Signet epistemic report

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

Source-and-claim boundary

The question is which Signet routes acquire or transform truth-apt content, check it, authorize later reliance and revoke that authorization. Scope and external exclusions are those of boundary; source identity and evidence layers are in its Source register. All new inspection below is SRC-1 implementation unless marked doctrine/design. No execution, model output, verified benchmark result or causal experiment is supplied. Model weights and external hosts remain outside the frozen source. Their truth checks, compliance, learning and exact model versions cannot be inferred from Signet's interfaces; see RT-CMP-inference-router.

The runtime is provisional, not exhaustive. Direct inspection found a consequential additional route: aggregate recall can synthesize and save semantic memory outside a Dreaming pass. This qualifies any reading of Dreaming as the exclusive semantic writer, but does not invalidate the runtime's explicitly bounded maintenance account. See EPI-RTE-aggregate-synthesis and EPI-RTE-aggregate-retention. Public claims assessed are EPI-CLM-source-connections, EPI-CLM-benchmark-accuracy, and the supplied tool-boundary claim RT-CLM-tool-boundary.

Coverage

Direct source reads and supplied findings are distinct. Available but unread operations below prevent a system-complete negative about acceptance, evaluators or semantic checks.

Entry point or operation	Source path	Coverage and records, or limit
CLI dispatch	SRC-1 surfaces/cli/src/cli.ts	Direct registration search: memory, knowledge, ontology, hook, source and Dream commands are shipped; command-body equivalence not inspected.
Pi lifecycle and hook delivery	SRC-1 integrations/pi/extension/src/index.ts, platform/daemon/src/hooks.ts	Supplied RT-RTE-pi-lifecycle, RT-BAP-host-context; no independent reinspection of budget selectors.
Direct remember and structured input	SRC-1 platform/daemon/src/routes/memory-routes.ts, platform/daemon/src/mcp/tools.ts	Supplied RT-RTE-direct-remember; direct MCP registration search confirms recall, aggregate options, remember and forget exposure. Full direct-ingest transaction not re-traced.
Dreaming proposal and operation checks	SRC-1 platform/daemon/src/pipeline/dreaming-operations.ts, platform/daemon/src/pipeline/dreaming-capabilities.ts	Direct citation/validation reads and capability-registration searches; supplied RT-RTE-dreaming-pass. Structural and hygiene exceptions prevent a universal quote requirement.
Memory head support and removal	SRC-1 platform/daemon/src/memory-head-owner.ts	Direct reads; EPI-RTE-head-support, EPI-RTE-head-publication. Full file-to-host reload beyond supplied lifecycle remains uninspected.
Aggregate recall	SRC-1 platform/daemon/src/aggregate-recall.ts, platform/daemon/src/routes/memory-routes.ts	Direct synthesis, guard, transaction and final-result reads; EPI-RTE-aggregate-synthesis, EPI-RTE-aggregate-guard, EPI-RTE-aggregate-retention, EPI-RTE-aggregate-delivery. Query-expansion and cache-refresh internals not fully inspected.
Deterministic normalization	SRC-1 platform/daemon/src/content-normalization.ts	Complete direct read; EPI-RTE-content-normalization. Native fallback selection visible, native implementation uninspected.
Supersession	SRC-1 platform/daemon/src/routes/memory-routes.ts, platform/daemon/src/transactions.ts, platform/daemon/src/memory-head-owner.ts	Direct endpoint, transaction and current-source reads; EPI-RTE-memory-supersession. Forget/recover endpoints were located, not traced through their consumers; prevents comprehensive restoration/withdrawal assurance.
Recall selection	SRC-1 platform/daemon/src/memory-search.ts	Direct selected response and access-tracking read; exhaustive ranking and visibility predicates uninspected. Neither access counters nor selected delivery prove endorsement.
MemoryBench evaluator	SRC-1 memorybench/src/orchestrator/phases/evaluate.ts, memorybench/src/judges/base.ts	Complete direct reads; EPI-RTE-benchmark-judgment, EPI-RTE-benchmark-checkpoint. Dataset/reference validity, actual judge implementations and protocol-specific evaluators remain uninspected.
Alternate executor	SRC-1 platform/daemon/src/inference-router.ts, platform/daemon/src/pipeline/acpx-dreaming-mcp.ts	Supplied RT-RTE-alternate-executor; external evaluator/tools opaque.
Other material included paths	SRC-1 platform/daemon/src/transcript-capture.ts, scripts/claim-trace-eval.ts, scripts/observer-belief-eval.ts, platform/daemon/src/mcp/tools.ts	Capture normalization, feedback, other registered graph tools, evaluation scripts, source imports, ontology review/recovery, history summarization, identity generation and other host adapters not fully inspected. Their presence is not absence of a truth check; no whole-system negative is supported.

