Zikkaron epistemic report

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

Source-and-claim boundary

This report asks which retained or delivered propositions Zikkaron checks and licenses for later reliance, rather than equating persistence with accepted knowledge. Scope, exclusions and source identity are see boundary; evidence identities are see Source register. All new source reads use SRC-1 at the frozen commit, with implementation evidence except explicitly labelled claims. Nothing was executed. There is no supplied observed-run or causal-experiment evidence.

The ordinary host-authored memory, seed and replay routes acquire content; its external warrant is unknown. Directly inspected counterpaths include semantic promotion that actually vetoes inconsistent clusters, recall that rewrites future memories from queries, and inferred reranking passages that add unsupported preferences. These prevent both a blanket “no checks” conclusion and a blanket “validated memory” conclusion. The runtime is a baseline, not the inspection boundary.

Entry point or operation	Source path under SRC-1	Coverage and record	Unassessed part / conclusion prevented
CLI dispatch: service, drain, restore, capture, context, stats, seed	zikkaron/__main__.py	Direct dispatch read; supplied RT-RTE-replay-context, RT-RTE-seed-admission, RT-RTE-trace-consolidation	Stats formatting is not a truth-acceptance gate.
MCP remember, recall, project context, anchor, checkpoint, restore; HTTP compact/capture	zikkaron/server.py, zikkaron/restoration.py	Supplied RT-RTE-remember-recall, RT-RTE-context-selection, RT-RTE-replay-context; direct recall/curation reads, EPI-RTE-recall-rewrite	Full CRDT conflict resolution and sensory-buffer lineage uninspected; no universal provenance guarantee.
MCP forget, validate_memory, rate_memory	zikkaron/server.py, zikkaron/storage.py, zikkaron/staleness.py, zikkaron/thermodynamics.py	Direct reads; EPI-RTE-memory-withdrawal, EPI-RTE-file-validation, EPI-RTE-feedback-update	Whole derived-state withdrawal is not established.
Background/forced consolidation and sleep	zikkaron/consolidation.py, zikkaron/curation.py, zikkaron/cls_store.py, zikkaron/sleep_compute.py, zikkaron/compression.py	Direct selected producers/checks/consumers; EPI-RTE-cls-consistency, EPI-RTE-cls-promotion, EPI-RTE-synthetic-facts, EPI-RTE-memory-compression	Astrocyte-domain transformations, full fractal clustering and sleep cluster summaries remain unassessed; prevents system-complete negative about acceptance or semantic preservation.
Insertion-time enrichment and profile extraction; callable belief derivation	zikkaron/storage.py, zikkaron/enrichment.py, zikkaron/profiles.py	Direct wiring and producer reads; EPI-RTE-index-enrichment, EPI-RTE-profile-beliefs	External ConceptNet responses and model outputs absent; no judgment of actual generated propositions.
Service reranking, implied passages, NLI, adversarial confidence	zikkaron/retrieval.py	Direct reads; EPI-RTE-inferred-ranking, supplied RT-RTE-context-selection	All retrieval signals not independently verified; relevance scores do not establish truth.
Prompt/session hooks and installed integration	zikkaron/hooks/prompt-recall.py, zikkaron/hooks/session-start-context.py, zikkaron/server.py	Supplied RT-RTE-hook-context, RT-RTE-install-integration; direct prompt stale-filter search	Host execution/injection/compliance unobserved.
MCP hierarchical recall/drill-down/navigation and hot/stale/process/stat resources	zikkaron/server.py, zikkaron/retrieval.py	Direct registration/wrapper reads; supplied RT-RTE-context-selection, RT-CMP-cognitive-map	Full hierarchy internals unassessed; wrappers alone establish delivery, not validation.
MCP create_trigger, remember-trigger extraction and checking	zikkaron/prospective.py, zikkaron/server.py	Direct reads; EPI-RTE-trigger-delivery	Full matching grammar unassessed; no truth license from firing.
MCP project story, narrative resource; sleep narrative production	zikkaron/narrative.py, zikkaron/sleep_compute.py, zikkaron/server.py	Direct selected reads; EPI-RTE-narrative-production	Detailed extraction fidelity unassessed.
MCP coverage/gaps and causal-chain; maintenance graph discovery	zikkaron/metacognition.py, zikkaron/causal_discovery.py, zikkaron/knowledge_graph.py, zikkaron/consolidation.py	Direct selected reads; EPI-RTE-coverage-advice, EPI-RTE-causal-outputs	PC event encoding, independence tests and orientation not fully inspected; prevents mathematical-correctness and causal-identification conclusions.
Rule admission, query and application	zikkaron/server.py, zikkaron/rules_engine.py	Supplied RT-RTE-rule-admission	Fixed retrieval policy is not a truth oracle or universal access barrier.
Benchmark ingestion, retrieval, generation, F1 and LLM judgment	benchmarks/run_locomo_jscore.py, benchmarks/run_longmemeval.py, benchmarks/run_locomo_ablation.py, benchmarks/run_benchmark_gpu.py, benchmarks/test_e_locomo.py, benchmarks/od_diagnostic.py	Supplied RT-RTE-evaluation-mode; direct LoCoMo judge/result reads, EPI-RTE-answer-judgment	Dataset/results/provider responses absent; non-LoCoMo evaluator details supplied rather than independently rechecked. No reproduced scores or component attribution.

Knowledge/warrant claims: see RT-CLM-write-gate, RT-CLM-context-replay, EPI-CLM-derived-facts, EPI-CLM-cls-consistency, EPI-CLM-benchmark-benefit. The latency claim RT-CLM-hook-budget does not warrant content and is not independently reassessed here.

Epistemic-object inventory

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
RT-RTE-remember-recall	Caller statements about tasks/projects	see record	SRC-1, zikkaron/server.py	Authorship and external truth warrant unknown; natural-language SQLite content is interpreted by host.
RT-RTE-seed-admission	Project scan summaries	see record	SRC-1, zikkaron/seed.py	External source files absent; templates are not per-claim validation.
RT-RTE-trace-consolidation	Tool counts and truncated summaries	see record	SRC-1, zikkaron/consolidation.py	Counts can describe recorded calls, not prove tool outcomes.
EPI-RTE-synthetic-facts	Modified-together assertions and dream relation suggestions	see record	SRC-1, zikkaron/curation.py, zikkaron/sleep_compute.py	Separate producers and conditions, shared ordinary memory consumer; no observed instances.
EPI-RTE-cls-promotion	Recurring-pattern schema	see record	SRC-1, zikkaron/cls_store.py	Word abstraction can change or destroy propositions.
EPI-RTE-recall-rewrite	Existing text plus current query, or new query-as-memory	see record	SRC-1, zikkaron/reconsolidation.py	A question may be non-truth-apt; assertions in queries remain unverified.
EPI-RTE-profile-beliefs	Profile attributes and derived preferences/traits	see record	SRC-1, zikkaron/profiles.py	SQLite symbolic fields contain natural-language values; evidence IDs do not prove entailment.
EPI-RTE-index-enrichment	Concepts, generated inferences; synthetic queries may be non-truth-apt	see record	SRC-1, zikkaron/enrichment.py	Generated content unavailable; form and truth-aptness vary by output.
EPI-RTE-inferred-ranking	Transient inferred preferences/traits	see record	SRC-1, zikkaron/retrieval.py	Used as ranker input, not necessarily delivered verbatim.
EPI-RTE-causal-outputs	Directed entity relations claiming causes/effects	see record	SRC-1, zikkaron/causal_discovery.py, zikkaron/knowledge_graph.py	Fixed symbolic consumers; causal truth unestablished.
EPI-RTE-coverage-advice	Claims about available coverage plus advice	see record	SRC-1, zikkaron/metacognition.py	Numerical coverage and confidence are proxies, not checked adequacy for every task.
EPI-RTE-narrative-production	Recorded counts, decisions, events and focus areas	see record	SRC-1, zikkaron/narrative.py	Extracted labels and truncations need not preserve qualification.
EPI-RTE-file-validation	File-change status, not memory truth	see record	SRC-1, zikkaron/staleness.py	Missing file reference can return valid.
EPI-RTE-feedback-update	None in numerical confidence/stability update	see record	SRC-1, zikkaron/thermodynamics.py	Usefulness differs from factual accuracy.
RT-OBJ-memory-rules, EPI-RTE-trigger-delivery	Policy/conditions need not be truth-apt; reminder text can contain assertions	see records	SRC-1, zikkaron/rules_engine.py, zikkaron/prospective.py	Symbolic parsing/matching licenses selection, not truth.
RT-OBJ-zikkaron-section, RT-OBJ-hooks-config	Primarily behavioral policy and executable configuration	see records	SRC-1, zikkaron/server.py	Installed policy is not use-derived accepted knowledge.
RT-RTE-replay-context	Caller checkpoint/protected assertions	see record	SRC-1, zikkaron/restoration.py	Protection is operational preference, not corroboration.
EPI-RTE-answer-judgment	Generated answers and correctness labels	see record	SRC-1, benchmarks/run_locomo_jscore.py	Dataset gold supplies reference; provider response and actual correctness unknown.
RT-CMP-embedding-engine, RT-CMP-nli-model, RT-CMP-comet-inferencer	Distributed-parametric internal content not individually identifiable	see records	SRC-1, zikkaron/embeddings.py, zikkaron/retrieval.py, zikkaron/enrichment.py	Truth-apt propositions in weights: not determinable from accessible artifacts.

