Sparks epistemic report

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

Source-and-claim boundary

This assessment asks what warrants reliance on content entering, changing or leaving Sparks. The boundary and Source register supply SRC-1 and the frozen whole-product scope. External harness scheduling, model inference and the personal vault remain outside the supplied evidence. No target program, test, service or model was executed. Implementation deductions below are not observed runs.

The supplied runtime orders the investigation but does not settle the epistemic interpretation. Direct inspection adds all seven collection extractors, their shared generated header, diagnostic implementations and their CLI consumers, task matching, affected-collection advice, status, archive/clear and finalization. It also checks CLI dispatch, the MCP registry and system-Git invocation against the supplied routes. No additional source identity is introduced.

Entry point or operation	SRC-1 source path	Coverage and conclusion limit
CLI dispatch; MCP operations	cmd/sparks/main.go; internal/mcp/serve.go	Direct dispatch/registry read; handler findings supplied by RT-RTE-mcp-tools. The registry exposes status, scan, describe, lint, affected, index, prepare/finalize, done, tasks add, brief and query; no separate semantic approval tool.
init; describe; installed instructions	internal/contract/contract.md; internal/core/agent.go	Direct full contract read; installation supplied by RT-RTE-contract-install. Loading/compliance not observed.
prepare; finalize; abort; raw archive/clear	cmd/sparks/ingest.go; internal/core/ingest_finalize.go; internal/inbox/archive.go	Direct reads, prepare-core details supplied by RT-RTE-ingest-cycle. Completion does not certify page coverage or truth.
scan; status	internal/core/scan.go; internal/core/status.go	Scan findings supplied by RT-RTE-maintenance-write; status read directly. Counts/metadata, not semantic assessments.
lint; fmt	internal/lint/lint.go; internal/frontmatter/frontmatter.go; cmd/sparks/lint.go; cmd/sparks/fmt.go	Direct reads; EPI-RTE-health-check. Structural and proxy targets only.
collections regen: all registered extractors	internal/collections/collections.go; internal/collections/header.go; internal/collections/quotes.go; internal/collections/bookmarks.go; internal/collections/books.go; internal/collections/reading_list.go; internal/collections/media.go; internal/collections/ideas.go; internal/collections/projects.go	Direct full reads. Seven transformation records below separate their material outputs. Persistence supplied by RT-RTE-maintenance-write.
index	internal/core/index.go	Direct main transformation/description-preservation read; RT-RTE-maintenance-write. Old descriptions are retained, not reassessed.
tasks add; done	internal/core/tasks.go	Direct add/matching/completion read; RT-RTE-maintenance-write. Target identification is checked, real-world completion is not.
affected	internal/core/affected.go	Direct read; EPI-RTE-affected-select. Advice is not complete dependency proof.
query; brief; view	internal/manifest/queries.go; internal/core/brief.go; internal/view/server.go	Brief read directly; query/view supplied by RT-RTE-vault-consume. Delivery is not successful synthesis.
Git hooks and development checks	internal/git/git.go; Makefile; .github/workflows/ci.yml	Git wrapper read directly; build/CI supplied by RT-RTE-product-build. User-configured hooks can reject a commit; their content and deployed policies are uninspected. No semantic guarantee is inferred from hook capability.

Public/design claims assessed are RT-CLM-work-split, RT-CLM-mechanical-integrity, EPI-CLM-revision-sources and EPI-CLM-stable-maturity. The latter two make particular epistemic consequences explicit rather than replace the supplied containers.

Missing personal-vault content prevents identifying actual accepted claims or checking their truth. Missing agent-session traces prevents establishing semantic production, criticism, acceptance, integration or actual downstream reliance. Missing comparisons prevents attributing quality, learning or other improvement to these mechanisms. Opaque provider state prevents assessing parameter updates, exact model fixity or endpoint substitution beyond RT-CMP-agent. These are separate limits, not one negative verdict. Available helper/test/view/query details not directly reread here remain supplied-runtime coverage, not access failures or an exhaustive independent semantic audit.

Epistemic-object inventory

The named object records own their identity, encoding, substrate and evidence; this table adds no competing object fields. Truth-apt scope concerns possible content, not evidence that any actual proposition is true.

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
EPI-OBJ-inbox-raw	Captured propositions and attributed source text	see record	SRC-1, internal/contract/contract.md	Actual source warrant unknown; capture is not verification.
EPI-OBJ-wiki-pages	Summaries, cross-source claims, Revision text, maturity/source declarations	see record	SRC-1, internal/contract/contract.md	No actual page or acceptance trace supplied.
EPI-OBJ-quotes	Quoted text and attribution	see record	SRC-1, internal/collections/quotes.go	Rendered output omits raw path held by parser.
EPI-OBJ-bookmarks	URL/description associations	see record	SRC-1, internal/collections/bookmarks.go	Same-line extraction need not preserve prose meaning.
EPI-OBJ-books	Alleged referenced book/source titles and contexts	see record	SRC-1, internal/collections/books.go	Attribution punctuation is not evidence of a book title.
EPI-OBJ-reading-list	Mostly directives; reported queue membership is truth-apt	see record	SRC-1, internal/collections/reading_list.go	Historical capture is not current endorsement.
EPI-OBJ-media	Title/raw-file/first-URL association	see record	SRC-1, internal/collections/media.go	First URL need not identify titled work.
EPI-OBJ-ideas	File title/status classification	see record	SRC-1, internal/collections/ideas.go	Last matching title/status wins; body omitted.
EPI-OBJ-projects	Project ship-status headings	see record	SRC-1, internal/collections/projects.go	Maturity does not entail shipping status.
EPI-OBJ-tasks	Directives plus recorded open/done state	see record	SRC-1, internal/core/tasks.go	Checked marker does not establish completed work.
RT-OBJ-contract	Primarily prescriptions; also design claims	see record	SRC-1, internal/contract/contract.md	Delivery is not interpretation/compliance.
RT-OBJ-ingests	Recorded lifecycle status, not semantic completion	see record	SRC-1, internal/core/ingest_finalize.go	Administrative state has no page-level acceptance receipt.
RT-CMP-agent	Not determinable in distributed-parametric state	see record	SRC-1, internal/contract/contract.go	Provider internals and actual outputs uninspected.

