Synto epistemic report

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

Source-and-claim boundary

The question is how Synto turns source notes into claims available for later reliance, and what its checks license. Scope and exclusions are those of the boundary; source identities and evidence layers are in its Source register. Every source finding below uses SRC-1 at the frozen commit. Implementation and doctrine/design are distinguished; no target execution was performed and no observed-run or causal evidence was supplied.

Model weights, human judgment, actual host compliance and external source truth are outside the inspected implementation. Their absence prevents conclusions about factual accuracy, reviewer competence, downstream benefit or parameter learning. A source ID, content hash, confidence number or published status does not alone supply those facts. No supplied runtime finding needs supersession: the new records below separate material functions inside its valid route containers.

Coverage

Direct reads supplement, rather than replace, the supplied runtime account. Available but uninspected operations are explicit below; this report makes no system-complete negative about all possible evaluation or knowledge production.

Entry point or operation	Source path (SRC-1)	Inspection and record coverage or limit
run, ingest, compile, watch; legacy compile	src/synto/cli.py, src/synto/pipeline/orchestrator.py, src/synto/pipeline/ingest.py, src/synto/pipeline/compile.py, src/synto/watcher.py	Supplied RT-RTE-pipeline-run; direct CLI dispatch search, ingest filters/occurrences/relations, writer prompt and draft cleanup. Legacy compile dispatch identified but its planning implementation not inspected here; prevents exhaustive production-route classification.
Source add, extraction, segmentation	src/synto/cli.py, src/synto/extractors/	Supplied boundary import account; direct CLI dispatch search only. Extraction fidelity and exact import warrant remain unassessed.
Source-type prompting, parsing	src/synto/pipeline/prompts/notes.md, src/synto/models.py, src/synto/structured_output.py	Notes prompt and AnalysisResult/SingleArticle directly read; parser wiring supplied by runtime. Other six prompt wordings uninspected here, preventing universal prompt-policy claims.
approve, verify, reject, review, unblock	src/synto/pipeline/compile.py, src/synto/pipeline/review.py, src/synto/cli.py	Publication/verification/rejection directly read; interactive review dispatch searched; RT-RTE-draft-publication. Human decision content not supplied.
query, save, synthesize	src/synto/pipeline/query.py, src/synto/engines.py	Core, saving and synthesis checks directly read; engine wiring supplied by RT-RTE-query-answer.
MCP answer and article primitives	src/synto/serve.py	Handler search; supplied RT-RTE-mcp-answer. Article visibility is operational disclosure authority, not epistemic endorsement.
MCP verbatim reads/search/passages/list	src/synto/serve.py	Direct read of registration, segment read, search and passage selection preamble; remainder supplied at container level only. Exact list/passages formatting not fully assessed; prevents complete delivery-fidelity verdict.
maintain, curation, alias operations, items, trace, find, doctor reconcile	src/synto/pipeline/maintain.py, src/synto/pipeline/lint.py, src/synto/cli.py	CLI dispatch searched; lint graph advisories directly read; merge/split supplied by RT-RTE-concept-curation. Alias/keep/reconcile mutation bodies and item extraction remain uninspected here; prevents exhaustive semantic-curation and authority claims.
undo, clean	src/synto/cli.py	Direct complete command reads; recovery/withdrawal qualifications below.
eval, compare, smoke/CI	src/synto/pipeline/eval.py, src/synto/compare/runner.py, scripts/, .github/	Complete offline evaluator directly read; comparison supplied by RT-RTE-compare-eval. Test/script definitions are not runs.
pack export, indexes and read-back	src/synto/pack_export.py, src/synto/indexer.py, src/synto/readers.py	Direct generate_index read distinguishes synthesis admission from ordinary Q&A persistence; supplied RT-RTE-pack-export, RT-BAP-pack-guidance and boundary. Exported instructions request source verification; host adherence unobserved.
setup, config, init, migrate, vault, report, support	src/synto/cli.py	Dispatch searched. Configuration can change model/prompt/admission settings; model mutability supplied by runtime components. Setup/migration correctness not assessed; counters and support are not evidence of accepted factual content.

Knowledge/warrant-facing claims are EPI-CLM-wiki-synthesis, EPI-CLM-quality-evaluation and EPI-CLM-staged-review.