Other numerical operative parts remain separate: see RT-CMP-gte-reranker, RT-CMP-flashrank-ranker, RT-CMP-cross-encoder, RT-CMP-doc2-query, RT-CMP-hopfield-memory, RT-CMP-hdc-encoder, RT-CMP-cognitive-map, RT-CMP-evaluation-models. Fixity, mutation and endpoint resolution are see each component record. Dense learned weights, rebuilt pattern matrices and use-updated transition state are not interchangeable evidence of learning or truth.

Authority-route ledger

Every row is an architectural finding, not an observed event. Detailed endpoint, control and read-back fields are see the cited declaring record. Repeated records below split functions rather than duplicate identities.

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
EPI-RTE-curation-merge	content transformation	implemented	EPI-RTE-curation-merge	truth-apt transformation: non-ampliative reshaping	Caller and existing text	Cosine threshold plus token Jaccard >0.5; similarity/dedup domain	remember after gate	Concatenated text and tags	Overwrite existing row	Preserves both input texts, not their external warrant	Changes future selected memory	Retriever; SQLite text; advisory to host; later calls	SRC-1, zikkaron/curation.py	none	none	Operation unobserved
EPI-RTE-contradiction-scan	check/evidence production	implemented	EPI-RTE-contradiction-scan	no content change	Potential lexical conflict	Similarity plus negation/action-set comparison	Callable helper; no ordinary production caller found	Conflict list/reason	Internal confidence consumer follows	Lexical heuristic only	Identifies rows for confidence change	Helper; returned/branch-local result; advisory to caller; this invocation	SRC-1, zikkaron/curation.py	none	Not semantic truth evaluation	No observed invocation
EPI-RTE-contradiction-scan	disposition/acceptance	implemented	EPI-RTE-contradiction-scan	non-truth-apt policy/content update: lower confidence	Old flagged row	Negation/action divergence branch	Immediately after flag	Confidence minus 0.2 or 0.1, floor 0.1	SQL mutation	Changed reliance proxy, not truth rejection	Alters later weighting/pruning eligibility	Metadata consumers; SQLite confidence; ranking/selection; later calls	SRC-1, zikkaron/curation.py	none	No resolution of which statement is true	Production caller unestablished
EPI-RTE-cls-consistency	check/evidence production	implemented	EPI-RTE-cls-promotion	no content change	Cluster member consistency	Negation-presence pair mismatch; lexical domain	Before schema generation	consistent boolean and conflict strings	Result consumed by promotion	Only equal negation-presence established; not semantic consistency	Can prevent schema promotion	CLS; returned dict; enforcing; this cycle	SRC-1, zikkaron/cls_store.py	EPI-CLM-cls-consistency	Lexical check weaker than semantic claim	Operation unobserved
EPI-RTE-cls-promotion	disposition/acceptance	implemented	EPI-RTE-cls-promotion	no content change	Cluster eligibility for semantic memory	Consistency boolean, schema nonempty, semantic similarity duplicate check	After pattern selection	Skip or insert schema	Enforcing operational admission	No established truth acceptance criterion for intended downstream use	Vetoes inconsistent or duplicate patterns	CLS insertion; code branch; enforcing; cycle	SRC-1, zikkaron/cls_store.py	EPI-CLM-cls-consistency	none	Operation unobserved
EPI-RTE-cls-promotion	content transformation	implemented	EPI-RTE-cls-promotion	truth-apt transformation: indeterminate	Selected episodic text to schema	Common-word threshold, stop-word removal; no semantic checker	Admitted clusters	Recurring-pattern text	Candidate generation	Semantic preservation/entailment not established	New text available to ordinary selection	Retriever; semantic row; advisory; later calls	SRC-1, zikkaron/cls_store.py	none	none	Need concrete cluster and proposition interpretation to classify each output
EPI-RTE-cls-promotion	retention	implemented	EPI-RTE-cls-promotion	no content change	Schema and derived_from links	Direct insertion and linking	Same cycle	Semantic memory plus retained episodes	Durable storage	Lineage exists without epistemic acceptance	Later dual-store recall may select schema	DualStoreCLS/retriever; rows/links; selection; later queries	SRC-1, zikkaron/cls_store.py	none	none	Operation unobserved
EPI-RTE-synthetic-facts	content transformation	implemented	EPI-RTE-synthetic-facts	truth-apt transformation: ampliative conjecture	Weighted co-occurrence to modified-together assertion	Weight >=10 and no exact duplicate	Memify cycle	Synthetic derived-fact text	Direct durable admission	Co-occurrence/heat weights do not entail modification frequency	Creates normal retrievable memory	Retriever; ordinary memory row; advisory; future calls	SRC-1, zikkaron/curation.py	EPI-CLM-derived-facts	Proxy is promoted into stronger assertion	Operation unobserved
EPI-RTE-synthetic-facts	content transformation	implemented	EPI-RTE-synthetic-facts	truth-apt transformation: indeterminate	Random unconnected pair to dream relation	Embedding similarity >0.7	Sleep dream replay	May-relate text plus graph link	Direct durable admission	Association candidate only; “may relate” has underspecified domain	Adds ordinary text and weak graph edge	Retriever/graph traversal; rows; selection/advisory; future calls	SRC-1, zikkaron/sleep_compute.py	none	none	Exact truth conditions of relation suggestion not specified
EPI-RTE-inferred-ranking	content transformation	implemented	EPI-RTE-inferred-ranking	truth-apt transformation: ampliative conjecture	Mention to preference assertion	Keyword rules, including camping; ignores negation	Open-domain FlashRank branch	Inferred passages	Rank input only	Counterexample demonstrates non-entailment	May improve maximum rank score of source memory	FlashRank; transient passage; ranking; current recall	SRC-1, zikkaron/retrieval.py	none	Inferred proposition stronger than source	Operation unobserved
EPI-RTE-inferred-ranking	check/evidence production	implemented	EPI-RTE-inferred-ranking	no content change	Query-memory relevance / textual entailment proxy	Cross-encoders and NLI truncated premise/hypothesis	Configured model-dependent reranking	Rank scores	Blended ordering, not factual veto	NLI score is not checked proof or source truth	Selects memory delivery; uncertainty flag does not itself abstain	Retriever; numerical scores; ranking; current call	SRC-1, zikkaron/retrieval.py	none	none	Model/class mapping and calibration unverified; dependency may fail
EPI-RTE-file-validation	check/evidence production	implemented	EPI-RTE-file-validation	no content change	Referenced file equality to saved hash	Hash comparison or missing-reference success	validate_memory or file watcher	valid/invalid status	Status return and mismatch branch	At most file freshness; no-reference valid is not truth validation	Supplies mismatch for metadata consumer	Caller/staleness logic; return/comparison; advisory/branching; current call	SRC-1, zikkaron/staleness.py, zikkaron/server.py	none	is_valid exceeds check domain if read as factual validity	Operation unobserved
EPI-RTE-file-validation	disposition/acceptance	implemented	EPI-RTE-file-validation	non-truth-apt policy/content update: stale flag and lower heat	File-mismatch-associated memories	Mismatch branch, plus watcher parent-directory selection	After file comparison	Stale flag and halved heat	Persisted metadata mutation	Withdraws freshness reliance, not a factual verdict	Can lower future selection and exclude in prompt hook	Prompt hook/retriever; metadata; filtering/ranking; future calls	SRC-1, zikkaron/staleness.py, zikkaron/hooks/prompt-recall.py	none	Not all read paths exclude stale rows	Operation unobserved
EPI-RTE-feedback-update	behavior/policy adaptation	implemented	EPI-RTE-feedback-update	non-truth-apt policy/content update: confidence and stability	Memory usefulness history	Caller boolean; useful_count/access_count after >3 accesses	rate_memory	Numerical confidence/stability	Persisted ranking/rewrite controls	Usefulness is not independent factual corroboration	Changes priority and susceptibility to future rewrite	Server sort/context packer/reconsolidation; DB scalars; ranking/selection; later calls	SRC-1, zikkaron/thermodynamics.py, zikkaron/server.py, zikkaron/reconsolidation.py	none	none	Operation unobserved
EPI-RTE-recall-rewrite	disposition/acceptance	implemented	EPI-RTE-recall-rewrite	no content change	Mismatch between query and retrieved memory	Embedding/directory/time/tag distance, stability/plasticity thresholds; protected veto	After current recall scoring	none/update/archive	Code decides rewrite; no fact acceptance	Thresholds license state change only	Rewrites future state or creates query memory	Reconsolidation; branch; enforcing; future calls	SRC-1, zikkaron/server.py, zikkaron/reconsolidation.py	none	none	Operation unobserved
