pi-self-learning epistemic report

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

Source-and-claim boundary

The question is whether session reflection and maintenance produce checked knowledge or instead retain model-generated candidates for later operational reliance. Scope and exclusions: see boundary. Evidence identities and layers: see its Source register. This member directly inspected SRC-1, the complete extensions/self-learning.ts, and README.md, at the frozen commit. Packaging and absence of supplied evaluations are taken from the boundary and runtime, not new execution. No providers, host internals, user memories, target executions or experiments were inspected.

Coverage below checks dispatch against the provisional runtime. No omitted material operation was found. The documented command forms represent eight command names: the two documented forms of /learning-model share one registration; the source has eight registerCommand calls. The boundary/runtime's count of nine is a nonconsequential counting defect, not a missing route. No tool registration exists in the inspected extension.

Entry point or operation	Source path	Coverage and record citations	Basis and limit
agent_end, /learning-now	extensions/self-learning.ts	RT-RTE-task-reflection	Direct: bounded transcript, abort filter, operational-signal classification, completion, repair, parser, writes. Host serialization internals excluded.
before_agent_start	extensions/self-learning.ts	RT-RTE-context-injection, RT-BAP-memory-policy	Direct: notes, file ordering, truncation, returned policy; host compliance unobserved.
/learning-month	extensions/self-learning.ts	RT-RTE-month-summary	Direct: daily selection, bounded model input, nonempty-output admission, overwrite.
/learning-redistill	extensions/self-learning.ts	RT-RTE-core-redistill	Direct: confirmation, repair, missing-ID fallback, prefix enforcement, collision merge, final writes.
/learning-toggle, /learning-model, /learning-model-global	extensions/self-learning.ts	RT-RTE-runtime-controls	Direct: branch and settings writes; branch model selection remains ineffective for completions.
/learning-daily, /learning-status	extensions/self-learning.ts	RT-RTE-runtime-controls	Direct: path/config diagnostics, not semantic evaluation or retained-claim withdrawal.
Index load/migration, update, rendering and ranking	extensions/self-learning.ts	EPI-RTE-index-migration, EPI-RTE-watchout-conversion, EPI-RTE-core-ranking	Direct: additional split of material functions within runtime containers.
Git, default-file initialization, timeout/error paths	extensions/self-learning.ts	RT-RTE-task-reflection, RT-RTE-core-redistill	Direct: best-effort history, no automatic rollback, provider cancellation unestablished.
Package loading and remaining tracked files	package.json, AGENTS.md, docs/example-settings.json, .gitignore	boundary	Supplied findings only; no separate evaluator entry point reported.

Consequential claims: EPI-CLM-prevention-scope, EPI-CLM-full-history, and RT-CLM-strict-consultation. No supplied output individuates an actual truth-apt learned proposition: its correctness, novelty, acceptance and later use are not determinable. No run evidence establishes model faithfulness or activation; no causal comparison establishes improvement or cross-project transfer. Provider parameters/versioning: see RT-CMP-reflection-model. Opaque provider state does not justify a claim of either provider learning or provider immutability.

Epistemic-object inventory

Fields owned by a linked record are given there, not redefined here. Candidate truth-apt content is distinguished from imperative policy even when both occupy the same string array.

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
EPI-OBJ-conversation-signals	Claims about past actions, failures and fixes; tool error classifications	see record	SRC-1, extensions/self-learning.ts	Host transcript truth and classification precision unchecked.
EPI-OBJ-reflection-items	Mistake/fix assertions if generated; imperative prevention advice can have no proposition	see record	SRC-1, extensions/self-learning.ts	Actual outputs unavailable; content-level classification not determinable.
EPI-OBJ-index-items	Retained assertion/rule text; numerical recurrence metadata	see record	SRC-1, extensions/self-learning.ts	Scores measure occurrence, not verified success.
EPI-OBJ-monthly-text	Summary assertions if generated	see record	SRC-1, extensions/self-learning.ts	No output or entailment comparison.
RT-OBJ-runtime-notes	Prefix of a candidate mistake/fix; may be incomplete	see record	SRC-1, extensions/self-learning.ts	Truncation can remove scope or negation; not a checked proposition.
RT-CMP-reflection-model	Not individually inspectable in parameters	see record	SRC-1, extensions/self-learning.ts	External model state; exact version not pinned.
RT-BAP-memory-policy	None in the static consultation imperatives themselves	see record	SRC-1, extensions/self-learning.ts	They prescribe reliance; they do not authenticate memory.