Epistemic-object inventory

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source and local anchor	Gap/limit
EPI-OBJ-source-segment	Imported source assertions	Original wording/evidence	See record	Truth of original source unknown; extraction lineage not fully inspected
EPI-OBJ-analysis-result	Summaries, concept/source associations, source-quality assessment	Source analysis	See record	Model output semantic fidelity unobserved
EPI-OBJ-relation	Subject–predicate–object assertion and evidence string	Graph/routing relation	See record	Endpoint existence does not establish asserted relationship
EPI-OBJ-single-article	Generated article assertions	Retained wiki knowledge	See record	No supplied article instance or reviewer evidence
EPI-OBJ-confidence	None as a factual claim about article correctness; deterministic source-count/quality indicator	Admission and quality signal	See record	Not a calibrated probability of truth
EPI-OBJ-query-answer	Answer/synthesis assertions	Answer using wiki content	See record	Entailment, novelty and truth not determinable without actual candidate and inputs
RT-CMP-fast, RT-CMP-heavy, RT-CMP-embed	Not determinable at individually localized claim level	Inference and optional vectors	See components	Provider-held distributed-parametric state; no inspected parameter update, immutable weights unpinned, mutable endpoints possible; internals uninspected
RT-BAP-rejection-guidance	Feedback may challenge claims; no feedback instance supplied	Revision guidance	See record	Cannot infer formulated theory criticism or successful revision from delivery
RT-BAP-pack-guidance	Policy rather than new factual knowledge	Later-agent reading instructions	See record	Delivery/compliance host-dependent

Authority-route ledger

The architectural statuses below concern inspected wiring, not operation. Fields owned by declaring records are referenced rather than recopied. Content relations concern the named edge, not everything in its container. Separate rows keep checking, disposition and retention apart.

