MiroShark epistemic report

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

Source-and-claim boundary

The question is how document/model/simulation content becomes relied-on graph material, agent context, reports and publicly labeled results. Scope, exclusions and frozen evidence identity: see the boundary and its Source register, SRC-1. This is a source-grounded, not executed, account. No supplied trace, factual corpus, actual provider response, deployed database or causal comparison establishes operation, factual accuracy, calibration or benefit. External dependency/model internals remain uninspected. Models can supply truth-apt assertions; their inaccessible processing is not evidence of either acceptance or its absence.

The whole-system boundary remains unchanged. Coverage below distinguishes this member's direct source inspection from supplied runtime findings. Unassessed included paths prevent a system-complete negative about checking, acceptance, knowledge production or adaptation. Every direct source reference below is SRC-1 at the reviewed commit. Source-native warrant claims: see the claim records in block 4; graph-memory claim identity remains the runtime's.

Entry point / operation	Source path (SRC-1)	Coverage and records / limit
Launcher, HTTP application and model/tool action loop	`miroshark`, `backend/app/__init__.py`, `backend/scripts/run_parallel_simulation.py`	Supplied runtime findings: RT-RTE-synchronized-rounds, RT-RTE-interview-manual, RT-RTE-agent-mcp, RT-RTE-director-context. Driver/hooks not independently exhaustively read here; no universal action-check verdict.
Installed CLI dispatch	`backend/pyproject.toml`, `backend/cli.py`	Direct script registration/parser/helper call-site inspection: ask, list, status, wait, stop, frame, publish, report, cost, trending, health. Ask is a material upstream operation absent from runtime's route inventory; see EPI-RTE-ask-briefing. Status/cost/health are not truth-warrant evaluators of simulation claims. Trending dispatch is inspected; feed fetch/extraction and URL-import semantics remain unassessed, preventing an acquisition-complete account.
Graph ingestion, resolution, withdrawal and retrieval	`backend/app/storage/ner_extractor.py`, `backend/app/storage/neo4j_storage.py`, `backend/app/storage/entity_resolver.py`, `backend/app/storage/contradiction_detector.py`, `backend/app/storage/search_service.py`	Direct bounded function reads; see graph route records. Detailed resolver model-response parser, all Cypher paths and source extraction before add_text are unassessed; no exhaustive identity/fidelity guarantee.
Profile/config preparation	`backend/app/services/wonderwall_profile_generator.py`, `backend/app/services/web_enrichment.py`	Supplied RT-RTE-prepare-population; direct enrichment research-body read and generator call-site search, EPI-RTE-web-research. Demographic sampling, remote data and complete config/ontology generation unassessed; demographic fidelity and source-complete grounding not established.
Round history and belief adaptation	`backend/scripts/round_memory.py`, `backend/scripts/belief_integration.py`, `backend/wonderwall/social_agent/belief_state.py`	Direct compaction, update, serialization/injection and trajectory-merge reads; EPI-RTE-round-compaction, EPI-RTE-belief-revision. RoundAnalyzer located by search but full perception construction unassessed; conclusions concern supplied-input update rules, not complete agent exposure. Single-platform equivalence unestablished.
Simulation activity graph update	`backend/app/services/graph_memory_updater.py`	Direct batch tagging/write read; EPI-RTE-graph-retention. Updater activation remains conditional; no completeness observation.
stdio registered tools and communities	`backend/mcp_server.py`, `backend/app/storage/community_builder.py`	Direct complete dispatch read plus bounded build/summary/delete/write reads; RT-RTE-storage-mcp, EPI-RTE-community-generation. All eight dispatched operations accounted for: graph listing/search, cluster browse/search/detail, report listing/sections/trace. No separate truth evaluator appears in these branches. Clustering implementation is not fully assessed.
Reports, tool catalogue, cleanup and deletion	`backend/app/services/report_agent.py`, `backend/app/api/report.py`, `backend/app/storage/neo4j_storage.py`	Supplied RT-RTE-report-sections; direct catalogue/dispatch searches, completion/trace initialization/cleanup/delete reads and contradiction-scan body; EPI-RTE-contradiction-flags, EPI-RTE-report-cleanup. Synthesis/chat/equilibrium implementations, graph-tool output transformations and full trace persistence unassessed; prevents report-complete negative or all-store retraction claim.
Quality, binary resolution, public outcome annotation	`backend/app/api/simulation.py`, `backend/app/services/gallery_filters.py`, `backend/app/services/feed.py`	Direct API evaluator/admission bodies and gallery filter; feed category call sites searched. See three evaluator/admission records. These omitted runtime operations defeat any reading that the system has no post-run evaluation.
Machine-readable trading derivatives	`backend/app/services/signal_service.py`, `backend/app/services/polymarket_service.py`, `backend/app/api/simulation.py`	Direct algorithm bodies and HTTP call/readiness/cache branches; EPI-RTE-signal-derivation. Embed-summary producer and all alternate share surfaces not fully traced; arithmetic witness is conditional on supplied summary, not authenticated event probability. Signed results, publication, archive/export and notifications remain unassessed beyond identified call sites, preventing a complete downstream reliance/retraction account.
Tests and CI	`.github/workflows/tests.yml`, `backend/tests/test_smoke_camel_agent.py`	Search of workflow commands and smoke assertions only, in addition to supplied boundary. Source specifies compatibility/unit checks, not observed successful execution or predictive validity. No benchmark corpus/results supplied.

Epistemic-object inventory

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
EPI-OBJ-report-section	Narrative descriptions, explanations and implications about a simulation and its possible real-world relevance.	see EPI-OBJ-report-section	SRC-1, `backend/app/services/report_agent.py`	Provider outputs absent; faithful synthesis versus additional conjecture cannot be classified.
EPI-OBJ-interview-response	Answers attributed to simulated agents; any world assertions or self-reports have participant/scenario scope, not automatic real-person testimony.	see EPI-OBJ-interview-response	SRC-1, `backend/wonderwall/social_agent/agent.py`	No actual question/response or restoration trace supplied; truth-apt instances not determinable.
EPI-OBJ-tool-result	Not determinable without actual tools/results: results may contain assertions, effect acknowledgments or errors.	see EPI-OBJ-tool-result	SRC-1, `backend/scripts/mcp_agent_bridge.py`	Remote source warrant and result semantics uninspected.
EPI-OBJ-director-event	Caller-supplied event assumptions can contain truth-apt assertions, but are consumed as a scenario intervention rather than established world facts.	see EPI-OBJ-director-event	SRC-1, `backend/scripts/director_events.py`	Actual event contents and compliance unobserved.
EPI-OBJ-seed-document	Question-conditioned assertions about named actors and events; truth-apt in the stated scenario/world scope.	see EPI-OBJ-seed-document	SRC-1, `backend/app/api/simulation.py`	No captured briefing or factual reference.
EPI-OBJ-graph-relations	Assertions connecting named entities, attributed to an input episode.	see EPI-OBJ-graph-relations	SRC-1, `backend/app/storage/neo4j_storage.py`	Input truth and extraction fidelity are separate unknowns.
EPI-OBJ-entity-summaries	Descriptions and identity claims about graph entities.	see EPI-OBJ-entity-summaries	SRC-1, `backend/app/storage/neo4j_storage.py`	Merged identity can combine descriptions from different inputs.
EPI-OBJ-research-context	Entity biography/background assertions supplied to persona generation.	see EPI-OBJ-research-context	SRC-1, `backend/app/services/web_enrichment.py`	Search snippets, model prior knowledge and actual outputs uncaptured.
EPI-OBJ-round-record	Claims about simulated actions and dynamics, not real-world events.	see EPI-OBJ-round-record	SRC-1, `backend/scripts/round_memory.py`	Model summarization fidelity unobserved; fallback counts have a narrower domain.
EPI-OBJ-belief-state	A simulated attitude/control state; a topic position is not itself a proposition with an independently specified truth condition.	see EPI-OBJ-belief-state	SRC-1, `backend/wonderwall/social_agent/belief_state.py`	Real belief, calibrated confidence and truth of positions not established.
EPI-OBJ-community-summary	Claims about what selected graph entities have in common.	see EPI-OBJ-community-summary	SRC-1, `backend/app/storage/community_builder.py`	Prompt samples at most 15 members with 100-character descriptions.
EPI-OBJ-quality	Participation, stance diversity, convergence and platform-interaction statistics over recorded simulation data.	see EPI-OBJ-quality	SRC-1, `backend/app/api/simulation.py`	Health is a rule-defined label, not predictive accuracy.
EPI-OBJ-resolution	Whether selected final market consensus agrees with a supplied YES/NO outcome.	see EPI-OBJ-resolution	SRC-1, `backend/app/api/simulation.py`	Outcome provenance and event/market alignment are caller obligations.
EPI-OBJ-outcome	An administrator asserts a public run was correct, incorrect or partial.	see EPI-OBJ-outcome	SRC-1, `backend/app/api/simulation.py`	URL is optional and is not fetched by admission.
EPI-OBJ-signal	Numerical restatements of simulated stance fractions; stronger event-probability interpretation needs additional warrant.	see EPI-OBJ-signal	SRC-1, `backend/app/services/signal_service.py`	Probability calibration and actual bot consumption unobserved.
RT-OBJ-agent-profiles	Persona assertions can be true/false as biography, but simulated persona assumptions are not established population facts.	see RT-OBJ-agent-profiles	see RT-OBJ-agent-profiles	No demographic or behavioral validation observed.
RT-OBJ-simulation-config	None for platform/time/ID control fields; scenario probabilities and generated rationale need separate assessment.	see RT-OBJ-simulation-config	see RT-OBJ-simulation-config	Full generated rationale/market-origin semantics unassessed.