Authority-route ledger

The ledger splits functions even where the runtime record is a valid container for several phases. All rows are implementation deductions, not observed executions. “No semantic license” means the check supplies no new warrant for truth; it does not imply all stored propositions are false. Evidence anchors and gaps apply locally. Retention here is not lifecycle integration after evidence-consuming acceptance.

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-task-reflection	content transformation	implemented	EPI-OBJ-conversation-signals	truth-apt transformation: acquisition/import	Branch messages into completion input	Code filters aborts/takes tail; regex classifies error text	Automatic gate or manual request	Selected transcript and bounded constraint signals	Input selection	Source warrant unknown; errors are reported signals, not authenticated facts	Supplies selected evidence to proposer	Completion model, user message, informational, current request	SRC-1, extensions/self-learning.ts	EPI-CLM-prevention-scope	none	Host conversion may alter representation; tail excludes older context.
RT-RTE-task-reflection	content transformation	implemented	EPI-OBJ-reflection-items	truth-apt transformation: indeterminate	Transcript to mistake/fix candidates	Selected model under project/global prompt; no independent oracle	Enabled reflection, then optional malformed-output repair	Candidate JSON/text	Proposal only before parser	No verified entailment, factuality or generality	Creates candidates eligible for admission	Parser, completion text, proposed content, this invocation	SRC-1, extensions/self-learning.ts	EPI-CLM-prevention-scope	none	Acquisition, reshaping and conjecture remain possible; outputs needed. Imperative-only outputs would instead be policy updates.
RT-RTE-task-reflection	check/evidence production	implemented	EPI-OBJ-reflection-items	no content change	JSON object and at least one nonempty string list	Deterministic parser, structural/string domain	Before writes; extraction/one model repair if needed	Parsed object or undefined	Enforcing structural test	Licenses usable strings only	Result controls write eligibility	reflectNow, parsed return, enforcing, current invocation	SRC-1, extensions/self-learning.ts	none	none	No truth test or enforced max array count; repair can change content.
RT-RTE-task-reflection	disposition/acceptance	implemented	EPI-OBJ-reflection-items	no content change	Admit parsed candidate or skip	Code consumes parser outcome; model proposes; no per-item human veto	Immediately after parsing/repair	Return invalid-output reason or proceed	Enforcing admission	Not evidence-consuming epistemic acceptance	Permits persistence of any qualifying strings	Filesystem writer, branch in code, permissive, current update	SRC-1, extensions/self-learning.ts	none	none	Structural admission is not acceptance for verified future prevention.
RT-RTE-task-reflection	retention	implemented	EPI-OBJ-reflection-items, RT-OBJ-runtime-notes	truth-apt transformation: non-ampliative reshaping	Parsed strings to dated bullets and short note	Deterministic append/format/prefix selection	After admission	Daily entry and first-item prefix	Durable append; FIFO process note	No additional truth warrant; note completeness degraded by truncation	Makes candidates available to later turns	File bundler and note hook, text, advisory, sessions/process respectively	SRC-1, extensions/self-learning.ts	EPI-CLM-full-history	none	No transcript provenance link retained per item; no transactional rollback.
EPI-RTE-watchout-conversion	content transformation	implemented	EPI-OBJ-index-items	non-truth-apt policy/content update: prefix mistakes with avoidance advice	Mistake string to watch-out	Code unconditionally adds Avoid:	Each admitted mistake	Normalized imperative-like text	Automatic update	A historical error claim alone does not warrant generalized advice	Installs watch-out in candidate index	Index renderer, text, advisory downstream, sessions	SRC-1, extensions/self-learning.ts	EPI-CLM-prevention-scope	none	Concrete counterexample in record; no claim of semantic preservation.
EPI-RTE-index-migration	lineage/freshness/recovery	implemented	EPI-OBJ-index-items	truth-apt transformation: acquisition/import	Missing/invalid index replaced from Markdown bullets	Code parses every bullet and assigns defaults	Index load fallback	New in-memory index, default file creation if missing	Recovery substitution	Imported source warrant unknown; counters do not recover prior evidence	Allows updates/redistill using migrated items	Index consumers, symbolic defaults/text, permissive, later writes	SRC-1, extensions/self-learning.ts	EPI-CLM-full-history	none	Bullet hierarchy and original chronology not recovered.
EPI-RTE-core-ranking	retention	implemented	EPI-OBJ-index-items	truth-apt transformation: non-ampliative reshaping	Text-key updates, rendered core and long-term files	Normalized lowercase keys; fixed hit/score arithmetic	Reflection update or final redistill write	Merged index and overwritten views	Automatic file rewrite	Frequency is not corroboration	Retains merged candidates, not all textual versions	Future load/bundle, index/files, permissive, sessions	SRC-1, extensions/self-learning.ts	EPI-CLM-full-history	none	Case-insensitive collision may discard meaningful distinctions.
