-linked lines, these summaries produce no echo hits. Therefore the duplicate/echo gates can clear distinct invented assertions without any source read by the linter. This refutes the literal implication that clearing the template test necessitates reading sources, not the utility of duplicate detection. It licenses no claim about actual writer behavior. Counterevidence considered: strict mode can fail depth advisories, but still adds no source-to-claim comparison.

EPI-RTE-audit-disposition

Label: Audit exit and CI disposition

Part of: RT-RTE-health-audit

Identity: findings to exit-code phase plus the shipped GitHub Actions consumer, not the checking phase. Python decides CLI success/failure. Link findings and compile duplicate/echo findings fail; compile depth advisories fail only with --strict. The shipped CI caller requests orphan/frontmatter checks and non-strict compile lint. GitHub consumes the step result within its audit job. This changes operational continuation/status, not truth endorsement or article retention. SRC-1, tools/check_links.py, tools/compile_lint.py, .github/workflows/funes-lint.yml, implementation.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: CLI exit 0/1 and CI step outcome if invoked by the workflow.
• Later read-back: inapplicable — no automatic retained instruction or article update is written by this phase.
• Delegated visibility: CI checks out repository files and receives subprocess status; no model delegate.
• Selection predicate: counted problems and strict flag; CI selects the non-strict compile policy.
• Invalidation or expiry: rerun on invocation/push/PR; prior result is not a lasting semantic approval.
• Activation or effect: nonzero command status signals a failing audit step; advisories remain visible without failure in ordinary CI.
• Evidence limits: no job trace or branch-protection configuration; no merge veto, rollback or writer compliance is established.

No answer oracle or semantic acceptance criterion exists in this phase. Owner is the code rule and CI runner; a human can change invocation/policy, but there is no per-claim human admission. Trigger is maintenance, with no article-changing proposal, deployment or rollback in these read-only scripts.

if args.strict: return 1 if (gate_problems or advisory_flags) else 0 return 1 if gate_problems else 0 --- tools/compile_lint.py

• name: Check Markdown links, orphans, and frontmatter run: python3 tools/check_links.py --orphans --front
• name: Check compile depth run: python3 tools/compile_lint.py --- .github/workflows/funes-lint.yml

EPI-RTE-fidelity-review

Label: Editorial source fidelity review

Part of: RT-RTE-health-audit

Identity: semantic evidence-production part of editorial health review, not automatic mechanical checking or subsequent fix admission. The linter design calls for adversarial LLM comparison of article and source; protocol assigns grounding/contradiction review to editorial work. Target/domain is fidelity of the compiled article to the source, not truth of every source assertion in the world. It supplies a candidate comparator but no fixed gold answer or calibrated oracle. No additional review runtime is shipped. SRC-1, tools/compile_lint.py, tools/README.md, protocol.md, doctrine/design. Closest later disposition is RT-RTE-health-audit.

• instruction conclusion status: claimed
• operation conclusion status: uninspected
• Immediate return: requested editorial finding/health report or criticism; no observed result supplied.
• Later read-back: requested dated health report may reach later review; no trace demonstrates reuse or changed reliance.
• Delegated visibility: uninspected — adversarial LLM review is named but no separate context/runner is implemented.
• Selection predicate: health request/periodicity and editorial judgment of duplicates, contradictions, grounding and gaps.
• Invalidation or expiry: inapplicable — no stored fidelity certificate or expiry contract specified.
• Activation or effect: requested findings can guide safe fixes or human proposals, not an enforcing semantic veto.
• Evidence limits: no paired source/article review result, decision record, criterion threshold or downstream acceptance instance.

The agent diagnoses; ordinary safe fixes are admitted by the Librarian, other changes are proposed to the human under the parent route. Proposed content persists only if subsequently edited/saved. Review guidance is source specificity, caveats, source links and repair rather than padding. Git/editorial redo afford recovery, not automatic dependent-claim withdrawal. No supplied theory-specific criticism names a challenged theory, test result and blame placement; the generic review mechanism does not establish those events or iteration.

Fidelity is the job of an adversarial LLM review that actually reads source and article together — a process step, not a regex. See tools/README.md. --- tools/compile_lint.py

contradiction, grounding, and gap review beyond them remains editorial. --- protocol.md

EPI-RTE-caption-reshape

Label: Caption plain text reshaping

Part of: RT-RTE-caption-fetch

Identity: local plain-text branch after acquisition, not remote selection or extraction. Python normalizes each snippet's whitespace, skips empty snippets, optionally prefixes its start timestamp and joins lines. Input is returned snippet text; output is the text fed to write_output, with track metadata separately printed to stderr. Within ordinary prose transcription this is non-ampliative reshaping, not generation of a new speech claim. SRC-1, tools/youtube_transcript.py, implementation.

Concrete implementation deduction, not execution: snippet text Water\n boils. becomes Water boils. through split and join. An all-whitespace snippet is skipped. With timestamps disabled, start/duration do not appear in the text body. Thus the speech wording is reshaped while timing context is removed. JSON instead retains the raw-data envelope; SRT/VTT formatting is delegated to the external package and its exact transformation is uninspected. Layout-sensitive meanings remain outside this bounded preservation claim.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: rendered text passed to stdout/file writer, with separate stderr metadata.
• Later read-back: possible through caller-saved/ingested output; no automatic library registration or later delivery.
• Delegated visibility: inapplicable — this branch is local serialization, not an agent call.
• Selection predicate: text format, nonempty normalized snippets, optional timestamps flag.
• Invalidation or expiry: inapplicable — no retained cache validity rule; caller can refetch/rewrite output.
• Activation or effect: transformed transcript becomes caller-visible content; no factual approval or wiki write.
• Evidence limits: upstream recognition/selection is external and unobserved; formatting cannot validate spoken propositions.

The caller selects format and source; code admits nonempty text without per-snippet human approval or an answer oracle. This is targeted extraction mode. Errors/recovery and overwrite behavior remain as on the supplied parent route, not a second admission policy.

text = " ".join(snippet.text.split()) if not text: continue --- tools/youtube_transcript.py

Claims

EPI-CLM-library-abstraction

Label: Maintained library abstraction and reuse

Claimed operation: preserve raw sources, abstract them into concepts/topics, reuse the wiki for cited outputs and ongoing coherence. SRC-1, README.md, doctrine/design. Distinct from RT-CLM-library-guarantees, which owns immutable-raw/citation guarantees: this record owns abstraction and later working-memory use. Neither source provides an observed abstraction run or causal efficacy measurement.

This project preserves the raw record, then abstracts it into concepts, topics, and usable outputs. --- README.md

The wiki is not the end product. It is the working memory the Librarian uses to answer questions, generate reports, produce routines, notice gaps, and keep the knowledge base coherent over time. --- README.md

EPI-CLM-compile-depth

Label: Compile depth gates and fidelity limits

Claimed operation/warrant: duplicate/echo gates detect failures of substantive compilation, with a stronger rationale that clearing template reuse requires source reading. The source itself explicitly disclaims automated factual-grounding checks and assigns fidelity to editorial comparison. SRC-1, tools/compile_lint.py, tools/README.md, doctrine/design. The historical shallow-compile episode is reported operation only. This is a distinct tool-warrant claim, not a duplicate of RT-CLM-library-guarantees. Implemented counterexample and bounded supported warrant are on EPI-RTE-mechanical-checks.

The only way to clear (a) is to write genuinely distinct summaries, i.e. to read the sources. --- tools/compile_lint.py

Deliberately NOT here: an automated grounding / fidelity check (does every proper noun / date in an article appear in its cited source?). --- tools/compile_lint.py