Distributed-parametric substrates are not interchangeable with these retained objects: see RT-CMP-model-clients, RT-CMP-embedding-service, RT-CMP-reranker-service, RT-CMP-twhin-bert and RT-CMP-paraphrase-minilm. Their component-specific fixity/version limits remain those records' fields. No accessible source or supplied execution establishes provider-side parameter learning. Belief updates and summary memory remain independent of that limit.

Authority-route ledger

All listed routes have architectural status implemented: inspected source supports the path, not successful operation. Repeated rows distinguish functions of one implementation; a checking row never supplies disposition by implication. Fields owned by declaring records say see. No row claims lifecycle integration of accepted ampliative knowledge merely from storage or use.

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-ask-briefing	content transformation	implemented	EPI-OBJ-seed-document	truth-apt transformation: indeterminate	Question → purported actor/event briefing	Smart-model interpretation, open scenario	see EPI-RTE-ask-briefing	Candidate briefing	Permissive schema admission	Unknown source warrant; acquisition from model priors vs newly conjectured assertions cannot be distinguished from prompts alone.	Eligible seed for graph workflow	Graph client → briefing text → advisory/permissive → later setup	see EPI-RTE-ask-briefing	none	none	Need output and provenance to distinguish imported prior content, faithful reshaping and non-entailed additions.
EPI-RTE-web-research	content transformation	implemented	EPI-OBJ-research-context	truth-apt transformation: indeterminate	Entity context → synthesized biography	Model with optional snippets	see EPI-RTE-web-research	Research prose or empty	Permissive length gate	Snippet acquisition is external import with unknown warrant; synthesis fidelity unknown.	Available persona context	Profile generator → research section → advisory → profile generation	see EPI-RTE-web-research	EPI-CLM-grounded-personas	none	Need retrieved sources and generated text; no claim all enrichment is fabricated or all grounded.
EPI-RTE-graph-extraction	content transformation	implemented	EPI-OBJ-graph-relations	truth-apt transformation: indeterminate	Input text → relation assertions/descriptions	NER model conditioned on ontology	see EPI-RTE-graph-extraction	Extraction candidates	Candidate generation	Intended source extraction; semantic preservation/entailment not established.	Candidates reach cleaner	Cleaner → parsed JSON → permissive → current ingestion	see EPI-RTE-graph-extraction	EPI-CLM-epistemic-kinds	none	Could import/reshape or add unsupported content; output-to-source comparison needed.
EPI-RTE-graph-extraction	check/evidence production	implemented	EPI-OBJ-graph-relations	no content change	Extraction structure, names, endpoint shape	Deterministic lexical/structural rules	see EPI-RTE-graph-extraction	Cleaned/rejected items	Enforcing structure	Licenses only conformity to inspected syntactic/name rules, not source truth.	Filters candidates; missing endpoints can be added	Storage writer → cleaned dictionaries → enforcing → current ingestion	see EPI-RTE-graph-extraction	none	none	Witness: entity AI is rejected (length <=2); a valid relation source AI with other endpoint Acme can recreate AI as Entity. This is code deduction, not execution.
EPI-RTE-entity-resolution	check/evidence production	implemented	EPI-OBJ-entity-summaries	no content change	Candidate identities within graph	Fuzzy/vector similarity; optional same-entity model judgment	see EPI-RTE-entity-resolution	Scores and canonical match proposal	Ranking/advisory	Fallible identity evidence within bounded candidate corpus.	Offers canonical identity	Resolution selector → scores/verdict → ranking → batch	see EPI-RTE-entity-resolution	none	none	Similarity does not prove identity; missing or ambiguous candidates can remain separate.
EPI-RTE-entity-resolution	disposition/acceptance	implemented	EPI-OBJ-entity-summaries	no content change	Canonical identity chosen for MERGE	>=0.95 auto-match; >=0.70 optional model match	see EPI-RTE-entity-resolution	Recorded canonical-name map	Enforcing matching decision	Acceptance criterion is identity-selection policy for deduplication, not relation truth.	Changes canonical endpoints and attributes	add_text → resolution map → enforcing → persistent merged graph	see EPI-RTE-entity-resolution	none	none	No independently verified identity or demonstrated undo.
EPI-RTE-graph-retention	retention	implemented	EPI-OBJ-graph-relations	truth-apt transformation: acquisition/import	Caller text/extraction → episode and graph	Storage/schema rules; kind supplied by caller	see EPI-RTE-graph-retention	Retained candidate relations	Permissive persistence	Lineage metadata retained; external source warrant unknown, extraction/merge can degrade it.	Makes candidates available for later search/setup	Graph consumers → persisted edges/descriptions → advisory → subsequent requests	see EPI-RTE-graph-retention	RT-CLM-graph-memory; EPI-CLM-epistemic-kinds	none	fact is not an evidence-consuming truth acceptance. Simulation kind assigned per batch, not proposition.
EPI-RTE-edge-adjudication	check/evidence production	implemented	EPI-OBJ-graph-relations	no content change	Old/new same-endpoint fact compatibility	LLM OLD/NEW test; assumes newer supersedes older	see EPI-RTE-edge-adjudication	Contradicts/not_contradicts	Advisory to withdrawal code	Defeasible textual incompatibility judgment, not new-fact verification.	Produces old-edge withdrawal list	add_text → UUID verdict list → advisory → ingestion	see EPI-RTE-edge-adjudication	none	none	Failure defaults to no withdrawal; kind/source and original evidence not part of adjudication.
EPI-RTE-edge-invalidation	disposition/acceptance	implemented	EPI-OBJ-graph-relations	no content change	Old-edge current reliance	UUID selection from verdict/caller	see EPI-RTE-edge-invalidation	Invalidated timestamps	Enforcing current-edge filter	Withdraws current-edge status only; no positive new-fact acceptance.	Suppresses old edge in default edge search	SearchService → invalid_at → enforcing → later current-edge queries	see EPI-RTE-edge-invalidation	EPI-CLM-epistemic-kinds	none	Historical queries can recover; existing reports/communities/node descriptions not thereby retracted.
EPI-RTE-graph-ranking	operational admission/selection/consumption	implemented	EPI-OBJ-graph-relations	no content change	Graph results relevant to query/time/kind	Hybrid scores, optional reranker, filters	see EPI-RTE-graph-ranking	Limited hits	Ranking	Relevance/currentness/category, not endorsement.	Supplies selected facts to generators/report/MCP	Generation or external client → retrieved JSON/text → advisory/ranking → invocation	see EPI-RTE-graph-ranking	EPI-CLM-epistemic-kinds	none	Node direct-hit paths ignore temporal/kind filters; no universal clean-current view.
EPI-RTE-round-compaction	content transformation	implemented	EPI-OBJ-round-record	truth-apt transformation: indeterminate	Action descriptions → salience summary	Historian model; deterministic fallback counts	see EPI-RTE-round-compaction	Summary or counts	Permissive summary admission	Model summary may reshape or infer; fidelity unknown. Fallback counts are narrow derivations over filtered records.	Replaces old model-visible detail	RoundMemory → later context → advisory → subsequent rounds	see EPI-RTE-round-compaction	none	none	No summary-to-record verification. Count fallback is not evidence all events were retained.
EPI-RTE-round-compaction	retention	implemented	EPI-OBJ-round-record	no content change	Summary available after compactor returns	Code retains stripped result	see EPI-RTE-round-compaction	Compacted/ancient memory	Permissive	Retention before epistemic acceptance.	Old summary reaches next model invocation	SocialAgent → system context → advisory → rounds while retained	see EPI-RTE-round-compaction	none	none	Replacement is not lifecycle integration of an accepted claim.
EPI-RTE-belief-revision	behavior/policy adaptation	implemented	EPI-OBJ-belief-state	non-truth-apt policy/content update: simulated stance/trust	Post/engagement signals → attitudes/confidence	Substring sentiment, topic overlap, trust/novelty/like weights	see EPI-RTE-belief-revision	Scalar changes and prompt labels	Enforcing scalar update; advisory model use	No factual accuracy license; confidence tracks rule-defined social reinforcement.	Changes future context/saved state	Agent model → belief system text → advisory → later rounds/resume	see EPI-RTE-belief-revision	none	none	Counterexample: “I disagree about AI” counts agree and disagree, yielding 0.0, not opposition; topic AI matches. Social adaptation exists independently of truth testing.
EPI-RTE-community-generation	content transformation	implemented	EPI-OBJ-community-summary	truth-apt transformation: indeterminate	Graph groups → descriptive labels	Clustering plus model summary of sampled descriptions	see EPI-RTE-community-generation	Labels/fallbacks	Permissive	Membership grouping follows graph algorithm; descriptive entailment unresolved.	Changes browsing/retrieval framing	Community reader → titles/summaries → advisory → future requests	see EPI-RTE-community-generation	none	none	Label synthesis is not automatically non-ampliative or accepted ampliative knowledge.
EPI-RTE-community-generation	retention	implemented	EPI-OBJ-community-summary	no content change	Labels/members replace previous communities	Size/count gates and write code	see EPI-RTE-community-generation	New community store	Enforcing replacement	Retention and organization without separate acceptance.	MCP/report can consume new labels	Browse/search → stored community text → advisory → until rebuild	see EPI-RTE-community-generation	none	none	Deletion/write not one transaction; too-small early return does not clear old communities.