EPI-RTE-core-ranking	operational admission/selection/consumption	implemented	EPI-OBJ-index-items	no content change	Which candidates occupy core	Score minus age penalty; hit/time tie-breaks; kind balance	Core rendering	Selected bounded view	Ranking	No factual or utility license from recurrence/recency	Promotes selected text into default bundled file	Foreground agent via core bundle, text, ranking then advisory, next turns	SRC-1, extensions/self-learning.ts	EPI-CLM-prevention-scope	none	Eviction from core is not rejection from index or retraction.
RT-RTE-month-summary	content transformation	implemented	EPI-OBJ-monthly-text	truth-apt transformation: indeterminate	Daily text to monthly summary	Selected model, bounded month input, unchanged summary prompt	Manual month request	Proposed summary	Proposal	Faithfulness unknown	Produces summary eligible for overwrite	Monthly writer, text, proposal, invocation	SRC-1, extensions/self-learning.ts	none	none	Reshaping versus novel assertion unresolved without input/output.
RT-RTE-month-summary	check/evidence production	implemented	EPI-OBJ-monthly-text	no content change	Nonempty text	Code trims concatenated text blocks	Before writer	Empty/nonempty	Enforcing minimal test	Only text presence	Determines continuation	Summary command, Boolean branch, enforcing, invocation	SRC-1, extensions/self-learning.ts	none	none	No section, source or semantic comparison.
RT-RTE-month-summary	disposition/acceptance	implemented	EPI-OBJ-monthly-text	no content change	Nonempty candidate admitted	Code; human initiates month, not content approval	After minimal test	Skip or admit	Enforcing admission	No epistemic acceptance established	Permits monthly overwrite	Writer, control flow, permissive, invocation	SRC-1, extensions/self-learning.ts	none	none	Cannot reject plausible false summary.
RT-RTE-month-summary	retention	implemented	EPI-OBJ-monthly-text	truth-apt transformation: non-ampliative reshaping	Admitted text into dated monthly wrapper	Code filesystem write	After admission	Same-month file replaced	Automatic overwrite	No additional warrant	Enables optional later bundle	File reader, Markdown, advisory, sessions	SRC-1, extensions/self-learning.ts	none	none	Later daily edits do not invalidate old summary.
RT-RTE-core-redistill	content transformation	implemented	EPI-OBJ-index-items	truth-apt transformation: indeterminate	Selected rules into generic rules	Model prompted to preserve meaning/remove specifics; code enforces kind prefix	Manual global request; optional coarse confirmation; per chunk	Rewritten or original text for missing IDs	Proposal and deterministic prefix editing	No demonstrated semantic preservation or scope transfer	Generates replacement candidates	Chunk parser/update code, JSON, proposed content, invocation	SRC-1, extensions/self-learning.ts	EPI-CLM-prevention-scope	none	Removing identifiers may change truth conditions; output pairs required.
RT-RTE-core-redistill	check/evidence production	implemented	EPI-OBJ-index-items	no content change	Parseable items with finite IDs and nonempty text	Code parser, not semantic evaluator; optional model repair	Each chunk before staging	Nonempty map or failed chunk	Enforcing syntax test	Structural usability only	Allows chunk processing	Redistill loop, map/undefined, enforcing, invocation	SRC-1, extensions/self-learning.ts	none	none	Does not enforce complete requested ID set.
RT-RTE-core-redistill	disposition/acceptance	implemented	EPI-OBJ-index-items	no content change	Stage chunk or abort; commit entire staged transformation unless dry-run	Code consumes parse result; human can cancel coarse start, not approve each rewrite	All chunks must finish before final write block	Abort, preview or admission	Enforcing staging gate	No evidence-consuming acceptance for cross-project use	Permits index/core/long-term replacement	Final write block, staged values, permissive, sessions	SRC-1, extensions/self-learning.ts	EPI-CLM-prevention-scope	none	Partial final writes have no automatic rollback; collisions merge histories.
RT-RTE-context-injection	operational admission/selection/consumption	implemented	EPI-OBJ-index-items, EPI-OBJ-monthly-text, RT-OBJ-runtime-notes	no content change	Selected retained text to foreground model	Code flags/counts/lexical dates/prefix caps; model later interprets	Enabled before-start hook	Hidden message and consultation policy returned	Advisory historical-evidence framing, imperative policy	Requested reliance, not authenticated evidence	Can shape later answers/consultation; no read/answer gate	see RT-BAP-memory-policy	SRC-1, extensions/self-learning.ts	RT-CLM-strict-consultation	Enforcement wording exceeds gate	Delivery unobserved; long-term supplied by path only; omitted qualifiers can change interpretation.
RT-RTE-runtime-controls	behavior/policy adaptation	implemented	None: settings and branch controls	non-truth-apt policy/content update: enabled/model configuration	Which reflection/context routes run and model is selected	Human command; code validates values and persists references	Explicit command, not evaluation feedback	Toggle/global model change; branch model stored but ignored by pickers	Deterministic settings gate	No truth judgment	Changes route eligibility/model resolution; maintenance bypasses enabled toggle	Hook/picker, symbolic settings, enforcing, later invocations	SRC-1, extensions/self-learning.ts	none	none	No autonomous policy revision from measured outcomes.