EPI-RTE-recall-rewrite	content transformation	implemented	EPI-RTE-recall-rewrite	truth-apt transformation: indeterminate	Old memory and current query	Append with update delimiter; truncate old text if long; archive branch inserts query	After selection	Query-augmented/new text	Durable mutation	Imported assertion warrant unknown; question need not be truth-apt	Later retrieval can expose query-derived text	Retriever; mutated/new row; advisory; future calls	SRC-1, zikkaron/reconsolidation.py	none	none	Depends on query content; archive branch does not delete original
EPI-RTE-index-enrichment	check/evidence production	implemented	EPI-RTE-index-enrichment	no content change	Candidate enrichment terms	FPA cosine similarity to original embedding	Enabled expander output, before assembly	Kept/rejected similarity result	Threshold comparison	Association threshold only	Supplies subset for assembly	Enrichment pipeline; kept list; enforcing subset; this insertion	SRC-1, zikkaron/enrichment.py	none	Does not validate claims	Missing embedding can bypass FPA
EPI-RTE-index-enrichment	disposition/acceptance	implemented	EPI-RTE-index-enrichment	no content change	Supplement retention eligibility	Enabled branches and FPA-kept terms; logic bypasses FPA	Before enriched-content assembly	Omitted or appended terms	Code admission	No evidence-consuming factual acceptance established	Selects supplements retained/encoded	Enrichment/storage; list assembly; enforcing; insertion	SRC-1, zikkaron/enrichment.py, zikkaron/storage.py	none	Logic structural label not semantic verification	Generated outputs absent
EPI-RTE-index-enrichment	content transformation	implemented	EPI-RTE-index-enrichment	truth-apt transformation: indeterminate	Content to index supplements	ConceptNet/COMET/Doc2Query/logic with optional similarity filtering	Caller supplies settings/embeddings	Enriched fields and dense encoding	Index/ranking influence	Association and model generation are not checked entailment; imported warrant unknown	Can alter later vector selection	Embedding/index consumers; enriched fields/vectors; ranking; future calls	SRC-1, zikkaron/enrichment.py, zikkaron/storage.py	none	none	Generated outputs absent; different supplement types remain distinct
EPI-RTE-profile-beliefs	content transformation	implemented	EPI-RTE-profile-beliefs	truth-apt transformation: ampliative conjecture	Activity mentions to enjoys assertions	Substring categories and first name; fixed confidence 0.6	Callable belief derivation	Belief row with evidence IDs	Direct persistence when called	Activity mention does not entail liking	Stored belief can enter recall after reranking	Retriever FTS belief search; row; ranking/advisory; future calls	SRC-1, zikkaron/profiles.py, zikkaron/retrieval.py	none	none	Production caller not found in scoped search; not evidence that this producer runs in ordinary insertion
EPI-RTE-causal-outputs	content transformation	implemented	EPI-RTE-causal-outputs	truth-apt transformation: ampliative conjecture	Co-occurrence plus order to caused_by relation	Threshold and temporal-order heuristic	Consolidation graph maintenance	Causal-labelled graph edge	Persisted graph organization	Temporal association does not entail causation	Changes graph available for traversal	Graph consumers; relation row; selection; later calls	SRC-1, zikkaron/knowledge_graph.py	none	none	Temporal helper and full downstream graph usage not independently verified
EPI-RTE-causal-outputs	check/evidence production	implemented	EPI-RTE-causal-outputs	truth-apt transformation: indeterminate	Event data to PC directed edges	PC procedure after minimum 5 variables/10 windows	Periodic discovery	DAG metadata and persisted directed edges	MCP causal-chain answers	Conditional statistical output only; causal assumptions unvalidated	Provides causes/effects to caller	get_causal_chain; DB edge query; advisory; later requests	SRC-1, zikkaron/causal_discovery.py, zikkaron/server.py	none	none	CI implementation/encoding/orientation partly uninspected; no causal experiment
EPI-RTE-coverage-advice	check/evidence production	implemented	EPI-RTE-coverage-advice	no content change	Stored-topic coverage and gaps	Density/entity/recency/confidence blend; thresholds 0.7/0.4	Explicit coverage/gap request	sufficient/partial/insufficient advice	Advisory result	Description of proxy coverage, not verified answer adequacy	Asks caller to proceed or investigate; no enforcing gate	Host; MCP result; advisory; current task	SRC-1, zikkaron/metacognition.py, zikkaron/server.py	none	none	Operation unobserved
EPI-RTE-memory-compression	content transformation	implemented	EPI-RTE-memory-compression	truth-apt transformation: indeterminate	Memory text to gist/tag	Sentence scores, first/last selection, code preservation, tag truncation	Maintenance compression	Archived original and replacement content	Future ordinary reads receive compressed text	Omitted qualifications can change meaning; no fidelity evaluator found in read methods	Changes text available to retrieval	Retriever; replacement row; advisory; future calls	SRC-1, zikkaron/compression.py	none	none	Need specific input and clause dependencies for semantic classification
EPI-RTE-narrative-production	content transformation	implemented	EPI-RTE-narrative-production	truth-apt transformation: indeterminate	Project memory window to narrative	Counts plus truncated extracted decisions/events/entities/hot topics	Sleep auto_narrate	Persisted summary	Story tool/resource delivery	Recorded-count premise bounded to DB, interpretation of decisions/focus unchecked	Later story text supplies selected history	Host; narrative tool/resource; advisory; later calls	SRC-1, zikkaron/narrative.py, zikkaron/sleep_compute.py	none	none	Operation unobserved
EPI-RTE-trigger-delivery	operational admission/selection/consumption	implemented	EPI-RTE-trigger-delivery	non-truth-apt policy/content update: reminders and matching conditions	Reminder type and matching context	Type enumeration; fixed matcher and firing-count deactivation	remember and explicit create_trigger	Persisted reminder and triggered output	Context-conditioned delivery	No factual license from trigger match	May surface previously retained reminder	Host; remember response; advisory; triggering call	SRC-1, zikkaron/prospective.py, zikkaron/server.py	none	none	Detailed match grammar not fully assessed
EPI-RTE-memory-withdrawal	lineage/freshness/recovery	implemented	EPI-RTE-memory-withdrawal	no content change	Memory and named dependents	Caller ID; direct SQL deletion	forget or pruning	Memory/archive/transition/vector deletion	Enforcing store removal	Withdrawal of source row, not retraction of every derived proposition	Stops future ordinary lookup of deleted row	Store readers; SQL deletion; enforcing; subsequent calls	SRC-1, zikkaron/storage.py, zikkaron/server.py	none	none	Profile/belief evidence IDs are not traversed by inspected delete_memory
EPI-RTE-answer-judgment	check/evidence production	implemented	EPI-RTE-answer-judgment	no content change	Generated answer versus dataset gold	F1 lexical comparison; LLM judge accepts essential meaning and partial answers	After generation in benchmark	F1 and CORRECT/WRONG or judge error	Appends sample; aggregates and optional JSON persistence	Dataset-relative answer evaluation only	Changes reported scores, not production memory acceptance	Benchmark report; sample/aggregate; measurement; evaluation run	SRC-1, benchmarks/run_locomo_jscore.py	EPI-CLM-benchmark-benefit	none	No observed result or independently frozen oracle; judge errors counted false
RT-RTE-remember-recall	operational admission/selection/consumption	implemented	RT-RTE-remember-recall	truth-apt transformation: acquisition/import	Caller content	Predictive novelty policy and bypasses; curator/direct insert	Explicit remember	Reject or persist	Early veto and insertion	Imported source warrant unknown; novelty is not truth	Controls which caller text persists	Retriever/host; DB and MCP; selection/advisory; later calls	SRC-1, zikkaron/server.py, zikkaron/predictive_coding.py	RT-CLM-write-gate	Gate is not storage-wide	Operation unobserved
RT-RTE-rule-admission	operational admission/selection/consumption	implemented	RT-OBJ-memory-rules	non-truth-apt policy/content update: retrieval rules	Symbolic policy	Parser/type validation then scoped rules application	Rule addition and later service recall	Error or filter/boost/penalty	Enforcing on covered retrieval path	Grammar validity only	Filters/reranks service results, not hooks/replay	RulesEngine; SQLite rules; enforcing/ranking; until changed	SRC-1, zikkaron/rules_engine.py, zikkaron/retrieval.py	none	none	See supplied route for bypass boundaries