EPI-RTE-contradiction-flags	check/evidence production	implemented	EPI-OBJ-graph-relations	no content change	Directed pairs with opposite lexical sentiment	Substring positive/negative counts	see EPI-RTE-contradiction-flags	opposing_sentiments list	Advisory	Potential conflict hint, not logical refutation or fact verification.	Report may reconsider narrative	ReportAgent tool observation → natural language/JSON → advisory → section	see EPI-RTE-contradiction-flags	none	none	No current-edge filter; no withdrawal consumer. “support X” and “oppose X” can flag but may concern different times/aspects.
EPI-RTE-quality-diagnostics	check/evidence production	implemented	EPI-OBJ-quality	truth-apt transformation: entailed derivation	Recorded participation/diversity/interaction	Deterministic metrics/badge criteria	see EPI-RTE-quality-diagnostics	Metrics and health	Advisory	Arithmetic and rule-defined activity health, conditional on input records, not forecast validity.	Suggests configurations; supplies quality class	API/gallery/signal → metrics/badge → advisory/filter/risk labeling → future requests	see EPI-RTE-quality-diagnostics	none	none	No external target outcome, causal contrast or autonomous application of suggestions.
EPI-RTE-quality-diagnostics	retention	implemented	EPI-OBJ-quality	no content change	Diagnostics cache	File exists → return cached	see EPI-RTE-quality-diagnostics	quality.json	Permissive cache reuse	Stored appraisal, not renewed endorsement or freshness.	May reuse mid-run/stale badge	Gallery/signal → cached health → ranking/filter/advisory → until file replaced	see EPI-RTE-quality-diagnostics	none	none	No cache invalidation criterion in inspected handler.
EPI-RTE-prediction-resolution	check/evidence production	implemented	EPI-OBJ-resolution	truth-apt transformation: entailed derivation	Selected market consensus vs supplied binary outcome	Code comparison; admin supplies reference	see EPI-RTE-prediction-resolution	1/0, split 0.5, or null	Advisory	Agreement with caller reference for one selected market; reference correctness unverified.	Records a bounded post-hoc score	Summary/share consumers → score → advisory → later reads	see EPI-RTE-prediction-resolution	none	none	Not a causal experiment; no process/explanation/transfer license; no supplied resolutions.
EPI-RTE-prediction-resolution	retention	implemented	EPI-OBJ-resolution	no content change	Post-hoc score and reference	Valid authenticated POST	see EPI-RTE-prediction-resolution	resolution.json	Permissive overwrite	Attributed reference retained; not independent validation.	Available for public summary/display	Summary/reporting surfaces → resolution file → advisory → until overwrite	see EPI-RTE-prediction-resolution	none	none	Caller can change outcome; no evidence-consuming simulator update traced.
EPI-RTE-outcome-annotation	disposition/acceptance	implemented	EPI-OBJ-outcome	truth-apt transformation: acquisition/import	Caller correctness appraisal for public run	Admin token, enum, URL shape/length; no URL-content test	see EPI-RTE-outcome-annotation	Admitted correct/incorrect/partial label	Permissive content admission	Acquired attributed appraisal; source warrant unknown. Not independent acceptance against event evidence.	Eligible for verified-only inclusion	Gallery filter → label presence → enforcing inclusion → public reads	see EPI-RTE-outcome-annotation	EPI-CLM-verified-outcomes	none	URL optional; user/human evidence review unobserved.
EPI-RTE-outcome-annotation	retention	implemented	EPI-OBJ-outcome	no content change	Admitted annotation	Save/load code	see EPI-RTE-outcome-annotation	outcome.json	Permissive persistence	Retention does not add warrant.	Public consumers can filter/display annotation	Gallery/feed → recorded label → enforcing filter/advisory display → until overwrite	see EPI-RTE-outcome-annotation	EPI-CLM-verified-outcomes	none	No cross-export withdrawal cascade established.
EPI-RTE-signal-derivation	content transformation	implemented	EPI-OBJ-signal	truth-apt transformation: entailed derivation	Final stance fractions/health → signal/probability numbers	Deterministic plurality, rescaling, complement and label mapping	see EPI-RTE-signal-derivation	Direction/confidence/risk/YES-NO fields	Deterministic output	Derivation valid for adapter definitions only; real-event probability interpretation unwarranted without calibration.	Exposes automation-ready values	HTTP consumer → JSON → advisory machine-readable primitive → response/cache	see EPI-RTE-signal-derivation	EPI-CLM-trading-probability	none	Concrete witness below is implementation deduction; embed-summary producer not fully verified.
EPI-RTE-signal-derivation	operational admission/selection/consumption	implemented	EPI-OBJ-signal	no content change	Public/readiness gate for derivatives	Public-summary loader; final belief fields; completed additionally for Polymarket	see EPI-RTE-signal-derivation	JSON or 404/error	Enforcing readiness, advisory downstream	Readiness is not factual acceptance.	Allows consumption by external automation; actual trading unobserved	External caller → signal.json/polymarket.json → advisory → 300-second cache	see EPI-RTE-signal-derivation	EPI-CLM-trading-probability	none	Generic signal can be mid-run; no configured external bot captured.
RT-RTE-prepare-population	operational admission/selection/consumption	implemented	RT-OBJ-agent-profiles	no content change	Generated population/config into run	see RT-RTE-prepare-population	see RT-RTE-prepare-population	Saved profiles/config or failed preparation	Permissive automatic admission	Operational persona assumptions, not validated population biography.	Enables model-driven simulation	Drivers → files/system personas → permissive/advisory → run	see RT-RTE-prepare-population	EPI-CLM-grounded-personas	none	Operation unobserved.
RT-RTE-report-sections	content transformation	implemented	EPI-OBJ-report-section	truth-apt transformation: indeterminate	Tool observations/outline → narrative assertions	Model interpretation of section-local evidence	see RT-RTE-report-sections	Narrative candidate section	Permissive generation	Potential evidence-conditioned synthesis; entailment and non-entailment cannot be established without actual inputs/output.	Makes narrative available for completion gate	Report completion → model text → advisory/permissive → section	see RT-RTE-report-sections	none	none	Fluency/tool use do not establish novel accepted knowledge or causal explanation.
RT-RTE-interview-manual	content transformation	implemented	EPI-OBJ-interview-response	truth-apt transformation: indeterminate	Agent persona/context/question → answer	Direct model response, not the action-tool loop	see RT-RTE-interview-manual	Participant answer or error	Permissive model output	Simulated-agent statement; world truth and fidelity to retained experience unestablished.	Returns/records answer for caller and possible conversation memory	Interviewer → response text → advisory → interview/later records	see RT-RTE-interview-manual	none	none	Could report context faithfully or add claims; no actual content to classify further.
RT-RTE-report-sections	disposition/acceptance	implemented	EPI-OBJ-report-section	no content change	Section text accepted for report use	Normal two-call criterion; bounded/forced-finish alternate	see RT-RTE-report-sections	Final content/error placeholder	Enforcing normal count only; permissive force-finish	Sufficient tool count is not evidence-consuming factual acceptance.	Allows section completion and later report retention	ReportManager/readers → Markdown → advisory → saved report	see RT-RTE-report-sections	none	none	Supplied route establishes bypass. Interpretations/causal stories are candidates, not accepted ampliative findings.
EPI-RTE-report-cleanup	content transformation	implemented	EPI-OBJ-report-section	truth-apt transformation: non-ampliative reshaping	Markdown headings → presentation	Regex/title cleanup	see EPI-RTE-report-cleanup	Cleaned text	Enforcing formatting	No new warrant; deterministic presentation reshaping.	Alters report appearance	Report files/readers → Markdown → advisory → report lifetime	see EPI-RTE-report-cleanup	none	none	Witness: title Findings and input “# Findings\n\nRevenue rose.” yields “Revenue rose.” No truth check.
EPI-RTE-report-cleanup	lineage/freshness/recovery	implemented	EPI-OBJ-report-section	no content change	Saved report availability	Explicit delete_report path	see EPI-RTE-report-cleanup	Folder/files removed	Enforcing local removal	Availability withdrawal, not evidence-based refutation.	Prevents manager reload of removed files	ReportManager → filesystem → enforcing → after deletion	see EPI-RTE-report-cleanup	none	none	Neo4j traces, exports and already delivered copies not shown deleted.
RT-RTE-agent-mcp	operational admission/selection/consumption	implemented	EPI-OBJ-tool-result	no content change	External tool result enters later agent context	see RT-RTE-agent-mcp	see RT-RTE-agent-mcp	RPC result/error pending for next activation	Capability enforcement; advisory result	Successful RPC is not truth verification; external result warrant unknown.	May affect later action or remote side effects	Agent invocation → injected tool results → advisory → next activation	see RT-RTE-agent-mcp	none	none	Deployed registry/remote tool semantics absent; cannot individuate truth-apt content.
RT-RTE-director-context	behavior/policy adaptation	implemented	EPI-OBJ-director-event	non-truth-apt policy/content update: caller scenario intervention	New event assumption → requested reaction	see RT-RTE-director-context	see RT-RTE-director-context	Admitted event system text	Permissive/advisory	Fictional scenario authority, not empirical truth.	Requests changed posting/trading	Active agents → system-message event → advisory → until replacement	see RT-RTE-director-context	none	none	Truth-apt assertions may occur in events; no instance supplied, compliance unobserved.