No lifecycle-integration route after epistemic acceptance was found in the complete inspected extension. The positive functions above include operational use and organization of candidates, not evidence-consuming acceptance under a named truth/prevention criterion. Git snapshots and display statistics add recovery/visibility, not evaluation. The supplied runtime retains their implementation passages; the new split records below retain additional load-bearing passages.

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-prevention-scope	see record	SRC-1, README.md, extensions/self-learning.ts; doctrine/design	Mistake/fix focus and reusable global rules	RT-RTE-task-reflection, RT-RTE-core-redistill, EPI-RTE-watchout-conversion	None supplied	No intervention/comparison, evaluator of future success, or transfer evidence	Implemented improvement-directed proposal and memory-use pathway	“Reusable” is requested, not warranted; prefix conversion can change function without semantic judgment.
EPI-CLM-full-history	see record	SRC-1, README.md, extensions/self-learning.ts; doctrine/design	Complete-history wording	EPI-RTE-core-ranking, EPI-RTE-index-migration, RT-RTE-core-redistill	None supplied	File history/commits unobserved; no automatic restore	Long-term renders all current merged index items	Not every reflection/version/evidence lineage; daily logs and optional Git are separate.
RT-CLM-strict-consultation	see record	SRC-1, README.md; doctrine/design	Imperative policy explicitly requests consultation	RT-RTE-context-injection	None supplied	Host/model compliance not measured	Consultation request is wired	No enforcement gate; does not establish truth of consulted material.

Bounded conclusion

The package acquires a selected account of coding activity, not authenticated task outcomes. It can construct and retain mistake/fix candidates and ask a later model to treat rendered memory as historical evidence. The structural evaluator's domain is JSON/string usability; the monthly evaluator checks text presence; redistill checks usable ID/text mappings. Their operative results control rejection, repair, continuation and writes. None supplies a truth, prevention-success or transfer criterion for epistemic acceptance. Starting maintenance or confirming a large batch is not per-content human acceptance (RT-RTE-task-reflection, RT-RTE-month-summary, RT-RTE-core-redistill).