Retention is not lifecycle integration after acceptance. The CLS links constitute implemented organization, but no truth acceptance of the abstracted proposition has been established; the retention row deliberately does not call it lifecycle integration. No system-complete absence follows because the named unassessed transformations remain in scope.

System-claim versus route comparison

Claim ID	Claimed operation or warrant	Claim source ID/anchor and evidence layer	Doctrine/design support	Implemented route IDs	Observed-run support	Causal support and design limits	Supported conclusion	Mismatch/unknown
RT-CLM-write-gate	see record	SRC-1, zikkaron/predictive_coding.py; doctrine/design	see record	RT-RTE-remember-recall, RT-RTE-seed-admission	None supplied	None supplied	Gate selects novelty on explicit path	Novelty neither truth nor invariant across writers.
RT-CLM-context-replay	see record	SRC-1, zikkaron/restoration.py; doctrine/design	see record	RT-RTE-replay-context, RT-RTE-hook-context	None supplied	None supplied	Selected retained context can be reconstructed	Preferential bounded selection, not full conversation or factual validation.
EPI-CLM-derived-facts	see record	SRC-1, zikkaron/curation.py; doctrine/design	see record	EPI-RTE-synthetic-facts	None supplied	Heat can boost weights without modifications; no controlled outcome comparison	Automatic conjecture production and retained-state adaptation	“Derived-fact” is stronger than demonstrated warrant; improved capacity unestablished.
EPI-CLM-cls-consistency	see record	SRC-1, zikkaron/cls_store.py; doctrine/design	see record	EPI-RTE-cls-consistency, EPI-RTE-cls-promotion	None supplied	No semantic-consistency validation	Real lexical veto before promotion	Equal negation flags do not imply consistent propositions.
EPI-CLM-benchmark-benefit	see record	SRC-1, README.md; reported operation	see record	RT-RTE-evaluation-mode, EPI-RTE-answer-judgment	No inspectable results supplied	Different reader models explicitly confound QA comparison; no independent component intervention supplied	Shipped evaluator and attributed reported scores	Neither reproduced benefit nor specific causal contribution established.

Bounded conclusion

Zikkaron implements durable acquisition, selection and reshaping of caller text, trace summaries, project scans and checkpoints. Those routes grant operational availability, not automatic epistemic authority: see RT-RTE-remember-recall, RT-RTE-trace-consolidation, RT-RTE-seed-admission, RT-RTE-replay-context. Acquired warrant is unknown, and transformations can degrade qualifications. Host delivery is advisory; rule filters enforce only their covered selection path, see RT-BAP-context-delivery, RT-BAP-retrieval-rules.

The inspected semantic-promotion check is consequential: a negation mismatch blocks promotion. Its criterion is lexical consistency of a cluster, intended use is admission of a generated semantic memory, and its scope is that cluster/cycle. It does not verify the schema's proposition against the world. Retaining a schema and connecting derived_from edges are therefore not evidence of accepted ampliative knowledge, see EPI-RTE-cls-consistency, EPI-RTE-cls-promotion. A separate contradiction detector can lower confidence, but production invocation was not found in the recorded search; its existence cannot be attributed to every remember call, see EPI-RTE-contradiction-scan.

The source produces conjectures, not merely storage artifacts: modification-frequency claims, mention-to-preference beliefs, transient inferred preferences, dream suggestions and causal-labelled graph structures. The concrete inferred-ranking counterexample establishes non-entailment on that route. Other model outputs and word abstractions remain locally indeterminate rather than automatically classified as conjecture or derivation. No inspected route demonstrates a new accepted ampliative proposition supported by an evidence-consuming decision with criterion, use and scope; the unassessed domain/sleep/graph internals prevent a system-complete negative. See EPI-RTE-synthetic-facts, EPI-RTE-profile-beliefs, EPI-RTE-inferred-ranking, EPI-RTE-causal-outputs, EPI-RTE-index-enrichment.

File validation supports a narrower check—reference equality/freshness—not truth. Usefulness feedback adapts numerical authority; recall can also rewrite retained text from a query before any later truth check. “Archive” preserves an archive and creates a new memory but does not remove the original in the inspected branch. Forget removes named rows without traversing profile/belief evidence references. These are independent authority and withdrawal limits, see EPI-RTE-file-validation, EPI-RTE-feedback-update, EPI-RTE-recall-rewrite, EPI-RTE-memory-withdrawal.

The benchmark has a genuine reference outcome: dataset gold supplied to F1 and the judge, not merely an unconstrained model opinion. Its operative consumer is the measurement/reporting path. That check licenses dataset-relative scores under a generous rubric, not general memory truth or an automatic deployed update. No observed result or causal comparison was supplied. Improved future capacity, exercised self-improvement, reliable reflection and theory-building iteration are separate unestablished properties, not consequences of the name memify. For memify, the boundary is the memory engine across consolidation cycles and later retrieval; the independent proposed objective is useful future retrieval. State updates and later consumers are wired, but objective-bearing outcome evidence selecting these updates and subsequent operation depending on them are not established, so the four-link exercised self-improvement attribution is not closed. See EPI-CLM-derived-facts, EPI-RTE-feedback-update, EPI-RTE-answer-judgment.

Shared records

Routes

EPI-RTE-curation-merge

Label: Curation merge

Part of: RT-RTE-remember-recall

SRC-1, zikkaron/curation.py, zikkaron/server.py. The curator is the ingestion subroute of the supplied remember container. Caller proposes content; code chooses merge/link/create. Merge concatenates text, unions tags and re-embeds; neither human veto nor expected-answer oracle is wired. Operating mode is open requests. Gate rejection precedes this part. No whole-update rollback is shown.

