AriGraph epistemic report

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

Source-and-claim boundary

The question is how AriGraph turns observations or paragraphs into material that later decisions rely on, and what its checks license. Scope and exclusions are see boundary; evidence identities and layers are see Source register. SRC-1 is the frozen repository. Direct source inspection supports implementation deductions, not fresh target execution. Supplied runtime records remain provisional but their selected forcing cases survive the direct reads below.

The principal public performance claim is see RT-CLM-memory-performance. Extraction and replacement instructions make narrower source-fidelity and conservative-maintenance claims, see EPI-CLM-concrete-extraction and EPI-CLM-conservative-replacement. The README's title uses “Learning Knowledge Graph World Models”; the exposed routes support experience-fed memory construction, not by that label alone validated explanatory models or improved capacity.

Coverage separates direct reads from supplied findings. No new source is introduced. An available but uninspected task or log is not an access gap.

Entry point or operation	Source path	Coverage and record citations	Limit and conclusion prevented
Game invocation, planner/action, exploration, navigation	pipeline_arigraph.py; agents/parent_agent.py; prompts/system_prompts.py	Direct complete reads; see RT-RTE-game-cycle, RT-RTE-graph-navigation	No fresh trace; current-default activation, reliable compliance and benefit unestablished.
Fact extraction, replacement, add/delete, clear and retrieve	graphs/contriever_graph.py; graphs/parent_graph.py; prompts/prompts.py	Direct complete reads; see RT-RTE-game-memory, EPI-RTE-triplet-parsing, EPI-RTE-removal-parsing	Warrant of external inputs and opaque model interpretations unresolved.
QA ingest, answer, metrics and local-model variant	musique_test_big.py; graphs/contriever_graph.py	Direct complete reads; see RT-RTE-qa-cycle, EPI-RTE-answer-metrics; local-agent detail supplied by RT-CMP-llama-agent	HotpotQA shape mismatch and graph-model selection defect bound reproduction claims. Dataset correctness unassessed.
Hypergraph exposed interface, deletion and event ranking	graphs/hypergraph.py; prompts/prompts.py	Direct complete reads; see RT-RTE-hypergraph-update	Interface implementation is not evidence of a shipped downstream consumer or deployment.
Intended removal evaluator	pipeline_graph_evaluation.py; graphs/lazy_graph.py	Evaluator directly read; LazyGraph finding supplied by RT-RTE-graph-evaluation	Undefined initial state and uncalled scoring defeat an operational-validation inference.
Embedding fact traversal and fixed-reference diagnostic	utils/retriever_search_drafts.py	Direct complete read; see EPI-RTE-reference-match, RT-RTE-game-memory	Similarity thresholds and reference membership do not certify truth.
Observation cleanup, normalization, episode ranking and auxiliary navigation	utils/utils.py	Direct bounded reads of lines 1–110, 129–193, 235–304; see EPI-RTE-observation-cleanup, EPI-RTE-triplet-parsing	Remaining reward, exit and thesis helpers are covered only by supplied runtime findings; no exhaustive helper correctness claim. Auxiliary navigation has an uninitialized curr_loc read; no root invocation found in the scoped search.
Environment feedback and reference facts	utils/textworld_adapter.py	Direct lines 1–100; see RT-RTE-game-cycle, RT-RTE-graph-evaluation	External TextWorld/task semantics not executed or independently checked; rest of rendering helper uninspected.
Summary/full-history/smart-RAG alternatives and human driver	pipeline_summary.py; pipeline_fullhist.py; pipeline_smartrag.py; interactive_human_games.py	Supplied findings plus direct scoped dispatch/update searches; see RT-RTE-summary-cycle, RT-RTE-full-history, RT-RTE-smart-rag, RT-RTE-human-play	No independently observed baseline fidelity, human policy or benefit.
Model/checkpoint and embedding interiors	utils/contriever.py; src/contriever.py; agents/llama_agent.py	Supplied RT-CMP-retriever, RT-CMP-gpt-agent, RT-CMP-llama-agent	Exact upstream weights and provider internals unavailable; no parametric knowledge acquisition or training-quality conclusion.
Authored reflexion instruction	prompts/prompts.py	Direct definition and scoped search corroborate RT-ABS-reflex-invocation	No named invocation; does not rule out opaque model reasoning.
Results, tasks, assets and training support	logs/; qa_data/; envs/; img/; src/utils.py; src/dist_utils.py	Coverage and exclusions see boundary; performance claim see RT-CLM-memory-performance	Not directly reassessed here; prevents aggregate efficacy, fixture fidelity and all-repository evaluation negatives.

Dispatch search at the frozen revision covered root Python files, agents/, graphs/, prompts/, utils/ and src/contriever.py, using argparse, click, register, hook, tools=, tool_calls, reflex_prompt, score calls and mutation/load terms. It found script entry points and TextWorld registration rather than an application tool registry in these paths. The source exposes callable graph/helper operations beyond the main driver; the fixed-reference diagnostic is therefore assessed separately. This is not a claim that all source-support or excluded bytecode paths lack evaluators. No runtime supersession is needed: the supplied combined routes remain valid containers for the parts declared here.