Epistemic-object inventory

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
RT-OBJ-pass-prompt	see record	see record	see record	Guidance requests interpretation; not an observed accepted claim.
EPI-OBJ-memory-entries	see record	see record	see record	Exact support is not semantic entailment.
EPI-OBJ-aggregate-evidence	see record	see record	see record	Retrieved content's original truth warrant unknown.
EPI-OBJ-aggregate-answer	see record	see record	see record	No actual output instance supplied; truth-apt scope is intended memory facts.
EPI-OBJ-normalized-content	see record	see record	see record	Storage representation differs from dedupe representation.
EPI-OBJ-judge-result	see record	see record	see record	Evaluation labels concern benchmark answers, not all stored propositions.

Authority-route ledger

Fields owned by declaring records are referenced rather than copied. Architectural status describes source support, not observed activation. For every row below, possible results, consumers, timing, license and limits are in the cited route record.

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-direct-remember	operational admission/selection/consumption	implemented	see route	truth-apt transformation: acquisition/import	see route	see route	see route	see route	see route	Import warrant unknown; schema validity is not source truth	see route	see route	see route	none	none	Human authorship unresolved; retention does not establish acceptance.
RT-RTE-dreaming-pass	content transformation	implemented	RT-OBJ-pass-prompt	truth-apt transformation: indeterminate	see route	see route	see route	see route	see route	Requested extraction lacks checked semantic entailment	see route	see route	see route	EPI-CLM-source-connections	none	No observed input/output pair distinguishes reshaping, derivation or conjecture.
RT-RTE-dreaming-pass	check/evidence production	implemented	see route	no content change	see route	see route	see route	see route	see route	Scoped exact-source matching, not truth adjudication	see route	see route	see route	RT-CLM-tool-boundary	none	Direct source reading confirms exceptions for structural and hygiene writes.
RT-RTE-pi-lifecycle	operational admission/selection/consumption	implemented	see route	no content change	see route	see route	see route	see route	see route	No additional epistemic license from startup delivery	see route	see route	see route	none	none	Host uptake unobserved; see RT-BAP-host-context.
EPI-RTE-content-normalization	content transformation	implemented	EPI-OBJ-normalized-content	truth-apt transformation: non-ampliative reshaping	see route	see route	see route	see route	see route	see route	see route	see route	see route	none	none	Confined to storage formatting; normalized hash equivalence need not preserve meaning.
EPI-RTE-head-support	check/evidence production	implemented	EPI-OBJ-memory-entries	no content change	see route	see route	see route	see route	see route	see route	see route	see route	see route	EPI-CLM-source-connections	none	Provenance check, not acceptance of the entry's truth.
EPI-RTE-head-publication	retention	implemented	EPI-OBJ-memory-entries	no content change	see route	see route	see route	see route	see route	see route	see route	see route	see route	EPI-CLM-source-connections	none	Operational commit acceptance is not evidence-consuming truth acceptance.
EPI-RTE-aggregate-synthesis	content transformation	implemented	EPI-OBJ-aggregate-answer	truth-apt transformation: indeterminate	see route	see route	see route	see route	see route	see route	see route	see route	see route	none	additional semantic route	Output semantics not observed; instruction alone does not establish non-ampliative transformation.
EPI-RTE-aggregate-guard	check/evidence production	implemented	EPI-OBJ-aggregate-answer	no content change	see route	see route	see route	see route	see route	see route	see route	see route	see route	none	none	Eligibility/style domain only.
EPI-RTE-aggregate-retention	retention	implemented	EPI-OBJ-aggregate-answer	no content change	see route	see route	see route	see route	see route	see route	see route	see route	see route	none	additional semantic route	Source links do not establish accepted ampliative knowledge or lifecycle integration after acceptance.
EPI-RTE-aggregate-delivery	operational admission/selection/consumption	implemented	EPI-OBJ-aggregate-answer	no content change	see route	see route	see route	see route	see route	see route	see route	see route	see route	none	save rejection need not block use	Final returned text is credential-redacted; source-memory metadata excludes ontology source IDs.
EPI-RTE-memory-supersession	lineage/freshness/recovery	implemented	see route	no content change	see route	see route	see route	see route	see route	see route	see route	see route	see route	none	none	Supersession changes eligibility, not demonstrated falsity of old content.
EPI-RTE-benchmark-judgment	check/evidence production	implemented	EPI-OBJ-judge-result	truth-apt transformation: indeterminate	see route	see route	see route	see route	see route	see route	see route	see route	see route	EPI-CLM-benchmark-accuracy	none	Model judgment or opaque protocol; no evaluator calibration supplied.
EPI-RTE-benchmark-checkpoint	retention	implemented	EPI-OBJ-judge-result	no content change	see route	see route	see route	see route	see route	see route	see route	see route	see route	EPI-CLM-benchmark-accuracy	none	Lab result, not product knowledge admission.