Authority-route ledger

Fields owned by the linked declaring records are incorporated as see; these rows separate consequential functions, including when the same runtime container implements several. All implementation rows denote inspected wiring, not exercised behavior. There is no observed route in the supplied evidence.

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-cycle	content transformation	implemented	EPI-OBJ-inbox-raw	truth-apt transformation: acquisition/import	Capture enters prepare result	see route; parser, not source-truth judge	see route	see route	delivery	Imported source warrant unknown	Offers entries for agent work	see route	SRC-1, internal/core/ingest.go	RT-CLM-work-split	none	Entry content is not endorsed.
EPI-RTE-semantic-revision	content transformation	doctrine only	EPI-OBJ-wiki-pages	truth-apt transformation: indeterminate	Captures/existing pages to derived prose	see record	see record	see record	protocol	Candidate summaries/syntheses, not licensed accepted claims	Requests direct edits and later cited synthesis	see record	SRC-1, internal/contract/contract.md	EPI-CLM-revision-sources	no enforcement inferred	Actual propositions needed to distinguish reshaping, derivation and conjecture.
EPI-RTE-semantic-revision	check/evidence production	doctrine only	EPI-OBJ-wiki-pages	no content change	Contradictions and source adequacy	see record	see record	see record	protocol	Neither contradiction resolution nor named acceptance standard established	Requests Revision text and cautious maturity	see record	SRC-1, internal/contract/contract.md	EPI-CLM-revision-sources	evidence-consuming acceptance unestablished	Both-source recording preserves disagreement, not adjudication.
RT-RTE-ingest-cycle	retention	implemented	EPI-OBJ-inbox-raw; EPI-OBJ-wiki-pages	no content change	Archive/clear/complete	see route	see route	see route	file and lifecycle mutation	Retained candidates, not acceptance	Clears inbox and enables later ingest	see route	SRC-1, internal/core/ingest_finalize.go	RT-CLM-mechanical-integrity	administrative completion only	Agent page writes precede and are not validated by finalize.
EPI-RTE-quotes-extract	content transformation	implemented	EPI-OBJ-quotes	truth-apt transformation: non-ampliative reshaping	Blockquote text and attribution	see record	see record	see record	renders candidate text	No independent warrant; raw-path lineage degraded	Offers quotations as generated page	see record	SRC-1, internal/collections/quotes.go	none	none	Attribution is copied, not authenticated.
EPI-RTE-bookmarks-extract	content transformation	implemented	EPI-OBJ-bookmarks	truth-apt transformation: indeterminate	URL plus residual same-line text	see record	see record	see record	renders association	No semantic equivalence licensed	Offers extracted links/descriptions	see record	SRC-1, internal/collections/bookmarks.go	none	none	URL trimming and context removal can change reference/meaning.
EPI-RTE-books-extract	content transformation	implemented	EPI-OBJ-books	truth-apt transformation: ampliative conjecture	Attribution to probable source/book	see record	see record	see record	renders conjectural label	Book identity not entailed by punctuation	Offers guessed title as book grouping	see record	SRC-1, internal/collections/books.go	none	deterministic does not mean entailed	Explicit book hints are a separate literal-import branch; see record.
EPI-RTE-reading-extract	content transformation	implemented	EPI-OBJ-reading-list	non-truth-apt policy/content update: extracted reading directives	Archived to-read hints to deduplicated queue	see record	see record	see record	renders queue	No current-interest warrant	Supplies retained reading suggestions	see record	SRC-1, internal/collections/reading_list.go	none	none	Count entails only distinct matched strings.
EPI-RTE-media-extract	content transformation	implemented	EPI-OBJ-media	truth-apt transformation: indeterminate	File title to first URL	see record	see record	see record	renders association	No verified work/link identity	Supplies navigation	see record	SRC-1, internal/collections/media.go	none	none	Need actual note to assess association.
EPI-RTE-ideas-extract	content transformation	implemented	EPI-OBJ-ideas	truth-apt transformation: non-ampliative reshaping	Selected title/status lines to groups	see record	see record	see record	renders selected fields	Carries supplied status, not a checked project fact	Supplies portfolio-like grouping	see record	SRC-1, internal/collections/ideas.go	none	none	Semantics limited to field extraction, not whole-note summary.
EPI-RTE-projects-extract	content transformation	implemented	EPI-OBJ-projects	truth-apt transformation: ampliative conjecture	Maturity to ship status	see record	see record	see record	renders status claim	Reviewed/stable does not warrant shipped	Supplies Shipped/Building/Exploring/Archived headings	see record	SRC-1, internal/collections/projects.go	EPI-CLM-stable-maturity	maturity/status mismatch	Requires extra equivalence premise absent from page contract.
RT-RTE-maintenance-write	retention	implemented	Seven collection objects above	no content change	Extractor result to persisted file	see route	see route	see route	enforcing overwrite on non-dry-run	No acceptance criterion consumed	Replaces generated pages without semantic approval	see route	SRC-1, internal/collections/collections.go	EPI-CLM-stable-maturity	auto-stable is not review	Generated header says stable with empty sources.
EPI-RTE-health-check	check/evidence production	implemented	EPI-OBJ-wiki-pages	no content change	Schema, graph, paths, timestamps, punctuation count	see record	see record	see record	diagnostic; text-mode exit signal	Licenses only named mechanical predicates	Can inform repair; no built-in finalize veto	see record	SRC-1, internal/lint/lint.go	RT-CLM-mechanical-integrity	health is not truth	No substantive support, contradiction or semantic completeness evaluator.
EPI-RTE-affected-select	operational admission/selection/consumption	implemented	Collection objects above	no content change	Changed paths to regen advice	see record	see record	see record	advisory selection	Heuristic dependency indication only	Can prioritize manual regeneration	see record	SRC-1, internal/core/affected.go	none	incomplete dependency witness	Books consumes quote files but empty-glob branch checks inbox archives.
RT-RTE-maintenance-write	check/evidence production	implemented	EPI-OBJ-tasks	no content change	Requested task identity among open lines	see route	see route	see route	exact/fuzzy candidate selector	Match confidence only; no completed-work evidence	Selects line or rejects missing/ambiguous match	see route	SRC-1, internal/core/tasks.go	none	none	Score gap is not outcome verification.
RT-RTE-maintenance-write	disposition/acceptance	implemented	EPI-OBJ-tasks	non-truth-apt policy/content update: close selected task	Selected open marker to checked marker	see route	see route	see route	enforcing write	Acceptance of target selection only, not claim of real-world success	Removes task from later open-task previews	see route	SRC-1, internal/core/tasks.go	none	operational disposition only	No second approval or task execution check.
RT-RTE-maintenance-write	content transformation	implemented	EPI-OBJ-wiki-pages	truth-apt transformation: non-ampliative reshaping	Existing descriptions and manifest to index	see route	see route	see route	rebuild/preserve	Source count describes listed paths, not evidential strength	Retains descriptions for later navigation	see route	SRC-1, internal/core/index.go	none	none	Path/title matching does not revalidate old descriptions.
RT-RTE-vault-consume	operational admission/selection/consumption	implemented	Wiki, task and generated-page objects above	no content change	Query/brief/view selection	see route	see route	see route	delivery/advisory	Selection not endorsement	Offers metadata, retained log/task text or bodies to consumer	see route	SRC-1, internal/core/brief.go	RT-CLM-work-split	none	Actual answer dependence unobserved; brief outputs only selected projections.
RT-RTE-ingest-cycle	lineage/freshness/recovery	implemented	RT-OBJ-ingests	no content change	Abort/failure recovery	see route	see route	see route	lifecycle transition/manual recovery	No withdrawal of false claims established	Abort unblocks next prepare; no content rollback	see route	SRC-1, cmd/sparks/ingest.go	none	abort is not retraction	Scan deletion removes metadata visibility, not truth-based repudiation.
RT-RTE-product-build	check/evidence production	implemented	Product sources and RT-OBJ-contract	no content change	Authored tests/vet/build/copy parity	see route	see route	see route	CI/process results	Program-specific assertions only	Possible build failure; merge/deployment veto unresolved	see route	SRC-1, .github/workflows/ci.yml	none	none	No observed passing tests, causal experiment or wiki-answer oracle.