Missing fresh execution evidence prevents current-default operation claims. Missing intervention/comparison evidence prevents causal attribution of performance to extraction, replacement, episodes or embeddings individually. External-input truth, exact model versions and model interpretation fidelity remain unknown. Accessible logs can support only their attributed sequences, see the runtime's game and QA records; no new log sampling was performed here.

Epistemic-object inventory

Each row identifies one material part; generic identity/form/substrate belong to the cited declaration. Natural-language propositions inside symbolic triplets must not be confused with numerical embeddings used to rank them.

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
EPI-OBJ-observation-text	Statements about visible game state or paragraph subject matter	see record	SRC-1, pipeline_arigraph.py	Source warrant unknown; game cleanup can discard text.
EPI-OBJ-triplets	Subject–relation–object propositions about an environment or QA subject	see record	SRC-1, graphs/contriever_graph.py	Semantic fidelity and certainty are not locally adjudicated.
EPI-OBJ-episodic-text	Retained observation/paragraph statements	see record	SRC-1, graphs/contriever_graph.py	Raw relative to graph extraction, not raw relative to all game preprocessing.
EPI-OBJ-outdated-triplets	A model assertion that an old fact is outdated/redundant	see record	SRC-1, utils/utils.py	Parsed deletion target omits replacement side and rationale.
EPI-OBJ-answer	QA answer and accompanying explanation	see record	SRC-1, musique_test_big.py	No checked entailment from retrieved evidence.
EPI-OBJ-answer-metrics	Token overlap and exact normalized answer equality	see record	SRC-1, musique_test_big.py	Lexical evaluator, not explanation or source-truth oracle.
RT-OBJ-plan	Goals are not truth-apt; reasons/progress can be	see record	SRC-1, pipeline_arigraph.py	Truth-apt subcontent not individually recorded or checked.
EPI-OBJ-thesises	Thesis sentences and retained event text	see record	SRC-1, graphs/hypergraph.py	No root consumer established; linked event/claim text differs from embeddings.
EPI-OBJ-summary	Model summary of states/outcomes plus strategic claims	see record	SRC-1, pipeline_summary.py	Fidelity and entailment indeterminate.
RT-CMP-retriever	None individuated in weights/embeddings	see record	SRC-1, utils/contriever.py	Numerical similarity is not a fact-truth estimate established by this evidence.
RT-CMP-gpt-agent, RT-CMP-llama-agent	Parametric truth-apt contents not determinable	see records	SRC-1, agents/parent_agent.py	Opaque model knowledge cannot be inventoried from returned text or identifiers.
RT-OBJ-system-prompts	Rules/goals primarily policy; factual examples are illustrative	see record	SRC-1, prompts/prompts.py	Static instructions do not establish experience-derived tested theory.

Authority-route ledger

Fields owned by canonical records say see. Splits below assess functions separately without redeclaring the same runtime container. Results are possible implementation results unless an explicitly cited record supplies bounded observed support. Epistemic authority is never inferred from delivery or storage alone.