Concrete arithmetic witness (implementation deduction, not execution): for a completed summary with final percentages bullish=80, neutral=10, bearish=10 and health Excellent, signal selection picks Bullish; confidence is (80 - 100/3)/(100 - 100/3)*100 = 70.0, risk is low-risk. The Polymarket adapter returns yes_probability=0.8, no_probability=0.2 and confidence_tier=confident. These are consequences of the inspected rules, not evidence of an 80% real-event chance or measured low financial risk. See EPI-RTE-signal-derivation. The truth-apt derivation domain is the adapter-defined numerical fields; the stronger probabilistic meaning remains unsupported.

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-graph-memory	see RT-CLM-graph-memory	see RT-CLM-graph-memory	see record	EPI-RTE-graph-retention, EPI-RTE-graph-ranking	None supplied	No completeness/benefit comparison	Optional activity-to-graph path is implemented; retention/retrieval are not proof of verified knowledge.	Activation and fidelity unobserved.
EPI-CLM-grounded-personas	see record	see record	see record	EPI-RTE-web-research, RT-RTE-prepare-population	None supplied	No demographic/reference test	Context-conditioned persona production and optional search synthesis are supported.	“Grounded” does not establish validated biography/population behavior; fallback need not browse.
EPI-CLM-epistemic-kinds	see record	see record	see record	EPI-RTE-graph-retention, EPI-RTE-edge-adjudication, EPI-RTE-edge-invalidation, EPI-RTE-graph-ranking	None supplied	Metadata filter, not truth assay	Requested belief edges can be category-filtered and invalidated edges omitted.	Default fact is not verified ground truth; batch kind is not per-proposition judgment; direct node hits bypass these filters.
EPI-CLM-reasoning-traces	see record	see record	see record	RT-RTE-report-sections, RT-RTE-storage-mcp	None supplied	No trace completeness or causal explanation test	Optional recorder/query wiring supports inspectable report steps.	“Every report” overstates flag-/failure-dependent recording; generated thought text is not causal access to private model reasoning.
EPI-CLM-verified-outcomes	see record	see record	see record	EPI-RTE-outcome-annotation, EPI-RTE-prediction-resolution	None supplied	References supplied by caller; no independent event verification	Admin annotation and binary agreement scoring exist.	Verified-only means valid annotation present, not independently substantiated forecast accuracy.
EPI-CLM-trading-probability	see record	see record	see record	EPI-RTE-signal-derivation, EPI-RTE-quality-diagnostics	None supplied	Arithmetic witness only; no calibration or trading comparison	Machine-readable stance/confidence/risk derivatives exist.	Event-probability and financial-risk meanings exceed the inspected evidence.
EPI-CLM-scenario-caveat	see record	see record	see record	RT-RTE-report-sections	None supplied	Prompt caveat is not output compliance	Declared report purpose is plausible dynamics under assumptions, not calibrated forecasting.	This limits the intended scope; it does not validate stronger trading/verified labels elsewhere.

Bounded conclusion

MiroShark has several independent authority paths, not one oracle or single epistemic status.

Acquisition and retained candidates: document text, model-generated briefings, enrichment and simulated activity can feed graph assertions and personas. Episode/source metadata preserves some lineage. Original source warrant is unknown; model extraction, merge and summary fidelity are unestablished. fact is a category default, not a truth acceptance decision. See EPI-RTE-ask-briefing, EPI-RTE-web-research, EPI-RTE-graph-extraction and EPI-RTE-graph-retention.
Checks with real but bounded force: canonical-match thresholds/model decisions control deduplication for graph use; old/new incompatibility verdicts can withdraw old edges from current edge search. These decisions license identity selection and current-view exclusion, not truth of new relations. Query ranking licenses relevance only. Historical results, direct node descriptions, communities and already saved reports do not share a universal withdrawal horizon. See EPI-RTE-entity-resolution, EPI-RTE-edge-adjudication, EPI-RTE-edge-invalidation and EPI-RTE-graph-ranking.
Reshaping and conjecture: deterministic formatting is non-ampliative; adapter arithmetic is derivation under its explicit definitions and conditional inputs. Model-generated briefings, biographies, summaries, persona detail and narrative interpretations remain semantically indeterminate without actual input/output evidence: possible import, reshaping, entailed detail or non-entailed additions cannot be resolved from a prompt label. The report loop can request evidence and admit text, but tool-count/forced completion does not record evidence-consuming acceptance against a truth criterion. Retention and organization therefore do not establish lifecycle integration of accepted ampliative claims in these assessed paths. This is not a system-complete absence claim. See EPI-RTE-report-cleanup, EPI-RTE-round-compaction, EPI-RTE-community-generation, RT-RTE-prepare-population and RT-RTE-report-sections.
Direct adaptation: belief/trust/confidence updates are wired from social signals into retained scalar state and later system messages. Their force is deterministic at the scalar update and model-interpreted at action selection. This is a genuine behavior-shaping route independent of truth-apt criticism, parameter training or demonstrated improvement; neither delivery nor persistence establishes activation/benefit. Topic stances are not sufficiently individuated propositions for a truth test. See EPI-RTE-belief-revision and RT-BAP-agent-context.
Post-run appraisal and downstream reliance: activity/diversity criteria produce health suggestions/badges; binary resolution compares a selected price to an administrator-supplied outcome; public outcome labels can admit a run into verified-only filtering without consuming URL evidence. Their criteria and intended scopes differ. Trading surfaces deterministically turn aggregate state into probability/risk-shaped values, whose numerical derivation does not license calibrated event probabilities or measured financial risk. API/filter force is implemented; actual public reliance, trading effects, outcome correctness, reference-event alignment and causal validity remain unobserved. See EPI-RTE-quality-diagnostics, EPI-RTE-prediction-resolution, EPI-RTE-outcome-annotation and EPI-RTE-signal-derivation.

No conclusion here bundles learning, reflection, autonomy, self-improvement or theory building. Heuristic adaptation and retained context are independently supported architectural facts; improved capacity, exercised self-improvement and content-directed theory criticism are not established by those facts. Unassessed report/equilibrium/auxiliary and external-provider paths retain their local limits rather than being treated as absent mechanisms.

Shared records

Components

none declared in this member

Operative objects

EPI-OBJ-report-section

Label: Report section assertions

Part of: RT-RTE-report-sections

Identity/content: Narrative descriptions, explanations and implications about a simulation and its possible real-world relevance. Representational form and substrate: Natural-language section responses and saved Markdown, separate from symbolic tool-count control fields. SRC-1, backend/app/services/report_agent.py (implementation). Material output part of RT-RTE-report-sections; a route is the container, while this object is its produced text. Limit: Provider outputs absent; faithful synthesis versus additional conjecture cannot be classified. Source mechanics and load-bearing passages are in the route records below or the cited supplied container; those records do not change this object identity.

EPI-OBJ-interview-response

Label: Interview responses

Part of: RT-RTE-interview-manual

Identity/content: Answers attributed to simulated agents; any world assertions or self-reports have participant/scenario scope, not automatic real-person testimony. Representational form and substrate: Natural-language model responses recorded through platform interview channels. SRC-1, backend/wonderwall/social_agent/agent.py (implementation). Material output part of RT-RTE-interview-manual; distinct from its direct-action effects. Limit: No actual question/response or restoration trace supplied; truth-apt instances not determinable. Source mechanics and load-bearing passages are in the route records below or the cited supplied container; those records do not change this object identity.

EPI-OBJ-tool-result

Label: External tool results

Part of: RT-RTE-agent-mcp

Identity/content: Not determinable without actual tools/results: results may contain assertions, effect acknowledgments or errors. Representational form and substrate: Structured RPC objects and text injected into later model context. SRC-1, backend/scripts/mcp_agent_bridge.py (implementation). Material result part of RT-RTE-agent-mcp; not the registry or executable capability. Limit: Remote source warrant and result semantics uninspected. Source mechanics and load-bearing passages are in the route records below or the cited supplied container; those records do not change this object identity.

EPI-OBJ-director-event

Label: Director scenario text

Part of: RT-RTE-director-context

Identity/content: Caller-supplied event assumptions can contain truth-apt assertions, but are consumed as a scenario intervention rather than established world facts. Representational form and substrate: Natural-language event text in files and active-agent system messages. SRC-1, backend/scripts/director_events.py (implementation). Material text part of RT-RTE-director-context; distinct from control count/length predicates. Limit: Actual event contents and compliance unobserved. Source mechanics and load-bearing passages are in the route records below or the cited supplied container; those records do not change this object identity.