No lifecycle-integration row is inferred from storage, indexing or links: the inspected routes do not establish an evidence-consuming acceptance decision for a substantive claim followed by integration. This is bounded to the listed binary paths and shipped protocol; inaccessible agent cognition and hypothetical user Git hooks are not assigned absent evaluators.

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-work-split	see record	SRC-1, internal/contract/contract.md, doctrine/design	see record	RT-RTE-ingest-cycle, RT-RTE-vault-consume, EPI-RTE-projects-extract, EPI-RTE-books-extract	none supplied	No agent-quality comparison	External agent owns open semantic writing; binary supplies deterministic processing	Literal claim that CLI does not understand content must not imply semantic neutrality: deterministic status/title heuristics add interpretive claims. No runtime record supersession required.
RT-CLM-mechanical-integrity	see record	SRC-1, internal/contract/contract.md, doctrine/design	see record	RT-RTE-maintenance-write, EPI-RTE-health-check, EPI-RTE-affected-select	none supplied	No reliability measurement	Selected mechanical mechanisms and diagnostics are wired	Not comprehensive admission; advisory dependency selection can omit a Books input change.
EPI-CLM-revision-sources	see record	SRC-1, internal/contract/contract.md, doctrine/design	see record	RT-RTE-contract-install delivers policy; EPI-RTE-semantic-revision is doctrine only	none supplied	No actual before/after contradiction trace	An explicit request to retain contradictions and both sources	Not implemented semantic comparison, acceptance or observed criticism.
EPI-CLM-stable-maturity	see record	SRC-1, internal/contract/contract.md, doctrine/design	see record	RT-RTE-maintenance-write, EPI-RTE-projects-extract	none supplied	No review record	Stable is a declared convention, checked only as an enum	Generated pages receive stable automatically; Projects reinterprets it as Shipped. Neither establishes source review or delivery.