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-observation-cleanup	content transformation	implemented	EPI-OBJ-observation-text	truth-apt transformation: acquisition/import	Environment feedback to extraction input	see route	see route	see route	Text supplied downstream	Input source warrant unknown; omission demonstrated	Limits what extraction sees	see route	SRC-1, utils/utils.py	none	none	Acquisition includes lossy preprocessing.
RT-RTE-qa-cycle	content transformation	implemented	EPI-OBJ-observation-text	truth-apt transformation: acquisition/import	Dataset paragraph to memory input	see route	see route	see route	Paragraph ingestion	Source warrant unknown; text retained	Paragraphs enter per-task graph	see RT-BAP-memory-decisions	SRC-1, musique_test_big.py	none	none	Dataset integrity unassessed.
RT-RTE-game-memory, RT-RTE-qa-cycle	content transformation	implemented	EPI-OBJ-triplets	truth-apt transformation: indeterminate	Input text to extracted propositions	Model interpretation, domain game/paragraph content; see routes	see routes	see routes	Parsed candidates can enter memory	Candidate extraction only; neither entailment nor non-entailment established generally	Creates facts later selected for decisions	see RT-BAP-memory-decisions	SRC-1, prompts/prompts.py	EPI-CLM-concrete-extraction	Fidelity not enforced	Per-output provenance/semantic comparison needed to distinguish faithful compression from conjecture.
EPI-RTE-triplet-parsing	content transformation	implemented	EPI-OBJ-triplets	truth-apt transformation: non-ampliative reshaping	Text fragments to normalized triplet candidates	see route	see route	see route	Syntactic filtering and normalized identity	No truth license; case/alias assumptions bound preservation	Malformed/empty fragments omitted; normalized duplicates suppressed	see route	SRC-1, utils/utils.py	none	none	Real entity distinctions can be collapsed by normalization.
RT-RTE-game-memory	check/evidence production	implemented	EPI-OBJ-outdated-triplets	no content change	Old candidates versus new facts: same-aspect redundancy/conflict	Model under conservative instruction; see route	see route	Replacement proposal	Advisory proposal until code consumes it	Model comparison, not independent truth check	Supplies removal candidates	see route	SRC-1, prompts/prompts.py	EPI-CLM-conservative-replacement	Narrow candidate pool	Directional/other excluded strings are not judged by this request.
EPI-RTE-removal-parsing	disposition/acceptance	implemented	EPI-OBJ-outdated-triplets	no content change	Proposal to deletion of matching old item	see route	see route	see route	Enforcing state deletion except guards	No evidence-consuming acceptance of replacement truth	Removes matching unprotected facts, potentially without installed replacement	see route	SRC-1, utils/utils.py	EPI-CLM-conservative-replacement	Replacement link discarded	Protected locations survive even a valid deletion proposal.
RT-RTE-game-memory, RT-RTE-qa-cycle	retention	implemented	EPI-OBJ-triplets, EPI-OBJ-episodic-text	no content change	Candidate/text to local stores	see routes	see routes	see routes	Admission without semantic acceptance	Retention grants availability, not warrant	Candidates remain retrievable until delete/reset	see routes	SRC-1, graphs/contriever_graph.py	none	none	Episodic text is not withdrawn with old triplet deletion.
RT-RTE-game-memory, RT-RTE-qa-cycle	operational admission/selection/consumption	implemented	EPI-OBJ-triplets, EPI-OBJ-episodic-text	no content change	Memory to context selection	see routes	see routes	see routes	Ranking then model-context delivery	Similarity/overlap supplies relevance only	Selected old facts/text become decision inputs	see RT-BAP-memory-decisions	SRC-1, utils/retriever_search_drafts.py	none	none	Omitted/new/recent filters change delivered subset; receipt is not activation.
RT-RTE-game-memory	content transformation	implemented	EPI-OBJ-triplets	truth-apt transformation: indeterminate	Changed location plus last action to reciprocal spatial claims	Fixed substring mapping; see route	see route	Two location edges	Direct addition and later path use	Conditional spatial inference only; environment motion semantics unverified	Enables route construction without semantic check	see RT-RTE-graph-navigation	SRC-1, graphs/contriever_graph.py	none	none	Teleports/ambiguous commands can defeat direction interpretation; no warranted-premise derivation claimed.
RT-RTE-graph-navigation	operational admission/selection/consumption	implemented	EPI-OBJ-triplets	no content change	Graph path to environment commands	see route	see route	see route	Enforcing command dispatch	Reachability computation is bounded by graph encoding, not world truth	Multiple commands before next planner check	see RT-BAP-action-dispatch	SRC-1, graphs/parent_graph.py	none	none	Feedback does not retrospectively validate all graph edges.
RT-RTE-game-cycle	behavior/policy adaptation	implemented	RT-OBJ-plan	non-truth-apt policy/content update: revised goals and priorities, with indeterminate factual reasons	New context to plan/action/exploration	see route	see route	see route	Whole-plan replacement, parser/dispatcher consumption	No independent acceptance of progress or reason claims	New guidance influences later requests; action parsing alone gates fallback	see RT-BAP-plan-decisions, RT-BAP-action-dispatch	SRC-1, pipeline_arigraph.py	none	Prompt policy not invariant	Independent learning/reflection/benefit not inferred.
RT-RTE-qa-cycle	content transformation	implemented	EPI-OBJ-answer	truth-apt transformation: indeterminate	Retrieved text/facts plus question to answer	see route	see route	see route	Answer returned and logged	Candidate answer; prompt asks abstention on doubt	Answer proceeds to metrics without truth gate	see route	SRC-1, musique_test_big.py	none	none	Needs per-answer premise/entailment analysis; fluency not warrant.
EPI-RTE-answer-metrics	check/evidence production	implemented	EPI-OBJ-answer, EPI-OBJ-answer-metrics	truth-apt transformation: entailed derivation	Normalized answer tokens versus dataset answer	see route	see route	see route	Diagnostic recording only	Lexical equality/overlap within normalization domain	Does not admit/reject/revise memory or next answer	see route	SRC-1, musique_test_big.py	RT-CLM-memory-performance	Metric is not explanation check	Derived count is warranted by code/input tokens, not truth of dataset answer.
RT-RTE-graph-evaluation	check/evidence production	doctrine only	EPI-OBJ-outdated-triplets	no content change	Predicted removed facts versus environment-fact differences	see route	see route	Intended precision/recall	No working score consumer established	Intended diagnostic, not product acceptance	No production update/gate	see route	SRC-1, pipeline_graph_evaluation.py	EPI-CLM-conservative-replacement	Executable defects	“Doctrine only” applies to functioning end-to-end scored validation; code fragments are present.
EPI-RTE-reference-match	check/evidence production	implemented	EPI-OBJ-triplets	truth-apt transformation: entailed derivation	Exact membership in fixed cooking list	see route	see route	see route	Print diagnostic only	Set/list membership, not general extraction fidelity	No gate or later behavioral consumer	see route	SRC-1, utils/retriever_search_drafts.py	none	none	No shipped invocation established by scoped search.
RT-RTE-hypergraph-update	content transformation	implemented	EPI-OBJ-thesises	truth-apt transformation: indeterminate	Text to thesis candidates	see route	see route	see route	Candidate production	Model fidelity unknown	Makes candidate theses available to update	see route	SRC-1, graphs/hypergraph.py	EPI-CLM-concrete-extraction	Unformatted template	Literal placeholder plus appended input; not demonstrated compliance.
RT-RTE-hypergraph-update	check/evidence production	implemented	EPI-OBJ-thesises	no content change	Existing versus new same-property claims	see route	see route	see route	Advisory model removal proposal	No independent truth evaluator	Supplies names to deletion	see route	SRC-1, prompts/prompts.py	EPI-CLM-conservative-replacement	none	Requirement for a replacement is prompt policy.
RT-RTE-hypergraph-update	disposition/acceptance	implemented	EPI-OBJ-thesises	no content change	Parsed names to removal	see route	see route	see route	Enforcing deletion/link cleanup	No epistemic acceptance of new thesis	Existing names removed without location guard	see route	SRC-1, graphs/hypergraph.py	EPI-CLM-conservative-replacement	Replacement side ignored	No transaction/rollback or replacement existence check.
RT-RTE-hypergraph-update	retention	implemented	EPI-OBJ-thesises	no content change	Candidates/events to dictionaries and links	see route	see route	see route	Local retention	Linked storage is not accepted-claim lifecycle integration	Later calls can retrieve theses/events	see route	SRC-1, graphs/hypergraph.py	none	none	Only API-level horizon established.
RT-RTE-hypergraph-update	operational admission/selection/consumption	implemented	EPI-OBJ-thesises	no content change	Thesis/event scores to returned text	see route	see route	see route	Ranking and return	Relevance ordering only	Caller can receive subset	see route	SRC-1, graphs/hypergraph.py	none	Ascending event order	No planner/action consumer established.
RT-RTE-summary-cycle	content transformation	implemented	EPI-OBJ-summary	truth-apt transformation: indeterminate	Observation/history/old summary to new account	see route	see route	see route	Whole-response assignment	Accuracy requested, not checked	New text is immediate decision context	see route	SRC-1, pipeline_summary.py	none	none	Summarization name alone cannot prove non-ampliative preservation.
RT-RTE-summary-cycle	retention	implemented	EPI-OBJ-summary	no content change	Summary response to later summary/decision calls	see route	see route	see route	Ungated retention/use	No acceptance criterion implemented	Previous account shapes next update	see route	SRC-1, pipeline_summary.py	none	none	No semantic rollback.
RT-RTE-full-history, RT-RTE-smart-rag	operational admission/selection/consumption	implemented	EPI-OBJ-observation-text	no content change	Retained observations/actions to decision context	see routes	see routes	see routes	Full delivery or ranked selection	History/relevance, not truth certification	Information and importance affect later decision inputs	see routes	SRC-1, pipeline_smartrag.py	none	none	Importance is an unvalidated policy score, not truth-apt world knowledge.
RT-RTE-game-memory, RT-RTE-qa-cycle	lifecycle integration	no route found within boundary	EPI-OBJ-triplets	no content change	Accepted proposition to evidence-linked organization/use	No prior evidence-consuming acceptance found in inspected updates	see routes for storage timing	No accepted-claim integration established	Storage/retrieval only	No added license	None distinct from pre-acceptance retention/use	see routes for actual memory use	SRC-1, graphs/contriever_graph.py	none	none	Bounded to these updates, not all possible external callers or model interiors.

