feat(competences): extract batch candidates

2026-05-29 11:25:00 +02:00
parent 4ba426c205
commit e8a0fb0e42
60 changed files with 18176 additions and 0 deletions
--- a/data/competences/observed/scroll_down_pdf_edge.yaml
+++ b/data/competences/observed/scroll_down_pdf_edge.yaml
@@ -0,0 +1,118 @@
+schema_version: 1
+id: scroll_down_pdf_edge
+name: Scroller vers le bas dans un PDF Edge
+version: 1
+learning_state: observed
+
+intent:
+  fr: faire defiler un document PDF vers le bas dans Microsoft Edge
+
+parameters: {}
+
+preconditions:
+  - id: edge_pdf_active
+    kind: active_window
+    any_of:
+      - process_active: msedge.exe
+
+methods:
+  - id: scroll_down_mouse
+    kind: scroll
+    primitive_ref: scroll_view
+    parameters:
+      direction: down
+      amount: 9
+      unit: lines
+    description: "Scroll vers le bas via molette souris dans un PDF Edge"
+    observed: true
+    trace_source: live_events.jsonl
+    trace_event_indices: [129, 130, 131, 133, 134, 135, 137, 138, 139]
+
+success_marker:
+  mode: all_of
+  timeout_ms: 5000
+  markers:
+    - kind: active_process_name_is
+      value: msedge.exe
+  supervised_requires:
+    - kind: ocr_contains
+      text: "contenu different apres scroll"
+      region_hint: document_body
+      evidence_state: hypothesis_offline
+      required_for: supervised_or_replay_verified
+
+failure_message_template:
+  intention: faire defiler le PDF vers le bas
+  attendu: le contenu visible doit changer apres le defilement
+  vu: "{observed_human_state}"
+  demande: indiquer si le document PDF actif peut defiler vers le bas
+
+chain_refs:
+  source_session: sess_20260318T010719_62a058
+  machine_id: DESKTOP-58D5CAC_windows
+  streaming_session_path: data/training/live_sessions/streaming_sessions/sess_20260318T010719_62a058.json
+  live_events_path: data/training/live_sessions/DESKTOP-58D5CAC_windows/sess_20260318T010719_62a058/live_events.jsonl
+  cleaned_segment:
+    status: documented_offline
+    keep_event_indices: [126, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140]
+    method_event_indices: [129, 130, 131, 133, 134, 135, 137, 138, 139]
+    success_event_indices: [140]
+    excluded_event_indices: [127, 128]
+    stop_before_event_index: 141
+    stop_before:
+      - subsequent_scroll_bursts
+      - heartbeat_without_window_metadata_later_in_session
+    ignored_between_method_and_success: []
+    notes:
+      - "Event #126 focus vers msedge.exe avec le PDF RapportS1 actif."
+      - "Events #129/#130/#131/#133/#134/#135/#137/#138/#139 sont des mouse_scroll dans msedge.exe."
+      - "Tous les events de methode ont delta [0, -1], ce qui prouve direction=down pour cette trace."
+      - "Events #132 et #136 sont des heartbeats sans metadonnees fenetre au milieu du burst."
+      - "Events #127/#128 sont un clic de positionnement et son action_result avant le burst scroll; ils sont exclus."
+      - "Event #140 est le premier mouse_scroll post-methode avec msedge.exe encore actif; il prouve la continuite active, pas le changement de contenu."
+  workflow_pipeline_id: null
+  graph_node_id: null
+  faiss_state_signatures: []
+  target_memory_keys: []
+  dashboard_knowledge_visible: false
+
+promotion:
+  candidate_requires:
+    - cleaned_segment_validated
+    - method_trace_present
+    - success_marker_defined
+    - failure_message_template_valid
+    - primitive_ref_satisfied
+    - scroll_trace_validated
+  supervised_requires:
+    - replay_verified_once
+    - success_marker_matched_after_action
+    - ocr_or_replay_verified_scroll_effect
+    - human_validation
+  stable_requires:
+    min_successes: 3
+    distinct_contexts: 3
+    max_unexplained_failures: 0
+  t2_known_gaps:
+    - id: scroll_effect_not_observed_offline
+      description: "La trace prouve les mouse_scroll et la fenetre active, mais pas le changement visuel du contenu PDF."
+      impact: "Le niveau T2 doit verifier que le viewport ou le texte visible change apres le scroll."
+      proposed_resolution: "Ajouter OCR runtime, screenshot diff ou marker visuel avant promotion supervised."
+      acted_by: Dom
+      acted_at: "2026-05-28T14:20:00+02:00"
+    - id: no_ocr_offline
+      description: "Aucune preuve OCR avant/apres scroll n'est disponible dans cette validation offline."
+      impact: "Le success_marker offline reste une preuve de continuite active, pas une preuve de contenu different."
+      proposed_resolution: "Verifier par OCR ou replay supervise avant promotion supervised."
+      acted_by: Dom
+      acted_at: "2026-05-28T14:20:00+02:00"
+
+generalisation:
+  seen_contexts: []
+  method_success_rate: {}
+  variance_log: []
+
+failure_log: []
+
+created_at: "2026-05-28T14:20:00+02:00"
+last_updated_at: "2026-05-28T14:20:00+02:00"