Bounded conclusion

Sparks acquires human capture and attributed text, retains archives and agent-written files, and supplies metadata, selected text and rendered bodies for later use. Acquisition preserves some text and source pointers but does not import demonstrated source warrant. Quote rendering drops raw-path lineage; generated frontmatter supplies no source paths. Index rebuilding preserves descriptions without a new support check. These are implemented persistence/consumption affordances, not observed reliance or accepted knowledge: see RT-RTE-ingest-cycle, RT-RTE-vault-consume and the collection records.

The code is not epistemically inert merely because it is deterministic. Counts and field selections can be derived within their parser domain, conditional on readable selected inputs and manifest state. They do not establish source truth or exhaustive coverage. More consequentially, EPI-RTE-books-extract guesses source identity from punctuation, and EPI-RTE-projects-extract converts maturity into shipping status. The counterexamples retained in those records establish non-entailment without target execution. Their output can be persisted and read without an intervening acceptance check. That operational route gives the generated claims exposure, not warrant.

The semantic protocol explicitly asks agents to compare captures with existing pages, preserve conflicting sources, choose cautious maturity and cite pages in answers. It neither specifies a reproducible acceptance criterion for a substantive claim and intended use nor provides an inspected model evaluator. Actual truth-apt outputs are unavailable, so the production relation of external-agent summaries/syntheses is indeterminate, not automatically ampliative. Evidence-consuming acceptance and post-acceptance integration remain unestablished, independently of the positive retention and revision-policy findings: EPI-RTE-semantic-revision.

Mechanical checks have narrow licenses: schema membership, graph/path availability, source-mtime comparisons and punctuation-count proxies. Their consumer paths are caller-visible issues/exit signals and selected brief revisit suggestions, not enforced semantic admission. Finalization may archive and clear content without consuming a clean health report. Task matching and completion have stronger operational force: a selected marker is changed and later brief previews omit it, but no evaluator checks whether the underlying work succeeded. Abort and scan deletion are administrative recovery/visibility changes, not truth-based retraction. User-configured Git hooks remain an opaque possible commit gate, not an attributed semantic evaluator.

Persistence can support later improvement but does not itself establish learning, reflection, autonomy or self-improvement. No source-supplied execution or causal comparison establishes increased future capacity or evidence-responsive change to Sparks' own behavior-determining organization. The shipped contradiction protocol is a criticism affordance; it is not a demonstrated iterative theory-building episode. Distributed-parametric learning and internal criticism remain uninspected, rather than absent. These conclusions are independent and bounded by the source and coverage above.

Shared records

Operative objects

EPI-OBJ-inbox-raw

Label: Inbox and raw capture

Source-native identity: inbox.md entries and raw/ files, including raw/inbox/YYYY-MM-DD.md. Representational form: natural-language capture; separators/dates are symbolic parsing structure. Storage substrate: Markdown files. Evidence: SRC-1, internal/contract/contract.md, internal/inbox/archive.go. No counterpart object in runtime; distinct from lifecycle rows RT-OBJ-ingests. Archive copies captured content by date; its truth is not checked.

Raw is the source of truth. Wiki is a derived view. Manifest tracks state for incremental operations. All three are plain files on disk (the manifest is a single SQLite file), versioned by git. --- internal/contract/contract.md

EPI-OBJ-wiki-pages

Label: Derived wiki pages

Source-native identity: wiki entity/concept/summary/synthesis bodies and their source/maturity frontmatter, including Revision text and descriptions later reused by index. Representational form: natural-language propositions plus symbolic metadata consumed by fixed filters. Storage substrate: Markdown files; frontmatter projections in SQLite. Evidence: SRC-1, internal/contract/contract.md, internal/core/index.go. Distinct from runtime's static RT-OBJ-contract, which requests these writes rather than containing their accumulated results. No actual page is supplied. The quote in EPI-CLM-revision-sources supplies the semantic write policy; the runtime's RT-RTE-vault-consume supplies later consumers.

EPI-OBJ-quotes

Label: Quotes collection

Source-native identity: generated wiki/collections/Quotes.md. Representational form: quoted natural language and copied attribution; counts are symbolic output. Storage substrate: Markdown. Evidence: SRC-1, internal/collections/quotes.go. Unlike RT-OBJ-contract, this is a derived content artifact, not static instructions. See EPI-RTE-quotes-extract for the load-bearing passage and loss of SourcePath.

EPI-OBJ-bookmarks

Label: Bookmarks collection

Source-native identity: generated wiki/collections/Bookmarks.md. Representational form: natural-language descriptions, symbolic URLs and grouped counts. Storage substrate: Markdown. Evidence: SRC-1, internal/collections/bookmarks.go. No counterpart runtime object; its write is inside RT-RTE-maintenance-write. See EPI-RTE-bookmarks-extract.

EPI-OBJ-books

Label: Books collection

Source-native identity: generated wiki/collections/Books.md. Representational form: natural-language title/context associations, symbolic grouping/counts. Storage substrate: Markdown. Evidence: SRC-1, internal/collections/books.go. No counterpart runtime object; compare maintenance container RT-RTE-maintenance-write. See EPI-RTE-books-extract.

