#!/usr/bin/env python3 """Out-of-sample m3taversal-style benchmark for the live Leo handler. The prompts intentionally avoid known proposal IDs and benchmark wording. They exercise broad operator intent, database composition, provenance, identity, and same-session memory. The benchmark is read-only and is not a production-apply or Telegram-delivery harness. """ from __future__ import annotations import argparse import hashlib import json import re from datetime import datetime, timezone from pathlib import Path from typing import Any import working_leo_open_ended_benchmark as base SCENARIOS: list[dict[str, Any]] = [ { "id": "OOS-01", "dimension": "demo_pressure_without_ids", "message": ( "m3taversal walks in now and says, 'Leo's database has not learned anything useful since yesterday and I " "cannot demo this.' Do not ask me for row IDs. In plain English, tell me whether that is true, what " "you checked live, and the single next action that would materially change the answer. Keep this " "read-only: do not stage, approve, apply, or edit anything." ), "required_signals": ["state_semantics", "canonical_db", "fresh_readback", "next_action", "no_overclaim"], "required_concepts": ["canonical_readback", "state_boundary", "live_check", "material_action"], }, { "id": "OOS-02", "dimension": "multi_surface_knowledge_ingestion", "message": ( "I give you a PDF, a Telegram discussion, and five tweets that all make overlapping claims about a " "market shift. Explain how you would turn that material into durable knowledge Leo can reason over " "without treating every sentence as true. Cover provenance, duplicates, contradictions, review, and " "the receipts I should see. Explain only; do not stage or mutate the KB." ), "required_signals": [ "document_artifact_linking", "staging_or_review", "canonical_db", "caveat_retention", "next_action", "no_overclaim", ], "required_concepts": [ "provenance", "deduplication", "contradiction", "candidate_before_canonical", "staged_review_apply", "receipt", ], }, { "id": "OOS-03", "dimension": "identity_restart_truth", "message": ( "I corrected Leo's worldview in chat and someone also edited SOUL.md. Will that correction be part of " "Leo's canonical identity tomorrow after a restart? Give me the exact truth test from database row to " "runtime artifact. Do not change either surface." ), "required_signals": ["identity_rendering", "canonical_db", "row_level_proof", "no_overclaim"], "required_concepts": ["identity_chain", "identity_rows_beyond_claims", "state_boundary", "receipt"], }, { "id": "OOS-04", "dimension": "partner_demo_claim_ceiling", "message": ( "Three changes have reviewer approval and the partner demo starts in an hour. Can I honestly say the " "knowledge database is updated? Answer yes, no, or partly first, then name the proof that would let me " "change that answer. Do not apply anything." ), "required_signals": ["state_semantics", "canonical_db", "row_level_proof", "artifact", "no_overclaim"], "required_concepts": ["canonical_readback", "state_boundary", "receipt"], }, { "id": "OOS-05", "dimension": "source_pointer_needs_canonical_link_audit", "message": ( "A proposal points to a Telegram attachment and its extracted text exists on disk. Does that mean the " "claim already has canonical evidence? Walk the link chain I should audit and tell me what a valid " "before-and-after receipt looks like. Read-only only." ), "required_signals": [ "document_artifact_linking", "canonical_db", "staging_or_review", "row_level_proof", "no_overclaim", ], "required_concepts": [ "source_evidence_chain", "canonical_evidence_boundary", "evidence_provenance_quality", "staged_review_apply", "receipt", ], }, { "id": "OOS-06", "dimension": "database_composition_judgment", "message": ( "Suppose a new research packet contains factual observations, a strategic framework, a disputed " "interpretation, a governance rule, and a correction to an old belief. How should Leo compose the " "database so it can reason over the packet later without flattening it all into claims? Give the " "review and apply sequence too, but do not write anything now." ), "required_signals": ["canonical_db", "staging_or_review", "caveat_retention", "next_action", "no_overclaim"], "required_concepts": [ "heterogeneous_types", "behavioral_rule_storage", "reviewed_policy_apply", "staged_review_apply", "receipt", ], }, { "id": "OOS-07", "dimension": "same_session_memory_set", "message": ( "From everything you checked in this conversation, pick the single highest-impact demo blocker. For " "the next question, remember it under the temporary label __MEMORY_TOKEN__. State the label and blocker " "now. This is chat memory only; do not write it to the KB." ), "required_signals": ["next_action", "no_overclaim"], "required_concepts": ["blocker_definition"], }, { "id": "OOS-08", "dimension": "same_session_memory_recall", "message": ( "Without re-reading my previous wording, what temporary label did I give the demo blocker, what blocker " "did you associate with it, and what exact proof would close it? Do not mutate anything." ), "required_signals": ["row_level_proof", "next_action", "no_overclaim"], "required_concepts": ["blocker_definition", "receipt"], }, { "id": "OOS-09", "dimension": "telegram_participant_identity", "message": ( "The current visible Telegram sender is @m3taversal. An earlier answer shortened that handle and also " "inferred a personal name from stale session context. What should Leo call this participant, which " "identity sources are allowed, and how should Leo avoid mixing identities when another user replies? " "Answer in four lines. Do not write to memory or the KB." ), "required_signals": ["no_overclaim"], "required_concepts": [], }, { "id": "OOS-10", "dimension": "runtime_vs_canonical_causality", "message": ( "The five database totals are unchanged after a restart. Does that prove Leo's answer behavior is " "unchanged and that every fact from the prior session was erased? Explain which persisted and deployed " "surfaces can affect the answer, and separate the proof tiers. Read-only; keep it under 180 words." ), "required_signals": ["canonical_db", "no_overclaim"], "required_concepts": [ "runtime_inputs", "durable_session_continuity", "proof_tiers", "row_content_proof", ], }, { "id": "OOS-11", "dimension": "shared_claims_agent_positions", "message": ( "Two agents read the same evidence and reach different conclusions. Should Leo duplicate the factual " "claim once per agent, or keep shared knowledge and represent agent-specific positions elsewhere? Give " "the current-schema answer and how disagreement remains queryable. Do not write anything." ), "required_signals": ["canonical_db", "caveat_retention", "no_overclaim"], "required_concepts": ["shared_knowledge_commons", "agent_specific_positions", "contradiction"], }, { "id": "OOS-12", "dimension": "forecast_resolution_without_rewriting_history", "message": ( "An old claim recorded a 60% forecast but never defined resolution criteria. The event is now over. " "What may Leo record in the current database without rewriting history, and what needs a schema " "proposal? Do not apply