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-pipeline-run	content transformation	implemented	EPI-OBJ-analysis-result, EPI-OBJ-single-article	truth-apt transformation: indeterminate	Source text → generated summaries/articles	Fast/heavy inference; schema parse, not truth oracle	Ingest/compile invocation	Candidate text and metadata	Generation and persistence	Candidate only; imported premises are not independently warranted	Makes content available for draft review	see RT-RTE-pipeline-run	see RT-RTE-pipeline-run	EPI-CLM-wiki-synthesis	none	Requested faithful extraction/synthesis; neither entailment nor non-entailment established for absent instances
EPI-RTE-candidate-filter	check/evidence production	implemented	EPI-OBJ-analysis-result, EPI-OBJ-relation	no content change	Candidate name support and relation endpoints	Code lexical/identity domain	Before canonical admission	Surviving concepts; resolvable endpoints	Enforcing local filter	Lexical occurrence/identity only	Blocks selected candidates, not every unsupported assertion	see EPI-RTE-candidate-filter	see EPI-RTE-candidate-filter	none	none	No observed run
EPI-RTE-occurrence-lineage	retention	implemented	EPI-OBJ-source-segment, EPI-OBJ-relation	no content change	Chunk/segment association retained	Code canonical-name survival and chunk membership	After analysis normalization	Occurrence rows and relation evidence	Durable linkage	Chunk association, not exact supporting passage proof	Changes later passage retrieval/ranking and graph context	see EPI-RTE-occurrence-lineage	see EPI-RTE-occurrence-lineage	none	none	Concepts attributed to every segment in chunk; confidence 1.0 does not certify local assertion support
EPI-RTE-confidence-score	check/evidence production	implemented	EPI-OBJ-confidence	no content change	Source count and best quality	Code arithmetic over fast-model structural quality label	Before draft writing	0–1 scalar	Threshold input	Only named proxy; not correctness calibration	Can hold drafts back when threshold enabled	see EPI-RTE-confidence-score	see EPI-RTE-confidence-score	EPI-CLM-staged-review	Proxy versus truth scope	No observed run
EPI-RTE-markdown-repair	content transformation	implemented	EPI-OBJ-single-article	truth-apt transformation: non-ampliative reshaping	Link syntax in generated text	Code known-title rules	Before draft persistence	Unknown links become plain text	Enforcing formatting	No added warrant for underlying proposition	Removes navigation, retains unsupported prose	see EPI-RTE-markdown-repair	see EPI-RTE-markdown-repair	none	none	Classification scoped to link unwrapping; not universal semantic preservation of all cleanup
RT-RTE-draft-publication	disposition/acceptance	implemented	EPI-OBJ-single-article	no content change	Draft → verified/published or rejected	Explicit human command or automatic threshold; code path/existence checks	After generation; verify optional	Status/write or deletion/feedback	Publication enforcing; feedback advisory	Recorded approval marker, but evidence-consuming factual acceptance not established	Published location admits later consumers; rejection removes draft	see RT-RTE-draft-publication	see RT-RTE-draft-publication	EPI-CLM-staged-review	Verification not required	No observed run
RT-RTE-pipeline-run	retention	implemented	EPI-OBJ-analysis-result, EPI-OBJ-single-article	no content change	Generated analysis/draft → disk/DB	Code after parse and preprocessing	During ingestion/compile	Persistent candidates	Storage	Retention is not acceptance	Later compilation/review/export can read candidates	see RT-RTE-pipeline-run	see RT-RTE-pipeline-run	none	none	No observed run
RT-RTE-query-answer	operational admission/selection/consumption	implemented	EPI-OBJ-single-article, EPI-OBJ-query-answer	no content change	Retained pages → model input	Fast relevance selector, code caps/optional graph expansion	Per question	Page body slices in answer prompt	Ranking and context admission	Selected context, not truth endorsement	Conditions answer generation on delivered slice	see RT-RTE-query-answer	see RT-RTE-query-answer	EPI-CLM-wiki-synthesis	none	Selected names do not prove every body loaded
RT-RTE-query-answer	content transformation	implemented	EPI-OBJ-query-answer	truth-apt transformation: indeterminate	Question plus selected pages → answer	Heavy model; structured parsing and sanitation	After selection	Answer text	Returned/advisory	Grounding request; no inspected entailment decision	Returns answer even if no matching page context	see RT-RTE-query-answer	see RT-RTE-query-answer	EPI-CLM-wiki-synthesis	none	No answer instance; cannot classify as entailed derivation or ampliative conjecture
EPI-RTE-synthesis-guard	check/evidence production	implemented	EPI-OBJ-query-answer	no content change	Synthesis source chain, duplicate identity and overwrite safety	Code path/hash/equality domain	On synthesis save/update	Refusal, keep existing or eligible write	Enforcing safety checks	No factual-answer license	Blocks synthesis-of-synthesis and edited overwrite	see EPI-RTE-synthesis-guard	see EPI-RTE-synthesis-guard	none	none	Safety and lineage checks are counterexamples to no-check claims, not semantic validation
EPI-RTE-answer-retention	retention	implemented	EPI-OBJ-query-answer	no content change	Answer → queries/synthesis files	Caller save flag; code save guards	After answer return	Published-tagged Q&A file or registered synthesis	Durable admission	No evidence-consuming factual acceptance established	Synthesis enters index selection; ordinary Q&A remains filesystem/export-accessible and resolvable if its title is supplied	see EPI-RTE-answer-retention	see EPI-RTE-answer-retention	EPI-CLM-wiki-synthesis	Bypasses draft publication	No observed run
RT-RTE-draft-publication	behavior/policy adaptation	implemented	RT-BAP-rejection-guidance	non-truth-apt policy/content update: rejection feedback supplied as revision guidance	Rejected draft → later writing prompt	Human feedback; code stores/delivers; rejection cap blocks	Later recompile	Feedback inclusion or blocked concept	Advisory model instruction plus enforcing cap	Correctness of criticism unestablished	Changes next writer input; unblock can restore eligibility	see RT-BAP-rejection-guidance	see RT-RTE-draft-publication	none	none	No observed successful correction, capacity improvement or own-method revision
EPI-RTE-offline-evaluation	check/evidence production	implemented	EPI-OBJ-single-article	no content change	Article existence, sources-list presence, links, INDEX schema	Deterministic structural evaluator	Explicit eval; comparison uses separate runs	Metric report	Advisory/reporting	Structural coverage only	Operator can inspect; no inspected automatic production adoption	see EPI-RTE-offline-evaluation	see EPI-RTE-offline-evaluation	EPI-CLM-quality-evaluation	Citation support is metadata coverage	No observed run
EPI-RTE-verbatim-source	operational admission/selection/consumption	implemented	EPI-OBJ-source-segment	no content change	Stored source text → MCP response	ID lookup/license gate; optional character cap	Tool invocation	Body, hash, source identity, truncation marker	Enforcing disclosure gate	Attributed stored text, not truth of source	Permits external agent to inspect original assertions	see EPI-RTE-verbatim-source	see EPI-RTE-verbatim-source	none	none	Storage text lineage to original bytes not fully inspected
RT-RTE-concept-curation	content transformation	implemented	EPI-OBJ-analysis-result	truth-apt transformation: indeterminate	Human-selected concept grouping → identities/source allocation	Human semantic decision, software preflight	Explicit merge/split	Changed routing and source assignments	Enforcing mutation	No semantic equivalence oracle established	Changes later compile/query context	see RT-RTE-concept-curation	see RT-RTE-concept-curation	none	none	No applied curation instance; alias/keep branches not fully inspected here
RT-RTE-pack-export	operational admission/selection/consumption	implemented	EPI-OBJ-single-article, EPI-OBJ-query-answer	no content change	Retained files → distributed pack	Exporter selection/copy rules	Explicit export	Pack, including optional unpublished trees	Filesystem delivery and requested guidance	Portable content is not accepted factual content	Later host may read/index/summarize	see RT-BAP-pack-guidance	see RT-RTE-pack-export	none	none	Host compliance unobserved; no export-wide draft exclusion
EPI-RTE-state-recovery	lineage/freshness/recovery	implemented	EPI-OBJ-single-article, EPI-OBJ-analysis-result	no content change	Undo/clean retained state	Human invocation; code identity-op guard	Explicit recovery/destruction	File revert or state/wiki deletion with seed	Enforcing local operation	Withdrawal/rebuild, not factual correction	Changes available files; DB may diverge after revert	see EPI-RTE-state-recovery	see EPI-RTE-state-recovery	none	none	Not distributed revocation of prior packs or returned answers

No lifecycle-integration row is asserted: publication, indexes and source links are operational organization, but evidence-consuming acceptance of a claim against a named factual criterion is not established for these inspected paths. This is a route-local limit, not a claim that a human cannot perform such review.