System-claim versus route comparison

Claim ID	Claimed operation or warrant	Claim source ID/anchor and evidence layer	Doctrine/design support	Implemented route IDs	Observed-run support	Causal support and design limits	Supported conclusion	Mismatch/unknown
EPI-CLM-source-connections	see record	see record	see record	RT-RTE-dreaming-pass, EPI-RTE-head-support, EPI-RTE-head-publication	none supplied	No intervention; structural/hygiene exceptions in inspected operations	Scoped provenance machinery exists for selected semantic/head writes	“Every connection” is not established as a whole-graph invariant; source links do not certify truth.
EPI-CLM-benchmark-accuracy	see record	see record	see record	EPI-RTE-benchmark-judgment, EPI-RTE-benchmark-checkpoint	Reported score only; no inspected run	No component variation or verified comparison	Shipped evaluation can score answers against references and retain results	97.6% not verified; answer success would not establish extraction mechanism or transfer.
RT-CLM-tool-boundary	see record	see record	see record	RT-RTE-dreaming-pass, RT-RTE-alternate-executor	see record	see record	Bounded Pi worker configuration and daemon checks	Neither a whole-system write restriction nor an epistemic oracle.

Bounded conclusion

Signet has implemented acquisition, model-guided interpretation, scoped provenance checking, retained source-linked projections and advisory host delivery. Direct remember imports caller text with unknown source warrant. The selected Dreaming operation checks and memory-head checks license exact correspondence to eligible source text, not truth or entailment of the model's claim. They can block publication for invalid provenance, scope, pass state or head freshness; they do not establish an accepted ampliative claim under a named truth criterion, intended use and scope. Structural and hygiene exceptions are real counterexamples to a universal episodic-quote rule. See RT-RTE-direct-remember, RT-RTE-dreaming-pass, EPI-RTE-head-support.

The TypeScript storage-normalization path can be understood without execution: it preserves the chosen sentence while changing newline/edge formatting. Its dedupe representation additionally collapses case and punctuation; equality in that representation is not a truth judgment. See EPI-RTE-content-normalization. No new entailed factual proposition from warranted premises is established by that path.

Aggregate recall is an independent model-based semantic-production and optional persistence route. Its prompt requests evidence-only stable facts; no output pair establishes whether a realized answer preserves, derives or adds claims. Its save gate checks eligible evidence classes and surface form, not whether the answer follows from the evidence. A surface-form rejection can still return an unsaved answer to the caller. Source links and embeddings make admitted text available for later reliance without demonstrating epistemic acceptance or post-acceptance lifecycle integration. Immediate delivery has advisory force; retention has durable recall potential. See EPI-RTE-aggregate-synthesis, EPI-RTE-aggregate-guard, EPI-RTE-aggregate-retention, EPI-RTE-aggregate-delivery.

Supersession is an implemented withdrawal-related route: the transaction records replacement and invalidates derived memories, and a superseded memory cannot pass the head's current-source eligibility check. This is a bounded downstream consequence, not proof of withdrawal from already delivered host context or every recall branch. See EPI-RTE-memory-supersession.

MemoryBench is a separate evidence-consuming evaluator of answer hypotheses against benchmark references or rubrics. Its parser checks response consistency, and its checkpoint consumer records correctness and completion, suppressing repeat evaluation on resume. This does not approve individual product memories or demonstrate that model judgment is reliable. No supplied observation verifies the public accuracy score or supports causal component attribution. See EPI-RTE-benchmark-judgment, EPI-RTE-benchmark-checkpoint.

Missing provider internals, execution traces, actual semantic outputs, dataset validation and unread material routes bound these conclusions separately. They prevent claims of exercised learning, successful self-improvement, theory-directed criticism or a system-complete absence of truth acceptance. Persistence and scheduled maintenance are architectural contributions, not substitutes for independent evidence of those properties. No single epistemic verdict follows.

Shared records

Components

none declared in this member

Operative objects

EPI-OBJ-memory-entries

Label: Memory head entries

Natural-language canonical_text is retained in SQLite memory_head_entries; symbolic entry IDs, revision membership and provenance JSON determine publication/support eligibility. Entries are intended truth-apt remembered facts within the agent scope. This is a material output part distinct from the input prompt RT-OBJ-pass-prompt; no supplied object denotes this entry set. SRC-1 platform/daemon/src/memory-head-owner.ts.

const support = (parsed as Record[]).filter( (item) => supportError(db, agentId, row.entryId, item, evidence) === null, ); if (support.length > 0) entries.push({ entryId: row.entryId, text: row.text, support }); --- platform/daemon/src/memory-head-owner.ts

EPI-OBJ-aggregate-evidence

Label: Aggregate recall evidence

Natural-language RecallResult.content with symbolic IDs, types and timestamps is an in-memory selected evidence array. Its intended role is premises for a stable memory note, not a new source of truth. Truth-apt facts are acquired from already retained content; their source warrant is unknown. Distinct from the guidance object RT-OBJ-pass-prompt and from direct admission RT-RTE-direct-remember: this is retrieved input for an additional route. SRC-1 platform/daemon/src/aggregate-recall.ts.

const evidence = uniqueEvidence(recalls.flatMap((result) => result.results)); const evidenceIds = evidence.map((row) => row.id); --- platform/daemon/src/aggregate-recall.ts