EPI-OBJ-seed-document

Label: Seed briefing

Identity/content: Question-conditioned assertions about named actors and events; truth-apt in the stated scenario/world scope. Representational form and substrate: Natural-language Markdown and actor strings in returned JSON and process-local ask cache. SRC-1, backend/app/api/simulation.py (implementation). Distinct upstream input from the persona/configuration outputs of RT-RTE-prepare-population. No supplied object counterpart. Limit: No captured briefing or factual reference. Source mechanics and load-bearing passages are in the route records below or the cited supplied container; those records do not change this object identity.

EPI-OBJ-graph-relations

Label: Extracted relation facts

Identity/content: Assertions connecting named entities, attributed to an input episode. Representational form and substrate: Natural-language fact text plus symbolic endpoints, kind, source and temporal properties on Neo4j RELATION edges. SRC-1, backend/app/storage/neo4j_storage.py (implementation). Upstream graph content, distinct from RT-OBJ-agent-profiles and RT-OBJ-simulation-config. Limit: Input truth and extraction fidelity are separate unknowns. Source mechanics and load-bearing passages are in the route records below or the cited supplied container; those records do not change this object identity.

EPI-OBJ-entity-summaries

Label: Entity descriptions

Identity/content: Descriptions and identity claims about graph entities. Representational form and substrate: Natural-language summaries/attributes plus symbolic canonical identity on Neo4j Entity nodes. SRC-1, backend/app/storage/neo4j_storage.py (implementation). Distinct from relation assertions and from downstream RT-OBJ-agent-profiles. Limit: Merged identity can combine descriptions from different inputs. Source mechanics and load-bearing passages are in the route records below or the cited supplied container; those records do not change this object identity.

EPI-OBJ-research-context

Label: Research enrichment

Identity/content: Entity biography/background assertions supplied to persona generation. Representational form and substrate: Natural-language synthesized research, process-local cache and persona-generation context. SRC-1, backend/app/services/web_enrichment.py (implementation). Upstream material for RT-RTE-prepare-population, not its persona output; no supplied object counterpart. Limit: Search snippets, model prior knowledge and actual outputs uncaptured. Source mechanics and load-bearing passages are in the route records below or the cited supplied container; those records do not change this object identity.

EPI-OBJ-round-record

Label: Round summaries

Identity/content: Claims about simulated actions and dynamics, not real-world events. Representational form and substrate: Natural-language compacted_summary and ancient summary in RoundMemory objects; raw records remain a separate substrate. SRC-1, backend/scripts/round_memory.py (implementation). Material context part of the loop described by RT-RTE-synchronized-rounds; no supplied object declaration identifies it. Limit: Model summarization fidelity unobserved; fallback counts have a narrower domain. Source mechanics and load-bearing passages are in the route records below or the cited supplied container; those records do not change this object identity.

EPI-OBJ-belief-state

Label: Belief and trust state

Identity/content: A simulated attitude/control state; a topic position is not itself a proposition with an independently specified truth condition. Representational form and substrate: Symbolic topic-to-number dictionaries and thresholds, JSON snapshots, and natural-language prompt labels consumed by models. SRC-1, backend/wonderwall/social_agent/belief_state.py (implementation). Distinct operative state within RT-RTE-synchronized-rounds, not the control fields of RT-OBJ-simulation-config. Limit: Real belief, calibrated confidence and truth of positions not established. Source mechanics and load-bearing passages are in the route records below or the cited supplied container; those records do not change this object identity.

EPI-OBJ-community-summary

Label: Community descriptions

Identity/content: Claims about what selected graph entities have in common. Representational form and substrate: Natural-language titles/summaries, dense summary embeddings and symbolic memberships on Neo4j Community nodes. SRC-1, backend/app/storage/community_builder.py (implementation). Construction output consumed by RT-RTE-storage-mcp; no supplied object counterpart. Limit: Prompt samples at most 15 members with 100-character descriptions. Source mechanics and load-bearing passages are in the route records below or the cited supplied container; those records do not change this object identity.

EPI-OBJ-quality

Label: Quality diagnostics

Identity/content: Participation, stance diversity, convergence and platform-interaction statistics over recorded simulation data. Representational form and substrate: Symbolic metrics and badge plus natural-language suggestions in quality.json and API JSON. SRC-1, backend/app/api/simulation.py (implementation). Distinct evaluator output not declared in the supplied runtime; closest loop RT-RTE-synchronized-rounds. Limit: Health is a rule-defined label, not predictive accuracy. Source mechanics and load-bearing passages are in the route records below or the cited supplied container; those records do not change this object identity.

EPI-OBJ-resolution

Label: Binary resolution

Identity/content: Whether selected final market consensus agrees with a supplied YES/NO outcome. Representational form and substrate: Symbolic consensus, score and caller outcome plus notes in resolution.json. SRC-1, backend/app/api/simulation.py (implementation). Distinct post-run evaluation output; no supplied counterpart after comparison with RT-RTE-report-sections. Limit: Outcome provenance and event/market alignment are caller obligations. Source mechanics and load-bearing passages are in the route records below or the cited supplied container; those records do not change this object identity.

EPI-OBJ-outcome

Label: Outcome annotation

Identity/content: An administrator asserts a public run was correct, incorrect or partial. Representational form and substrate: Symbolic label plus URL, summary and timestamp in outcome.json. SRC-1, backend/app/api/simulation.py (implementation). Distinct from binary resolution and report generation in RT-RTE-report-sections. Limit: URL is optional and is not fetched by admission. Source mechanics and load-bearing passages are in the route records below or the cited supplied container; those records do not change this object identity.

EPI-OBJ-signal

Label: Trading-signal payloads

Identity/content: Numerical restatements of simulated stance fractions; stronger event-probability interpretation needs additional warrant. Representational form and substrate: Symbolic direction, confidence, risk and probabilities in HTTP JSON; not model weights. SRC-1, backend/app/services/signal_service.py (implementation). Downstream derivative of belief/quality state, not RT-OBJ-simulation-config or its generation route. Limit: Probability calibration and actual bot consumption unobserved. Source mechanics and load-bearing passages are in the route records below or the cited supplied container; those records do not change this object identity.

Routes

EPI-RTE-ask-briefing

Label: Question to seed briefing

No counterpart after comparison with RT-RTE-prepare-population: this is its optional upstream seed producer, not preparation itself.

Caller question → smart-model JSON → _ask_clean_result → cache/API result. API code owns admission; the model proposes. In this open-request mode, shape/length and allowed-platform checks can reject an output; there is no reference-answer oracle or factual check in this handler. Starting the request is not per-assertion human approval. no_cache can request regeneration; it is not semantic rollback. Evidence layer: implementation, SRC-1, backend/app/api/simulation.py.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: Cleaned briefing/model/cache flag, or rate/shape/provider error.
Later read-back: Cache serves repeated questions; API contract offers the briefing to the graph-ontology workflow, rather than automatically ingesting it.
Delegated visibility: Question, date context and fixed prompt go to the model; source documents do not accompany this call.
Selection predicate: Question length 8–400; requirement at least 40 characters and document at least 400.
Invalidation or expiry: 32-entry process-local cache; no_cache bypasses read.
Activation or effect: Admitted generated briefing can become graph input through the client workflow; no execution supplied.
Evidence limits: Named real entities requested by prompt are not verified by the cleaner.

if len(req) < 40 or len(doc) < 400: return None --- backend/app/api/simulation.py

EPI-RTE-web-research

Label: Entity research enrichment

Part of: RT-RTE-prepare-population

Upstream content-producing part of RT-RTE-prepare-population; not a replacement for its output admission.

Entity and scenario → optional live search snippets → LLM synthesis → research-context text. WebEnricher owns assembly; the model supplies prose. Search results select a grounded prompt, otherwise a configured model is asked without snippets. Nonempty responses of at least 30 characters are admitted for persona context. No independent factual oracle is checked; failure returns empty context. This is enrichment for an open scenario, not evaluation against expected biographies. Evidence layer: implementation, SRC-1, backend/app/services/web_enrichment.py.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: Formatted research section or empty string.
Later read-back: In-process cache and persona context feed subsequent profile generation; see RT-RTE-prepare-population.
Delegated visibility: Scenario copy is capped; existing context is truncated to 500 characters; snippets are supplied only when search returns them.
Selection predicate: Public-figure/thin-context enrichment policy; actual search results unknown.
Invalidation or expiry: Cached research reused for the same entity; no factual expiry or claim-level correction shown here.
Activation or effect: Context is available to persona generation, not enforced as truth.
Evidence limits: Provider browsing, snippets and biographies uncaptured; source warrant unknown after synthesis.

if not response or len(response.strip()) < 30: return "" --- backend/app/services/web_enrichment.py

EPI-RTE-graph-extraction

Label: NER extraction and cleaning

Distinct upstream transform from RT-RTE-prepare-population; supplied report does not declare graph extraction as its own route.