The separately assessed contradiction helper is see EPI-RTE-contradiction-scan; it is not an automatic step established in this merge route.

• Immediate return: Action/memory ID dictionary.
• Later read-back: Merged text reaches ordinary recall.
• Delegated visibility: Host receives memory result, not a new delegated agent context.
• Selection predicate: Similarity and Jaccard overlap.
• Invalidation or expiry: Ordinary forgetting, curation or decay; no transaction-wide rollback evidenced.
• Activation or effect: Persisted merge mutation wired.
• Evidence limits: No runs or semantic equivalence check supplied.
• implementation conclusion status: wired
• operation conclusion status: uninspected

merged_content = existing["content"] + "\n" + new_content --- zikkaron/curation.py

EPI-RTE-contradiction-scan

Label: Callable contradiction scan

SRC-1, zikkaron/curation.py. Distinct from the ingestion route RT-RTE-remember-recall and its merge part EPI-RTE-curation-merge: this is a separately callable check with its own confidence-mutation consumer. Search for detect_contradictions( across SRC-1 zikkaron/ found the definition and test use, not a production caller. Similar memories (>=0.7) are checked for unequal negation presence or divergent action-token sets. Program returns possible conflicts and immediately lowers old confidence by 0.2 or 0.1. It does not decide which proposition is true. For “Alice uses Python” and “Alice uses Java”, identical action sets and no negation produce no flag if the pair is selected; their intended scopes could nevertheless conflict. Conversely unrelated negation can flag a pair. This is a predicate deduction, not execution. Caller supplies candidate and embedding; code decides flags/mutations without human veto or expected-answer oracle. Mode is callable open input. Recovery is later metadata change/deletion, not a demonstrated rollback.

• Immediate return: Possible-conflict dictionaries with old memory content and reason.
• Later read-back: Mutated confidence is available to later ranking, pruning and coverage consumers if the helper is invoked.
• Delegated visibility: Caller receives conflict list; no independent critic agent.
• Selection predicate: Similarity >=0.7, negation-presence difference or action-set divergence.
• Invalidation or expiry: Later confidence updates/deletion can change effect; no automatic reversal shown.
• Activation or effect: Callable evaluator and internal confidence mutation implemented; ordinary remember invocation not established.
• Evidence limits: No run; absence of a production caller in this scoped search does not prove no external caller.
• implementation conclusion status: afforded
• operation conclusion status: uninspected

if new_has_negation != old_has_negation: contradictions.append({ "memory_id": mem_id, "content": old_content, "similarity": sim, "reason": "negation_mismatch", }) --- zikkaron/curation.py

old_confidence = mem.get("confidence", 1.0) self._storage._conn.execute( "UPDATE memories SET confidence = ? WHERE id = ?", (max(old_confidence - 0.2, 0.1), mem_id), ) --- zikkaron/curation.py

EPI-RTE-cls-consistency

Label: CLS cluster consistency check

Part of: RT-RTE-trace-consolidation

SRC-1, zikkaron/cls_store.py. Distinct check stage inside the runtime maintenance container; it precedes, and is not the same referent as, semantic promotion. Program evaluates each cluster pair for unequal negation presence; returns consistent only when no such pair is found. The promotion consumer skips inconsistent clusters. No human review or answer oracle. A cluster containing “Alice uses Python” and “Alice uses Java” passes this lexical check regardless of whether their project/version scopes actually conflict. Conversely unrelated negated and positive statements can fail. The counterexamples are static deductions about the inspected predicate, not executed tests.

• Immediate return: Consistency boolean and conflict strings.
• Later read-back: inapplicable — checker returns a current-cycle result; subsequent promotion owns persistence.
• Delegated visibility: Promotion code receives result; no delegated model.
• Selection predicate: Negation-presence inequality in any pair.
• Invalidation or expiry: inapplicable — result recomputed for each cluster.
• Activation or effect: Veto wired through continue in promotion.
• Evidence limits: Checks lexical polarity, not contradiction logic or source truth.
• implementation conclusion status: wired
• operation conclusion status: uninspected

if has_neg_a != has_neg_b: contradictions.append( --- zikkaron/cls_store.py

EPI-RTE-cls-promotion

Label: CLS semantic promotion

Part of: RT-RTE-trace-consolidation

SRC-1, zikkaron/cls_store.py, zikkaron/consolidation.py. A material maintenance part distinct from its lexical checker. Idle/forced cycles call CLS. Cluster selection proposes episodes; code can veto inconsistency, empty schema and similar existing semantic memory. It admits a generated schema, sets semantic type, links source episodes and retains them. Guidance is common-word thresholds, stop-word deletion and at most 15 phrase words; no answer oracle or per-schema human approval. This licenses operational admission, not epistemic acceptance. No rollback is shown for partial insertion/linking.

• Immediate return: Pattern/promotion/skipped counters.
• Later read-back: Semantic rows can enter dual-store query; derived_from links retain episode association.
• Delegated visibility: Retriever receives stored schema; host receives selected output.
• Selection predicate: Recurring pattern, lexical consistency, nonempty abstraction and duplicate threshold.
• Invalidation or expiry: Ordinary deletion/maintenance; original episodes are not deleted by promotion.
• Activation or effect: Schema insertion and linking wired.
• Evidence limits: No actual cluster/output; removing negation as a stop word prevents a blanket semantic-preservation claim.
• implementation conclusion status: wired
• operation conclusion status: uninspected

consistency = self.check_consistency(cluster_mems) if not consistency["consistent"]: stats["skipped_inconsistent"] += 1 continue --- zikkaron/cls_store.py

EPI-RTE-synthetic-facts

Label: Memify facts and dream suggestions

Part of: RT-RTE-trace-consolidation

SRC-1, zikkaron/curation.py, zikkaron/sleep_compute.py, zikkaron/consolidation.py. Distinct production part within maintenance, split by producer in the ledger. Memify proposes a modified-together assertion for co_occurrence weight >=10, exact-deduplicates, embeds and inserts it with derived/auto-generated tags. Earlier reweight can add 0.5 for hot endpoints and weight >=5, so crossing the threshold need not mean ten observed modifications. Co-occurrence does not entail modification frequency. Dream replay proposes “may relate” text for randomly selected unconnected memories with similarity >0.7 and adds a weak edge. Both directly admit candidates under fixed heuristics, without human review, answer oracle or evidence-consuming truth decision. Open idle/forced maintenance, not a controlled improvement curriculum. Later pruning removes cold/low-confidence unused memories; it is not falsification of their propositions. No whole-cycle rollback shown.

• Immediate return: Derived/insight counters.
• Later read-back: Ordinary memory retrieval and graph selection can use inserted candidates.
• Delegated visibility: Selected text later reaches host; no autonomous critic or delegated agent.
• Selection predicate: Memify weight/duplicate checks; dream random pair/connectivity/similarity checks.
• Invalidation or expiry: Forget, decay and ordinary pruning; no semantic retraction protocol evidenced here.
• Activation or effect: Automatic durable conjecture production wired.
• Evidence limits: No observed candidate, acceptance or capacity gain; dream truth conditions underspecified.
• implementation conclusion status: wired
• operation conclusion status: uninspected

derived_content = f"{src_name} and {tgt_name} are frequently modified together" --- zikkaron/curation.py

if avg_heat > 0.7 and weight >= 5.0: # Both entities are hot AND relationship is established -> boost new_weight = weight + 0.5 --- zikkaron/curation.py

content = f"Dream connection: {summary_a} may relate to {summary_b}" --- zikkaron/sleep_compute.py

EPI-RTE-inferred-ranking

Label: Inferred passages and ranking checks

Part of: RT-RTE-context-selection

SRC-1, zikkaron/retrieval.py. A material part of service context selection. Deterministic trace: input “Alice dislikes camping.” matches the capitalized-subject expression, contains camping, and yields “Alice enjoys outdoor activities, nature, camping, hiking, and national parks.” No negation/preference check intervenes. Disliking camping does not entail enjoying it; this establishes ampliative conjecture rather than merely unproven entailment. In open-domain FlashRank mode, inferred text is scored as an extra passage and maximum passage score is mapped to the original memory. The conjecture changes ranking, not necessarily returned content.