EPI-OBJ-aggregate-answer

Label: Aggregate recall answer

Natural-language model text exists first in memory and may enter SQLite memories as a semantic row. Symbolic sourceType, source links and visibility govern its retention and later eligibility. Its intended truth-apt domain is stable facts relevant to the recall question; no concrete answer instance is supplied. This is not the maintenance prompt or its pass output: compare RT-OBJ-pass-prompt, RT-RTE-dreaming-pass. SRC-1 platform/daemon/src/aggregate-recall.ts.

why: "aggregate recall", project, importance: 0.75, type: "semantic", tags: "aggregate,recall", --- platform/daemon/src/aggregate-recall.ts

EPI-OBJ-normalized-content

Label: Normalized memory content

NormalizedMemoryContent carries natural-language storageContent and a separate symbolic dedupe/hash representation (normalizedContent, hashBasis, contentHash) in memory. Storage content can carry caller propositions; the hash is not a proposition about their truth. A material normalization part of the supplied direct-admission route, not a duplicate retained-memory object. SRC-1 platform/daemon/src/content-normalization.ts.

Part of: RT-RTE-direct-remember

readonly storageContent: string; readonly normalizedContent: string; readonly hashBasis: string; readonly contentHash: string; --- platform/daemon/src/content-normalization.ts

EPI-OBJ-judge-result

Label: Benchmark judge result

Symbolic binary score/label/pass and natural-language explanation are held in a result object then a run checkpoint. The truth-apt judgment is whether a benchmark answer meets the reference/rubric for its question, not whether Signet's graph is true. No counterpart exists in supplied runtime records: RT-RTE-dreaming-pass expressly has no expected-answer oracle in its maintenance seam. SRC-1 memorybench/src/judges/base.ts, memorybench/src/orchestrator/phases/evaluate.ts.

if ((parsed.score === 1) !== (parsed.label === "correct")) { throw new Error("Judge response score and label disagree") } --- memorybench/src/judges/base.ts

Routes

EPI-RTE-content-normalization

Label: Content normalization

Part of: RT-RTE-direct-remember

The caller's text enters the TypeScript normalization functions, producing storage text and a separate hash basis; fixed code owns progression. SRC-1 platform/daemon/src/content-normalization.ts. This is the formatting part of admission, not another acquisition route. With native loading unsuccessful, concrete input " Alice moved.\r\n " becomes storage "Alice moved.", normalized text and hash basis "alice moved", and SHA-256 of that basis. This deduction uses implementation, not target execution. The storage edge is non-ampliative reshaping for this input: the same sentence remains. The dedupe edge intentionally discards case and terminal punctuation and therefore cannot certify semantic equivalence for arbitrary propositions (for example quoted case-sensitive identifiers).

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: storageContent, normalizedContent, hashBasis and contentHash.
• Later read-back: uninspected — normalization feeds ingestion, but downstream dedupe consumers were not completely traced.
• Delegated visibility: caller text is available to fixed code; no model judgment occurs here.
• Selection predicate: native implementation if loaded; otherwise the inspected TypeScript branch.
• Invalidation or expiry: current-call result; retained hashes depend on ingestion, not an expiry in this function.
• Activation or effect: changes text representation and hash input; does not establish truth or source endorsement.
• Evidence limits: concrete deduction confined to TypeScript branch; native behavior uninspected.

Evaluator/target/domain: deterministic newline, whitespace, case and punctuation rules over strings, not factual verification. Force: returned representation for admission/dedupe code in the current call. Epistemic license: none added. No human approval, oracle, novel proposition or epistemic acceptance is involved.

function tsNormalizeContentForStorage(content: string): string { return content.replace(CRLF, "\n").trim(); } --- platform/daemon/src/content-normalization.ts

const lowered = storageContent.toLowerCase(); return lowered.replace(COLLAPSE_WHITESPACE, " ").replace(TRAILING_PUNCTUATION, "").trim(); --- platform/daemon/src/content-normalization.ts

EPI-RTE-head-support

Label: Memory head support check

Part of: RT-RTE-dreaming-pass

Model-selected entry/support pairs → daemon supportError → admissible pair or error. The program checks quote occurrence in rendered canonical scoped evidence; source lookup checks current memory eligibility and source completion. This is the checking part of head maintenance, separate from publication. SRC-1 platform/daemon/src/memory-head-owner.ts; object EPI-OBJ-memory-entries.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: null for valid support or INVALID_PROVENANCE error.
• Later read-back: committedEntries rechecks retained provenance and omits entries with no valid support; unverifiable metadata is separately flagged.
• Delegated visibility: model supplies entry text and support references/quotes; daemon reads complete rendered sources in the agent scope.
• Selection predicate: supported source-ref classes, nonempty quote, eligible current source, exact substring occurrence.
• Invalidation or expiry: deleted, superseded, stale or archived memory sources fail currentSource; incomplete source captures fail completion eligibility.
• Activation or effect: commitEntries rejects failed support; committedEntries uses checks to select carried entries.
• Evidence limits: exact occurrence does not test whether entry text is entailed by the quote, whether the quote is true, or whether conflicting evidence exists.