EPI-OBJ-reading-list

Label: Reading List collection

Source-native identity: generated wiki/collections/Reading List.md. Representational form: natural-language reading directives and source labels; symbolic queue count. Storage substrate: Markdown. Evidence: SRC-1, internal/collections/reading_list.go. No counterpart runtime object; compare RT-RTE-maintenance-write. See EPI-RTE-reading-extract.

EPI-OBJ-media

Label: Media collection

Source-native identity: generated wiki/collections/Media.md. Representational form: natural-language titles plus symbolic links/counts. Storage substrate: Markdown. Evidence: SRC-1, internal/collections/media.go. No counterpart runtime object; compare RT-RTE-maintenance-write. See EPI-RTE-media-extract.

EPI-OBJ-ideas

Label: Ideas collection

Source-native identity: generated wiki/collections/Ideas.md. Representational form: natural-language selected titles/status and symbolic grouping/path labels. Storage substrate: Markdown. Evidence: SRC-1, internal/collections/ideas.go. No counterpart runtime object; compare RT-RTE-maintenance-write. See EPI-RTE-ideas-extract.

EPI-OBJ-projects

Label: Projects collection

Source-native identity: generated wiki/collections/Projects.md. Representational form: natural-language ship-status headings derived from symbolic frontmatter. Storage substrate: Markdown. Evidence: SRC-1, internal/collections/projects.go. No counterpart runtime object; compare RT-RTE-maintenance-write. See EPI-RTE-projects-extract.

EPI-OBJ-tasks

Label: Live Tasks collection

Source-native identity: wiki/collections/Tasks.md. Representational form: natural-language directives with symbolic open/done markers. Storage substrate: Markdown. Evidence: SRC-1, internal/core/tasks.go. Not a regenerated collection and not an ingest row. Runtime has no counterpart object; its mutation/consumption routes are RT-RTE-maintenance-write and RT-RTE-vault-consume.

lines[idx] = strings.Replace(lines[idx], "- [ ]", "- [x]", 1) --- internal/core/tasks.go

The selector checks line identity, not work completion. A unique case-insensitive substring wins; otherwise one fuzzy match or a lead exceeding the next score by five wins. Missing/close matches reject the completion request. The caller proposes and code decides the write; no per-match human review or answer oracle is required. This is open task maintenance with manual edit/Git recovery, not an improvement experiment.

Routes

EPI-RTE-semantic-revision

Label: Agent semantic revision

Part of: RT-RTE-ingest-cycle

This is the protocol-directed semantic part, not a second declaration of the complete ingest cycle. Endpoints/progression: prepare output and pages → external agent comparison/writing → wiki/index/log/task edits. Owner: agent under human request and shipped guidance. Context: source text, aliases and existing content the agent chooses to read. Effects requested: summaries/syntheses, contradiction records and cautious maturity, followed by finalization. Evidence: SRC-1, internal/contract/contract.md; passages in EPI-CLM-revision-sources and EPI-CLM-stable-maturity.

• implementation conclusion status: claimed
• operation conclusion status: uninspected
• Immediate return: Requested agent-written pages and eventual answer; no supplied instance.
• Later read-back: Requested persisted wiki/index/log material can reach later page reads, queries and briefs; actual dependence unobserved.
• Delegated visibility: uninspected — enclosing harness, not Sparks, determines model/tool visibility.
• Selection predicate: Agent is asked to inspect relevant pages/aliases and contradictions; no programmatic semantic selector.
• Invalidation or expiry: Revision sections preserve conflict; no truth-based withdrawal procedure or mandatory expiry is specified here.
• Activation or effect: Requested direct filesystem writing and later cited synthesis; compliance not observed.
• Evidence limits: Doctrine is implemented as delivered prose, not an implemented semantic evaluator; actual outputs cannot be classified without content/trace evidence.

Admission: human initiates; external agent proposes and directly writes under prose guidance; ordinary filesystem permissions and external human edits can intervene. Sparks finalize cannot veto a contradiction. No expected-answer oracle, named substantive acceptance criterion, accepted scope or evidence-consuming acceptance record is provided. Open maintenance, not a curriculum. Recovery follows external edits/Git and administrative ingest recovery. Persistence records candidate/revision text, not necessarily accepted theory. Theory-builder conditions 1–4: each is uninspected for actual theories, because no whole theory instance, content-dependent decision, criticism/revision trace or retained-next-round use is supplied. Addressability is afforded by ordinary editable prose, not demonstrated for a theory; learning and reflective/autonomous theory-building are not inferred from these requested steps.

EPI-RTE-quotes-extract

Label: Quotes extraction

Part of: RT-RTE-maintenance-write

Endpoints/progression: configured quote-file glob → blockquote parser → rendered Quotes body. Owner: deterministic extractor invoked by caller. Natural-language text and attribution are retained while parser structure is removed. Evidence: SRC-1, internal/collections/quotes.go. Distinct material part: this extracts quotations rather than other collection fields.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Markdown body with quote count, text and attribution or literal missing-attribution marker.
• Later read-back: Regenerate can persist the body for later page readers; actual use unobserved.
• Delegated visibility: inapplicable — extractor has no delegation; enclosing caller receives its result.
• Selection predicate: Configured glob, sorted readable files, blockquotes beginning with greater-than plus space, following attribution/blank/EOF boundaries.
• Invalidation or expiry: Regeneration overwrites output; source edits require another regeneration.
• Activation or effect: Body is offered to the dispatcher for optional persistence; no source-truth check.
• Evidence limits: Missing/unreadable files can be skipped; scanner errors are not surfaced here; completeness is not established.