System-claim versus route comparison

Claim ID	Claimed operation or warrant	Claim source and evidence layer	Doctrine/design support	Implemented routes	Observed-run support	Causal support/design limits	Supported conclusion	Mismatch/unknown
EPI-CLM-wiki-synthesis	Synthesized, current wiki; query/synthesis using pages	See claim	Source prompts and answer grounding instructions	See RT-RTE-pipeline-run, RT-RTE-query-answer, EPI-RTE-answer-retention	None supplied	No measured accuracy, novelty or benefit	Generated and retained wiki/answer candidates, bounded context reload	“Synthesis” does not establish produced accepted ampliative claims; freshness not endorsement
EPI-CLM-quality-evaluation	Offline quality/citation support	See claim	Offline structural scope explicitly documented elsewhere in README	See EPI-RTE-offline-evaluation	None supplied	No proposition-level citation support test	Structural indicators and reports	Citation label broader than actual nonempty-sources-list predicate if read semantically
EPI-CLM-staged-review	Confidence and draft → verified → published lifecycle	See claim	Optional explicit review step	See RT-RTE-draft-publication, EPI-RTE-confidence-score	None supplied	Human review quality unknown; automatic publishing available	Optional staging plus proxy threshold	Not mandatory sequential verification; score not truth probability

Bounded conclusion

Synto acquires externally authored notes and can deliver stored source paragraphs with identity, hash and optional truncation indication. Retained segments reach later MCP client invocations: this is read-back even though the read handler does not write knowledge. Delivery does not establish answer saving or downstream model use. This preserves access to attributed assertions, not their truth; original extraction fidelity remains unassessed. Chunk-to-segment attribution can overinclude passages: association is not exact sentence support. See EPI-RTE-verbatim-source and EPI-RTE-occurrence-lineage.

The fast/heavy stages produce truth-apt candidate summaries, relationships, articles and answers. Their semantic relation is indeterminate without candidate/input instances. Faithful extraction is requested, but model novelty or entailed derivation is not established. Deterministic unwrapping of an unknown link preserves its displayed assertion while removing navigation; source-count scoring derives only an operational indicator. Neither licenses source truth or article correctness. See EPI-RTE-markdown-repair, EPI-RTE-confidence-score and RT-RTE-pipeline-run.

There are real implemented checks: schema parsing, lexical candidate filters, closed relation endpoints, synthesis-chain/duplicate/edit guards, lint and offline structural tests. Their domains are not factual acceptance. Publication, verified labels and automatic thresholds grant operational availability rather than demonstrated evidence-consuming acceptance. Retention and index refresh therefore do not establish lifecycle integration after epistemic acceptance. Saved synthesis bypasses the draft gate and enters later index selection (the generated catalogue includes at most 25 synthesis entries). Ordinary Q&A saving writes files but does not automatically enumerate them into that index. Both retained forms remain available through filesystem/export paths; the recursive title resolver can load an ordinary Q&A file if its title is supplied. These are distinct consumer routes, not an absolute absence of Q&A recall. Exported packs may also contain drafts and raw material. The contrary cases were checked before limiting the conclusion: human review is possible, safety checks are present, but neither makes factual checking universal. See EPI-RTE-candidate-filter, EPI-RTE-synthesis-guard, EPI-RTE-offline-evaluation, RT-RTE-draft-publication and RT-RTE-pack-export.

Rejection can remove a draft, retain criticism and change later writer input; the cap can block compilation until unblocked. This is a durable behavioral-authority path independent of whether the feedback itself is truth-apt. Learning benefit, formulated content-directed theory criticism, autonomous theory building and exercised self-improvement do not follow from storage/delivery; no concrete criticism/revision episode or improvement comparison was supplied. Provider-held parameters remain separately opaque and unpinned as stated in the runtime components. See RT-BAP-rejection-guidance and RT-CMP-heavy.

Undo and clean change local availability but do not retract distributed packs, reverse all DB state or warrant corrected knowledge. Evaluation reports support operator inspection; no inspected automatic winner-installation consumer turns comparison scores into production policy. Actual accuracy, future capacity improvement and causal benefit remain unobserved, without negating the implemented candidate, check, retention and consumption routes. See EPI-RTE-state-recovery and RT-RTE-compare-eval.