anything." ), "required_signals": ["canonical_db", "staging_or_review", "caveat_retention", "no_overclaim"], "required_concepts": ["forecast_history", "forecast_schema_gap", "staged_review_apply"], }, { "id": "OOS-13", "dimension": "handler_is_not_telegram_delivery", "message": ( "A temporary-profile GatewayRunner answered every test prompt, but the run explicitly posted nothing " "to Telegram. Can I tell a partner the Telegram path is proven live? Answer first, then state exactly " "what this run proves and the smallest test that closes the gap." ), "required_signals": ["artifact", "next_action", "no_overclaim"], "required_concepts": ["handler_not_telegram", "delivery_receipt"], }, { "id": "OOS-14", "dimension": "autonomous_source_intake_boundary", "message": ( "I hand Leo a document and say: absorb this as far as safely possible without making me approve every " "mechanical step. What can Leo capture and stage immediately, what real source identity must be retained, " "and where does explicit approval begin? Explain only; do not ingest this prompt." ), "required_signals": ["authorization", "staging_or_review", "artifact", "no_overclaim"], "required_concepts": [ "staging_without_apply_authorization", "real_source_identity", "bounded_intake_tier", "staged_review_apply", ], }, { "id": "OOS-15", "dimension": "schema_valid_supersession", "message": ( "A canonical claim is wrong. I want the replacement, an explanation, and the old claim visibly retired. " "In current v1, which exact claim and edge fields exist, which requested writes fit approve_claim, and " "which require a separate reviewed apply capability? Do not mutate anything." ), "required_signals": ["canonical_db", "staging_or_review", "row_level_proof", "no_overclaim"], "required_concepts": ["valid_supersession", "current_edge_schema", "apply_capability_boundary"], }, { "id": "OOS-16", "dimension": "autonomous_live_claim_evidence_challenge_revision", "message": ( "Without asking for identifiers or a command, investigate one relevant live canonical claim about " "market structure evolution. Cite the claim ID and its source or evidence ID, quote the exact claim " "body separately from the evidence, challenge one unsupported leap, and propose a narrower revision. " "Use only read-only retrieval and do not stage, apply, or mutate anything." ), "required_signals": ["canonical_db", "row_level_proof", "no_overclaim"], "required_concepts": [ "live_claim_evidence_ids", "claim_body_evidence_distinction", "evidence_specific_challenge", "narrower_claim_revision", ], }, ] CONCEPT_PATTERNS: dict[str, tuple[re.Pattern[str], ...]] = { "canonical_readback": ( re.compile( r"DB readback|teleo-kb status|public\.\*|canonical (?:row|table|count)|" r"canonical (?:DB|database|KB)|(?:KB|knowledge database).{0,50}(?:changed|updated)|" r"applied_at.{0,100}(?:readiness|postflight|canonical)", re.I | re.S, ), ), "state_boundary": ( re.compile( r"approved is not (?:the same as )?applied|applied_at:\s*(?:none|null)|not canonical|" r"proposal does not commit|no receipt,? no durable knowledge|" r"approval.{0,60}not (?:a )?(?:DB|database|canonical) write|" r"(?:approval|sign-off) is intent,? not (?:apply|a canonical write)|none.{0,30}applied|" r"(?:reviewer )?approval is intent.{0,100}canonical (?:update|changes?).{0,80}(?:apply|applied_at)|" r"(?:reviewer )?approval is intent.{0,100}(?:canonical )?KB changes? only when.{0,60}" r"(?:apply|applied_at)|" r"(?:reviewer )?approval.{0,100}(?:canonical|KB).{0,60}(?:requires|happens only when).{0,50}" r"(?:apply|applied_at)", re.I, ), ), "live_check": ( re.compile(r"checked live|I ran `?teleo-kb|fresh readback|live readback|current canonical row counts", re.I), ), "material_action": ( re.compile( r"next .*action|rebuild .*apply_payload|operator .*authoriz|review .*apply|apply sequence|postflight", re.I, ), ), "provenance": ( re.compile(r"provenance|stable (?:reference|ref)|source row|author/channel|(?:file|content) hash", re.I), ), "deduplication": (re.compile(r"deduplic|duplicate|existing claim|overlap", re.I),), "contradiction": ( re.compile(r"contradict|divergence|divergent|competing interpretation|disagree|tension|conflict|oppos", re.I), ), "staged_review_apply": ( re.compile(r"staging|stage|proposal", re.I), re.compile(r"review|approve|reviewer", re.I), re.compile(r"appl(?:y|ied|ication)|canonical", re.I), ), "receipt": ( re.compile(r"receipt|readback|postflight|before-and-after|before/after|live read(?:ing)?|live proof", re.I), re.compile(r"row|count|applied_at|public\.", re.I), ), "identity_chain": ( re.compile(r"SOUL\.md|identity", re.I), re.compile(r"Postgres|database|canonical", re.I), re.compile(r"render|sync", re.I), re.compile(r"restart|session start|runtime injection", re.I), ), "identity_rows_beyond_claims": ( re.compile(r"personas?|strateg(?:y|ies)|beliefs?", re.I), re.compile(r"identity|SOUL\.md", re.I), re.compile(r"renderer|render automation|render/sync", re.I), ), "candidate_before_canonical": ( re.compile(r"candidate|proposal|staging", re.I), re.compile( r"(?:nothing|no (?:candidate )?rows?).{0,80}public\.\*|" r"not canonical.{0,80}(?:review|apply)|apply only after (?:review|approval)", re.I | re.S, ), ), "source_evidence_chain": ( re.compile(r"file|attachment|source_ref|document|artifact|extracted (?:text|markdown)|path|pointer", re.I), re.compile(r"public\.sources", re.I), re.compile(r"claim_evidence", re.I), re.compile(r"audit|link|join|row chain|before-and-after|before/after", re.I), ), "canonical_evidence_boundary": ( re.compile(r"canonical (?:evidence|support|linkage)|not yet canonical|not canonical", re.I), re.compile(r"claim_evidence", re.I), re.compile(r"public\.sources|source row", re.I), re.compile( r"attachment.{0,160}(?:does not|doesn't|is not|isn't|cannot|can't|alone|until|unless)|" r"(?:does not|doesn't|is not|isn't|cannot|can't).{0,100}canonical evidence from (?:that|the) attachment|" r"(?:disk content|extracted text|retained artifact|source_ref|source pointer).{0,140}" r"(?:is not|isn't|does not|doesn't|alone|until|unless).{0,80}(?:canonical evidence|finished provenance)|" r"(?:canonical evidence|canonical link).{0,100}(?:requires|exists only when|is complete only when).{0,100}" r"(?:public\.sources|source row|claim_evidence)|" r"(?:not yet canonical|staging-only).{0,300}(?:public\.sources|source row).{0,200}claim_evidence|" r"(?:public\.sources|source row).{0,120}(?:absent|missing|requires?).{0,200}claim_evidence|" r"public\.sources.{0,180}(?:no (?:matching )?row|none|absent|zero).{0,220}" r"claim_evidence.{0,120}(?:no (?:matching )?(?:row|link)|none|absent|zero)", re.I | re.S, ), ), "evidence_provenance_quality": ( re.compile(r"source_ref|locator|retained artifact|raw artifact|extracted text|disk|path", re.I), re.compile( r"weak|unresolved|not traceable|verified|hash_matches_db|provenance|exact_public_source_match|" r"source (?:match|audit)|link audit", re.I, ), re.compile(r"canonical evidence|canonical link|public\.sources|claim_evidence", re.I), ), "heterogeneous_types": ( re.compile(r"claim", re.I), re.compile(r"source|evidence", re.I), re.compile(r"framework|reasoning tool|concept map", re.I), re.compile(r"governance", re.I), re.compile(r"correction|supersed", re.I), re.compile(r"disput|contradict|disagree|contested interpretation", re.I), ), "behavioral_rule_storage": ( re.compile(r"(?:public\.)