NLI is a separate check: target is whether the first 512 memory characters support a question-to-statement hypothesis, domain is model-predicted textual entailment. Code assumes probability index 2, then blends with cross-encoder scores. It neither proves the premise true nor records a truth-acceptance decision. Adversarial checking computes abstain, but the inspected recall consumer only copies confidence and logs uncertainty; it does not branch on abstain. Numerical evidence therefore licenses ranking/advice, not enforced abstention.

• Immediate return: Scored selected memories; inferred text is internal ranker supply.
• Later read-back: inapplicable — this subroute does not persist inferred passages; outer recall separately changes access/rewrite state.
• Delegated visibility: Ranker sees extra inferred passage; host sees original selected memory after remaining packing.
• Selection predicate: Open-domain branch, configured models; maximum variant score, NLI blend and later selection.
• Invalidation or expiry: inapplicable — passage/score state is current-call only.
• Activation or effect: Rank-score effect wired; activation or benefit unobserved.
• Evidence limits: Model calibration, class mapping and external artifacts unverified; deterministic deduction uses implementation only.
• implementation conclusion status: wired
• operation conclusion status: uninspected

if any(word in lower for word in ("camping", "campfire", "meteor shower", "hiking", "forest", "mountains")): _add(f"{subject} enjoys outdoor activities, nature, camping, hiking, and national parks") --- zikkaron/retrieval.py

hypothesis = _question_to_statement(query) pairs = [(m["content"][:512], hypothesis) for m in memories] scores = self._nli_model.predict(pairs) # Shape: (n, 3) for [contradiction, neutral, entailment] --- zikkaron/retrieval.py

if self._settings.ADVERSARIAL_DETECTION_ENABLED and result_memories: adv_info = self._detect_adversarial(result_memories) for mem in result_memories: mem["_retrieval_confidence"] = adv_info["confidence"] if adv_info["is_uncertain"]: logger.debug( "Low retrieval confidence (%.3f), score_gap=%.3f", adv_info["confidence"], adv_info["score_gap"] ) --- zikkaron/retrieval.py

EPI-RTE-file-validation

Label: File freshness validation

SRC-1, zikkaron/server.py, zikkaron/staleness.py, zikkaron/hooks/prompt-recall.py. No supplied counterpart covers this exposed operation's evaluator domain. validate_memory calls the detector when initialized; its target is saved file-hash equality. Missing memory fails, missing hash or unresolved reference succeeds, changed hash lowers heat and flags stale. MCP fallback has different path resolution: it hashes directory_context and similarly treats missing saved hash as valid. The file watcher can flag all parent-directory memories as well as matching hashes. The prompt hook filters stale rows. This is operative freshness metadata, not acceptance of the natural-language memory. Success does not authenticate authorship or source correctness.

• Immediate return: MCP memory_id/is_valid/reason dictionary.
• Later read-back: Heat and stale flags affect future consumers; prompt FTS/vector SQL require is_stale=0.
• Delegated visibility: Caller sees status; later host receives filtered context.
• Selection predicate: Current/saved hash comparison; missing-reference exceptions above.
• Invalidation or expiry: No automatic truth certification or all-path retraction; successful unchanged check does not itself clear all stale state in read method.
• Activation or effect: Mismatch metadata mutation and prompt exclusion wired.
• Evidence limits: Only selected consumers verified; direct directory context does not filter by stale flag in inspected storage query.
• implementation conclusion status: wired
• operation conclusion status: uninspected

if not file_hash: return {"valid": True, "reason": "no file reference"} --- zikkaron/staleness.py

if current_hash != file_hash: new_heat = memory["heat"] / 2.0 self._storage.update_memory_heat(memory_id, new_heat) self._storage.update_memory_staleness(memory_id, True) return {"valid": False, "reason": "file changed"} --- zikkaron/staleness.py

EPI-RTE-feedback-update

Label: Usefulness confidence and stability update

SRC-1, zikkaron/server.py, zikkaron/thermodynamics.py, zikkaron/reconsolidation.py, zikkaron/metacognition.py. Distinct exposed rate_memory operation, not a duplicate of supplied generic context selection. Caller proposes was_useful; human versus automated authorship unresolved. Code accepts it without independent truth review, records counts and updates stability. After more than three accesses confidence becomes useful_count/access_count. It is consumed by server heat*confidence sorting and metacognitive chunk scoring; stability changes rewrite thresholds. No expected answer oracle, experiment, per-rating veto or rollback. Mode is open feedback; guidance is fixed count/stability formula. A useful false claim and a useless true claim are possible, so this is not truth calibration.

• Immediate return: Rated status, counters, confidence and stability.
• Later read-back: Persisted scalars affect later ranking, coverage advice and reconsolidation susceptibility.
• Delegated visibility: Caller receives rating result; no delegated judging context.
• Selection predicate: Existing memory and supplied boolean; count threshold for confidence.
• Invalidation or expiry: Later feedback/forget modifies or removes state; no rating reversal protocol shown.
• Activation or effect: Durable behavior-determining metadata update wired.
• Evidence limits: No observed helpfulness, accuracy or capacity comparison; callers not authenticated as evaluators in wrapper.
• implementation conclusion status: wired
• operation conclusion status: uninspected

if access_count > 3: confidence = useful_count / access_count else: confidence = mem.get("confidence", 1.0) --- zikkaron/thermodynamics.py

EPI-RTE-recall-rewrite

Label: Query-driven reconsolidation

Part of: RT-RTE-context-selection

SRC-1, zikkaron/server.py, zikkaron/reconsolidation.py. Part of the supplied service context-selection loop, after current scoring. Server calls reconsolidate(memory ID, query, empty directory). Program computes weighted embedding/directory/time/tag mismatch, changes plasticity, then vetoes rewriting for protected memories or low mismatch. Moderate mismatch appends current query, truncating old text when necessary; high mismatch archives old content and inserts current query with old tags. Contrary to an extinction/replacement reading, the inspected archive branch does not delete or disable the original. The current response remains the already-selected dictionaries; future reads see changed DB state. No fact oracle or per-change human veto; guidance is fixed thresholds plus use-derived stability. Archive is recovery material, not evidence an automatic rollback is invoked.

• Immediate return: Internal action/mismatch/new-content dictionary, ignored by recall wrapper.
• Later read-back: Mutated or new memories are available on later calls; current scored return is not rebuilt.
• Delegated visibility: Host current return and future DB contents differ; no delegated agent.
• Selection predicate: Protected flag, mismatch, stability and plasticity thresholds.
• Invalidation or expiry: Later edits/forget; archive retention depends on original deletion and storage lifecycle.
• Activation or effect: Durable query-fed update wired, semantic benefit unobserved.
• Evidence limits: Queries can be non-truth-apt or false; directory argument is empty, not authenticated project identity.
• implementation conclusion status: wired
• operation conclusion status: uninspected

_reconsolidation.reconsolidate(m["id"], query, "") --- zikkaron/server.py

EPI-RTE-index-enrichment

Label: Index-time enrichment

SRC-1, zikkaron/storage.py, zikkaron/enrichment.py. Distinct caller-conditioned insertion subroute spanning the supplied remember/seed/evaluation containers; not part of exactly one supplied route. Requires settings, embedding engine, embedding and minimum length; normal direct insert callers need not supply these. Separate parts are ConceptNet acquired concepts (external warrant unknown), COMET generated inference, Doc2Query questions (not necessarily truth-apt), and fixed logic expansion fragments. Optional FPA uses similarity, not proposition validation. Code directly persists supplements/model-name metadata and updates the memory embedding column using enriched text; then the inspected end of insertion inserts the original local embedding variable into the vector table. Therefore changed ranking from that particular enriched vector is consumer-dependent, not universally guaranteed. No per-supplement human approval, expected-answer oracle or whole-insert rollback. Guidance is enabled expanders/filter thresholds, not outcome criticism.