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-memory-performance	see record	SRC-1, README.md; doctrine/design and reported operation	see record	see RT-RTE-game-cycle, RT-RTE-qa-cycle	Bounded supplied sequences, see routes; not aggregate replication	No independent intervention assessed; component effects not isolated	Memory-to-decision architecture wired; reported scores remain attributed	QA branch/configuration defects and uninspected aggregate evidence prevent reproduction and causal endorsement.
EPI-CLM-concrete-extraction	see record	SRC-1, prompts/prompts.py; doctrine/design	see record	see RT-RTE-game-memory, EPI-RTE-triplet-parsing, RT-RTE-hypergraph-update	Supplied game/QA candidate sequences, not general fidelity checks	No source-fidelity intervention or semantic validator	Requests concrete facts and modal hypotheses; creates structured candidates	Parser accepts syntax, not source support or calibrated uncertainty.
EPI-CLM-conservative-replacement	see record	SRC-1, prompts/prompts.py; doctrine/design	see record	see RT-RTE-game-memory, EPI-RTE-removal-parsing, RT-RTE-hypergraph-update	Actual correctness of replacement unobserved in handed samples	Intended evaluator defective, see RT-RTE-graph-evaluation; not a negative about all logs	Model judgment has operational withdrawal force, with exact-match and game-location guards	Replacement linkage/semantic criterion not enforced; QA does not invoke removal.