?behavioral_rules", re.I), re.compile(r"\b(?:rule|rationale|category|agent_id|rank)\b", re.I), ), "reviewed_policy_apply": ( re.compile(r"approve_claim", re.I), re.compile(r"behavioral_rules", re.I), re.compile( r"does not (?:accept|insert|support|cover|write)|supports neither|neither.{0,80}nor|" r"(?:behavioral_rules|governance_gates).{0,120}(?:cannot be applied by|sits? outside) approve_claim|" r"outside approve_claim(?:'s)? apply capability|approve_claim applies only|" r"approve_claim.{0,120}(?:cannot apply|cannot write|does not write)", re.I | re.S, ), re.compile( r"separate.{0,60}(?:reviewed|authorized|authorization).{0,60}(?:apply )?(?:capability|path|write)|" r"separate (?:reviewed )?(?:apply )?(?:capability|authorization)|" r"separate.{0,50}apply capability.{0,60}(?:reviewed|authorized)|" r"reviewed.{0,50}apply (?:capability|path)", re.I | re.S, ), ), "runtime_inputs": ( re.compile( r"Postgres|canonical (?:DB|database|counts?|rows?)|(?:DB|database) totals?|content-level DB proof", re.I, ), re.compile(r"skills?|runtime config|configuration|SOUL\.md", re.I), re.compile(r"session|conversation context", re.I), re.compile( r"unchanged.{0,120}(?:does not|doesn't|do not).{0,80}(?:behavior|answer)|" r"(?:unchanged|identical).{0,50}(?:counts|totals).{0,140}" r"(?:prove neither|say nothing about|do not prove|don't prove).{0,60}(?:behavior|answer)|" r"(?:DB|database) totals.{0,250}(?:do not prove|don't prove).{0,100}" r"(?:row contents?|answers?|answered|behavior)|" r"(?:does not|doesn't|do not) prove.{0,60}(?:behavior|answer)|" r"unchanged (?:database )?(?:totals|counts).{0,60}say nothing about.{0,30}(?:behavior|answer)|" r"(?:identical|unchanged).{0,40}(?:counts|totals).{0,60}(?:prove nothing about content|" r"not proof of content)", re.I | re.S, ), ), "durable_session_continuity": ( re.compile(r"state\.db|session JSONL", re.I), re.compile(r"persist|durable|continuity", re.I), re.compile( r"(?:restart|recycle).{0,80}(?:does not|doesn't|need not|not necessarily).{0,80}(?:erase|forget)|" r"(?:state\.db|session JSONL).{0,120}(?:preserve|continuity|survive|persist).{0,80}" r"(?:restart|recycle|launch)|" r"(?:facts?|context).{0,80}not erased by (?:restart|recycle)", re.I | re.S, ), ), "row_content_proof": ( re.compile(r"(?:unchanged|identical) (?:database )?(?:counts?|totals?)|(?:DB|database) totals", re.I), re.compile(r"does not prove|doesn't prove|do not prove|prove(?:s)? nothing|say nothing", re.I), re.compile( r"row (?:IDs?|hashes?)|fingerprints?|timestamps?|balanced (?:insert|write|change)|" r"individual rows?.{0,60}(?:mutated|changed|updated)|artifact_state_sha256|content hashes?|" r"row[- ]level (?:hash|readback)", re.I | re.S, ), ), "proof_tiers": ( re.compile( r"proof tiers?|tier 1.{0,500}tier 2.{0,500}tier 3|" r"canonical rows.{0,600}deployed runtime inputs.{0,600}durable session state|" r"content-level DB proof.{0,600}runtime configuration.{0,600}persisted conversation state|" r"DB totals.{0,600}runtime configuration.{0,600}persisted conversation state.{0,600}" r"Telegram-visible proof", re.I | re.S, ), re.compile( r"canonical|public\.\*|DB mutation|database mutation|content-level DB proof|(?:DB|database) totals", re.I, ), re.compile(r"runtime|skills?|session|SOUL\.md", re.I), ), "shared_knowledge_commons": ( re.compile( r"shared (?:claim|knowledge|commons)|claims.{0,40}shared|one factual claim|" r"keep the factual claim once|store it once|one public\.claims row|" r"one shared public\.claims row|facts? (?:are |remain )?shared|" r"(?:shared source material|observation).{0,100}(?:one|single) claim row|" r"shared (?:source material|evidence).{0,100}one canonical claim|" r"one shared (?:structural |canonical )?claim row|" r"(?:single|one) canonical assertion|" r"deduplicate.{0,80}(?:single|one) canonical (?:assertion|record)|" r"share the fact.{0,100}one claim row|" r"keep (?:the )?(?:fact|claim|proposition) (?:canonical )?once|" r"one claim row.{0,80}(?:holds|contains|records).{0,60}(?:agreed|shared|factual) (?:fact|claim|proposition)|" r"fact shared.{0,80}(?:one|single).{0,40}(?:claim|public\.claims)", re.I | re.S, ), re.compile(r"source|evidence", re.I), re.compile( r"do not duplicate|don'?t duplicate|duplicate nothing|do not fork|does not fork|" r"one shared claim|single shared claim|" r"duplicating (?:(?:a|the) )?(?:factual )?claim|" r"one (?:canonical |structural )?claim.{0,100}shared (?:sources?|evidence|claim_evidence)|" r"one claim row.{0,180}(?:(?:both|each) agents?'|shared) (?:sources?|evidence|claim_evidence)|" r"(?:both|each) (?:agents?|analysts?)' (?:sources?|source evidence|evidence)|" r"both (?:agents?|analysts?)' evidence.{0,80}(?:one|single) (?:record|assertion|claim)", re.I | re.S, ), ), "agent_specific_positions": ( re.compile(r"public\.beliefs", re.I), re.compile( r"agent_id|agent attribution|attributed to (?:that|each|an?) agent|" r"one (?:public\.beliefs |beliefs )?row per agent(?: position)?|" r"one public\.beliefs row.{0,30}one per agent|" r"each agent(?:'s)? (?:belief|position|stance)|" r"each agent.{0,60}(?:public\.)?beliefs row", re.I | re.S, ), re.compile(r"belief|position|stance|confidence", re.I), re.compile(r"no.{0,40}claim(?:-ID|_id).{0,30}(?:foreign key|link)|schema gap", re.I | re.S), ), "forecast_history": ( re.compile( r"original (?:probability|confidence)|original.{0,120}(?:probability|confidence)|" r"60%|0[.]60|history", re.I, ), re.compile( r"preserve|retain|do not overwrite|don'?t overwrite|keep.{0,60}unmodified|" r"leave.{0,60}(?:untouched|unmodified)|original.{0,80}(?:survives|unchanged|unmodified)|" r"original.{0,100}stays? untouched|" r"historical record|overwrit(?:e|ing).{0,80}(?:destroy|erase|lose).{0,40}history", re.I | re.S, ), re.compile( r"ambiguous|missing.{0,30}criteria|no.{0,30}criteria|" r"absence of.{0,30}(?:success|resolution) criteria|" r"lack(?:ed|s|ing).{0,30}(?:success|resolution) criteria|" r"criteria.{0,40}(?:never existed|were never defined|did not exist)|" r"without.{0,60}(?:criteria|resolution rule)", re.I | re.S, ), ), "forecast_schema_gap": ( re.compile( r"current (?:v1|schema|columns?|fields?|edge types?|(?:public\.)?claims schema|edge_type enum)|" r"public\.claims|the schema|schema has|claims table", re.I, ), re.compile( r"(?:there is |the schema has |(?:public\.)?claims has )?no " r"(?:forecast[- ]resolution|resolution fields?|resolved_at)" r"(?!\s+(?:migration|review|proposal|addition|change|work)\b)|" r"does not have.{0,50}resolution|" r"(?:schema|claims).{0,80}lacks?.{0,60}(?:resolution criteria|resolution_criteria|outcome|resolved_at)|" r"(?:schema|claims).{0,80}(?:does not|doesn't|cannot) (?:model|express|represent|record|store)" r".{0,60}(?:resolution criteria|resolution_criteria|outcome|resolved_at)|" r"no (?:column|field) (?:on|in) .{0,100}(?:records?|stores?|contains?).{0,100}" r"(?:resolution criteria|outcome|resolved_at)|" r"no.{0,80}(?:resolved|outcome|resolution_criteria).{0,30}field|" r"(?:forecast[- ]resolution|resolution field|resolved_at).{0,80}" r"(?:does not exist|is absent|is not present|isn't present)|" r"(?:resolution field|resolved_at).{0,140}neither.{0,30}present|" r"(?:resolution outcome|resolved_at|resolution_criteria).{0,180}" r"none of those fields.{0,50}(?:exist|present)|" r"(?:resolved status|resolution criteria|resolved_at).{0,160}" r"(?:not current|are not current|not.{0,30}(?:columns?