Target/domain: source identity/eligibility and verbatim correspondence, not the candidate proposition's truth. Evaluator is fixed daemon code before write or carry-forward selection. The model proposes; code vetoes; no per-entry human decision or expected-answer oracle is established. The returned error enforces a provenance condition for the current commit; null permits further processing, not unconditional publication. Epistemic authority is bounded correspondence to eligible source text. Behavioral path: supportError result → commitEntries/committedEntries → enforced publication or carried-entry selection, current pass and later head reads. The source interpretation remains indeterminate between reshaping, entailed derivation and conjecture.

if (rendered === null || !rendered.includes(quote)) return { code: "INVALID_PROVENANCE", error: entry ${entryId} quote is not exact scoped evidence }; return null; --- platform/daemon/src/memory-head-owner.ts

return source?.completed === true ? source : null; --- platform/daemon/src/memory-head-owner.ts

EPI-RTE-head-publication

Label: Memory head publication

Part of: RT-RTE-dreaming-pass

Model entry set → authorized content-pass head commit → revision/entry tables and pending publication. Code owns disposition and persistence after EPI-RTE-head-support; this record concerns retained membership, not a second truth check. SRC-1 platform/daemon/src/memory-head-owner.ts, platform/daemon/src/pipeline/dreaming-capabilities.ts; object EPI-OBJ-memory-entries.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: accepted commit, NOOP, or rejection such as PASS_NOT_AUTHORIZED, STALE_HEAD or INVALID_HEAD.
• Later read-back: committedEntries reconstructs the last published supported entry set for later passes; exact host reload after publication uninspected.
• Delegated visibility: model submits complete entries, not direct SQLite statements; daemon sees pass/head revisions and retained support.
• Selection predicate: running scoped incremental-content pass, matching base revision, input budget and support; unchanged current hash can return NOOP before revalidating supplied support.
• Invalidation or expiry: omitted active entries become removed; an empty replacement is blocked if carried entries still have valid or unverifiable support.
• Activation or effect: commitEntries creates revision/entries; head metadata and pending publication change on success; content-pass finalization governs durability as supplied runtime establishes.
• Evidence limits: code's accepted commit is operational acceptance of a publication request, not acceptance of factual truth against a substantive criterion.

Trigger/change/guidance: scheduled or requested maintenance submits retained memory entries under the tool's complete-set and budget instructions. Model proposes; daemon vetoes for authorization, stale revision, invalid support and clearing rules. No human per-entry approval or answer oracle appears here. Recovery: stale rejection requires a fresh proposal; supplied runtime distinguishes staged head finalization from earlier committed ontology operations. Criterion/intended use/scope: eligible provenance and publication constraints for agent-scoped MEMORY.md, not a truth criterion. Epistemic license adds no warrant beyond eligible source correspondence. Operational path: fixed head writer → revision membership and publication queue → retained memory projection for later reads; enforcing now, potentially advisory at host consumption. This is retention, not established post-epistemic-acceptance integration.

if (head?.is_current === 1 && currentHash === contentHash) return { ok: true, code: "NOOP", revision, hash: contentHash, changed: false, changedIds: [] }; --- platform/daemon/src/memory-head-owner.ts

if (retained.has(old.entry_id)) continue; db.prepare( "UPDATE memory_head_entries SET status='removed', last_revision=?, updated_at=? WHERE agent_id=? AND entry_id=?", ).run(nextRevision, now, input.agentId, old.entry_id); --- platform/daemon/src/memory-head-owner.ts

EPI-RTE-aggregate-synthesis

Label: Aggregate recall synthesis

Selected recall evidence → rendered prompt → inference-router execute → trimmed answer/null. Model proposes candidate facts and code handles nonempty output. Distinct from RT-RTE-dreaming-pass: invoked by aggregate recall, no Dreaming pass/tool loop required. Compare shared model-selection machinery RT-CMP-inference-router; no new distributed-parametric component identity is established. SRC-1 platform/daemon/src/aggregate-recall.ts; inputs EPI-OBJ-aggregate-evidence, output EPI-OBJ-aggregate-answer.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: answer string or null with routed usage; downstream failure falls back to retrieved evidence.
• Later read-back: answer can enter optional aggregate retention and later recall; synthesis alone does not persist it.
• Delegated visibility: model receives query plus selected content, IDs, creation dates and types, not all underlying sources or full graph.
• Selection predicate: evidence exists and router available; prompt asks for evidence-only stable facts; routed execution returns nonempty text.
• Invalidation or expiry: 30-second timeout supplied to router; failed/empty synthesis yields no generated answer.
• Activation or effect: creates a candidate for caller delivery and optional storage, outside Dreaming maintenance.
• Evidence limits: implemented request for grounded synthesis does not establish compliance, entailment or non-entailment of an actual answer.

Relation: truth-apt transformation indeterminate; remaining possibilities are non-ampliative reshaping, entailed derivation and ampliative conjecture. Need actual source/output propositions and checked interpretation to distinguish them. Epistemic license: none newly established by fluency or router success; source warrant unknown. Operational path: model output → aggregateRecall → optional save and immediate API result, advisory to a caller. Operating mode is open recall, no expected-answer oracle or human per-answer approval. Guidance persists as shipped prompt; model result may persist separately. No accepted ampliative knowledge-production result is established.