Bounded conclusion

AriGraph acquires external game/paragraph statements with unknown source warrant. Game preprocessing can omit text before memory receives it; QA directly supplies paragraph text. Its semantic store holds normalized extracted candidates, and its episodic store preserves the supplied text and fact associations. That is retention and later operational reliance, not by itself an evidence-consuming acceptance decision. See EPI-RTE-observation-cleanup, RT-RTE-game-memory, RT-RTE-qa-cycle.

The extraction prompts request faithful concrete facts and marked hypotheses. Model extraction, summary generation, spatial inference and answering remain indeterminate transformations at the semantic edge: no general checked entailment or positive demonstration of non-entailment is available for their outputs. Parsing/normalization are inspectable reshaping; they do not repair model meaning. Numerical similarity ranks relevance, not truth. Exact premises suffice to deduce the metrics computed on token lists and the fixed-reference membership count, but those deductions warrant only lexical results within the specified domain. See EPI-RTE-triplet-parsing, EPI-RTE-answer-metrics, EPI-RTE-reference-match.

The game maintenance model evaluates redundancy/conflict against selected old/new candidates. Its decision has operational withdrawal authority through the parser and graph deletion method. The criterion is requested same-aspect replacement with retention under uncertainty, intended for current game facts; no independent check establishes that the judgment is correct or that the named replacement exists. Exact old-item membership and location protection are enforcing guards, not semantic acceptance. QA's no-removal path is a counterexample to universal stale-fact reconciliation. Deleting a semantic fact does not delete all episodic text containing it. Hypergraph uses another removal path without the location guard and removes thesis links; its deployment and downstream decision consumer remain unestablished. See EPI-RTE-removal-parsing, RT-RTE-hypergraph-update.

Selected memory reaches later model user messages as informational context within one game attempt or QA task. Plans change priorities in retrieval/exploration/action requests, and graph paths directly dispatch multiple environment commands. These are consequential behavioral-authority paths, not evidence that models relied on each fact, that any explanation is warranted, or that retained changes improved capacity. Parse fallback blocks some malformed actions but not well-formed off-list commands. Environment score/completion assesses task outcomes, not correctness of every producing component or general transfer. See RT-BAP-memory-decisions, RT-BAP-plan-decisions, RT-BAP-action-dispatch.

QA metrics compare answers after generation to a local reference. Their only implemented consequence is diagnostic accumulation/logging; they do not reject memory, revise a plan or train a model. The fact-replacement evaluation script does not establish a working scored validation route. The separately exposed fixed-reference diagnostic prints membership, likewise without a later consequential consumer. Neither licenses a system-complete claim that no evaluation exists: task contents, remaining logs and external model interiors were not exhaustively assessed. See EPI-RTE-answer-metrics, RT-RTE-graph-evaluation, EPI-RTE-reference-match.

No accepted ampliative claim or post-acceptance lifecycle integration is established by the inspected memory updates. Experience-fed writes and model-revised plans do establish a behavior-shaping pathway, independently of the unused named reflexion prompt. Learning benefit, content-directed theory criticism, reflective theory building and exercised self-improvement require separate evidence and are not inferred from memory replacement, static prompts or performance labels. Exact parametric contents, upstream weight identity and provider-internal updates are not determinable; wrapper inference paths alone do not settle those questions. See RT-ABS-reflex-invocation, RT-CMP-gpt-agent, RT-CMP-retriever, RT-CMP-llama-agent.

Shared records

Operative objects

EPI-OBJ-observation-text

Label: Observation and paragraph text

Part of: RT-RTE-game-memory

This is the text-input part of that route, not its graph-update process. Source-native identity: observation passed to extraction; QA supplies paragraph_text to the same extraction API. Form: natural-language statements, alongside driver-added task/step text in the game. Substrate: Python strings. It overlaps the text inputs in RT-RTE-qa-cycle, not the answer or graph itself. Lineage and warrant: environment/fixture or dataset supplied, then game cleanup/annotation; truth of these inputs is not independently checked. SRC-1, pipeline_arigraph.py, musique_test_big.py.

observation = observation.split("$$$")[-1] observation = observation_processing(observation) --- pipeline_arigraph.py

triplets, episodic = graph.update_without_retrieve(text["paragraph_text"], [], log) --- musique_test_big.py

EPI-OBJ-triplets

Label: Triplets retained as semantic facts

Part of: RT-RTE-game-memory

This is the retained semantic-content part of the route; supplied QA route shares the graph class but is a different caller. Source-native identity: self.triplets, with each normalized subject/object and label. Form: symbolic list/dictionary structure holding natural-language proposition labels; downstream models interpret serialized facts. Substrate: local Python list plus numerical embedding caches, which are not the proposition content. Lineage: model extraction and deterministic spatial additions. No independent acceptance event is stored. Distinct from EPI-OBJ-episodic-text: deletion/retrieval operate differently. SRC-1, graphs/contriever_graph.py.