Source warrant is unknown at transcript import. Model transformation, truncation and loss of item-to-transcript lineage do not preserve a demonstrated warrant. No inspected premise set plus checked inference establishes a derived substantive claim. Individual model outputs could be imports, compressions, conjectures or policy-only instructions; without outputs their relation remains indeterminate, not automatically ampliative. Deterministic mistake-to-watch-out conversion does establish a policy transformation, including for truth-apt historical inputs (EPI-RTE-watchout-conversion).

Frequency and recency select visibility, not corroboration. Text-key merging preserves counters but can erase distinctions and version lineage. Long-term memory represents the current merged index, not the full evidence history. No source-supported semantic rejection/retraction route removes a false belief from every downstream copy: overwriting a core item need not change daily/monthly text or process notes (EPI-RTE-core-ranking, EPI-RTE-index-migration, EPI-CLM-full-history). This is a bounded implementation finding over the inspected file, not a claim that users cannot manually edit memory.

Operational authority is real but bounded: update code admits retained candidates; rankers determine prominence; returned hidden text and system policy can influence the foreground model on later turns/sessions. No semantic acceptance is needed for these paths. Actual consultation, changed decisions and benefit remain unobserved (RT-BAP-memory-policy). No model-weight update is implemented by this package; inaccessible provider behavior stays unresolved (RT-CMP-reflection-model).

Learning, reflection, theory construction and self-improvement remain separate. The theory account in RT-RTE-task-reflection supports localized-rule affordance and wired trace-to-memory-to-context organization, not observed content-directed criticism, iterative theory revision or improved capacity. For extension-owned memory/context organization across sessions, avoiding repeated mistakes is an independently stated objective. Evidence-responsive update and a later behavioral-authority path are wired; causal dependence of subsequent operation on that change (self-improvement link 4) and successful improvement are unestablished. This supports a dispositional improvement-directed pathway, not exercised or successful self-improvement. Manual control changes are policy updates, not learning from an evaluator's result.

Shared records

Operative objects

EPI-OBJ-conversation-signals

Label: Selected conversation and execution signals

Part of: RT-RTE-task-reflection

This object is a material input part of the runtime route, not a duplicate route. Source-native identity: conversationText and executionSignals. Form: natural-language serialized assertions plus symbolically selected error classifications; substrate: transient strings. The host supplies branch messages; code excludes recognized cancellations, takes a tail, and classifies error-tool text. Original message truth and conversion fidelity remain unknown. SRC-1, extensions/self-learning.ts, implementation.

const conversationText = serializeConversation(convertToLlm(messages as any)); --- extensions/self-learning.ts

if (!isError) continue;

if (PERMISSION_DENIED_PATTERN.test(text)) { push(Tool ${toolName} failed with a permission denial: ${textPreview}); continue; } --- extensions/self-learning.ts

EPI-OBJ-reflection-items

Label: Parsed reflection and daily text

Part of: RT-RTE-task-reflection

Source-native identity: LearningReflection, parsed, daily Markdown. Form: natural-language string items in symbolic arrays/sections; substrate: completion response, process object and daily files. This is the admitted candidate content part of the runtime route; RT-OBJ-runtime-notes is a distinct shortened derivative. Claimed role is identifying what went wrong/how fixed. Neither array requires a truth-apt sentence or evidence link. SRC-1, extensions/self-learning.ts, implementation; parser passages also retained in the parent record.

lines.push("### What went wrong"); if (reflection.mistakes.length === 0) lines.push("- (none)"); for (const item of reflection.mistakes) lines.push(- ${item}); --- extensions/self-learning.ts

EPI-OBJ-index-items

Label: Core index items and rendered views

Source-native identity: CoreIndexRecord, core/index.json, core/CORE.md, long-term-memory.md. Form: natural-language rules and assertions, symbolic keys/kinds/counters/timestamps; substrate: JSON/Markdown disk files. Material output/input across RT-RTE-task-reflection and RT-RTE-core-redistill, therefore not part of exactly one supplied record. No supplied operative-object counterpart; runtime notes are separate process prefixes. Claimed role is durable/ranked learning history. Text is not linked per item to supporting events. SRC-1, extensions/self-learning.ts, implementation.