Shared records

Components

none declared in this member

Operative objects

EPI-OBJ-source-segment

Label: Source segment

Imported natural-language paragraph held in SQLite source_segments, with symbolic segment/source IDs and content hash. It can contain truth-apt source assertions. Distinct from model components; a material input/read object along RT-RTE-mcp-answer rather than a duplicate route. SRC-1, src/synto/serve.py (read_source_segment).

EPI-OBJ-analysis-result

Label: AnalysisResult

Natural-language summary and concept names plus symbolic quality/language fields; persisted analysis and source associations are in DB/generated source pages. RT-RTE-pipeline-run consumes/produces this object; different record kind, not a duplicate. SRC-1, src/synto/models.py and src/synto/pipeline/ingest.py.

EPI-OBJ-relation

Label: Relation

Part of: RT-RTE-pipeline-run

Subject/predicate/object labels and natural-language evidence with symbolic confidence/source segment binding, persisted in SQLite. Candidate assertions can affect graph-expanded querying. A material part of RT-RTE-pipeline-run as a cross-kind output; not the inference component or full pipeline. SRC-1, src/synto/pipeline/ingest.py (_extract_and_persist_relations).

EPI-OBJ-single-article

Label: SingleArticle

Natural-language Markdown assertions, title and tags returned by heavy inference; draft/published Markdown and DB metadata retain the result. Object distinct from RT-RTE-pipeline-run and RT-RTE-draft-publication route containers. SRC-1, src/synto/models.py and src/synto/pipeline/compile.py.

EPI-OBJ-confidence

Label: Confidence

Symbolic numeric indicator in draft frontmatter/DB. Source-count and best structural-quality calculation, not a calibrated correctness probability. Distinct scalar output of a part of RT-RTE-pipeline-run and input to RT-RTE-draft-publication. SRC-1, src/synto/pipeline/compile.py (_compute_confidence).

EPI-OBJ-query-answer

Label: QueryAnswer

Natural-language Markdown answer/title in current-call result; save paths retain Markdown and symbolic source lists/hashes. It is the output object, not a duplicate of RT-RTE-query-answer. SRC-1, src/synto/pipeline/query.py (_query_core, run_query, _save_synthesis_new).

Object evidence (implementation/schema; model instruction descriptions are doctrine/design):

summary: str = Field(description="2-3 sentence summary in the note's language") --- src/synto/models.py

description="Source quality: high=well-structured, medium=usable, low=noise" --- src/synto/models.py

title: str content: str = Field( description="Full markdown body with [[wikilinks]] inline (no frontmatter)" ) --- src/synto/models.py

Routes

EPI-RTE-candidate-filter

Label: Concept and relation filtering

Part of: RT-RTE-pipeline-run

SRC-1, src/synto/pipeline/ingest.py: code checks model-proposed concept lexical support (low/medium sources, including validated aliases) and relation endpoint resolvability. Python decides/vetoes; high-quality sources are a counterexample to universal lexical filtering. No answer oracle. Open ingestion, not a controlled experiment. These are checks before durable admission, not evidence-consuming decisions about article truth.

• Immediate return: Filtered candidate list and eligible relation writes.
• Later read-back: Surviving names and relation rows affect later compilation/routing; rejected raw relations remain in relation_candidates.
• Delegated visibility: Models supply proposals; code sees source body/path, quality and known identities.
• Selection predicate: Non-noise names; low/medium lexical evidence or alias evidence; relation endpoints canonicalizable.
• Invalidation or expiry: Reingestion/identity changes may replace associations; no truth expiry established.
• Activation or effect: Blocks selected candidate admission by fixed predicates.
• Evidence limits: No observed candidate or semantic evaluation; relation predicate/evidence truth not established.
• implementation conclusion status: wired
• operation conclusion status: uninspected

if result.quality in ("low", "medium") and not has_evidence: --- src/synto/pipeline/ingest.py

if subject is None or object_ is None: --- src/synto/pipeline/ingest.py

db.insert_relation_candidates(result.relations, seg_id) --- src/synto/pipeline/ingest.py

EPI-RTE-occurrence-lineage

Label: Chunk occurrence and relation retention

Part of: RT-RTE-pipeline-run

SRC-1, src/synto/pipeline/ingest.py: normalized concepts are attributed to every segment in their analysis chunk, with occurrence confidence 1.0; relation evidence and segment identity are retained separately. Code owns writes after normalization/endpoint filtering; no per-row human approval or semantic oracle. This is retained provenance metadata, not lifecycle integration after factual acceptance. Its consumer overlaps RT-RTE-mcp-answer and graph-expanded RT-RTE-query-answer but its persistence is within the pipeline only.