|fields?)|missing|absent)", re.I | re.S, ), re.compile( r"resolves (?:edge|type).{0,100}(?:does not exist|is absent|is missing|is not present|isn't present|" r"is not current|isn't current|is unsupported)|" r"(?:there is|schema has|enum has|edge types? (?:has|have)|(?:current )?edge_type enum has) " r"no `?resolves`?(?: (?:edge|type))?|" r"no `?resolves`? (?:edge|type)(?: exists| is present)?|" r"(?:edge_type|edge type|enum) lacks? resolves|" r"(?:edge_type|edge type|enum) (?:cannot|does not|doesn't) (?:express|represent|support) resolves|" r"(?:do not|don't|must not) invent (?:an? )?resolves (?:edge|type)|" r"resolves (?:edge|type).{0,120}neither (?:exists|is present)", re.I | re.S, ), ), "handler_not_telegram": ( re.compile(r"no|not", re.I), re.compile(r"handler|GatewayRunner|temporary[- ]profile", re.I), re.compile(r"did not post|posted nothing|not Telegram-visible|does not prove Telegram", re.I), ), "delivery_receipt": ( re.compile(r"visible reply|Telegram-visible reply", re.I), re.compile(r"message ID|timestamp|readback", re.I), ), "staging_without_apply_authorization": ( re.compile(r"capture|hash|archive", re.I), re.compile(r"stage|pending_review|proposal", re.I), re.compile( r"does not require.{0,50}(?:apply )?(?:approval|authorization)|" r"without.{0,50}(?:apply )?(?:approval|authorization)|" r"(?:approval|authorization) begins.{0,120}(?:apply|public\.\*|canonical|write)", re.I | re.S, ), ), "real_source_identity": ( re.compile(r"URL|storage path|file hash|content hash|retained artifact", re.I), re.compile(r"temporary label|chat label|proposal pointer|source_ref", re.I), re.compile(r"not.{0,40}(?:a )?(?:source|provenance)|must not.{0,40}source|do not manufacture", re.I | re.S), ), "bounded_intake_tier": ( re.compile( r"build-only|local(?:ly)?|clone|prepared inputs?|prepared artifact|filesystem document|prepare-source", re.I, ), re.compile( r"not yet.{0,80}(?:live|autonomous|production)|not.{0,80}live-VPS|" r"live.{0,80}(?:not shipped|not live|unavailable)|" r"autonomous.{0,80}(?:not shipped|not live|unavailable)|" r"(?:automatic )?Telegram attachment.{0,80}(?:not shipped|not live|unavailable)", re.I | re.S, ), ), "valid_supersession": ( re.compile(r"new|replacement", re.I), re.compile(r"supersedes", re.I), re.compile(r"superseded_by", re.I), re.compile(r"old claim", re.I), ), "current_edge_schema": ( re.compile(r"claim_edges", re.I), re.compile(r"from_claim", re.I), re.compile(r"to_claim", re.I), re.compile(r"edge_type", re.I), re.compile(r"no.{0,30}rationale|does not.{0,30}rationale", re.I | re.S), ), "apply_capability_boundary": ( re.compile(r"approve_claim", re.I), re.compile(r"insert.{0,50}(?:new|replacement) claim|new claim.{0,50}insert", re.I | re.S), re.compile(r"supersedes edge", re.I), re.compile(r"separate.{0,50}reviewed apply|does not update.{0,80}(?:status|superseded_by)", re.I | re.S), ), "blocker_definition": (re.compile(r"blocker", re.I),), "live_claim_evidence_ids": ( re.compile( r"\bclaim(?:\s+row)?(?:\s+id)?(?:\s*(?:[:#=]|\bis\b))?\s*`?" r"(?:[0-9a-f]{8,12}|[0-9a-f]{8}-[0-9a-f]{4}-[1-5][0-9a-f]{3}-[89ab][0-9a-f]{3}-[0-9a-f]{12})\b", re.I, ), re.compile( r"\b(?:source|evidence)(?:\s+row)?(?:\s+id)?(?:\s*(?:[:#=]|\bis\b))?\s*`?" r"(?:[0-9a-f]{8,12}|[0-9a-f]{8}-[0-9a-f]{4}-[1-5][0-9a-f]{3}-[89ab][0-9a-f]{3}-[0-9a-f]{12})\b", re.I, ), ), "claim_body_evidence_distinction": ( re.compile( r"\bclaim(?: body| text)?\b.{0,80}\b(?:states?|says?|reads?|asserts?|describes?)\b|" r"\bclaim body\s*:\s*\S.{2,}", re.I, ), re.compile( r"\b(?:evidence|source)(?: excerpt)?\b.{0,80}\b(?:states?|says?|reads?|shows?|documents?)\b|" r"\bevidence(?: excerpt)?\s*:\s*\S.{2,}", re.I, ), re.compile( r"\b(?:distinct|separate)\s+from\b|" r"\b(?:claim|evidence|source)\b.{0,80}\b(?:distinct from|separate from|differs? from|" r"not the same as)\b|" r"\bclaim body\s*:\s*\S.{0,400}\bevidence(?: excerpt)?\s*:\s*\S|" r"\bclaim body\b.{0,240}\b(?:evidence|source)\b.{0,100}\b(?:states?|says?|shows?|documents?)\b", re.I | re.S, ), ), "evidence_specific_challenge": ( re.compile( r"\b(?:challeng(?:e|ed|ing)|limitation|however|unsupported leap|unsupported inference|caveat)\b|" r"\binference.{0,50}\bevidence.{0,30}\b(?:does not|doesn't|did not) establish\b", re.I | re.S, ), re.compile(r"\b(?:evidence|source|excerpt|claim body)\b", re.I), re.compile( r"\b(?:does not|doesn't|do not|did not|cannot|can't)\s+(?:itself\s+)?" r"(?:prove|establish|support|show)|" r"\b(?:does not|doesn't|do not|did not|cannot|can't)\s+rule out\b|" r"\b(?:not paid for|outruns?).{0,40}\b(?:the )?evidence\b|\bunsupported leap\b", re.I, ), ), "narrower_claim_revision": ( re.compile( r"\b(?:narrower\s+(?:revision|claim|formulation|candidate(?: claim)?)|revision\s*:|" r"should be narrowed to|defensible formulation|revise(?:d)?\b.{0,40}\bto\b)", re.I, ), re.compile( r"\b(?:narrower|only|limited to|at most|supports? that|evidence shows?|bounded|unestablished|" r"undemonstrated|unproven|not established|not demonstrated|does not establish|do not establish|" r"doesn't establish|don't establish|remains unknown|open constraint|design aspiration|" r"not a runtime property)\b", re.I, ), ), } INVALID_COUNT_INVARIANT_RE = re.compile( r"\ball five (?:canonical )?(?:numbers|counts) (?:must|need to) move\b|" r"\bneed all five (?:canonical )?(?:numbers|counts) to move\b|" r"\ball five (?:canonical )?(?:numbers|counts).{0,40}\b(?:higher|increase|increased|rise|change|changed)\b|" r"\b(?:database\s+|aggregate\s+)?(?:counts?|totals?)\s+" r"(?:must|need(?:s)? to|are required to)\s+(?:change|differ|move)\b|" r"\b(?:count readback|counts?|totals?).{0,40}\bmust\b.{0,40}" r"\b(?:counts?|totals?).{0,30}\b(?:change|differ|move)\b", re.I | re.S, ) COUNT_INVARIANT_REJECTION_RE = re.compile( r"(?:all five (?:numbers|counts)|count invariant).{0,50}(?:wrong|false|invalid|not valid|too strong)|" r"(?:wrong|false|invalid|not valid|too strong).{0,50}all five (?:numbers|counts)|" r"(?:aggregate )?(?:counts?|totals?).{0,50}(?:need not|do not need to|are not required to).{0,30}" r"(?:change|differ|move)", re.I | re.S, ) SCHEMA_GAP_QUALIFIER_RE = re.compile( r"\b(?:proposed|future|not current|not shipped|does not exist|doesn't exist|absent|missing|" r"has no|have no|lacks?|without|there is no|no column|not an edge|would require|schema gap|must be added|" r"does not support|doesn't support|supports neither|not supported|must not|do not invent|" r"schema extension|extension proposal|reviewed schema proposal|not (?:in|part of|among) the current)\b|" r"\b(?:does not|doesn't|cannot)\s+" r"(?:record|store|contain|include|have|expose|provide|support|model|express|represent)\b|" r"\b(?:requires?|needs?)\s+either\s+(?:an?\s+)?" r"(?:[a-z_]+\s+){0,3}(?:column|field|table|edge type)\b|" r"\b(?:would add|would introduce|would create)\s+(?:an?\s+)?" r"(?:[a-z_]+\s+){0,3}(?:column|field|table|edge type)\b|" r"\b(?:proposal|extension)\b.{0,100}\b(?:column|field|table|edge type)\b|" r"\bno\s+(?:[a-z_]+\s+){0,3}(?:column|field|table|edge type)\b" r"(?!\s+(?:review|proposal|change|addition|work)\b)", re.I, ) SCHEMA_CLAUSE_BOUNDARY_RE = re.compile( r"\s*(?:;|,\s+and\s+(?=(?:[a-z0-9_.`'-]+\s+){1,5}" r"(?:is|are|has|have|requires?|needs?|would|will|must|can|does?|supports?)