"Write one concise atomic memory note from the memory evidence below.", "Use only the evidence.", --- platform/daemon/src/aggregate-recall.ts

const text = result.value.text.trim(); return { answer: text.length > 0 ? text : null, usage: usageStage("synthesis", result.value), }; --- platform/daemon/src/aggregate-recall.ts

EPI-RTE-aggregate-guard

Label: Aggregate save guard

Generated answer and selected evidence → fixed eligibility/style functions → booleans consumed by aggregate saving. This is a distinct check with no counterpart in the supplied direct remember or maintenance route; compare RT-RTE-direct-remember. SRC-1 platform/daemon/src/aggregate-recall.ts; candidate EPI-OBJ-aggregate-answer.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: booleans for evidence eligibility and answer saveability.
• Later read-back: inapplicable — checks return current-call booleans rather than retained evaluation artifacts.
• Delegated visibility: fixed code inspects row identity/visibility/scope and output text; no human evaluator.
• Selection predicate: nonempty evidence all ontology claims or global unscoped linkable memories; answer at least 12 trimmed characters, not refusal or conversational framing.
• Invalidation or expiry: recomputed for the current save attempt.
• Activation or effect: false prevents persistence in this branch, not immediate answer delivery.
• Evidence limits: checks do not compare the answer's propositions with the evidence.

Target/domain: storage eligibility of evidence and surface-form suitability of memory text. Program is evaluator; model generates candidate; no truth oracle or human per-candidate gate is present. Epistemic license: none regarding factual correctness. Operational path: booleans → aggregate save conditional → permissive/enforcing retention gate for this call only. Counterexample to a truth-acceptance reading: an invented third-person sentence longer than 12 characters that avoids forbidden framing would pass the answer-text condition; this is a deduction about the guard, not a fabricated execution. Evidence eligibility is an independent condition.

function aggregateAnswerCanBeSaved(text: string): boolean { const trimmed = text.trim(); return trimmed.length >= 12 && !isInsufficientAggregateAnswer(trimmed) && !isConversationalAggregateAnswer(trimmed); } --- platform/daemon/src/aggregate-recall.ts

isOntologyClaimRow(row) || (isLinkableSourceMemoryRow(row) && row.visibility === "global" && row.scope === null), --- platform/daemon/src/aggregate-recall.ts

EPI-RTE-aggregate-retention

Label: Aggregate memory retention

Eligible synthesis → optional transaction/dedupe → semantic memory with source links and query hint → optional embedding. Fixed code admits/persists model text; not a maintenance-pass effect. Distinct from RT-RTE-direct-remember because producer, guard and source linkage differ; distinct from RT-RTE-dreaming-pass because no pass required. SRC-1 platform/daemon/src/aggregate-recall.ts; object EPI-OBJ-aggregate-answer.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: stored/deduplicated row and saved/deduped flags, or an unsaved result.
• Later read-back: saved semantic rows receive source links, query hints and an attempted embedding for later retrieval; exact complete recall eligibility uninspected.
• Delegated visibility: fixed transaction consumes generated answer, source identities, agent/project and chosen visibility; model does not write SQLite directly.
• Selection predicate: saving defaults on unless either save flag is false; evidence and answer guards must pass; duplicates and unique-constraint races have distinct results.
• Invalidation or expiry: source links support derived-memory invalidation; cache-refresh/dedupe internals not completely inspected.
• Activation or effect: transaction persists text and linkage; unsuccessful embedding does not undo saved memory.
• Evidence limits: saved status certifies retention, not truth acceptance or successful later retrieval/use.

Trigger/proposal: aggregate recall requests a generated stable note; model proposes text, program decides saving, caller can disable it. No per-answer human approval or reference oracle. Rejection/recovery: guard rejection returns unsaved content; unique conflicts resolve duplicates or return unsaved; embedding failure is logged after save. Guidance is the evidence-only synthesis prompt. Intended use/scope: future agent/project memory retrieval; admission criteria are save flag, row eligibility and surface form. Epistemic authority: source-linked retained candidate with unknown warrant, not accepted ampliative knowledge. Behavioral path: ingest/source-link/query-hint writes → memory retrieval machinery → potential later advisory content, durable horizon. Linking before epistemic acceptance remains retention/lineage, not established lifecycle integration after acceptance.

const saveAggregate = params.saveAggregate !== false && params.save_aggregate !== false; --- platform/daemon/src/aggregate-recall.ts

linkAggregateEvidenceSources(db, id, evidenceSources, agentId, now); linkAggregateQueryHint(db, id, agentId, params.query, now); saved = true; return loadAggregateMemory(db, id); --- platform/daemon/src/aggregate-recall.ts

log.warn("memory", "Aggregate recall memory saved without embedding", { --- platform/daemon/src/aggregate-recall.ts