Admission/recovery: caller selects regeneration; parser generates; dispatcher may reject extractor/write failures or skip persistence on dry-run. No human content approval or answer oracle; open maintenance. Manual edit/source repair and regeneration provide recovery; output edits are overwritten. Count is entailed only for parsed records. Source warrant remains unknown, and lineage is degraded: QuoteRecord has SourcePath, but the renderer emits Text and Attribution without that path.

if r.Attribution != "" { fmt.Fprintf(&out, "— %s\n\n", r.Attribution) } else { out.WriteString("— (no attribution)\n\n") } --- internal/collections/quotes.go

EPI-RTE-bookmarks-extract

Label: Bookmarks extraction

Part of: RT-RTE-maintenance-write

Endpoints/progression: configured raw link files → URL regex matches → per-file grouped URL/description body. Owner: extractor; context/state is selected file text. Effects: candidate link associations. Evidence: SRC-1, internal/collections/bookmarks.go.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Markdown links and residual same-line descriptions, counts and filename headings.
• Later read-back: Dispatcher can persist Bookmarks for later readers; actual consumption unobserved.
• Delegated visibility: inapplicable — no internal delegate.
• Selection predicate: Sorted glob files; every URL match on each scanned line.
• Invalidation or expiry: Next regeneration replaces body; no content expiry.
• Activation or effect: Renders cleaned URLs and text associations without semantic check.
• Evidence limits: Unreadable files skipped; punctuation stripping and URL removal need not preserve meaning/reference.

Admission/recovery: explicit caller request, deterministic extraction and dispatcher write; errors/dry-run can prevent persistence, not reject misleading associations. No answer oracle or required human review; open maintenance with manual source repair/regeneration. For a line with multiple URLs, each output retains the other URL as part of its description; this is not evidence of a semantically accurate summary. Warrant unknown; source filename grouping preserves partial lineage, not an evidential citation check.

clean := strings.TrimRight(u, trailingPunct) desc := strings.TrimSpace(strings.Replace(line, u, "", 1)) desc = strings.TrimLeft(desc, "-*> ") --- internal/collections/bookmarks.go

EPI-RTE-books-extract

Label: Books extraction

Part of: RT-RTE-maintenance-write

Endpoints/progression: default quote stream and archived book hints → probable-title heuristic/literal hint extraction → grouped Books body. Owner: extractor. Context includes quote text, attribution and archive paths. Evidence: SRC-1, internal/collections/books.go. Distinct from quote rendering: this assigns a source/book identity and truncates long quote context.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Sorted title headings, quote snippets or archive-path contexts and distinct-title count.
• Later read-back: Dispatcher can persist Books, supplying later page readers; actual reliance unobserved.
• Delegated visibility: inapplicable — no internal delegate.
• Selection predicate: Quotes uses default Quotes glob, not its configured override; attribution segment after first matching separator in ordered slash/comma/em-dash alternatives; archive lines matching book prefix provide literal titles.
• Invalidation or expiry: Replacement on regeneration; no canonical identity or expiry check.
• Activation or effect: Renders inferred titles as referenced books/sources without independent validation.
• Evidence limits: Quote parse failures and archive read failures can be ignored; heuristic name classification is not book verification.

Concrete implementation deduction, not execution: attribution Ada, London matches comma-space and yields London. A raw author/location attribution is compatible with that input and need name no book. The renderer nevertheless puts London under a book heading/count. Thus the inferred book/source proposition is not entailed by the input, even though the string transformation is deterministic. In contrast, an explicit book: Dune line is literal supplied book content, not a new heuristic title conjecture. Snippet truncation at 80 bytes and omitted quote raw paths limit context/lineage. Count is over distinct title strings, not verified distinct books.

Admission/recovery: explicit regeneration; code proposes and writes without human title approval, expected-answer oracle or source-identity check. Open maintenance. Dispatcher errors/dry-run can stop a write; manual source changes/regeneration recover. No evidence-consuming acceptance decision intervenes.

for _, sep := range []string{" / ", ", ", " — "} { if i := strings.Index(attr, sep); i > 0 { return strings.TrimSpace(attr[i+len(sep):]) } } --- internal/collections/books.go

fmt.Fprintf(&out, "%d distinct book(s) referenced.\n\n", len(sortedTitles)) --- internal/collections/books.go

EPI-RTE-reading-extract

Label: Reading queue extraction

Part of: RT-RTE-maintenance-write

Endpoints/progression: archive to-read lines → exact-text deduplication → Reading List queue. Owner: extractor; directives and archive paths become rendered queue entries. Evidence: SRC-1, internal/collections/reading_list.go.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Queue Markdown, count and first retained archive path per exact text.
• Later read-back: Persisted queue can guide later reading requests; actual behavioral activation unobserved.
• Delegated visibility: inapplicable — no internal delegate.
• Selection predicate: Sorted raw/inbox archives, case-insensitive prefix, exact case-sensitive captured-text deduplication.
• Invalidation or expiry: Regeneration recomputes from archives; no read/done removal criterion.
• Activation or effect: Converts historical directives into a displayed queue, without current-intent check.
• Evidence limits: Unreadable archives skipped; queue is not evidence of current endorsement or unread status.