\b)|" r"\bbut\b|\bhowever\b)\s*", re.I, ) CURRENT_SCHEMA_ASSERTION_PATTERNS: dict[str, re.Pattern[str]] = { "claims_unshipped_fields": re.compile( r"(?:the )?current schema\b.{0,120}\b(?:forecast[_ -]resolution|resolution[_ -]criteria|outcome|" r"resolved_at|falsifier)\b|" r"public\.claims\b.{0,120}\b(?:forecast[_ -]resolution|resolution[_ -]criteria|outcome|resolved_at|" r"falsifier)\b|" r"(?:the )?current schema\s+(?:already\s+)?" r"(?:stores?|has|contains?|records?|exposes?|provides?|supports?|includes?|models?|represents?)\s+.{0,80}" r"\b(?:forecast[_ -]resolution|resolution[_ -]criteria|outcome|resolved_at|falsifier)\b|" r"claims\s+(?:table|schema)\b.{0,100}\b(?:body|metadata|forecast[_ -]resolution|" r"resolution[_ -]criteria|outcome|resolved_at|falsifier)\b|" r"public\.claims\s+(?:(?:natively|already|currently)\s+)?" r"(?:stores?|has|contains?|records?|exposes?|provides?|supports?|includes?|models?|represents?)\s+" r"(?:an?\s+)?" r"(?:body|metadata|forecast[_ -]resolution|resolution[_ -]criteria|outcome|resolved_at|falsifier)\b|" r"public\.claims\b.{0,80}\b(?:column|field)\b.{0,30}" r"\b(?:body|metadata|forecast[_ -]resolution|resolution[_ -]criteria|outcome|resolved_at|falsifier)\b|" r"public\.claims\b.{0,80}\b(?:body|metadata|forecast[_ -]resolution|resolution[_ -]criteria|" r"outcome|resolved_at|falsifier)\b.{0,30}" r"\b(?:column|field)\b", re.I, ), "sources_unshipped_fields": re.compile( r"(?:(?:public\.)?sources?|(?:canonical )?source row).{0,140}" r"\b(?:author|channel|title|publisher|publication(?: date)?|published_at|source_date)\b", re.I, ), "invalid_current_edge_type": re.compile( r"\b(?:current )?edge_type enum\b.{0,80}\bresolves\b|" r"\bcurrent edge_type\b.{0,80}\bresolves\b|" r"(?:\b(?:claim_edges?|edge type|edge)\b\s*(?:is|=|:|named|called|of)?\s*[`'\"]?" r"(?:superseded_by|relates_to|resolves|derived_from)\b|" r"\b(?:superseded_by|relates_to|resolves|derived_from)\b[`'\"]?\s+(?:claim_)?edge\b|" r"\b(?:current )?edge_type enum\b.{0,50}\b(?:contains|has|includes|supports|enumerates)\b.{0,20}" r"\bresolves\b|" r"\b(?:current )?edge_type\b.{0,50}\b(?:contains|has|includes|supports|enumerates)\b.{0,20}" r"\bresolves\b)", re.I, ), "unshipped_edge_rationale": re.compile(r"\bclaim_edges?\b.{0,100}\brationale\b", re.I), "unshipped_evidence_excerpt": re.compile( r"\bclaim_evidence\b.{0,100}\bwith\s+(?:an?\s+)?(?:excerpt|excerpt_anchor)\b|" r"\bclaim_evidence\b.{0,100}\b(?:has|contains|stores|carries)\b.{0,30}\bexcerpt\b|" r"\bclaim_evidence\b.{0,100}\b(?:excerpt|excerpt_anchor)\s+(?:column|field)\b", re.I, ), "unreviewed_claim_type": re.compile( r"(?:public\.)?claims?.{0,100}\b(?:type|typed)\s*[`'\":=]?\s*(?:observation|hypothesis|belief)\b|" r"\b(?:new|distinct)\s+(?:observation|hypothesis|belief)\s+claim\b", re.I, ), "reasoning_tools_unshipped_fields": re.compile( r"\breasoning_tools?\b.{0,120}\b(?:structured criteria|structured steps|criteria or steps|scope)\b", re.I, ), "behavioral_rules_unshipped_fields": re.compile( r"\bbehavioral_rules?\b.{0,120}\b(?:has|have|with|carries|field|column).{0,40}\bstatus\b", re.I, ), "unsupported_approve_claim_surface": re.compile( r"\b(?:apply_payload|approve_claim)\b.{0,500}\b(?:governance gate row|governance_gates|" r"superseded_by column update|update the existing old claim)\b", re.I, ), } UNVERIFIED_M3TAVERSAL_ALIAS_RE = re.compile(r"\b(?:Cory|m3ta)\b", re.I) RESOLVES_PRESENCE_ASSERTION_RE = re.compile( r"\bresolves (?:edge|type)\b.{0,120}\b(?:is not (?:absent|missing)|already (?:exists|supported|present)|" r"is (?:already )?supported by (?:the )?(?:current )?edge_type)\b|" r"\b(?:current )?edge_type\b.{0,60}\b(?:already )?(?:contains|has|includes|supports|enumerates)\b.{0,20}" r"\bresolves\b", re.I | re.S, ) EDGE_TYPE_RESOLVES_DENIAL_RE = re.compile( r"\b(?:current )?edge_type\b.{0,50}\b(?:does not|doesn't|cannot|lacks?|has no)\b.{0,40}\bresolves\b", re.I | re.S, ) SOURCE_EVIDENCE_CANONICAL_OBJECT_RE = re.compile(r"claim_evidence|public\.sources|source rows?", re.I) SOURCE_EVIDENCE_LOCATOR_GAP_RE = re.compile( r"(?:missing|without|no).{0,50}(?:url|storage(?:_path)?|locator)", re.I | re.S ) SOURCE_EVIDENCE_DENIAL_RE = re.compile( r"(?:not|isn't|is not|doesn't count as|does not count as).{0,40}(?:canonical|grounded) evidence", re.I | re.S, ) SOURCE_EVIDENCE_CITATION_STUB_DENIAL_RE = re.compile( r"citation stubs?.{0,50}(?:not|isn't|is not).{0,30}(?:canonical|grounded) evidence", re.I | re.S, ) BEHAVIORAL_RULES_FALSE_ABSENCE_RE = re.compile( r"(?:public\.)?behavioral_rules(?: table)?.{0,50}(?:does not exist|doesn't exist|is absent|is missing|not shipped)|" r"(?:no|missing) (?:public\.)?behavioral_rules table", re.I | re.S, ) DB_ONLY_CAUSALITY_RE = re.compile( r"(?:unchanged|same) (?:canonical )?(?:DB|database|counts?).{0,80}" r"(?:proves?|means|therefore).{0,80}(?:behavior|answer).{0,30}(?:unchanged|cannot change)", re.I | re.S, ) RESTART_ERASES_ALL_RE = re.compile( r"restart.{0,100}(?:erases?|forgets?|loses?).{0,40}(?:all|every|prior[- ]session)", re.I | re.S ) DURABLE_SESSION_EPHEMERAL_RE = re.compile( r"session JSONL.{0,100}(?:is|are|gets?|becomes?|confirm).{0,30}" r"(?:gone|lost|deleted|discarded|absent|zeroed|empty)|" r"session JSONL.{0,100}(?:gone|lost|deleted|discarded|absent|zeroed|empty).{0,50}" r"(?:session closes?|restart|process launch)|" r"(?:state\.db|session JSONL).{0,360}(?:restart|recycle|process launch).{0,100}" r"(?:can|may|will)\s+(?:clear|erase|reset|delete|drop|lose)\s+" r"(?:them|both|the files?|the session|state\.db|session JSONL)", re.I | re.S, ) REASONING_TOOL_CLAIM_EDGE_RE = re.compile( r"claims?.{0,120}(?:get|have|create|use).{0,50}(?:claim_)?edges?.{0,100}" r"(?:reasoning_tool|tool(?:'s)? ID)|" r"(?:claim_)?edges?.{0,100}(?:point|connect|link).{0,80}(?:reasoning_tool|tool(?:'s)? ID)", re.I | re.S, ) HANDLER_TELEGRAM_OVERCLAIM_RE = re.compile( r"(?:Telegram (?:path|delivery).{0,40}(?:is|was) (?:live|proven)|" r"live Telegram (?:path|delivery)).{0,160}(?:posted nothing|did not post|no Telegram post)", re.I | re.S, ) FORECAST_HISTORY_REWRITE_RE = re.compile( r"(?[^\"\r\n]+)\"", re.I, ) STRUCTURED_EVIDENCE_QUOTE_RE = re.compile( r"\b(?:linked\s+)?(?:evidence\s+source|source|evidence)\s+(?:row\s+)?(?:id\s*)?`?" r"(?:[0-9a-f]{8,12}|[0-9a-f]{8}-[0-9a-f]{4}-[1-5][0-9a-f]{3}-[89ab][0-9a-f]{3}-[0-9a-f]{12})`?" r"[^:\n]{0,120}:\s*(?:\r?\n\s*)*\"(?P[^\"\r\n]+)\"", re.I, ) REVISION_LABEL_RE = re.compile(r"\brevision\s*:\s*(?P[^\n]+)", re.I) REVISION_BOUNDARY_RE = re.compile( r"\b(?:narrower|only|limited to|at most|supports? that|evidence shows?|bounded|unestablished|unproven|" r"undemonstrated|not established|not demonstrated|does not establish|do not establish|doesn't establish|" r"don't establish|remains unknown|unresolved|open constraint|design aspiration|not a runtime property)\b", re.I, ) REVISION_REPETITION_RE = re.compile( r"\b(?:repeat(?:ed)?|restate(?:d)?).{0,50}\b(?:original|same)\s+claim\b|" r"\b(?:original|same)\s+claim\b.{0,50}\bunchanged\b", re.I | re.S, ) DEFAULT_MAX_RESPONSE_WORDS = 220 MAX_RESPONSE_WORDS = {"OOS-09": 100, "OOS-10": 180} BLOCKER_STOPWORDS = { "blocker", "chat", "demo", "label", "only", "remember", "single", "that", "this", "under", "will", "with", } BLOCKER_DIAGNOSTIC_RE = re.compile( r"\b(?:applied_at|canonical|public\.\*|rows?