if triplet not in self.triplets: self.triplets.append(triplet) self.triplets_emb[self.str(triplet)] = self.get_embedding_local(self.str(triplet)) --- graphs/contriever_graph.py

EPI-OBJ-episodic-text

Label: Episodic observation text

Part of: RT-RTE-game-memory

Source-native identity: keys in self.obs_episodic, values carrying extracted-fact strings and observation embedding. Form: natural-language observation content, symbolic fact associations and separate dense numerical rank inputs. Substrate: Python dictionary and tensors; snapshots are local lists. This text-content part differs from graph facts because fact deletion does not withdraw the dictionary key or earlier text. Same class stores QA paragraphs, see RT-RTE-qa-cycle. Source warrant is unknown. SRC-1, graphs/contriever_graph.py.

obs_embedding = self.retriever.embed(observation) obs_value = [new_triplets_str, obs_embedding] self.obs_episodic[observation] = obs_value --- graphs/contriever_graph.py:77-79

EPI-OBJ-outdated-triplets

Label: Parsed outdated triplets

Part of: RT-RTE-game-memory

Source-native identity: predicted_outdated, derived from model-returned replacement text. Form: symbolic triplet targets naming natural-language facts, not retained justification or new fact identity. Substrate: temporary Python list and textual logs of original response. This is a material proposal part of the supplied update container, distinct from accepted replacements, which are not recorded as such. SRC-1, graphs/contriever_graph.py, utils/utils.py; passage see EPI-RTE-removal-parsing.

EPI-OBJ-answer

Label: QA answer text

Part of: RT-RTE-qa-cycle

Source-native identity: answer returned by get_answer. Form: natural-language explanation plus direct answer; no locally verified proof structure. Substrate: Python string and textual log. Lineage: question, retrieved facts/text and opaque model knowledge. Distinct from metrics and from the supplied route that generates/scores it. SRC-1, musique_test_big.py.

answer = get_answer(agent_qa, question, subgraph, episodic) log("AGENT ANSWER: " + str(answer)) log("TRUE ANSWER: " + str(task["answer"])) --- musique_test_big.py

EPI-OBJ-answer-metrics

Label: Answer metrics accumulators

Part of: RT-RTE-qa-cycle

Source-native identity: trueP, pred_len, true_len, EM, aggregate prec, rec, f1, em. Form: symbolic numerical counts/ratios/equality decisions. Substrate: Python lists and computed scalars, rendered to logs. The metric object is distinct from answer text and dataset oracle. It carries lexical comparison results, not semantic acceptance. SRC-1, musique_test_big.py; passage see EPI-RTE-answer-metrics.

EPI-OBJ-thesises

Label: Hypergraph theses and events

Part of: RT-RTE-hypergraph-update

Source-native identity: self.thesises, self.events, linked self.entities. Form: natural-language thesis/event names, symbolic hash-indexed parent/child links, and distinct numerical embeddings. Substrate: Python dictionaries/object fields. This is the retained content part of the supplied route, not a duplicate route or an established planner consumer. Event text is retained alongside generated claims; thesis deletion changes links but retains event objects. SRC-1, graphs/hypergraph.py.

if event_id not in self.events: self.events[event_id] = Event(event, event_emb, thesises_ids) else: self.events[event_id].add_children(thesises_ids) --- graphs/hypergraph.py

EPI-OBJ-summary

Label: Summary baseline text

Part of: RT-RTE-summary-cycle

Source-native identity: summary, model-replaced situation account. Form: natural-language compressed statements and strategic guidance; substrate: Python string/logs. Lineage: prior summary, observations/history and model output; preserved meaning unresolved. Different checks/consumers from semantic triplets. Supplied route supplies direct assignment and return passages; this object is its text part, not a second summary process. SRC-1, pipeline_summary.py, prompts/system_prompts.py; evidence see supplied route.

Routes

EPI-RTE-observation-cleanup

Label: Game observation cleanup

Part of: RT-RTE-game-cycle

Endpoints: environment feedback → delimiter stripping → observation_processing → step/task annotation → extraction/decision context. Owner: game driver and deterministic helper. SRC-1, pipeline_arigraph.py, utils/utils.py. Distinct part of the supplied container, not its entire environment/decision cycle. Concrete source deduction, without target execution: Room contains apple > take apple becomes Room contains apple because >.*$ under DOTALL removes the suffix. A statement after that marker also disappears; hence cleanup does not establish full source-content preservation. Renaming Recipe #1 to Recipe also collapses identity under a task-local alias assumption.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Cleaned natural-language string, then driver annotations.
• Later read-back: Cleaned text is retained in graph episodes/recent history for later model invocations, not the removed suffix.
• Delegated visibility: inapplicable — local synchronous preprocessing.
• Selection predicate: Suffix after last $$$; regex suffix removal from >; literal spelling substitutions.
• Invalidation or expiry: Original omitted text is not restored by this function; attempt history/reset bounds later use.
• Activation or effect: Changes the available text delivered to extraction and decisions; individual behavioral effect unobserved.
• Evidence limits: Deterministic deduction establishes text loss, not that omitted text is materially true or false in an observed task.