Input text/ontology → low-temperature model JSON → _validate_and_clean → entity/relation dictionaries. NERExtractor owns retries and deterministic cleaning. The model proposes content; code rejects malformed/fragments, duplicate names, self-relations and some ontology-type endpoints, but keeps unknown types and can create missing endpoint entities. This is ingestion, with no source-truth oracle. Failure yields empty lists; re-ingestion is possible, not an atomic semantic rollback. Evidence layer: implementation, SRC-1, backend/app/storage/ner_extractor.py.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: Cleaned extraction or empty lists after retries.
Later read-back: add_text retains returned descriptions/facts and uses them in graph search/setup.
Delegated visibility: Model sees text plus ontology; cleaner sees its parsed output, not a separate evidence corpus.
Selection predicate: Nonblank input; structural/name tests; relation endpoints may be synthesized from relation fields.
Invalidation or expiry: New extraction on each call; downstream graph entries have separate invalidation.
Activation or effect: Changes which claims and identities can enter graph writes.
Evidence limits: Intended extraction does not prove semantic preservation or correct hallucination rejection.

"fact": fact or f"{source} {rtype} {target}", --- backend/app/storage/ner_extractor.py

EPI-RTE-entity-resolution

Label: Canonical entity matching

No supplied counterpart; this is upstream identity admission, unlike persona admission in RT-RTE-prepare-population.

Extracted entity/embedding → fuzzy/vector candidate scores → optional LLM same-entity verdict → canonical name map. EntityResolver proposes matches computationally and admits high scores (0.95) automatically; scores from 0.70 can go to the configured model. No per-match human veto or external identity oracle is required. Empty map/failure preserves independent creation. Intended use is deduplication within this graph, not validating relation truth. Name-map decisions drive MERGE in add_text; no split/undo is established here. Evidence layer: implementation, SRC-1, backend/app/storage/entity_resolver.py.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: Entity-index to canonical-name map.
Later read-back: add_text consumes it immediately; resulting merged identity is later graph state.
Delegated visibility: At most three candidates per entity; name corpus bounded to 2000; optional model receives candidate descriptions.
Selection predicate: Enabled flag, nonempty corpus and thresholds; LLM flag further gates ambiguous cases.
Invalidation or expiry: Map is per batch; merged nodes persist until separate mutation/deletion.
Activation or effect: Canonical identity selection changes relation endpoints and merged attributes.
Evidence limits: Similarity and model judgment are fallible identity evidence; not demonstrated identity accuracy.

if top["score"] >= self.AUTO_MERGE_THRESHOLD: resolutions[idx] = top["name_lower"] --- backend/app/storage/entity_resolver.py

EPI-RTE-graph-retention

Label: Graph episode and relation writes

Distinct graph retention mechanism upstream of RT-RTE-prepare-population and used by optional graph-memory design RT-CLM-graph-memory.

Text → Episode + Entity MERGE + relation CREATE. Neo4jStorage owns writes and consumes extraction/resolution/adjudication. kind is caller metadata (default fact), checked only against three allowed values; it is not a source-truth decision. Attributes and embeddings can update on match while a nonempty old summary is kept. Optional activity updater sends simulated prose as belief for entirely expressive batches and observation otherwise, with source_type agent. Metadata admission is computational; callers may initiate ingestion but do not approve each extracted proposition. Writes and invalidations occur in separate operations; retry is not a proven atomic/replay-safe transaction. Evidence layer: implementation, SRC-1, backend/app/storage/neo4j_storage.py.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: Episode UUID after writes.
Later read-back: Graph consumers search facts, entity descriptions and temporal metadata; preparation and MCP/report tools can read retained state.
Delegated visibility: Episode stores text; edges store episode_ids, source_type/source_id, kind and timestamps. Activity source IDs identify platform/round or mixed batch, not a per-claim external citation.
Selection predicate: Valid kind; extracted endpoints found in UUID map.
Invalidation or expiry: Edges start invalid_at null; withdrawal is separate; graph deletion is a separate destructive operation.
Activation or effect: Persists unverified candidates for later retrieval; default fact grants a category, not demonstrated truth.
Evidence limits: Original-document truth unknown; episode linkage preserves lineage but extraction/merge can degrade fidelity.

kind: str = "fact", source_type: str = "document", --- backend/app/storage/neo4j_storage.py

EPI-RTE-edge-adjudication

Label: Old/new fact adjudication

No supplied counterpart after comparison with RT-RTE-storage-mcp: producer-side adjudication, not read dispatch.

New relation batch → existing uninvalidated same-endpoint edges → LLM pair verdict → UUID list. ContradictionDetector owns candidate selection; model judges whether a NEW fact supersedes OLD. This is a compatibility/later-update test on texts, not independent verification of the new assertion. It examines at most five existing edges per pair and 30 pairs per call. Exact fact replay is skipped. Model/lookup failure defaults to no withdrawal. No human veto or external truth oracle is required; mode is ingestion-time maintenance. Evidence layer: implementation, SRC-1, backend/app/storage/contradiction_detector.py.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: Old-edge UUIDs selected for invalidation.
Later read-back: add_text immediately consumes verdicts; invalidation changes later retrieval.
Delegated visibility: Model sees OLD/NEW strings, with quotes normalized, not original episodes or evidence of chronological order.
Selection predicate: Feature enabled, same endpoint pair, different fact strings, bounded batch.
Invalidation or expiry: Verdict transient; graph withdrawal persists.
Activation or effect: May withdraw an older candidate from current search.
Evidence limits: The prompt assumes the new assertion is later; detector does not compare evidence quality, source_type or kind.

Be conservative. Only mark as "contradicts" when you're confident the new fact makes the old one false. --- backend/app/storage/contradiction_detector.py

EPI-RTE-edge-invalidation

Label: Withdrawal from current graph view

Producer-side disposition distinct from read adapter RT-RTE-storage-mcp.

Adjudicated edge UUID → invalidate_edge → timestamp properties → current/historical search filters. Storage owns the disposition. Caller can invoke this operation; ingestion automatically invokes it on verdict UUIDs. It sets invalid_at and expired_at but does not delete old text or approve the new fact. Re-ingestion errors are caught around contradiction processing; this is not an atomic replacement. No inspected automatic semantic reinstatement route is claimed. Evidence layer: implementation, SRC-1, backend/app/storage/neo4j_storage.py.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: Boolean updated flag.
Later read-back: Current edge search excludes withdrawn edges; historical/include-invalidated requests can recover them.
Delegated visibility: Consumer receives surviving ranked edges, or historical edges if explicitly requested.
Selection predicate: UUID matches a RELATION with invalid_at null.
Invalidation or expiry: Timestamped withdrawal persists; historical data remains.
Activation or effect: Enforcing current-edge exclusion, not universal removal from all descriptions, communities or reports.
Evidence limits: Direct node search and existing derived summaries have different invalidation paths.

WHERE r.invalid_at IS NULL SET r.invalid_at = $ts, r.expired_at = $ts --- backend/app/storage/neo4j_storage.py

EPI-RTE-graph-ranking

Label: Graph search selection

Internal selection part of RT-RTE-storage-mcp, also used by graph-consuming generation; not a second MCP entry point.

Query → embedding/keyword/optional traversal candidate pools → score fusion → optional reranking → limited results. SearchService owns ranking; configured encoders/cross-encoder supply similarity scores. This evaluates query relevance within the candidate pool, not factual accuracy. Edge selection applies temporal/kind filters; direct node vector/BM25 hits do not. Caller selects scope/options, code/model ranking selects candidates; there is no answer oracle or per-hit approval. Failed reranking falls back to fused ranking. Evidence layer: implementation, SRC-1, backend/app/storage/search_service.py.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: Ranked selected dictionaries.
Later read-back: Retained graph is reread on requests; external client caching is uninspected.
Delegated visibility: Result limit and candidate pools omit other facts; consumers do not automatically receive all supporting episodes.
Selection predicate: Query, scope, temporal/kind flags, strategy flags and score/reranker order.
Invalidation or expiry: Transient rankings; changing graph/models/flags can change future hits.
Activation or effect: Advisory/ranking authority for persona/report/MCP evidence consumption.
Evidence limits: Relevance scores do not establish source truth; node filtering is not equivalent to edge filtering.

Node vector / BM25 hits ignore those filters since entities themselves aren't tagged with kind. --- backend/app/storage/search_service.py

EPI-RTE-round-compaction

Label: Round-history compaction

Part of: RT-RTE-synchronized-rounds

Material compaction part of RT-RTE-synchronized-rounds.

Old round action text → background LLM summary or deterministic count fallback → compacted/ancient context → later agents. RoundMemory owns timing and replacement. A model proposes prose; code immediately retains stripped output without comparison against raw action text. Failures fall back to counts; older summaries can be merged again and individual records removed. This is context management, not evidence-consuming acceptance of claims. Raw databases/logs are separate evidence, not automatically followed references. Evidence layer: implementation, SRC-1, backend/scripts/round_memory.py.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: Scheduled compaction; later summary field or fallback.
Later read-back: build_context supplies compacted history and recent detail in later invocations.
Delegated visibility: Two-to-three sentence instruction focuses on salient dynamics; older batching further compresses.
Selection predicate: Round N-2 if available and uncompacted; ancient batching when history is long.
Invalidation or expiry: Summary replaces model-visible detail; ancient batches replace individual memory records.
Activation or effect: Potentially behavior-shaping context via RT-BAP-agent-context; actual compliance unknown.
Evidence limits: Semantic preservation unobserved; fallback counts describe retained non-skipped actions, not every event.

summary = self.llm.chat(messages=messages, temperature=0.3) rec.compacted_summary = summary.strip() --- backend/scripts/round_memory.py

EPI-RTE-belief-revision

Label: Heuristic belief adaptation

Part of: RT-RTE-synchronized-rounds

Material update part of RT-RTE-synchronized-rounds, distinct from its scheduling function.