• Immediate return: Outer storage returns memory ID.
• Later read-back: Enrichment columns persist; numerical consumers may see updated DB embedding, but initial vector insert uses original variable.
• Delegated visibility: Host normally sees original content, not proof of supplements; models receive source text.
• Selection predicate: Flags, minimum length, model availability and similarity filtering.
• Invalidation or expiry: Caller updates/deletion; exact provider content not frozen.
• Activation or effect: Persistence/encoding wiring inspected; actual expansion and consumer-specific use unobserved.
• Evidence limits: Cannot individuate model-produced propositions without outputs; syntax/association terms do not imply sound logical inference.
• implementation conclusion status: wired
• operation conclusion status: uninspected

if enrichment_data.get("enriched_content") and embeddings_engine is not None: new_embedding = embeddings_engine.encode_document_enriched( memory["content"], enrichment_data["enriched_content"] ) --- zikkaron/storage.py

similarity = float(np.dot(original_vec, text_vec)) if similarity >= threshold: kept.append(text) --- zikkaron/enrichment.py

EPI-RTE-profile-beliefs

Label: Profile extraction and callable beliefs

SRC-1, zikkaron/profiles.py, zikkaron/storage.py, zikkaron/retrieval.py. A mixed production/consumption chain not wholly contained in one supplied route. Settings-enabled insertion calls ProfileExtractor and persists attributes/evidence IDs. BeliefDeriver separately maps activity substrings to enjoys assertions and fixed 0.6 confidence; trait/preference patterns and multi-profile summaries have their own fixed confidences/evidence references. A mention such as “Alice dislikes camping” is insufficient warrant for enjoys-outdoor-category. Search for derive_from_memory/derive_from_profiles across zikkaron/ found definitions and tests, not a production invocation; ordinary insertion therefore does not establish belief generation. Existing belief rows are nevertheless consumed by FTS and merged after cross-encoder reranking. Software proposes/admit values; no human approval or gold oracle. Current source specifies no evidence-consuming truth acceptance or production rollback.

• Immediate return: Profile extraction side effect / callable belief insertion; outer insert returns memory ID.
• Later read-back: Profile and belief FTS rows become synthetic recall results with confidence-based scores.
• Delegated visibility: Host can receive structured-value summaries; no delegated critic.
• Selection predicate: Extraction patterns/category substring match; retrieval FTS/confidence.
• Invalidation or expiry: Update/delete methods and ordinary store lifetime; inspected memory deletion does not traverse evidence IDs.
• Activation or effect: Profile insertion and belief consumer wired; production belief producer invocation unestablished.
• Evidence limits: Model-free extraction still requires semantic interpretation; source IDs are lineage, not confirmation.
• implementation conclusion status: wired
• operation conclusion status: uninspected

belief_content = f"enjoys {label} (evidence: {', '.join(activities)})" --- zikkaron/profiles.py

EPI-RTE-causal-outputs

Label: Causal-labelled graph outputs

Part of: RT-RTE-trace-consolidation

SRC-1, zikkaron/knowledge_graph.py, zikkaron/causal_discovery.py, zikkaron/consolidation.py, zikkaron/server.py. A maintenance subroute with separately exposed read consumer. Temporal heuristic assigns caused_by to sufficiently weighted co-occurrences ordered by a helper; association plus temporal ordering does not entail causation. PC discovery is separate: after >=5 variables and >=10 windows it calls pc_algorithm and persists directed edges for resolvable entities; periodic consolidation invokes it after an event counter reaches 50. The MCP causal-chain wrapper returns the PC engine's graph answer, not necessarily heuristic graph edges. These consumers are kept separate. Program proposes and admits graph updates without human review, expected-answer oracle or verified causal intervention. Full CI/orientation implementation and assumptions remain unassessed; the claim is not that PC cannot identify causes under suitable assumptions. No update-wide rollback or old-edge withdrawal policy established.

• Immediate return: Counters/DAG metadata or causes/effects dictionary.
• Later read-back: Directed edges persisted for later causal-chain requests; heuristic relations remain graph state.
• Delegated visibility: Host receives graph answer; no autonomous causal experiment.
• Selection predicate: Heuristic co-occurrence/order; PC minimum data and resulting edges.
• Invalidation or expiry: Deletion/supersession of old causal edges not traced; no current-completeness guarantee.
• Activation or effect: Discovery/storage/query wiring present; no observed or causal validation.
• Evidence limits: PC algorithm internals and observational assumptions not fully inspected; cannot certify encoding or identify true causal edges.
• implementation conclusion status: wired
• operation conclusion status: uninspected

dag = self.pc_algorithm(data, variable_names)

Store directed edges in causal_dag_edges table

--- zikkaron/causal_discovery.py

EPI-RTE-coverage-advice

Label: Metacognitive coverage advice

SRC-1, zikkaron/metacognition.py, zikkaron/server.py. New exposed operation omitted from the supplied principal routes. The check target is available topic coverage, not whether a specific answer is true. Code blends density, entity coverage, recency and mean confidence; >=0.7 yields sufficient/advice to proceed, >=0.4 partial, otherwise insufficient. Gap detection emits descriptions and suggested investigations. MCP returns the advice; no inspected wrapper enforces investigation or prevents answering. It is evidence about local proxies, not a measured guarantee of adequate knowledge. Current-run check, no proposal admission/rollback and no expected-answer oracle.

• Immediate return: Coverage/confidence/suggestion/gaps or gap list.
• Later read-back: inapplicable — advisory result is not persisted by these wrappers.
• Delegated visibility: Host receives advice; no delegated inspection task automatically launched.
• Selection predicate: Query/directory, retrieval proxies and thresholds.
• Invalidation or expiry: inapplicable — each request recomputes current advice.
• Activation or effect: Advice delivery wired; changed host behavior unobserved.
• Evidence limits: Density/recency and usefulness-derived confidence do not prove factual completeness.
• implementation conclusion status: wired
• operation conclusion status: uninspected

if overall >= 0.7: suggestion = "sufficient" detail = "Proceed with confidence — strong knowledge coverage." --- zikkaron/metacognition.py

EPI-RTE-memory-compression

Label: Gist and tag compression

Part of: RT-RTE-trace-consolidation

SRC-1, zikkaron/compression.py, zikkaron/consolidation.py. Maintenance part, not a new acceptance route. Code chooses compression by schedule, archives original, selects high-score sentences with first/last and code blocks, replaces content/embedding, then optionally generates a shortened tag representation. No human approval, truth oracle or semantic equivalence checker is shown in read extraction/update methods. This is extraction/reshaping in form, but proposition preservation is indeterminate without clause-level interpretation of actual input: omitted conditions can change truth. Source archive/original_content affords reconstruction; neither proves automatic retrieval uses the original. Admission is automatic according to schedule and compression level; no cycle-wide rollback.

• Immediate return: Gist/tag string and cycle counters.
• Later read-back: Ordinary consumers receive replacement text; archive/original_content are separate recovery material.
• Delegated visibility: Host later sees selected compressed memory, not necessarily original.
• Selection predicate: Schedule/level plus sentence scores and length target.
• Invalidation or expiry: Further compression and deletion; delete_memory also removes associated archives.
• Activation or effect: Content/embedding replacement wired.
• Evidence limits: Full compression schedule/decompression consumers unassessed; no lossless guarantee.
• implementation conclusion status: wired
• operation conclusion status: uninspected

Archive original content

self._storage.insert_archive({ "original_memory_id": memory_id, "content": original_content, "embedding": memory.get("embedding"), "archive_reason": "compression", }) --- zikkaron/compression.py

EPI-RTE-narrative-production

Label: Narrative production and story read

Part of: RT-RTE-trace-consolidation

SRC-1, zikkaron/narrative.py, zikkaron/sleep_compute.py, zikkaron/server.py. Distinct sleep-produced state and story/resource consumers, compared with the runtime maintenance container. Sleep invokes auto_narrate; active directories without recent narrative get memory-window counts and bounded extracted decisions/events/entities/focus text. Code directly inserts narrative entries, and story retrieval concatenates summaries chronologically. No human approval, expected answer oracle or truth acceptance; guidance is fixed extraction/time-window policy. A count is entailed only about recorded DB entries selected by this window, not actual work or successful outcomes. Detailed extracted propositions remain indeterminate. Recovery is later stored entries/operator action; no automatic revision loop shown.