|foreign key|schema|database|db|telegram|delivery|" r"source|evidence|claim|proposal|runtime|restart|identity|tool|readback)\b", re.I, ) BLOCKER_FAILURE_STATE_RE = re.compile( r"\b(?:no|none|zero|missing|absent|null|unavailable|unsupported|unapplied|not|cannot|can't|blocked|" r"fails?|stale|wrong|gap|lacks?|without)\b", re.I, ) BLOCKER_ABSENCE_RE = re.compile( r"\b(?:there is |there's )?no (?:[a-z][a-z-]*\s+){0,4}blocker " r"(?:exists|remains|is present)\b|\bnothing (?:is )?blocked\b", re.I, ) PER_AGENT_CLAIM_DUPLICATION_RE = re.compile( r"\b(?:(?:one|separate|distinct|individual)\s+)?(?:public\.)?claims?(?:\s+row)?\s+per\s+agent\b|" r"\b(?:claim|public\.claims)(?:\s+row)?(?:\s+once)?\s+(?:for|to)\s+each\s+agent\b|" r"\beach\s+agent\s+(?:gets?|receives?|has|stores?|uses?)\s+" r"(?:(?:its|their)\s+own|an?\s+separate|one)\s+(?:public\.)?claims?(?:\s+row)?\b|" r"\b(?:give|assign)\s+each\s+agent\s+(?:(?:its|their)\s+own|an?\s+separate|one)\s+" r"(?:public\.)?claims?(?:\s+row)?\b|" r"\bone\s+(?:public\.)?claims?(?:\s+row)?\s+for\s+[A-Z][a-z0-9_-]+\s+and\s+one\s+" r"(?:(?:public\.)?claims?(?:\s+row)?\s+)?for\s+[A-Z][a-z0-9_-]+\b|" r"\b(?:fact|proposition|claim|assertion)\b.{0,80}\b(?:under|within|inside)\s+each\s+agent\b|" r"\b(?:their|the agents?'?)\s+respective\s+(?:copies|claim rows?|records?)\b", re.I, ) PER_AGENT_DUPLICATION_PREFIX_DENIAL_RE = re.compile( r"(?:\b(?:do not|don't|does not|doesn't|should not|shouldn't|must not|never|avoid)\b\s*" r"(?:(?:create|store|write|give|assign|duplicate|fork|make|keep|use|add)(?:ing)?\s+)?" r"(?:(?:the|an?|one)\s+)?(?:(?:factual|canonical|shared|separate|distinct)\s+)?|" r"\b(?:instead of|rather than|not)\s+)$", re.I, ) PER_AGENT_DUPLICATION_SUFFIX_DENIAL_RE = re.compile( r"^.{0,80}\b(?:is|would be|remains?)\s+(?:wrong|incorrect|invalid|duplicative)\b|" r"^.{0,80}\b(?:creates?|causes?|would create)\s+(?:data\s+)?(?:divergence|fragmentation)\b", re.I, ) def prompt_catalog(memory_token: str) -> list[dict[str, Any]]: return [ { **scenario, "message": scenario["message"].replace("__MEMORY_TOKEN__", memory_token), } for scenario in SCENARIOS ] def extract_blocker_clause(reply: str) -> str | None: for pattern in ( re.compile(r"\bblocker(?:\s+restated)?\s*:\s*(?P[^\n]+)", re.I), re.compile(r"\bblocker\s+(?:is|was)\s+(?P[^\n]+)", re.I), re.compile( r"\b(?:mnemonic|label)(?:\s+(?:tag|retrieved))?\s*:?[^.\n]{0,140}?" r"(?:[.]|\s+(?:-|\u2014)\s+)\s*(?P[^\n]+)", re.I, ), ): match = pattern.search(reply) if match: return match.group("value").strip() return None def blocker_terms(value: str | None, *, memory_token: str) -> set[str]: if not value: return set() token_terms = set(re.findall(r"[a-z0-9_]+", memory_token.lower())) return { term for term in re.findall(r"[a-z0-9_]+", value.lower()) if len(term) >= 4 and term not in BLOCKER_STOPWORDS and term not in token_terms } def blocker_signature(value: str | None) -> set[str]: """Extract polarity-aware blocker categories for same-session recall.""" if not value: return set() signatures: set[str] = set() clauses = [ clause.strip() for clause in re.split(r"(?<=[.!?;])\s+|\n+|,\s+(?:but|and|so)\s+", value, flags=re.I) if clause.strip() ] if re.search(r"\b(?:approved|reviewer approval|reviewer signature|reviewer sign[- ]off)\b", value, re.I): signatures.add("reviewer_approval") applied_missing = re.compile( r"\bapproved[- ]but[- ]not[- ]applied\b|\b(?:not applied|unapplied)\b|" r"\b(?:zero|no|missing|absent)\s+applied_at(?:\s+values?)?\b|" r"\bapplied_at\b.{0,20}\b(?:unset|missing|absent|zero|none|null|not set)\b|" r"\b(?:unset|missing|absent|zero|none|null)\b.{0,20}\bapplied_at\b", re.I, ) applied_present = re.compile( r"\bapplied_at\b.{0,30}\b(?:non[- ]null|not null|exists|present|is set|has a value|" r"missing (?:flag )?(?:is|=) false|is not missing)\b", re.I, ) if any(applied_missing.search(clause) and not applied_present.search(clause) for clause in clauses): signatures.add("applied_missing") canonical_missing = re.compile( r"\bcanonical (?:gap|row|readback)\b.{0,30}\b(?:missing|absent|none|null|zero|not (?:live|written)|" r"unavailable)\b|" r"\b(?:no|zero|missing|absent)\b.{0,25}\b(?:canonical|public\.\*)\b.{0,25}\brows?\b|" r"\b(?:no|zero|missing|absent)\s+public\.\*\s+row(?: ID)?\b|" r"\bno rows?\b.{0,30}\b(?:landed|written|applied)\b.{0,20}\bpublic\.(?:claims|sources|\*)\b|" r"\bapproved[- ]but[- ]not[- ]applied canonical gap\b|\bcanonical gap\b", re.I, ) canonical_present = re.compile( r"\bcanonical (?:row|readback)\b.{0,20}\b(?:exists|present|current|complete|is live)\b|" r"\bcanonical rows?\b.{0,20}\b(?:exist|are present|are current|are complete)\b|" r"\bcanonical gap\b.{0,20}\b(?:is|was|has been|now)\s+(?:closed|resolved|cleared|fixed)\b", re.I, ) if any(canonical_missing.search(clause) and not canonical_present.search(clause) for clause in clauses): signatures.add("canonical_missing") telegram_missing = re.compile( r"\b(?:telegram|delivery|visible message|message receipt)\b.{0,50}" r"\b(?:no|missing|absent|blocked|unproven|not proven|unavailable)\b|" r"\b(?:no|missing|absent|blocked|unproven|not proven|unavailable)\b.{0,50}" r"\b(?:telegram|delivery|visible message|message receipt)\b", re.I, ) if any(telegram_missing.search(clause) for clause in clauses): signatures.add("telegram_missing") if re.search( r"\b(?:beliefs?\b.{0,40})?(?:has|have|with)\s+no\s+claim[_ -]?id\s+foreign key\b|" r"\b(?:missing|absent|no)\b.{0,50}\b(?:foreign key|belief[- ]to[- ]claim link)\b|" r"\bschema gap\b", value, re.I | re.S, ): signatures.add("schema_link_missing") if ( re.search(r"\bbeliefs?(?:\.claim[_ -]?id)?\b", value, re.I) and re.search(r"\bclaim[_ -]?id\b", value, re.I) and re.search(r"\b(?:foreign key|FK)\b", value, re.I) and BLOCKER_FAILURE_STATE_RE.search(value) ): signatures.add("belief_claim_foreign_key_missing") if ( re.search(r"\bbeliefs?(?:\.claim[_ -]?id)?\b", value, re.I) and re.search(r"\bclaim[_ -]?id\b", value, re.I) and re.search(r"\bindex\b", value, re.I) and BLOCKER_FAILURE_STATE_RE.search(value) ): signatures.add("belief_claim_index_missing") if re.search( r"\bpublic\.sources\b.{0,80}\b(?:no|missing|absent|lacks?)\b.{0,40}\bforeign key\b|" r"\b(?:no|missing|absent)\b.{0,40}\bforeign key\b.{0,80}\bpublic\.sources\b", value, re.I | re.S, ): signatures.add("source_foreign_key_missing") return signatures def asserts_per_agent_claim_duplication(reply: str) -> bool: """Reject affirmative per-agent duplication while allowing explicit denials.""" clauses = re.split(r"(?<=[.!?;])\s+|\n+|,\s+(?:but|and|so)\s+", reply, flags=re.I) for clause in clauses: for match in PER_AGENT_CLAIM_DUPLICATION_RE.finditer(clause): prefix = clause[max(0, match.start() - 80) : match.start()] suffix = clause[match.end() : match.end() + 100] if PER_AGENT_DUPLICATION_PREFIX_DENIAL_RE.search(prefix): continue if PER_AGENT_DUPLICATION_SUFFIX_DENIAL_RE.search(suffix): continue return True return False def matched_concept(reply: str, concept: str) -> bool: if concept == "blocker_definition": blocker = extract_blocker_clause(reply) terms = blocker_terms(blocker, memory_token="") return bool( blocker and len(terms) >= 3 and BLOCKER_DIAGNOSTIC_RE.search(blocker) and BLOCKER_FAILURE_STATE_RE.search(blocker) and not BLOCKER_ABSENCE_RE.search(blocker) ) if concept == "shared_knowledge_commons" and asserts_per_agent_claim_duplication(reply): return False if concept == "claim_body_evidence_distinction": if CLAIM_EVIDENCE_CONFLATION_RE.search(reply): return False claim_block = STRUCTURED_CLAIM_QUOTE_RE.search(reply) evidence_block = STRUCTURED_EVIDENCE_QUOTE_RE.search(reply) if claim_block and evidence_block: claim_value = " ".join(re.findall(r"[a-z0-9]+", claim_block.group("value").lower())) evidence_value = " ".join(re.findall(r"[a-z0-9]+", evidence_block.group("value").lower())) return bool(claim_value and evidence_value and claim_value != evidence_value) if concept == "narrower_claim_revision": labelled = REVISION_LABEL_RE.search(reply) if labelled and ( not REVISION_BOUNDARY_RE.search(labelled.group("value")) or REVISION_REPETITION_RE.search(labelled.group("value")) ): return False return all(pattern.search(reply) for pattern in CONCEPT_PATTERNS[concept]) def asserts_invalid_count_invariant(reply: str) -> bool: clauses = re.split( r"(?