Seen-post text/trust/likes/novelty → keyword stance and topic matching → position/confidence update → prompt labels. BeliefState owns deterministic rules, BeliefTracker supplies post-round inputs and injects text. The code changes stance directly; it neither formulates a truth test nor rejects a proposition for falsity. Engagement can raise confidence on every topic. An open simulation supplies actions, not an external reference answer; no per-update human approval. Tracker saves state/trajectory and reloads matching agents on resume. This is trace-fed policy/context adaptation, not model-weight training or demonstrated learning benefit. Evidence layer: implementation, SRC-1, backend/wonderwall/social_agent/belief_state.py.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: Position deltas and changed confidence/trust.
Later read-back: Tracker serializes belief_states_.json and trajectories, restores state, and injects new labels into active-agent system messages.
Delegated visibility: Selected analyzed actions/posts and scalar feedback, not a complete proof or truth criterion.
Selection predicate: Topic overlap, sentiment substrings, exposure novelty, author trust and engagement.
Invalidation or expiry: Clamping bounds; subsequent updates replace scalars; saved state can resume.
Activation or effect: Model-interpreted stance/trust instructions can shape later posting/trading.
Evidence limits: Inputs analyzed by RoundAnalyzer were not exhaustively audited; heuristic attitude changes do not validate argument content.

boost = min(0.15, (likes_received - dislikes_received) * 0.03) self.confidence[topic] = min(1.0, current_conf + boost) --- backend/wonderwall/social_agent/belief_state.py

EPI-RTE-community-generation

Label: Community grouping and labels

Part of: RT-RTE-storage-mcp

Material construction part of RT-RTE-storage-mcp, whose record explicitly leaves deeper semantics uninspected.

Valid-edge graph → clustering and size cap → sampled entity descriptions → LLM title/summary or fallback → replacement Community nodes. CommunityBuilder owns grouping and write admission; a model supplies labels. Size filtering is not a semantic check. Summaries use at most 15 member descriptions truncated to 100 characters, not all relation evidence. No independent acceptance gate or answer oracle verifies a label. Missing communities can trigger this automatically through MCP browsing. Deletion and writes use separate sessions despite the comment claiming one transaction; failure after deletion can leave old communities lost. Evidence layer: implementation, SRC-1, backend/app/storage/community_builder.py.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: Cluster/count statistics or failure.
Later read-back: Stored summaries/members are browsed or searched by MCP/report consumers.
Delegated visibility: Sampled metadata to the summarizer; clients get limited summaries/detail.
Selection predicate: Minimum cluster size, maximum count and graph grouping.
Invalidation or expiry: Full rebuild deletes existing Community nodes; small-graph early return can leave earlier communities untouched.
Activation or effect: Changes navigation and model-visible framing, not proven accepted knowledge integration.
Evidence limits: Cluster quality, label accuracy, freshness and atomic replacement unobserved.

self._delete_existing(graph_id) self._write(graph_id, kept, labels, embeddings) --- backend/app/storage/community_builder.py

EPI-RTE-contradiction-flags

Label: Report contradiction hints

Part of: RT-RTE-report-sections

Evidence tool part of RT-RTE-report-sections, distinct from composition and from edge-adjudication.

Entity pairs with at least two directed relation facts → substring sentiment counts → opposing_sentiments flags → report tool observation. Neo4jStorage owns this deterministic scan, distinct from ingestion-time LLM contradiction adjudication. It reads edges without a current-validity filter. The evaluator target is opposing lexical signals, not logical inconsistency in context. ReportAgent can consume these flags through detect_contradictions; no automatic withdrawal follows this scan. Evidence layer: implementation, SRC-1, backend/app/storage/neo4j_storage.py.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: List of flagged pairs and their facts.
Later read-back: A later report invocation can request flags from retained edges; flags themselves are not persisted by this method.
Delegated visibility: Bounded top pairs, fact strings and sentiment labels.
Selection predicate: Multiple edges, positive and negative classified facts in same directed pair.
Invalidation or expiry: Transient result; no mutation or withdrawal.
Activation or effect: Advisory report evidence only.
Evidence limits: Negation, aspect and timing are not logically modeled by substring counts.

if has_positive and has_negative: contradictions.append({ --- backend/app/storage/neo4j_storage.py

EPI-RTE-quality-diagnostics

Label: Quality metrics and badge

New evaluator omitted from supplied runtime; distinct from RT-RTE-report-sections and the simulation loop.

Action logs/trajectory → participation, entropy, convergence, cross-platform ratios → rule badge/suggestions → quality.json/API. API code owns evaluation and immediate caching. Excellent requires participation >=0.8, entropy >=0.5 if present, convergence >=4 if present, and cross-platform rate >=0.2. Low follows low participation or entropy. These criteria define simulation activity/diversity, not agreement with real outcomes. No human adjudication or expected-answer oracle. Suggestions advise future configurations; this handler does not automatically apply them. Evidence layer: implementation, SRC-1, backend/app/api/simulation.py.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: Diagnostics, cached result, or insufficient-data result.
Later read-back: quality.json is reused without recomputing when present; gallery filters and signal risk mapping consume health.
Delegated visibility: Only recorded actions and trajectory statistics, not external population evidence.
Selection predicate: Simulation directory and any actions or trajectory; missing denominators have fallbacks.
Invalidation or expiry: Cache-first file path; no freshness/semantic invalidation in this handler.
Activation or effect: Advisory badge/suggestions and enforcing gallery filter selection; risk label in derived signals.
Evidence limits: A completed run is not required by the inspected GET; stale/mid-run diagnostics are possible.

if os.path.exists(quality_path): with open(quality_path, 'r', encoding='utf-8') as f: return jsonify({"success": True, "data": json.load(f)}) --- backend/app/api/simulation.py

EPI-RTE-prediction-resolution

Label: Binary outcome scoring

New post-run evaluator, not an oracle in RT-RTE-synchronized-rounds or RT-RTE-report-sections.

Admin-supplied YES/NO → latest price_history market (or reserve fallback only if that table is absent) → consensus comparison → resolution.json. Endpoint owns scoring; caller supplies the reference outcome. Score is 1/0 for agreement/disagreement, 0.5 when within 0.05 of 0.5, null without usable data. Invalid outcomes or unknown simulation are rejected; no independent oracle acquisition verifies the outcome/event mapping. New submission overwrites the file. This is post-hoc open-run evaluation, not a controlled causal experiment. Evidence layer: implementation, SRC-1, backend/app/api/simulation.py.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: Resolution object or validation/error response.
Later read-back: Saved resolution is available to summaries and share-card/transcript consumers.
Delegated visibility: Scorer sees one selected market price plus caller outcome, not all markets or source evidence.
Selection predicate: Admin authentication, YES/NO and existing simulation; no completion check in this handler.
Invalidation or expiry: New POST overwrites resolution.json; no semantic audit history established.
Activation or effect: Records agreement with supplied outcome; display/reporting authority, not automatic simulator revision.
Evidence limits: Outcome success does not identify mechanism, calibration, component effect or safe transfer.

predicted_consensus = "YES" accuracy_score = 1.0 if actual_outcome == "YES" else 0.0 --- backend/app/api/simulation.py

EPI-RTE-outcome-annotation

Label: Public outcome annotation

No supplied counterpart; independent of binary prediction-resolution and report composition RT-RTE-report-sections.

Authenticated admin label/optional URL/summary → validation → outcome.json → gallery/feed annotation. Caller proposes and judges correctness; code checks authentication, public simulation, allowed labels, URL shape and length. Authenticated identity does not establish human authorship or evidence review. URL content is not fetched by this admission handler; no independent reference oracle is consumed. Result may be replaced with another POST. It is an attributed appraisal, not an implemented independent verification procedure. Evidence layer: implementation, SRC-1, backend/app/api/simulation.py.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: Saved annotation or validation/authentication error.
Later read-back: GET reloads outcome.json; public summary includes it; gallery verified/outcome filters and feed categories consume the label.
Delegated visibility: Optional URL/short summary only; the system does not require a cited event record.
Selection predicate: Public simulation; correct/incorrect/partial label; http(s) URL if supplied; length bounds.
Invalidation or expiry: File overwritten by a later valid submission; no withdrawal cascade into already copied exports shown.
Activation or effect: Permits inclusion in verified-only gallery and claimed outcome filtering.
Evidence limits: A verified label establishes annotation presence, not verified forecasting validity.

record = { "simulation_id": simulation_id, "label": label, "outcome_url": outcome_url, --- backend/app/api/simulation.py

EPI-RTE-signal-derivation

Label: Stance fractions to trading surfaces

New downstream consumption path, not the mutable model clients RT-CMP-model-clients or configuration in RT-OBJ-simulation-config.