type CoreIndexRecord = { key: string; text: string; kind: CoreKind; hits: number; score: number; firstSeen: string; lastSeen: string; }; --- extensions/self-learning.ts

EPI-OBJ-monthly-text

Label: Monthly summary text

Part of: RT-RTE-month-summary

Source-native identity: summary, monthly/YYYY-MM.md. Natural-language content in Markdown on disk; generated from selected daily files and optional later context. Distinct from daily candidates and index views in source and admission criterion. No supplied object counterpart. Model truth/faithfulness not inspected. SRC-1, extensions/self-learning.ts, implementation and embedded doctrine/design.

"Return markdown with these sections:", "- Wins", "- Recurring issues", "- Most important learnings", --- extensions/self-learning.ts

Routes

EPI-RTE-watchout-conversion

Label: Mistake to watch-out conversion

Part of: RT-RTE-task-reflection

Endpoints: parsed mistakes → updateCoreFromReflection → normalized anti-pattern entries → index/rendering. Owner is deterministic extension code, not another model judgment. This splits a material policy conversion from the parent's broader proposal/admission/write route. The model proposes mistakes; code decides conversion without human approval or expected-answer oracle. Mode is open session reflection; admitted mistakes trigger it under unchanged prevention guidance. Recovery is external editing/conditional Git, not automatic semantic rollback.

• Immediate return: Updated file paths from parent update call.
• Later read-back: Watch-outs enter core/long-term and later foreground context or consultation.
• Delegated visibility: No separate delegation; renderers receive resulting text/kind.
• Selection predicate: Every admitted mistake string; whitespace normalized.
• Invalidation or expiry: Exact normalized-key merging/ranking, not semantic falsification or expiry.
• Activation or effect: Conversion and downstream writes wired; model interpretation unobserved.
• Evidence limits: Implementation deduction, no target execution or observed content.
• implementation conclusion status: wired
• operation conclusion status: uninspected

SRC-1, extensions/self-learning.ts, implementation:

...reflection.mistakes.map((text) => ({ text: Avoid: ${text}, kind: "antiPattern" as const })), --- extensions/self-learning.ts

Concrete symbolic trace: admissible input {"mistakes":["The command failed yesterday."],"fixes":[]} survives parseReflection; conversion produces Avoid: The command failed yesterday. with kind antiPattern. Normalization preserves those words, and learningKey lowercases the key. The historical proposition does not entail an avoidance policy, and adding a prefix does not turn it into a sound imperative. Rendering strips that prefix under Watch-outs. This deduction disproves a universal meaning-preserving-reshaping account of mistake conversion without claiming this input occurred. It establishes no new checked truth-apt conclusion. Potential localized theories depend on actual prose; the parent theory account's independent limits remain applicable.

EPI-RTE-core-ranking

Label: Core index merging and ranking

Source-native endpoints: admitted reflection/rewrite → normalized-key merge → score/hit/time updates → ranked balanced core and all-index long-term render. This spans RT-RTE-task-reflection and RT-RTE-core-redistill; it is a shared deterministic maintenance stage, not a duplicate of either full route. Code owns decisions. No human judges particular entries; no answer oracle compares scores to truth or outcome. Trigger is text admission/rewrite, not measured improvement. It changes visibility and retained organization under unchanged selection rules.

• Immediate return: Rendered file paths and updated index; redistill also exposes counts.
• Later read-back: Selected core is bundled by default; all current merged items appear in long-term file referenced by policy.
• Delegated visibility: Renderers see item text/kind and metadata, not original supporting sessions.
• Selection predicate: Normalized lowercase text keys; score/age ordering, hits/time tie-breaks, kind balancing and size limit.
• Invalidation or expiry: Core exclusion does not delete index entries; rewrites replace text/keys and merge collisions. No semantic invalidation.
• Activation or effect: File rewrite/ranking wired; foreground reliance unobserved.
• Evidence limits: Scores show repetition/recency only; no correctness or benefit measurement. Writes are not transactional; optional Git permits only conditional manual recovery.
• implementation conclusion status: wired
• operation conclusion status: uninspected