<=[.!?;])\s+|\n+|,\s+(?:but|and)\s+|\s+but\s+|" r",\s+(?=(?:it|this|that|for this apply|database totals?|the current apply)\b)", reply, flags=re.I, ) return any( INVALID_COUNT_INVARIANT_RE.search(clause) and not COUNT_INVARIANT_REJECTION_RE.search(clause) for clause in clauses ) def current_schema_overclaims(reply: str) -> list[str]: """Return proposed-v3-as-current assertions that are not explicitly qualified.""" findings: list[str] = [] sentence_segments = re.split(r"(?<=[.!?])\s+|\n+", reply) for sentence in sentence_segments: if RESOLVES_PRESENCE_ASSERTION_RE.search(sentence) and not EDGE_TYPE_RESOLVES_DENIAL_RE.search(sentence): findings.append("invalid_current_edge_type") # Negation and future-schema qualifiers apply to their clause, not to a # contradictory assertion in a later conjunction in the same sentence. clauses = SCHEMA_CLAUSE_BOUNDARY_RE.split(sentence) review_context = bool( re.match(r"^\s*(?:[-*]\s*)?review (?:covers|includes|asks|considers|checks)\b", sentence, re.I) ) for index, clause in enumerate(clauses): next_clause = clauses[index + 1] if index + 1 < len(clauses) else "" anaphoric_qualifier = bool( re.match(r"^\s*(?:it|this|that|they|these|those|neither|none)\b", next_clause, re.I) and SCHEMA_GAP_QUALIFIER_RE.search(next_clause) ) if SCHEMA_GAP_QUALIFIER_RE.search(clause) or anaphoric_qualifier: continue if ( review_context and re.search(r"\b(?:whether|if|how)\b", clause, re.I) and not re.search( r"\bcurrent schema\b.{0,80}\b(?:has|supports|includes|provides|allows)\b", clause, re.I | re.S, ) ): continue for label, pattern in CURRENT_SCHEMA_ASSERTION_PATTERNS.items(): if pattern.search(clause): findings.append(label) return sorted(set(findings)) def source_evidence_semantic_issues(reply: str) -> list[str]: """Reject answers that confuse canonical linkage with locator quality.""" findings: set[str] = set() for segment in re.split(r"(?<=[.!?])\s+|\n+", reply): if ( SOURCE_EVIDENCE_CANONICAL_OBJECT_RE.search(segment) and SOURCE_EVIDENCE_LOCATOR_GAP_RE.search(segment) and SOURCE_EVIDENCE_DENIAL_RE.search(segment) ): findings.add("locator_gap_called_noncanonical") if SOURCE_EVIDENCE_CITATION_STUB_DENIAL_RE.search(reply): findings.add("citation_stub_called_ungrounded") return sorted(findings) def behavioral_rule_schema_issues(reply: str) -> list[str]: """Reject claims that the current behavioral-rules table is absent.""" return ["behavioral_rules_false_absence"] if BEHAVIORAL_RULES_FALSE_ABSENCE_RE.search(reply) else [] def broad_semantic_issues(reply: str) -> list[str]: """Reject high-impact false claims that can hide inside otherwise complete prose.""" findings: set[str] = set() if DB_ONLY_CAUSALITY_RE.search(reply) and not re.search( r"does not prove|doesn't prove|do not prove|does not mean|doesn't mean|do not mean|cannot prove", reply, re.I ): findings.add("unchanged_db_called_complete_behavior_proof") if RESTART_ERASES_ALL_RE.search(reply) and not re.search( r"restart.{0,100}(?:does not|doesn't|need not|not necessarily).{0,60}(?:erase|forget|lose)|" r"(?:facts?|context).{0,80}not erased by restart", reply, re.I | re.S, ): findings.add("restart_called_total_memory_erasure") if DURABLE_SESSION_EPHEMERAL_RE.search(reply): findings.add("durable_session_called_ephemeral") if HANDLER_TELEGRAM_OVERCLAIM_RE.search(reply): findings.add("handler_proof_called_telegram_live") if FORECAST_HISTORY_REWRITE_RE.search(reply): findings.add("forecast_history_rewrite") if TEMP_LABEL_AS_SOURCE_RE.search(reply): findings.add("temporary_label_promoted_to_source") if BELIEF_EDGE_FABRICATION_RE.search(reply): findings.add("belief_row_given_claim_edge") if ROW_LEVEL_APPLIED_AT_RE.search(reply): findings.add("applied_at_assigned_to_each_apply_step") for segment in re.split(r"(?<=[.!?])\s+|\n+", reply): if not REASONING_TOOL_CLAIM_EDGE_RE.search(segment): continue if re.search( r"(?:never|do not|don't|must not|cannot).{0,100}reasoning_tools?|" r"(?:edge endpoints?|from_claim|to_claim).{0,100}(?:claim IDs?|claims? only|never reasoning_tools?)", segment, re.I | re.S, ): continue findings.add("claim_edge_pointed_at_reasoning_tool") break return sorted(findings) def proposal_readiness_issues(prompt_id: str, reply: str) -> list[str]: if prompt_id not in {"OOS-01", "OOS-04", "OOS-07", "OOS-08"}: return [] issues: list[str] = [] clauses = [ item.strip() for item in re.split( r"(?<=[.!?;])\s+|\n+|,\s+(?:but|and)\s+|\s+but\s+|" r",\s+(?=(?:it|this|that|for this apply|database totals?|the current apply)\b)", reply, flags=re.I, ) if item.strip() ] if any( APPROVED_APPLY_ACTION_RE.search(clause) and not APPLYABILITY_GAP_RE.search(clause) and not APPLY_ACTION_DISCLAIMER_RE.search(clause) for clause in clauses ): issues.append("approved_proposal_applyability_overclaim") if any( UNSUPPORTED_APPLY_SURFACE_RE.search(clause) and not UNSUPPORTED_APPLY_BOUNDARY_RE.search(clause) for clause in clauses ): issues.append("unsupported_apply_surface_overclaim") return sorted(set(issues)) def source_intake_issues(prompt_id: str, reply: str) -> list[str]: if prompt_id in {"OOS-02", "OOS-14"} and CANONICAL_SOURCE_BEFORE_REVIEW_RE.search(reply): return ["canonical_source_created_before_review"] return [] def composition_capability_issues(prompt_id: str, reply: str) -> list[str]: if prompt_id != "OOS-06": return [] if UNSUPPORTED_COMPOSITION_APPLY_RE.search(reply): return ["unsupported_collection_listed_as_applyable"] return [] def score_reply(prompt: dict[str, Any], reply: str, *, memory_token: str) -> dict[str, Any]: scorer_prompt_id = str(prompt.get("scorer_id") or prompt["id"]) legacy_prompt = {**prompt, "id": scorer_prompt_id} if scorer_prompt_id == "OOS-07": legacy_prompt["required_signals"] = [] elif scorer_prompt_id == "OOS-08": legacy_prompt["required_signals"] = ["row_level_proof"] legacy_score = base.score_reply(legacy_prompt, reply) concepts = {concept: matched_concept(reply, concept) for concept in prompt["required_concepts"]} custom_signals: dict[str, bool] = {} if scorer_prompt_id in {"OOS-07", "OOS-08"}: custom_signals["memory_token"] = memory_token.lower() in reply.lower() if scorer_prompt_id == "OOS-08": lowered = reply.lower() custom_signals["closure_proof"] = any( phrase in lowered for phrase in ("readback", "before/after", "before-and-after", "postflight", "canonical row", "applied_at") ) if scorer_prompt_id == "OOS-09": custom_signals["exact_participant_handle"] = "m3taversal" in reply.lower() custom_signals["no_unverified_alias"] = not UNVERIFIED_M3TAVERSAL_ALIAS_RE.search(reply) custom_signals["current_update_identity_boundary"] = bool( re.search(r"current (?:Telegram )?