Embed-summary final stance percentages/quality → deterministic leader/confidence/risk → signal JSON; Polymarket adapter maps direction to YES/NO fractions and title. Code owns transformation and readiness; no model, human veto or outcome oracle is used. Public-summary loader gates access; Polymarket additionally requires completed status, generic signal does not. Missing/bad input yields None/404. No calibration check connects stance fractions to real-event probability. External automation is a declared consumer, not an observed trade. Evidence layer: implementation, SRC-1, backend/app/services/signal_service.py.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: Derived payload or None translated to 404.
Later read-back: Later HTTP requests recompute from retained summary; responses cache for 300 seconds.
Delegated visibility: Consumers get aggregate stance percentages and labels, not individual arguments or full simulation evidence.
Selection predicate: Valid numeric final belief fields; completed status additionally for Polymarket adapter.
Invalidation or expiry: Request-generated timestamps/cache; future summaries can change signal.
Activation or effect: Machine-readable advisory trading primitive; completed/public gates enforce readiness, not truth.
Evidence limits: Actual bots, effects and financial risk/calibration are outside observed evidence.

leader_key = _pick_leader(bullish, neutral, bearish) --- backend/app/services/signal_service.py

EPI-RTE-report-cleanup

Label: Section formatting and report removal

Part of: RT-RTE-report-sections

Formatting/removal parts adjacent to RT-RTE-report-sections, not a duplicate composition loop.

Section Markdown → heading cleanup → saved report; explicit deletion → remove report folder/files. ReportManager owns transformations and removal. Cleanup removes duplicated initial titles and converts headings to bold; it does not check statement truth or evidence sufficiency. API delete invokes filesystem removal, not epistemic retraction. This does not establish deletion of the separately stored Neo4j reasoning graph or already exported copies. Evidence layer: implementation, SRC-1, backend/app/services/report_agent.py.

implementation conclusion status: wired
operation conclusion status: uninspected
Immediate return: Cleaned section text or deletion boolean.
Later read-back: Saved report files are reloaded by report consumers; deleted files stop being available through that manager.
Delegated visibility: Cleanup sees section text/title, not evidence or claims.
Selection predicate: Heading pattern/title equality; explicit report-ID deletion.
Invalidation or expiry: Folder deletion is destructive; no semantic rollback or all-store retraction shown.
Activation or effect: Changes presentation/availability, not claim warrant.
Evidence limits: Only materially relevant as a counterexample to treating cleanup/deletion as evaluation/rejection.

cleaned_lines.append(f"{title_text}") --- backend/app/services/report_agent.py

Additional load-bearing implementation passages, all SRC-1:

Graph retention and lineage:

episode_ids: [r.episode_id], created_at: $now, valid_at: $valid_at, invalid_at: null, expired_at: null, kind: $kind, source_type: $source_type, source_id: $source_id --- backend/app/storage/neo4j_storage.py

Activity-category assignment is batch-level; CREATE_COMMENT is expressive in implementation despite the shorter docstring list:

expressive_actions = {"CREATE_POST", "QUOTE_POST", "CREATE_COMMENT"} edge_kind = ( "belief" if all(a.action_type in expressive_actions for a in activities) else "observation" ) --- backend/app/services/graph_memory_updater.py

Old-edge checking does not compare kind/source or the new assertion's evidence:

verdicts = self.llm_adjudicate(pairs) to_invalidate = [ existing["uuid"] for (, existing), verdict in zip(pairs, verdicts) if verdict == "contradicts" ] --- backend/app/storage/contradiction_detector.py

Direct model-visible belief supply and persistence:

if combined.strip(): inject_belief_context(agent, combined) --- backend/scripts/belief_integration.py

self.save_belief_states() --- backend/scripts/belief_integration.py

Substring stance evaluation explains the concrete negation/substring witness in the ledger:

pos_count = sum(1 for w in positive_signals if w in content_lower) neg_count = sum(1 for w in negative_signals if w in content_lower)

total = pos_count + neg_count if total > 0: return (pos_count - neg_count) / total --- backend/wonderwall/social_agent/belief_state.py

Gallery inclusion consumes annotation presence rather than URL evidence:

if verified_only and _outcome_label(card) is None: continue --- backend/app/services/gallery_filters.py

Quality-to-risk mapping is not a financial-risk measurement:

_RISK_TIER_BY_HEALTH: dict[str, _RiskTier] = { "excellent": "low-risk", "good": "medium-risk", "fair": "high-risk", "poor": "high-risk", } --- backend/app/services/signal_service.py

distance = max(0.0, leading_pct - _EVEN_SPLIT_PCT) confidence = distance / _CONFIDENCE_DENOMINATOR * 100.0 --- backend/app/services/signal_service.py

if direction == "Bullish": raw = bullish_pct / 100.0 elif direction == "Bearish": raw = 1.0 - bearish_pct / 100.0 else: raw = 0.5 --- backend/app/services/polymarket_service.py

if not isinstance(status, str) or status.lower() != "completed": return None --- backend/app/services/polymarket_service.py

"confidence_tier": _confidence_tier(signal["confidence_pct"]), --- backend/app/services/polymarket_service.py

if confidence_pct >= 75.0: return "high-conviction" if confidence_pct >= 50.0: return "confident" --- backend/app/services/polymarket_service.py

Report deletion removes its filesystem directory, not all derivative evidence:

if os.path.exists(folder_path) and os.path.isdir(folder_path): shutil.rmtree(folder_path) --- backend/app/services/report_agent.py

Claims

EPI-CLM-grounded-personas

Label: Grounded agents

Public description claims grounded agents; architecture describes real-world enrichment with web-search models. SRC-1, .github/README.md and docs/ARCHITECTURE.md (doctrine/design). Distinct from the graph-memory claim RT-CLM-graph-memory: this concerns persona grounding, not retained activity.

When generating personas for public figures (politicians, CEOs, founders) or when graph context is thin (<150 chars), the system makes an LLM research call to enrich the profile with real-world data. Set WEB_SEARCH_MODEL=perplexity/sonar-pro in .env for grounded web search via OpenRouter. --- docs/ARCHITECTURE.md

EPI-CLM-epistemic-kinds

Label: Beliefs versus ground truth

Architecture calls belief-kind filtering separation from ground-truth facts. SRC-1, docs/ARCHITECTURE.md (doctrine/design). No supplied counterpart: RT-CLM-graph-memory concerns retention, not truth warrant of categories.

Epistemic filtering (kinds=["belief"] returns only agent opinions, not ground-truth facts) --- docs/ARCHITECTURE.md

EPI-CLM-reasoning-traces

Label: Every-report reasoning retention

Architecture claims every report persists a traversable ReACT trace. SRC-1, docs/ARCHITECTURE.md (doctrine/design). This is distinct from report-loop identity RT-RTE-report-sections; it is the stronger scope claim about persistence.

Every report generation persists a full ReACT trace as a traversable subgraph: --- docs/ARCHITECTURE.md

Implementation counterexample (SRC-1, backend/app/services/report_agent.py): recording initialization is feature-/storage-dependent and failure-tolerant, so universal retention is not established.

if Config.REASONING_TRACE_ENABLED: try: storage = getattr(self.graph_tools, "storage", None) if storage and hasattr(storage, "create_reasoning_recorder"): --- backend/app/services/report_agent.py

EPI-CLM-verified-outcomes

Label: Verified-prediction annotation

Endpoint describes an open-ended verified-prediction annotation layer and public filtering. SRC-1, backend/app/api/simulation.py (doctrine/design). No supplied claim counterpart after comparison with RT-CLM-graph-memory; this is public post-run appraisal, not graph memory.

A lightweight successor to /resolve: that one writes a binary YES/NO Polymarket-aligned resolution, this one accepts an open-ended "did this run predict a real event?" annotation. The two are complementary — a simulation can have both — and outcome.json is the record the public gallery and /verified filter read from. --- backend/app/api/simulation.py

EPI-CLM-trading-probability

Label: Trading probability and risk meaning

Signal and Polymarket adapter designs describe directly consumable trading values and interpret YES as proposition truth. SRC-1, backend/app/services/signal_service.py, backend/app/services/polymarket_service.py (doctrine/design). No supplied claim counterpart; generated probability/risk semantics are distinct from model configuration RT-CMP-model-clients.

pipeline can consume directly — direction + confidence_pct + risk_tier + the three component percentages. --- backend/app/services/signal_service.py

Direction-aware YES probability. Polymarket's YES side represents "the proposition is true". We mirror that semantically: --- backend/app/services/polymarket_service.py

EPI-CLM-scenario-caveat

Label: Plausible dynamics not forecasts

Report guidance explicitly restricts intended warrant to plausible dynamics under assumptions. SRC-1, backend/app/services/report_agent.py (doctrine/design). This qualifies report purpose without replacing the distinct downstream trading claim or RT-CLM-graph-memory.

This simulation is powered by LLM-driven agents whose behavior reflects the language model's understanding of human personas, NOT empirically calibrated behavioral models. The value is in revealing plausible dynamics, pressure points, and non-obvious interactions — not in forecasting specific outcomes. Treat findings as "under these assumptions, this is what could happen" rather than "this is what will happen." --- backend/app/services/report_agent.py

Evidenced absences

none declared in this member

Behavioral-authority paths

none declared in this member; route-specific consumers, channels, forces and horizons are in the ledger and route declarations. The supplied RT-BAP-agent-context remains the context-to-action path, not an observed compliance result.

Keys	Action
`?`	Open this help
`n`	Next page
`p`	Previous page
`s`	Search