• Immediate return: Inserted entry/counters or story string.
• Later read-back: Narrative entries persist and reach get_project_story and narrative resource.
• Delegated visibility: Host receives concatenated summaries rather than all source rows.
• Selection predicate: Directory, created-at window, active heat and narrative interval; limited lists.
• Invalidation or expiry: New time windows/entry selection; withdrawal propagation unassessed.
• Activation or effect: Sleep production and later read wired.
• Evidence limits: No actual narrative; extraction fidelity and completeness not verified.
• implementation conclusion status: wired
• operation conclusion status: uninspected

entry_id = self._storage.insert_narrative_entry(entry) entry["id"] = entry_id return entry --- zikkaron/narrative.py

stats["narrative"] = self._narrative.auto_narrate() --- zikkaron/sleep_compute.py

EPI-RTE-trigger-delivery

Label: Prospective reminder delivery

SRC-1, zikkaron/prospective.py, zikkaron/server.py. Part of remember includes automatic future-phrase extraction and checking of prior reminders; explicit create_trigger is an additional producer. This spans more than one supplied referent. Program validates trigger type, directly persists reminder/condition/directory, matches context, increments counts and deactivates after count limit. Caller or regex proposes text; code admits it without per-content human approval. No gold oracle or fact check. Triggering is operational matching, not claim confirmation. Guidance is fixed phrase/type/match policy; no automatic rollback beyond deactivation is shown.

• Immediate return: Created reminder ID or triggered reminder material in remember flow.
• Later read-back: Persisted reminders can match later remember contexts.
• Delegated visibility: Host receives triggered reminder through response, not a separately tasked agent.
• Selection predicate: Valid type; context match; firing count.
• Invalidation or expiry: Count-based deactivation; detailed matching grammar unassessed.
• Activation or effect: Persistence and later condition-driven selection wired.
• Evidence limits: Reminder can contain claims or instructions; no factual license follows.
• implementation conclusion status: wired
• operation conclusion status: uninspected

_prospective.auto_create_from_content(content, context) --- zikkaron/server.py

if triggered_memories: memory["triggered_prospective_memories"] = [ {"id": pm["id"], "content": pm["content"]} for pm in triggered_memories ] --- zikkaron/server.py

EPI-RTE-memory-withdrawal

Label: Memory deletion and dependent-state limits

SRC-1, zikkaron/server.py, zikkaron/storage.py. Exposed forget is distinct from runtime's remember/recall container; memify pruning is an alternate caller of the same deletion primitive. Caller proposes ID; code checks existence and deletes. delete_memory explicitly removes archives, transitions and vector rows before the memory; it does not traverse profile/belief evidence_memory_ids. Thus retaining derived evidence IDs does not establish dependency retraction. No expected-answer oracle or human per-content veto; deliberate caller deletion is not a recorded factual rejection. Guidance is caller ID or pruning thresholds. No automatic undo; archive deletion can remove recovery material.

• Immediate return: Deleted/not_found status.
• Later read-back: Deleted memory no longer reaches ordinary lookup; independent profile/belief rows can remain queryable.
• Delegated visibility: Host receives status; later consumers see surviving state.
• Selection predicate: ID existence or caller pruning predicate.
• Invalidation or expiry: Deletion is terminal absent external copy; no automatic rollback.
• Activation or effect: Scoped removal wired, not universal epistemic withdrawal.
• Evidence limits: No concurrent-failure test or complete dependency cleanup audit; no whole-system retraction guarantee.
• implementation conclusion status: wired
• operation conclusion status: uninspected

self._conn.execute("DELETE FROM memories WHERE id = ?", (memory_id,)) self._conn.commit() --- zikkaron/storage.py

EPI-RTE-answer-judgment

Label: Benchmark answer judgment

Part of: RT-RTE-evaluation-mode

SRC-1, benchmarks/run_locomo_jscore.py. Check part of supplied evaluation container; generation, judging and deployed memory changes are not conflated. Operator supplies data/configuration and model/provider. Gold answer is an explicit expected-answer oracle for F1 and judge, withheld from generation in supplied runtime trace. Judge compares generated answer with gold under a generous essential-meaning/partial-answer rubric. F1 is lexical similarity under category handling, not independent truth proof. Driver maps CORRECT to true, exceptions to false, appends samples, aggregates and optionally saves JSON. Result consumer is measurement/reporting, not production-memory or model-weight admission. This is bounded benchmark mode; retries/fallbacks are not rollback of a product update. Provider fixity/caching limits are see RT-CMP-evaluation-models.

• Immediate return: F1, correctness boolean and reasoning in samples and aggregate report.
• Later read-back: Optional output file can be read by operator; no automatic production-policy read-back established.
• Delegated visibility: Generation sees question/selected context; judge sees question/gold/generated answer.
• Selection predicate: Question subset/category, f1-only flag, selected generation/judge configuration.
• Invalidation or expiry: Operator controls result lifetime; evaluator failures become false scores.
• Activation or effect: Scoring/reporting wired, observed scores absent.
• Evidence limits: No dataset/results/provider traces; answer scores do not identify mechanism effects or transfer.
• implementation conclusion status: wired
• operation conclusion status: uninspected

j_correct = judgment["label"].upper() == "CORRECT" j_reasoning = judgment.get("reasoning", "") --- benchmarks/run_locomo_jscore.py

Claims

EPI-CLM-derived-facts

Label: Memify derived facts and self-improvement

SRC-1, zikkaron/curation.py; doctrine/design. Distinct claim from RT-CLM-write-gate: automatic synthetic fact production and improvement during maintenance, not explicit ingestion novelty. Its implemented route is EPI-RTE-synthetic-facts. The source calls the cycle self-improvement; this does not establish an improvement objective measured against outcomes or exercised improvement. The transformation is stronger than its weighted co-occurrence premise.

"""Generate synthetic derived-fact memories for high-weight entity pairs.""" --- zikkaron/curation.py

"""Run the full memify self-improvement cycle. --- zikkaron/curation.py

EPI-CLM-cls-consistency

Label: CLS consistency verification

SRC-1, zikkaron/cls_store.py; doctrine/design. No supplied claim counterpart after comparison with RT-CLM-write-gate; this targets semantic-promotion consistency rather than novelty. The lexical implementation in EPI-RTE-cls-consistency does not establish the unqualified semantic claim.

"""Verify cluster members don't contradict each other. --- zikkaron/cls_store.py

EPI-CLM-benchmark-benefit

Label: Reported benchmark and knowledge-update benefit

SRC-1, README.md; reported operation and doctrine/design. No supplied claim counterpart: RT-RTE-evaluation-mode is the executable evaluation container, not this public outcome claim. README reports LoCoMo Recall@10 86.8%, BEAM overall 0.404 and attributes update/contradiction abilities to thermodynamics/reconsolidation/metadata. It explicitly discloses different readers, preventing attribution of the entire QA contrast to retrieval. No output artifacts or causal intervention were supplied. EPI-RTE-answer-judgment supports a measurement procedure, not these reported numbers.

One important caveat: the BEAM paper's baselines use GPT-4.1-nano as the reader model. We use Claude Opus 4.6, which is substantially more capable. The instruction-following result (0.750 with a retrieval MRR of just 0.086) demonstrates that reader quality contributes to QA scores independently of retrieval quality. Our retrieval metrics (MRR 0.422, Recall@10 0.353) are model-independent and directly comparable to the paper's baselines. --- README.md

Behavioral-authority paths

none declared in this member; delivery and rule authority are see RT-BAP-context-delivery, RT-BAP-retrieval-rules. Route declarations above identify distinct scalar-update and measurement consumers without redefining those paths.