EPI-RTE-aggregate-delivery

Label: Aggregate answer delivery

Aggregate result → saved or unsaved row → credential-redacted single-row API response. Program consumes synthesis even where save guards fail. No counterpart after comparison with RT-RTE-pi-lifecycle: explicit recall return is distinct from unsolicited startup injection. SRC-1 platform/daemon/src/aggregate-recall.ts; object EPI-OBJ-aggregate-answer.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: one redacted content row with aggregate flags, queries and memory-source IDs; absent synthesis instead returns retrieved evidence.
• Later read-back: inapplicable — this delivery itself changes no retained content; saving is the separate retention route.
• Delegated visibility: caller receives rendered answer and metadata, not the full evidence array; metadata filters ontology sources out of sourceMemoryIds.
• Selection predicate: successful nonempty synthesis; a save guard failure does not exclude the answer from delivery.
• Invalidation or expiry: response belongs to the current request; caller persistence/expiry is uninspected.
• Activation or effect: can supply generated text to model/tool callers without truth acceptance or persistence.
• Evidence limits: actual caller use, credential-redaction semantic effects and further host filtering uninspected.

Target/condition: deliver available row, not approve its truth. Force: advisory API/tool content to requesting consumer in the current call. Epistemic license: no increment over the generated candidate. Model caller can decide how to rely on it, but that decision is not established by this source. This is a counterexample to treating failed saveability as rejection from consequential use.

} else { row = unsavedAggregateResult(answer, key, project); } --- platform/daemon/src/aggregate-recall.ts

results: row ? [{ ...row, content: redactCredentials(row.content) }] : [], --- platform/daemon/src/aggregate-recall.ts

sourceMemoryIds: evidenceSources .filter((source) => source.sourceKind === "memory") .map((source) => source.sourceId), --- platform/daemon/src/aggregate-recall.ts

EPI-RTE-memory-supersession

Label: Memory supersession

Explicit supersede request → permission/scope/freeze checks → transaction records replacement/version/history and invalidates derived memories → head current-source check can exclude old evidence. Caller proposes and fixed code decides structural admissibility; supplied RT-RTE-direct-remember names mutation alternatives but does not declare this endpoint/consumer chain. Distinct identity, no supersession of that valid container. SRC-1 platform/daemon/src/routes/memory-routes.ts, platform/daemon/src/transactions.ts, platform/daemon/src/memory-head-owner.ts.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: superseded/idempotent status and versions or missing/self/scope/frozen error.
• Later read-back: head support lookup requires superseded_by IS NULL; derived-memory invalidation is invoked transactionally.
• Delegated visibility: caller provides old/new IDs, optional reason and actor; daemon reads both scopes, state and versions.
• Selection predicate: authorized mutation, existing different IDs in matching agent/project/scope/visibility and admissible prior supersession state.
• Invalidation or expiry: old memory receives superseded_by/time/reason and version increment; dependents invalidated; head support eligibility changes.
• Activation or effect: enforces replacement lineage and prevents superseded memory from serving as current head support.
• Evidence limits: not proof of old claim falsity or new claim truth; every recall selector, recovery reversal and already delivered context were not traced.

Operating mode: explicit open mutation; human authorship/approval unresolved from API identity. Oracle: none established. Admission/rejection is structural, not epistemic. Recovery: transaction status and history retained; full reversal mechanics uninspected. Epistemic authority: changed eligibility, no factual verdict. Behavioral path: supersession state → currentSource predicate → enforced rejection of old source for future head support; bounded future publication effect, not retroactive removal from external windows. Source content is not rewritten by this transition. Guidance is endpoint fields and supplied reason.

invalidateDerivedMemoriesForMemoryInTx(db, input.memoryId, existing.agent_id ?? "default", input.changedAt); --- platform/daemon/src/transactions.ts:688-735

"SELECT 1 FROM memories WHERE id=? AND agent_id=? AND is_deleted=0 AND superseded_by IS NULL AND stale_at IS NULL AND visibility != 'archived'", --- platform/daemon/src/memory-head-owner.ts

EPI-RTE-benchmark-judgment

Label: Benchmark answer judgment

Completed answer hypothesis → benchmark protocol.evaluateQuestion or judge.evaluate with question/reference/hypothesis → score/pass/explanation. The orchestrator owns scheduling; model/protocol judges candidate answers. Distinct from maintenance checks in RT-RTE-dreaming-pass; this is the shipped lab evaluation seam. SRC-1 memorybench/src/orchestrator/phases/evaluate.ts, memorybench/src/judges/base.ts; object EPI-OBJ-judge-result.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: score/pass/explanation and optional metrics/usage; parsing can reject malformed score/label/inconsistency.
• Later read-back: scores/explanations feed checkpoint retention, not product memory admission in this inspected phase.
• Delegated visibility: standard judge receives question, question type, ground truth, hypothesis and optional provider prompts; preference questions label reference a rubric; protocol branch is opaque here.
• Selection predicate: targeted question with completed answer, available hypothesis and evaluation not completed; separate derived-only scoring unless reusing product answer.
• Invalidation or expiry: failure recorded; phase throws with instruction to fix and resume; completed entries skipped on resume.
• Activation or effect: result labels answer correct/incorrect and supplies lab status; no graph rewrite or knowledge acceptance is implemented by this phase.
• Evidence limits: judge implementation/model selection, protocol internals, reference correctness, calibration and execution uninspected.