Admission/recovery: caller requests, code selects/writes via dispatcher, no per-entry human approval or oracle; open maintenance. Errors/dry-run affect persistence; source edit/regeneration recover. The queue directives are not truth-apt assertions; the count derives only from selected unique strings. A rendered path preserves partial lineage.

if text == "" || seen[text] { continue } seen[text] = true entries = append(entries, entry{text: text, source: rel(v.Root, path)}) --- internal/collections/reading_list.go

EPI-RTE-media-extract

Label: Media extraction

Part of: RT-RTE-maintenance-write

Endpoints/progression: configured media files → filename/H1 and first URL → linked Media body. Owner: extractor; effects are navigation associations, not verified work identities. Evidence: SRC-1, internal/collections/media.go.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Raw-file wikilinks with titles and optional cleaned first-URL links.
• Later read-back: Dispatcher can persist Media for later readers.
• Delegated visibility: inapplicable — no internal delegate.
• Selection predicate: Sorted configured glob; title starts as filename, matching H1 can replace while title equals filename; first regex URL wins.
• Invalidation or expiry: Replaced by regeneration; no link-health expiry.
• Activation or effect: Publishes file/title/URL associations without checking whether URL identifies the titled work.
• Evidence limits: Unreadable files skipped; actual inputs needed to establish semantic preservation.

Admission/recovery: caller request, deterministic rendering, dispatcher errors/dry-run; no human per-link approval or oracle. Open maintenance with manual source repair/regeneration. Raw-file links preserve navigation, not source warrant; first URL could be a review or unrelated link, so association remains indeterminate without actual content.

if firstURL == "" { if m := urlRE.FindString(line); m != "" { firstURL = strings.TrimRight(m, trailingPunct) } } --- internal/collections/media.go

EPI-RTE-ideas-extract

Label: Ideas extraction

Part of: RT-RTE-maintenance-write

Endpoints/progression: configured idea files → selected title/status → grouped Ideas body with source paths. Owner: extractor. Evidence: SRC-1, internal/collections/ideas.go. This selects fields; it does not synthesize the idea's body.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Title/path entries grouped by normalized status, including custom statuses.
• Later read-back: Persisted Ideas can feed later page reads.
• Delegated visibility: inapplicable — no internal delegate.
• Selection predicate: Sorted glob; last H1 and last status line encountered override prior values; filename and uncategorized fallbacks.
• Invalidation or expiry: Replacement on regeneration; status is not independently refreshed.
• Activation or effect: Exposes supplied title/status fields and count, without a status-truth check.
• Evidence limits: Comment says first H1, but loop repeatedly assigns; body and interpretation context are omitted.

Admission/recovery: explicit regeneration; code selects and dispatcher persists, no oracle or per-item human review. Open maintenance; errors/dry-run affect writes, source editing/regeneration recover. Source path is preserved, source warrant unknown. Non-ampliative classification is limited to reporting the selected field values, not to claiming a comprehensive or faithful note summary.

if strings.HasPrefix(line, "# ") { title = strings.TrimSpace(strings.TrimPrefix(line, "# ")) } if strings.HasPrefix(strings.ToLower(line), "status:") { status = strings.TrimSpace(line[len("status:"):]) } --- internal/collections/ideas.go

EPI-RTE-projects-extract

Label: Projects extraction

Part of: RT-RTE-maintenance-write

Endpoints/progression: manifest entity pages with project tag → group by maturity → Projects ship-status headings and links. Owner: extractor; symbolic metadata is interpreted into natural-language status. Evidence: SRC-1, internal/collections/projects.go, contrasted with internal/contract/contract.md.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Tracked-project count and headings Building, Shipped, Exploring, Archived.
• Later read-back: Dispatcher can persist Projects; future readers see status headings with links, not an acceptance receipt.
• Delegated visibility: inapplicable — no internal delegate.
• Selection predicate: Type entity, case-insensitive project tag; fixed maturity order and label map.
• Invalidation or expiry: Regeneration replaces output; changes in manifest depend on scan freshness.
• Activation or effect: Publishes ship-status claims from maturity alone.
• Evidence limits: No ship-date, deployment result, human status approval or independent status evaluator is consulted.

Concrete implementation deduction: an entity with tag project and maturity stable is placed under Shipped. Under the shipped page contract, stable means well-sourced, reviewed and unlikely to change. A reviewed, well-sourced description of an unshipped project satisfies that definition. Shipping therefore does not follow from the input. This is a deterministic ampliative status conjecture, not entailed derivation. The same map also interprets historical as Archived; none of these labels receives an outcome check here.

Admission/recovery: caller requests; code assigns labels and dispatcher writes without a second human gate or expected-answer oracle. Open maintenance; failures/dry-run can prevent writing, manual metadata/source repair and regeneration recover. The deliberate design mapping does not establish empirical equivalence. See EPI-CLM-stable-maturity.

maturityLabel := map[string]string{ "working": "Building", "stable": "Shipped", "seed": "Exploring", "historical": "Archived", } --- internal/collections/projects.go