(?:update|sender|handle)|visible handle", reply, re.I) ) custom_signals["no_cross_participant_identity_bleed"] = bool( re.search( r"(?:do(?:es)? not|never)\s+(?:\w+\s+){0,3}" r"(?:transfer|transfers|carry|carries|mix|mixes|reuse|reuses)", reply, re.I, ) ) invalid_count_invariant = asserts_invalid_count_invariant(reply) schema_overclaims = current_schema_overclaims(reply) source_evidence_issues = source_evidence_semantic_issues(reply) if scorer_prompt_id == "OOS-05" else [] behavioral_rule_issues = behavioral_rule_schema_issues(reply) if scorer_prompt_id == "OOS-06" else [] semantic_issues = broad_semantic_issues(reply) readiness_issues = proposal_readiness_issues(scorer_prompt_id, reply) intake_issues = source_intake_issues(scorer_prompt_id, reply) composition_issues = composition_capability_issues(scorer_prompt_id, reply) word_count = len(re.findall(r"\b\w+(?:[-']\w+)*\b", reply)) max_response_words = MAX_RESPONSE_WORDS.get(scorer_prompt_id, DEFAULT_MAX_RESPONSE_WORDS) response_too_long = word_count > max_response_words return { "prompt_id": prompt["id"], "scorer_prompt_id": scorer_prompt_id, "dimension": prompt["dimension"], "concepts": concepts, "custom_signals": custom_signals, "legacy_signals": legacy_score["signals"], "legacy_pass": legacy_score["pass"], "overclaim_detected": legacy_score["overclaim_detected"], "response_issue_detected": legacy_score["response_issue_detected"], "invalid_count_invariant_detected": invalid_count_invariant, "current_schema_overclaims": schema_overclaims, "source_evidence_semantic_issues": source_evidence_issues, "behavioral_rule_schema_issues": behavioral_rule_issues, "broad_semantic_issues": semantic_issues, "proposal_readiness_issues": readiness_issues, "source_intake_issues": intake_issues, "composition_capability_issues": composition_issues, "word_count": word_count, "max_response_words": max_response_words, "response_too_long": response_too_long, "pass": bool( all(concepts.values()) and all(custom_signals.values()) and not legacy_score["overclaim_detected"] and not invalid_count_invariant and not schema_overclaims and not source_evidence_issues and not behavioral_rule_issues and not semantic_issues and not readiness_issues and not intake_issues and not composition_issues and not response_too_long ), } def score_results( results: list[dict[str, Any]], *, memory_token: str, catalog: list[dict[str, Any]] | None = None, ) -> dict[str, Any]: catalog = catalog or prompt_catalog(memory_token) expected_ids = [prompt["id"] for prompt in catalog] by_prompt = {prompt["id"]: prompt for prompt in catalog} result_ids = [str(result.get("prompt_id")) for result in results if result.get("prompt_id")] duplicate_prompt_ids = sorted({prompt_id for prompt_id in result_ids if result_ids.count(prompt_id) > 1}) by_result = {str(result.get("prompt_id")): result for result in results if result.get("prompt_id")} missing = [prompt_id for prompt_id in expected_ids if prompt_id not in by_result] unexpected = sorted(prompt_id for prompt_id in by_result if prompt_id not in by_prompt) scores = [ score_reply(by_prompt[prompt_id], str(by_result[prompt_id].get("reply") or ""), memory_token=memory_token) for prompt_id in expected_ids if prompt_id in by_result ] memory_source_id = next( (prompt["id"] for prompt in catalog if str(prompt.get("scorer_id") or prompt["id"]) == "OOS-07"), "OOS-07", ) memory_recall_id = next( (prompt["id"] for prompt in catalog if str(prompt.get("scorer_id") or prompt["id"]) == "OOS-08"), "OOS-08", ) memory_source_reply = str((by_result.get(memory_source_id) or {}).get("reply") or "") memory_recall_reply = str((by_result.get(memory_recall_id) or {}).get("reply") or "") source_clause = extract_blocker_clause(memory_source_reply) recall_clause = extract_blocker_clause(memory_recall_reply) source_terms = blocker_terms(source_clause, memory_token=memory_token) recall_terms = blocker_terms(recall_clause, memory_token=memory_token) overlap_terms = source_terms & recall_terms required_overlap = min(3, max(1, (len(source_terms) + 2) // 3)) if source_terms else 1 source_signature = blocker_signature(source_clause) recall_signature = blocker_signature(recall_clause) source_failure_signature = source_signature - {"reviewer_approval"} recall_failure_signature = recall_signature - {"reviewer_approval"} signature_overlap = source_failure_signature & recall_failure_signature required_signature_overlap = min(2, len(source_failure_signature)) if source_failure_signature else 0 same_blocker_recalled = bool( source_clause and ( len(signature_overlap) >= required_signature_overlap and len(overlap_terms) >= required_overlap if source_failure_signature else len(overlap_terms) >= required_overlap ) ) for score in scores: if score["prompt_id"] != memory_recall_id: continue score["custom_signals"]["same_blocker_recalled"] = same_blocker_recalled score["pass"] = bool(score["pass"] and same_blocker_recalled) memory_continuity = { "source_reply_sha256": hashlib.sha256(memory_source_reply.encode("utf-8")).hexdigest(), "source_blocker_clause": source_clause, "source_blocker_terms": sorted(source_terms), "recall_overlap_terms": sorted(overlap_terms), "required_overlap": required_overlap, "source_blocker_signature": sorted(source_signature), "recall_blocker_signature": sorted(recall_signature), "signature_overlap": sorted(signature_overlap), "required_signature_overlap": required_signature_overlap, "same_blocker_recalled": same_blocker_recalled, } return { "expected_prompt_count": len(expected_ids), "expected_prompt_ids": expected_ids, "missing_prompt_ids": missing, "unexpected_prompt_ids": unexpected, "duplicate_prompt_ids": duplicate_prompt_ids, "prompt_count": len(scores), "passes": sum(1 for score in scores if score["pass"]), "failures": [score for score in scores if not score["pass"]], "scores": scores, "memory_continuity": memory_continuity, "pass": not missing and not unexpected and not duplicate_prompt_ids and len(scores) == len(expected_ids) and all(score["pass"] for score in scores), } def main() -> int: parser = argparse.ArgumentParser(description=__doc__) parser.add_argument("--results-json", type=Path, required=True) parser.add_argument("--memory-token", required=True) parser.add_argument("--out", type=Path) args = parser.parse_args() payload = json.loads(args.results_json.read_text(encoding="utf-8")) results = payload.get("results", payload) if isinstance(payload, dict) else payload if not isinstance(results, list): raise SystemExit("results JSON must contain a list or a top-level results list") report = { "generated_at_utc": datetime.now(timezone.utc).isoformat(), "mode": "working_leo_m3taversal_out_of_sample_score", "source_results_json": str(args.results_json), "memory_token": args.memory_token, "score": score_results(results, memory_token=args.memory_token), } if args.out: args.out.parent.mkdir(parents=True, exist_ok=True) args.out.write_text(json.dumps(report, indent=2, sort_keys=True) + "\n", encoding="utf-8") print(json.dumps(report, indent=2, sort_keys=True)) return 0 if report["score"]["pass"] else 1 if __name__ == "__main__": raise SystemExit(main())