• Immediate return: DB occurrence replacement and relation upserts.
• Later read-back: MCP concept passages and graph neighbors can reuse these bindings later.
• Delegated visibility: Chunk input segment IDs and returned concept names; relations carry model evidence strings.
• Selection predicate: Only names surviving canonical normalization; relations require known endpoints.
• Invalidation or expiry: Clears source occurrences before replacement; source/identity changes govern future associations.
• Activation or effect: Retains attribution that can rank/select later source passages and context.
• Evidence limits: Chunk association overincludes possible evidence; no supplied attribution correctness measurements.
• implementation conclusion status: wired
• operation conclusion status: uninspected

for sid in seg_ids: seg_to_concepts.setdefault(sid, set()).update(canon_here) --- src/synto/pipeline/ingest.py

db.upsert_concept_occurrences([occ(name=c, confidence=1.0) for c in sorted(concepts)], sid) --- src/synto/pipeline/ingest.py

EPI-RTE-confidence-score

Label: Source-count quality proxy

Part of: RT-RTE-pipeline-run

SRC-1, src/synto/pipeline/compile.py: deterministic code computes min(1, 0.25 × resolved source count + best quality bonus). Quality means structural usability in AnalysisResult, not source truth. It evaluates a proxy for draft admission, not propositions. Example deduction from inspected code: four low-quality resolved sources give 1.0; three high-quality resolved sources also give 1.0. Those scores can satisfy an enabled threshold without checking any article sentence. This is arithmetic over metadata, not entailed derivation of article facts.

• Immediate return: Numeric scalar passed to draft writer.
• Later read-back: Retained confidence is read by review/publication thresholds and article consumers.
• Delegated visibility: Code sees source paths/DB quality, not claim-level supporting text in this function.
• Selection predicate: Best available quality, default low; count of resolved sources.
• Invalidation or expiry: Recompile recalculates; changing sources does not establish historical calibration.
• Activation or effect: Can enforce caller-selected minimum through publication branch; not a universal gate.
• Evidence limits: No calibration or factual-outcome comparison supplied.
• implementation conclusion status: wired
• operation conclusion status: uninspected

_QUALITY_BONUS = {"high": 0.25, "medium": 0.1, "low": 0.0} --- src/synto/pipeline/compile.py

best = _source_quality_summary(source_paths, db) return min(1.0, len(source_paths) * 0.25 + _QUALITY_BONUS.get(best, 0.0)) --- src/synto/pipeline/compile.py

EPI-RTE-markdown-repair

Label: Unknown-link cleanup

Part of: RT-RTE-pipeline-run

SRC-1, src/synto/pipeline/compile.py: _write_draft calls _strip_unknown_wikilinks before retention. Deterministic example: with known_titles=["Known"] and no stem map, input "A causes [[Missing]]." becomes "A causes Missing." The predicate keeps known/source links; otherwise display or target text is returned. This scoped non-ampliative reshaping removes a navigation assertion but neither tests nor retracts the causal assertion. The example is a code deduction, not target execution; other cleanup steps are not assumed globally semantics-preserving.

• Immediate return: Rewritten Markdown body.
• Later read-back: Draft/published text and later query/export receive the cleaned body.
• Delegated visibility: Code sees generated text, resolvable titles and source targets.
• Selection predicate: Unknown target outside source prefix and known names.
• Invalidation or expiry: Recomputed on draft writing; original model text not guaranteed separately retained.
• Activation or effect: Removes broken-link syntax while preserving ordinary display text.
• Evidence limits: No truth evaluator or observed generation; semantic safety of every cleanup helper unassessed.
• implementation conclusion status: wired
• operation conclusion status: uninspected

if target.lower().startswith("sources/") or target.lower() in known: return match.group(0) return display or f"{target}{fragment}" --- src/synto/pipeline/compile.py

body = _strip_unknown_wikilinks(body, (resolvable_titles or []) + source_targets, stem_to_title) --- src/synto/pipeline/compile.py

EPI-RTE-synthesis-guard

Label: Synthesis safety checks

Part of: RT-RTE-query-answer

SRC-1, src/synto/pipeline/query.py: Python rejects synthesis sources that resolve to synthesis pages, recognizes duplicate normalized questions, and refuses in-place overwrite of hash-detected edits. The caller selects duplicate strategy; code can veto; humans can edit files but no factual reviewer is required. These checks target source-chain policy, identity and edit preservation, not answer truth. Keep-existing may return an old file before chain checks. Open question-saving mode; no reference-answer oracle.