EPI-RTE-health-check

Label: Mechanical health checks

Part of: RT-RTE-maintenance-write

Endpoints/progression: explicit lint/fmt request → manifest/filesystem parsing and predicates → issues/exit result; brief separately consumes selected lint issue kinds as revisit paths. Owner: program for predicates, caller/agent for semantic repair. Targets and domains: frontmatter required fields/enums/dates; path existence; manifest graph references and alias collisions; updated date versus listed-source mtime; punctuation-count thinness proxy. Evidence: SRC-1, internal/lint/lint.go, internal/frontmatter/frontmatter.go, cmd/sparks/fmt.go, cmd/sparks/lint.go, internal/core/brief.go. Unlike the parent mutation container, this record owns only diagnostic production and its advisory consumers, not collection writes.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Issue list/counts; fmt and lint text mode fail on issues; lint JSON returns serialization result without issue-count failure.
• Later read-back: inapplicable — diagnostics do not retain reports; brief recomputes selected signals from retained files/manifest.
• Delegated visibility: inapplicable — no internal delegate; CLI/MCP caller receives findings.
• Selection predicate: Lint's eight named checks; fmt walks chosen/default Markdown glob; brief keeps stale/thin issues and computes seed-orphans separately.
• Invalidation or expiry: Results describe current read/manifest state; scan freshness is a precondition, no automatic freshness endorsement.
• Activation or effect: Caller can act on issues; brief offers revisit paths. No automatic repair, substantive acceptance or finalize veto is implemented by these checks.
• Evidence limits: Source existence and metadata validity do not establish source truth, adequate support, successful contradiction resolution or completeness; skipped read/DB errors can hide findings.

Admission comparison: caller proposes a diagnostic request; fixed predicates decide findings; external user/agent decides repairs. No expected semantic answer oracle or mandatory human content acceptance. Open maintenance. Check failure does not roll back files. The frozen lint CLI has no fix flag; fmt reports rather than rewrites. This qualifies the prose safe-fix claim, without denying implemented diagnostics.

Concrete deduction: outside headings/fences, body False. False. False. reaches punctuation count three and avoids thin-page warning, regardless of truth. Conversely a correct two-sentence definition can be flagged. Thus thinness is a length proxy, not a semantic-quality evaluator. A listed path can exist while supporting no page assertion; dead-source check establishes availability only. Generated collection headers set sources empty, while Validate reports empty sources: generation alone is not even a clean diagnostic result.

if len(fm.Sources) == 0 { issues = append(issues, "missing required field: sources (must list raw/ paths)") } --- internal/frontmatter/frontmatter.go

switch r { case '.', '!', '?': count++ } --- internal/lint/lint.go

case "stale-pages": r.Stale = append(r.Stale, iss.Path) case "thin-pages": r.Thin = append(r.Thin, iss.Path) --- internal/core/brief.go

EPI-RTE-affected-select

Label: Affected collection advice

Endpoints/progression: caller → changed manifest paths since last completed ingest → fixed mapping → affected collection names/reasons. Owner: code for mapping, caller for regeneration. Context/state: manifest change state and configured/default globs. This is distinct from RT-RTE-maintenance-write: it advises which maintenance writes to request, rather than doing those writes. It shares transport with RT-RTE-mcp-tools, not identity. Evidence: SRC-1, internal/core/affected.go, checked against internal/collections/books.go.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Affected names and path reasons, or empty set/error.
• Later read-back: inapplicable — this function does not retain advice; future calls reread changed-path state.
• Delegated visibility: inapplicable — no internal delegate; result goes to requesting CLI/MCP client.
• Selection predicate: Path-prefix matches against glob prefix; Projects checks wiki/entities changes; other empty-glob derived collections check raw/inbox changes.
• Invalidation or expiry: Recomputed per call relative to last completed ingest; stale manifest limits input completeness.
• Activation or effect: Advisory collection selection; no regeneration is invoked here.
• Evidence limits: Prefix advice is not semantic dependency validation; downstream caller obedience and regenerated freshness are unobserved.

Counterexample to completeness: Books reads default raw/quotes as well as inbox archives, but its empty-glob affected branch triggers only on inbox changes. A quote-only path change can affect Books content without selecting Books here. This implementation deduction bounds incremental-regeneration claims. Caller proposes the query, program decides suggestions; there is no expected-output oracle, required human confirmation or improvement trigger. Open maintenance; no persisted change or rollback.

if anyMatches(changed, func(p string) bool { return strings.HasPrefix(p, "raw/inbox/") }) { reason[name] = []string{"raw/inbox/* changed"} } --- internal/core/affected.go

Claims

EPI-CLM-revision-sources

Label: Preserve conflicting sources

Part of: RT-CLM-mechanical-integrity

Claimed operation: agents reconcile new and existing content by noting contradictions and both sources, without silent overwrite. This makes the particular semantic obligation explicit within the supplied invariants claim. Source: SRC-1, internal/contract/contract.md, doctrine/design. Installed prose can request this, but EPI-RTE-semantic-revision remains doctrine only as a semantic procedure.

1. If an existing wiki page contradicts a new entry, append a ## Revision section noting the contradiction and both sources. Never silently overwrite. --- internal/contract/contract.md

EPI-CLM-stable-maturity

Label: Stable maturity meaning

Claimed operation/warrant: stable pages are well-sourced, reviewed and unlikely to change; new pages should normally be seed. Source: SRC-1, internal/contract/contract.md, doctrine/design. Distinct from RT-CLM-mechanical-integrity: this is a meaning/reliance convention for metadata, not a generic integrity guarantee. The binary validates enum membership, not review. Header generation assigns stable without a review decision; Projects applies a different operational interpretation.

• stable — well-sourced, reviewed, unlikely to change --- internal/contract/contract.md

• Set maturity: seed on new pages unless clearly well-sourced. --- internal/contract/contract.md

title: %s type: collection maturity: stable sources: [] created: %s updated: %s --- internal/collections/header.go