SRC-1, extensions/self-learning.ts, implementation:

function learningKey(text: string): string { return normalizeLearningText(text).toLowerCase(); } --- extensions/self-learning.ts

const recencyPenalty = ageDays * 0.05; return item.score - recencyPenalty; --- extensions/self-learning.ts

const repetitionBonus = Math.min(1, existing.hits * 0.08); existing.score += incrementBase + repetitionBonus; --- extensions/self-learning.ts

existing.hits += item.hits; existing.score += item.score; --- extensions/self-learning.ts

A counterexample to corroboration: repeating the same unsupported output raises hits and score by this code; no independent observation need occur. Case-normalized equality also cannot prove equivalence of case-sensitive code identifiers. This is symbolic ranking/merge, not claim-directed criticism; theories that may occur in retained prose are not independently evaluated here.

EPI-RTE-index-migration

Label: Core index migration fallback

Source-native endpoints: loadCoreIndex → missing/invalid-index fallback → ensureCoreFile/parseCoreItems → default index records. Extension code owns recovery; local Markdown supplies content without human approval at load time. Overlaps RT-RTE-task-reflection and RT-RTE-core-redistill, not contained in exactly one. No supplied standalone counterpart. This imports prior text rather than producing verified claims. There is no answer oracle or semantic evaluation; the trigger is structural load failure/missing index, operating mode deterministic recovery under unchanged defaults.

• Immediate return: In-memory version-1 index; can create missing default core file.
• Later read-back: Subsequent update/redistill can save migrated items and rendered views for future contexts.
• Delegated visibility: Index consumers receive bullets plus invented default counters/current timestamps, not lost index history.
• Selection predicate: Valid index items win; fallback takes every trimmed - bullet except (none yet).
• Invalidation or expiry: A later valid index takes priority; no source-age check or semantic invalidation.
• Activation or effect: Recovery/migration wired; actual migrations unobserved.
• Evidence limits: Earlier source/provenance unavailable; defaults do not prove historical frequency, timing or endorsement. No automatic restoration of corrupted index.
• implementation conclusion status: wired
• operation conclusion status: uninspected

SRC-1, extensions/self-learning.ts, implementation:

const corePath = ensureCoreFile(root); const migrated = parseCoreItems(readFileSync(corePath, "utf-8")).map((text) => ({ key: learningKey(text), text, kind: text.toLowerCase().startsWith("avoid:") ? ("antiPattern" as const) : ("learning" as const), hits: 1, score: 1, firstSeen: new Date().toISOString(), lastSeen: new Date().toISOString(), })); --- extensions/self-learning.ts

Countercase: rendered Watch-outs omit Avoid:; fallback migration classifies them as learning unless the remaining text itself begins with that prefix. Therefore migration does not guarantee preservation of kind or complete index history. It supplies structural recovery, not epistemic acceptance or a theory-criticism cycle.

Claims

EPI-CLM-prevention-scope

Label: Mistake prevention and cross-project reuse

Claimed operation: extract mistakes/fixes and, in global mode, generate reusable rules. No identical supplied claim record; RT-RTE-task-reflection retains the prompt's prevention focus and RT-RTE-core-redistill retains preservation intent. This claim concerns intended prevention/reuse, distinct from RT-CLM-strict-consultation. SRC-1, README.md and embedded prompt in extensions/self-learning.ts, doctrine/design, not measured benefit.

• storage.mode: "global" distills reflections into cross-project, reusable action rules and avoids project-specific identifiers. --- README.md

"- Distill each item into a cross-project rule that is reusable in any repository.", --- extensions/self-learning.ts

EPI-CLM-full-history

Label: Full history in long-term memory

Claimed operation: long-term file holds full history. No counterpart among supplied claims; related runtime routes are RT-RTE-task-reflection and RT-RTE-core-redistill. SRC-1, README.md, doctrine/design; implementation renders current merged index, not an append-only evidence ledger.

1. writes full history to long-term-memory.md, --- README.md

const sorted = sortedIndexItems(index); const learnings = sorted.filter((x) => x.kind === "learning"); const antiPatterns = sorted.filter((x) => x.kind === "antiPattern");

const content = [ "# Long-term Memory", --- extensions/self-learning.ts