• Immediate return: Save eligibility, existing-file return or exception.
• Later read-back: Source paths/hashes and saved answer are retained only through save route; existing file can remain.
• Delegated visibility: Code reads resolved source types, question hash and existing file/body hash.
• Selection predicate: Duplicate strategy, synthesis source exclusion, content hash conflict.
• Invalidation or expiry: New question/source resolution per call; manual edits block update rather than validate truth.
• Activation or effect: Can prevent write or keep existing; no factual authority is granted.
• Evidence limits: No executed save or recovery trace; hash consistency does not show grounding.
• implementation conclusion status: wired
• operation conclusion status: uninspected

if any(_is_synthesis_source(config, db, path) for path in resolved_sources): raise SynthesisChainError("Synthesis sources cannot include another synthesis page") --- src/synto/pipeline/query.py

if existing.content_hash and existing.content_hash != _body_hash(existing_body): --- src/synto/pipeline/query.py

EPI-RTE-answer-retention

Label: Saved answer admission

Part of: RT-RTE-query-answer

SRC-1, src/synto/pipeline/query.py and src/synto/indexer.py: caller flags admit model-written answer text through two distinct paths. Ordinary Q&A is written to wiki/queries/ with published metadata but no article DB registration in _save_query; refreshing the index does not enumerate that tree. Synthesis is registered as an article with kind synthesis and enters the generated index, which shows up to 25 newest synthesis entries; it uses separate guards. Both forms are filesystem/export-accessible (see RT-RTE-pack-export). _find_page_unchecked recursively scans wiki Markdown frontmatter titles, so an unindexed Q&A can still load if its title reaches the resolver. The registration/index distinction agrees with the supplied peer MEM-OBJ-synthesis and MEM-OBJ-index. Flags authorize automatic content admission, not per-content review; the generating model proposes, code writes, caller can decline saving. Recovery uses keep-existing/update/suffix strategies and edit guards rather than draft rejection. Prompt guidance requests answering from wiki text. No inspected evidence-consuming factual acceptance decision precedes saving.

• Immediate return: Path/resolution or save exception.
• Later read-back: Registered synthesis enters later index selection, subject to the 25-entry catalogue cap. Ordinary Q&A files are not automatically enumerated into that index; they remain filesystem/export-accessible and can be loaded by the recursive title resolver when a matching title is supplied. No actual later consumption was observed.
• Delegated visibility: Stored answer, selected page names, synthesis source hashes; selected names may exceed loaded bodies.
• Selection predicate: save/synthesize flags and synthesis guard outcomes.
• Invalidation or expiry: Question deduplication and source hashes track identity/freshness, not endorsement; no factual expiry.
• Activation or effect: Saves both forms without the draft publication gate; synthesis changes indexed answer context, while ordinary Q&A changes filesystem/export availability and remains title-resolvable without automatic index admission.
• Evidence limits: No saved answer instance; status published is operational metadata.
• implementation conclusion status: wired
• operation conclusion status: uninspected

status="published", kind="synthesis", question_hash=question_hash, --- src/synto/pipeline/query.py

"Synthesis pages capture prior answers; consider them when relevant " "but never prefer them over a fresh concept page. " --- src/synto/pipeline/query.py

"status": "published", } write_note(path, meta, answer) --- src/synto/pipeline/query.py

write_note(path, meta, answer) append_log(config, f"query | {question[:60]}") generate_index(config, db) return path --- src/synto/pipeline/query.py

visible_entries = synthesis_entries[:25] --- src/synto/indexer.py

for md in config.wiki_dir.rglob("*.md"): --- src/synto/pipeline/query.py

EPI-RTE-offline-evaluation

Label: Offline structural checks

Part of: RT-RTE-compare-eval

SRC-1, src/synto/pipeline/eval.py: deterministic evaluator compares expected concept existence, nonempty sources metadata, INDEX schema and link resolution. Fixture author supplies expected concepts, not an answer oracle for truth; expected_contains is loaded but term_recall is skipped. Operating mode is offline structural evaluation; no production-policy successor selection is implemented here. A concrete deduction: any article with sources=["missing.md"] counts in citation coverage because only list type and length are checked; its body and the source contents are not compared. This does not establish actual corpus behavior.

• Immediate return: EvalResult and text/JSON metric report.
• Later read-back: Reports/comparison artifacts can be inspected later; no inspected automatic installation consumer.
• Delegated visibility: Reader supplies published article frontmatter/body and index; fixture supplies expected concepts.
• Selection predicate: All listed articles and fixture concepts; missing metrics skipped in aggregate.
• Invalidation or expiry: A later invocation recomputes; report itself confers no expiry or acceptance.
• Activation or effect: Reporting force; no universal publication veto or knowledge rewrite.
• Evidence limits: No observed scores, answer quality or causal component effect; live mode unimplemented.
• implementation conclusion status: wired
• operation conclusion status: uninspected

sources = article.frontmatter.get("sources", []) if isinstance(sources, list) and len(sources) > 0: with_sources += 1 --- src/synto/pipeline/eval.py

term_recall = None --- src/synto/pipeline/eval.py

"metrics_skipped": ["term_recall"], --- src/synto/pipeline/eval.py

EPI-RTE-verbatim-source

Label: Verbatim segment delivery

Part of: RT-RTE-mcp-answer

SRC-1, src/synto/serve.py: a separate primitive in the MCP container looks up a source segment, checks license permission and returns stored text/identity/hash with truncation status. This is not the synthesized answer subroute. ID and policy checks decide disclosure; no evaluator decides original source truth. Source search ranks by BM25 and concept passages use retained occurrence bindings; ranking does not confer epistemic endorsement.

