spb/ai-atlas

- fetch: validate_destination() on every request and redirect hop (max 5, manual loop): non-http(s), localhost/.local/.internal, RFC1918,
  loopback, link-local (metadata), CGNAT, IPv6 loopback/link-local/ULA, unspecified → FetchError("blocked destination")
- facts: facts_to_json / facts_from_json (tagged datetimes) for quarantined runs
- writer: run_id/source_key/is_first_run; a claim supersedes only when tier <= current or same source AND same extractor (LLM never
  overwrites deterministic from the same URL); licence/openness/modalities/status/kind/org_kind normalised via the ontology with value_raw
  + taxonomy_mappings, same assertion in another spelling re-encoded in place (no event); license_key claim; results get config_key,
  trust_level, variant, run_group, extractor, effort config for variant refs, one current row per (model, benchmark, metric, config_key),
  out-of-range scores → confidence low + anomaly; events get recorded_at, is_backfill (classify_backfill), group_key, deterministic
  importance; family/canonical hints materialised (family_id, canonical_id, artifact_kind, member_of_family, artifact_of); NEW_ARTIFACT 0,
  NEW_MODEL_FAMILY 1; conflicting claims recorded once; derived writers never open review items and mark their events backfill
- resolver: model/artifact compatible lookups; keep_separate decisions honoured; digit-collapsing aliases require the same variant_key;
  first_seen_hint → first_seen_at = min(now, hint); resolve_variant() folds evaluator-only effort variants onto the canonical model
- connector: run_id/source_key/first-run flag into the writer; --url reprocess rebuilds Targets from documents.meta._target (key, doc_type,
  meta, needs_llm, entity persisted at first fetch); quarantine for tier ≥ 2 connectors (rolling median baseline of the last 5 full runs:
  new entities > max(50, 1.3×), prices > 3×, entities < 0.8× / results < 0.5× on full extraction) → quarantined_runs + review item, nothing written
- services/events: classify_backfill, group_key_for, importance_for · services/anomalies: record / run_checks / list_anomalies
- merge: modes merge|alias|variant|family_member, keep_separate refusal, org-type and model/artifact pairs, dedupe/config_key recompute with
  collision handling, one current result per key, embeddings/sources/domains/review_queue/family_id/canonical_id re-pointed,
  resolution_decisions + admin_audit_log rows
- handlers: LLM extraction writer carries the snapshot's run_id

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

- models: FAMILY_ORGS / official_orgs / is_official_org; effort_config() and base_name() shared by the writer and the canonicalizer;
  effort suffixes chain ("-thinking-64k-high-effort"), LiveBench budgets ("32k thinking", "(no thinking)"); "max"/"fast"/"instant" are model
  tiers (Qwen3-Max, GPT-5.1-Codex-Max, Grok 4.1 Fast), no longer effort markers; family_release_hint reads "gpt-5-4" as 5.4 and never a size (38B)
- taxonomy: normalize_property(entity_type, prop, value) → (canonical, value_raw, mappings); unknown values kept, mapped to NULL
- benchmarks: RUN_KEYS (release, index_version, version, dataset_revision) leave config_key so a newer run replaces the older one; they are
  condition keys for comparability (partially comparable)
- tests/test_ontology.py: licence table, openness derivation, taxonomy, 22 model names, comparability trio, metric bounds/trust

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>