pattern = r">.*$" cleaned_text = re.sub(pattern, "", text, flags=re.DOTALL) --- utils/utils.py

EPI-RTE-triplet-parsing

Label: Triplet parsing and normalization

Part of: RT-RTE-game-memory

Endpoints: extraction response → process_triplets → clear_triplet → duplicate-aware add_triplets. Owner: helper and graph code. QA invokes the same functions. SRC-1, utils/utils.py, graphs/contriever_graph.py. This is the syntactic transformation/admission part, not a semantic evaluator. Concrete deduction: Apple, is in, I; bad yields one structured candidate ['Apple', 'I', {'label': 'is in'}]; bad fails the three-comma-field requirement. clear_triplet maps exact I to inventory and lowercases names, yielding ['apple', 'inventory', {'label': 'is in'}]. A lowercase i is not that exact alias before lowercasing. No execution was used. Normalized identity assumes case is immaterial and aliases refer to the same entity; otherwise fidelity can degrade.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Parsed list, then normalized candidates and local insertions.
• Later read-back: Inserted fact strings become later retrieval candidates and model context.
• Delegated visibility: inapplicable — local parsing and graph mutation.
• Selection predicate: Exactly three comma-separated fields, nonempty trimmed values; free relation skipped before normalization; normalized duplicate membership checked.
• Invalidation or expiry: Malformed fragments omitted; stored fact survives until matching deletion or reset.
• Activation or effect: Enforces shape/identity admission and determines later candidate population, not factual truth.
• Evidence limits: Commas/colons inside content can change parsing; semantic interpretation and source fidelity are not checked.

Proposer: extraction model; computational decider: parser/adder; human: initial prompts/configuration only. Veto: malformed/empty fragments and exact duplicate/free checks. Oracle: none. Mode: bounded attempt/task. Trigger/change: each supplied text; guidance requests concrete structured knowledge and marked hypotheses. Admission is immediate; no semantic rollback, reset is recovery. What persists: normalized facts, not rationale or evidence-consuming acceptance.

if len(triplet.split(",")) != 3: continue --- utils/utils.py:26-40

if triplet[0] == "I": triplet = ("inventory", triplet[1], triplet[2]) if triplet[1] == "I": triplet = (triplet[0], "inventory", triplet[2]) --- utils/utils.py

EPI-RTE-removal-parsing

Label: Replacement proposal to deletion

Part of: RT-RTE-game-memory

Endpoints: model replacement response → parse_triplets_removing → exact old-target list → delete_triplets. Owner: helper and ContrieverGraph. SRC-1, utils/utils.py, graphs/contriever_graph.py. This is disposition of proposed removal targets, distinct from the model comparison and new-fact insertion phases of the supplied route.

Concrete implementation deduction: [[apple, is in, room -> nonsense]] passes the arrow/two-side and left-three-field tests and returns the old apple/room/is in target. The parser never validates nonsense or checks it against incoming candidates. If that old triplet exists and both endpoints are not protected locations, the deletion method removes it. This counterexample defeats a local invariant that a justified new replacement must be installed before withdrawal. No target execution is needed.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Parsed old-target list; deletion mutates triplets/cache without a returned acceptance record.
• Later read-back: Remaining facts reach later retrieval; withdrawn facts no longer do through the semantic store, while associated episodic text can survive.
• Delegated visibility: inapplicable — local model response consumption.
• Selection predicate: Arrow pair and three left-side fields; exact existing triplet membership; skip when both endpoints belong to supplied locations.
• Invalidation or expiry: Removes matched facts and their fact embeddings; reset clears local stores; no transactional restoration.
• Activation or effect: Software-enforced withdrawal changes the later candidate pool; correctness/benefit unobserved.
• Evidence limits: Replacement meaning and existence unverified; protected or nonmatching targets survive; original text logged is not an acceptance rationale consumed later.

Proposer: maintenance model. Decider/veto: code membership and location guard, no per-item human approval. Oracle: none. Operating mode: within-game maintenance. Trigger: new facts; guidance: conservative same-aspect duplication/conflict, retaining uncertainty. Proposed change: replace old facts; operative result: delete named old targets and separately add extraction outputs. Recovery: attempt reset only, not semantic rollback. Persistence: altered fact pool, not validated old→new lineage.

splitted_pair = pair.split("->") if len(splitted_pair) != 2: continue first_triplet = splitted_pair[0].split(",") --- utils/utils.py

if triplet[0] in locations and triplet[1] in locations: continue if triplet in self.triplets: self.triplets.remove(triplet) self.triplets_emb.pop(self.str(triplet)) --- graphs/contriever_graph.py

EPI-RTE-answer-metrics

Label: QA answer lexical scoring

Part of: RT-RTE-qa-cycle

Endpoints: answer/reference → direct-answer suffix → lowercasing/punctuation trimming/token lists → set-valued matching-word count and list equality → aggregate ratios/logging. Owner: QA driver. SRC-1, musique_test_big.py. This evaluator part differs from answer generation and paragraph admission in the supplied route. Target/domain: lexical correspondence of direct answer to dataset reference under these exact normalization rules, not reasoning validity or premise truth.