• Immediate return: Dictionary with body, source identity/hash and truncation flag, or tool error.
• Later read-back: Previously imported and persisted source segments reach later MCP client invocations as stored text, with optional truncation. This is a read-only memory read-back, not a Synto answer-saving write; downstream model use remains host-dependent and unobserved. See the earlier import object MEM-OBJ-source-documents and later delivery MEM-RTE-source-passages.
• Delegated visibility: External caller receives stored paragraph, optionally clipped with ellipsis; models need not generate it.
• Selection predicate: Requested segment ID, existence and source license gate; search/passage alternatives select differently.
• Invalidation or expiry: Source/license policy changes govern availability; runtime notes process-cached legacy mode.
• Activation or effect: Permits or blocks source disclosure; gives a consumer attributed source text.
• Evidence limits: Extraction-to-original fidelity not fully inspected; no external reader compliance or truth evidence.
• implementation conclusion status: wired
• operation conclusion status: uninspected

body: str = row["text"] truncated = False --- src/synto/serve.py

body = body[:cap] + "…" truncated = True --- src/synto/serve.py

EPI-RTE-state-recovery

Label: Undo and clean state

Distinct identity: no supplied route isolates this operation; comparisons are stated below.

SRC-1, src/synto/cli.py: distinct recovery/destructive operations absent from the supplied route inventory, not duplicates of RT-RTE-concept-curation. An operator requests Git revert or clean. Undo refuses identity-op batches unless forced because Git cannot reverse DB mutations; clean preserves an INDEX rebuild seed then removes DB/wiki while keeping raw notes. Human invocation/confirmation admits withdrawal; software guards known inconsistency. These are operational recovery paths, not truth-directed rejection or complete rollback.

• Immediate return: Reverted commit messages, refusal, or deleted-state report.
• Later read-back: Reverted files/remaining DB or rebuild seed affect future ingest/compile.
• Delegated visibility: Operator sees warnings and selected targets; no model judgment is involved.
• Selection predicate: Last N auto-commits; identity-op guard unless force; configured DB/wiki paths for clean.
• Invalidation or expiry: Changes local availability immediately; raw notes/rebuild seed persist.
• Activation or effect: Withdraws local generated material or reverts files; DB may remain divergent after undo.
• Evidence limits: No observed rollback; cannot revoke already exported packs or external answers.
• implementation conclusion status: wired
• operation conclusion status: uninspected

if identity_commits and not force: --- src/synto/cli.py

seed_path = generate_index_json(config, db) --- src/synto/cli.py

targets = [ ("state DB", config.state_db_path), ("wiki/", config.wiki_dir), ] --- src/synto/cli.py

Claims

EPI-CLM-wiki-synthesis

Label: Synthesized current wiki

Distinct identity: no claims are declared in the supplied runtime. SRC-1, README.md, doctrine/design: the public rationale describes synthesis, cross-reference and currency; this is not observed knowledge quality. Compare RT-RTE-pipeline-run and RT-RTE-query-answer.

a personal knowledge base where the model doesn't just store what you tell it — it synthesizes, cross-references, and keeps everything current. You add raw material; it does the bookkeeping. --- README.md

EPI-CLM-quality-evaluation

Label: Offline quality evaluation

Distinct identity: no supplied claim counterpart. SRC-1, README.md, doctrine/design. Public “citation support” is bounded by the inspected metadata check, not source-entailment assessment. README also explicitly calls the mechanism offline structural evaluation, so the limited implementation is consistent with that narrower characterization. See EPI-RTE-offline-evaluation.

Quality evaluation. synto eval scores your wiki offline: concept coverage, citation support, link resolution, INDEX.json validity. Run it before a pack export or in CI. --- README.md

• synto eval — offline structural evaluation (coverage, citation support, link resolution) --- README.md

EPI-CLM-staged-review

Label: Confidence and staged lifecycle

Distinct identity: no supplied claim counterpart. SRC-1, README.md, doctrine/design. The three-state wording must not be read as enforced sequential verification; the same text allows an explicit staged step rather than requiring it. See RT-RTE-draft-publication and EPI-RTE-confidence-score.

Confidence scores. Each compiled draft gets a confidence score (0–1). Approve selectively or set a threshold — drafts below it stay in .drafts/ for manual review. --- README.md

Three-state lifecycle. Drafts move through draft → verified → published. --- README.md

Evidenced absences

none declared in this member

Behavioral-authority paths

none declared in this member; new route records state their distinct consumers, channels, forces and horizons. The supplied RT-BAP-rejection-guidance and RT-BAP-pack-guidance remain canonical for those paths.