Target/domain: answer fit to benchmark question/reference or preference rubric, not extraction truth, mechanism identification or general future reliability. Oracle: benchmark supplies groundTruth; judge uses it in standard branch. Parser checks output structure/consistency only, not factual accuracy of the judgment. Model-judged truth relation is indeterminate absent actual outputs and evaluator validity; fixed mapping of label to pass is symbolic. Human role in reference creation and review is uninspected. Operational mode is bounded benchmark evaluation, not open maintenance. Epistemic license: attributed reference-relative judgment only. Behavioral path: evaluator → checkpoint consumer → lab completion and report data, current run/resume horizon; no inspected product acceptance consumer. No exact judge component identity/fixity is established; the supplied routed-maintenance model record must not be silently reused as this judge.

const result = await judge.evaluate({ question: question.question, questionType: question.questionType, groundTruth: question.groundTruth, hypothesis: candidate, providerPrompts: provider?.prompts, }) --- memorybench/src/orchestrator/phases/evaluate.ts

const groundTruthLabel = isPreference ? "Rubric" : "Ground Truth Answer" --- memorybench/src/judges/base.ts

EPI-RTE-benchmark-checkpoint

Label: Benchmark evaluation checkpoint

Evaluation result → checkpointManager.updatePhase → completed score/label/explanation, or failed error → resume selector. Fixed orchestrator owns persistence requests and continuation. Distinct from supplied runtime records after comparison with RT-RTE-dreaming-pass; lab state is not maintenance history. SRC-1 memorybench/src/orchestrator/phases/evaluate.ts; object EPI-OBJ-judge-result.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: per-question ID/duration/label after checkpoint update, or an exception on evaluation failure.
• Later read-back: next evaluation invocation reads phase status and skips completed evaluations.
• Delegated visibility: checkpoint receives score/pass/label/explanation and optional metrics/details/usage; logger receives result/progress.
• Selection predicate: successful result records completed regardless of correct/incorrect; errors record failed and interrupt.
• Invalidation or expiry: failed phase can be resumed; checkpoint serialization/revision internals uninspected.
• Activation or effect: completed result changes continuation and prevents repeat scoring; incorrect is a completed evaluation, not a product-memory veto.
• Evidence limits: checkpoint storage internals and later release/autoresearch gates uninspected; no claim of verified durable run artifact or product improvement.

Target/criterion/use/scope: lab phase completion after valid evaluator return for a benchmark question. This is result retention and workflow disposition, not truth acceptance of the underlying Signet memory. Consumer/channel/force/horizon: updatePhase and getPhaseStatus, symbolic status, enforcing resume selection within that run. No semantic content update occurs here. Recovery is failed status plus resume instruction. No separate human approval is shown.

return status !== "completed" && answerStatus === "completed" && hypothesis --- memorybench/src/orchestrator/phases/evaluate.ts

const label = product.passed ? "correct" : "incorrect" checkpointManager.updatePhase(checkpoint, question.questionId, "evaluate", { status: "completed", score: product.score, passed: product.passed, label, explanation: product.explanation, --- memorybench/src/orchestrator/phases/evaluate.ts

Claims

EPI-CLM-source-connections

Label: Source linked connections

Signet claims automatic memory creation and structured maintenance with every connection linking to source. Doctrine/design, SRC-1 README.md. Distinct from RT-CLM-tool-boundary: this concerns provenance coverage, not tool restriction. The claim does not by itself assert an independently verified truth criterion. Selected implemented support is the Dreaming provenance seam and head support/publication, but structural/hygiene exceptions and uninspected graph paths prevent a universal invariant conclusion.

Signet creates memories automatically from your transcripts, imported files, and other sources. In the background, a process called "dreaming" builds and maintains a structured map of the people, projects, facts, and relationships in your history. Every connection links back to its source, so you can see where it came from. --- README.md

EPI-CLM-benchmark-accuracy

Label: Reported benchmark accuracy

Public reported operation, not an observed-run source: Signet reports its latest tracked MemoryBench average. SRC-1 README.md. No supplied runtime claim has this referent; closest comparator RT-CLM-tool-boundary concerns worker configuration, not answer performance. The inspected evaluator gives a possible measurement route, not verified score, unbiased judgment or causal component effect.

Signet's latest tracked MemoryBench run averages 97.6% LongMemEval answer accuracy. Keeping your memory local shouldn't mean you have to settle for weak recall. See Benchmarks for the methodology, scoring note, and run workflow. --- README.md

Evidenced absences

none declared in this member

Behavioral-authority paths

none declared in this member; each new route states its consumer, channel, force and horizon; supplied host/maintenance paths remain RT-BAP-host-context and RT-BAP-maintenance-tools.