Concrete deduction: reference a a and candidate direct answer a a yield equality true and true_P = 1 while both token lengths are 2. Thus even exact normalized equality does not entail an overlap ratio of one for repeated tokens. This bounds the interpretation of the script's precision/recall/F1 labels. Missing Direct answer: leaves the whole response as the candidate suffix. Empty totals can produce invalid numerical ratios; no guard is implemented.

• implementation conclusion status: wired
• operation conclusion status: observed
• Immediate return: No returned score object; appends counts/equality and logs running aggregates.
• Later read-back: inapplicable — accumulators feed aggregate diagnostics, not later answer/memory requests.
• Delegated visibility: inapplicable — local scoring, not worker feedback.
• Selection predicate: Last direct-answer delimiter, exact normalized token membership/equality; no semantic-equivalence evaluator.
• Invalidation or expiry: Lists initialized per run; no revision/removal of prior metric entries.
• Activation or effect: Metrics affect running diagnostic output only; no successor-selection or memory gate.
• Evidence limits: Observed status is limited to the attributed QA sequence quoted in the supplied route; fresh operation/configuration and aggregate correctness uninspected.

Proposer: answer model; metric evaluator: code; oracle: local task['answer'], consumed only after generation, not given to the answer producer. Human contribution: dataset/configuration, not per-answer adjudication. Mode: bounded evaluation batch. No metric-fed improvement trigger, revision admission, rejection or rollback is wired. Explanation and evidence-truth targets remain outside the lexical check's license.

true_P = len({word for word in answer if word in true_answer}) trueP.append(true_P), pred_len.append(len(answer)), true_len.append(len(true_answer)) EM.append(answer == true_answer) --- musique_test_big.py

EPI-RTE-reference-match

Label: Fixed cooking reference diagnostic

Source-native identity: callable eval_triplets in utils/retriever_search_drafts.py. No counterpart after comparison with supplied RT-RTE-graph-evaluation: this counts fixed-string membership, whereas that route intends before/after environment removal scoring. Endpoints: caller triplet strings → hardcoded reference_full membership list → printed count. Owner: helper function; context/state: cooking reference list and supplied candidate list. SRC-1, utils/retriever_search_drafts.py.

• implementation conclusion status: afforded
• operation conclusion status: uninspected
• Immediate return: Printed Found count; no returned evaluation or mutated graph.
• Later read-back: inapplicable — no installed result or later consumer identified in scoped callers.
• Delegated visibility: inapplicable — local helper diagnostic.
• Selection predicate: Exact string membership for each reference entry; list contains repeated reference strings, so count is entry-weighted rather than unique-fact coverage.
• Invalidation or expiry: Temporary list/count discarded after call; fixed references remain source constants.
• Activation or effect: Diagnostic printing only; no consequential admission or behavioral gate found.
• Evidence limits: No root/helper call found by scoped eval_triplets search; neither actual execution nor completeness/truth of the reference list independently assessed.

Target: membership of reference strings in the caller's triplet list. Evaluator: Python membership; oracle: author-supplied cooking string list. Proposer: external caller supplies candidates; human authored references. Mode: standalone diagnostic, not a feedback optimization loop. Trigger: explicit call; no installed change, rejection or rollback. A concrete deduction: passing an empty list yields all zeros and printed zero coverage, without establishing why facts were absent or whether extraction failed.

is_found = [int(r in triplets) for r in reference_full] print(f"Found {sum(is_found)}/{len(is_found)} from reference_full") --- utils/retriever_search_drafts.py

Claims

EPI-CLM-concrete-extraction

Label: Concrete knowledge and hypothesis marking

Claimed operation: extraction organizes concrete knowledge from input into triplets or thesis sentences, marking assumptions as hypotheses. Source: SRC-1, prompts/prompts.py, doctrine/design. Distinct from RT-CLM-memory-performance: a content-production policy, not a comparative outcome claim. No equivalent supplied claim record after comparison.

Length of your triplet should not be more than 7 words. You should extract only concrete knowledges, any assumptions must be described as hypothesis. --- prompts/prompts.py

EPI-CLM-conservative-replacement

Label: Conservative same-aspect replacement

Claimed operation: preserve distinct/complementary facts and uncertain old facts; replace only redundant/conflicting same-aspect facts. Hypergraph additionally requests an identifiable newer replacement. Source: SRC-1, prompts/prompts.py, doctrine/design. Distinct from RT-CLM-memory-performance; related implementation container is RT-RTE-game-memory, not another claim declaration.

If there is uncertainty about whether a triplet should be replaced, prioritize retaining the existing triplet over replacing it. --- prompts/prompts.py

For every thesis you want to replace you must find replacement from new thesises. If there is no clear replacement, you must not replace existing thesis. --- prompts/prompts.py