"""W15 structural tests — generic synthetic fixtures only. No semantic correctness assertions. No scenario-specific token expectations in the methodology — fixtures use abstract IDs (FP_T / FP_NT / BGT / BGN / S1_Shot1 / S2_Shot1 / Z1 / Z2 / Z3 / U_living / U_room_a / U_entry). Tests cover: source-evidence bundle, evidence_refs resolver, non-target fp copy-through, target bg coverage, no image API guard, render-safety guard, consistency audit sub-checks, HTML section order. The methodology grep on the source script uses per-char split tokens so this test file itself is exempt from that grep. """ from __future__ import annotations import json import re import sys from pathlib import Path import pytest _REPO_ROOT = Path(__file__).resolve().parents[3] _SCRIPTS_DIR = _REPO_ROOT / "backend" / "scripts" if str(_SCRIPTS_DIR) not in sys.path: sys.path.insert(0, str(_SCRIPTS_DIR)) def _synthetic_w12c_candidates() -> dict: return { "FP_T": { # target "fp_id": "FP_T", "group_id_pointer": "Gt", "candidate_diagram_t2i_prompt": "baseline target plan", "candidate_key_elements": [], "candidate_numbered_elements": [ {"number": 1, "label": "z1", "category": "area", "position_hint": "north", "zone_id_pointer": "Z1"}, ], "candidate_camera_recommendations": [ {"bg_id": "BGT", "sub_location": "Z1", "camera_position": "near #1", "camera_height": "1.6m", "lens_hint": "35mm", "framing_notes": ""}, ], "reconciliation_notes_vs_production": "", }, "FP_NT": { # non-target, copy-through "fp_id": "FP_NT", "group_id_pointer": "Gn", "candidate_diagram_t2i_prompt": "baseline non-target plan", "candidate_key_elements": [], "candidate_numbered_elements": [ {"number": 1, "label": "n1", "category": "area", "position_hint": "south", "zone_id_pointer": "Z2"}, ], "candidate_camera_recommendations": [], "reconciliation_notes_vs_production": "", }, } def _synthetic_w12c_per_bg() -> dict: return { "BGT": { "bg_id": "BGT", "group_id": "Gt", "fp_id": "FP_T", "applies_to_shots": ["S1_Shot1"], "floor_plan_ref_role": "layout_only", "use_numbered_elements": [1], "ignore_numbered_elements": [], "camera_axis_used": "ax", "camera_axis_source": "w11_axis", "visible_zone_scope": ["Z1"], "prior_bg_ref_role": "none", "render_prompt_appendix": "", "final_prompt_assembly_preview": "", }, "BGN": { "bg_id": "BGN", "group_id": "Gn", "fp_id": "FP_NT", "applies_to_shots": ["S2_Shot1"], "floor_plan_ref_role": "layout_only", "use_numbered_elements": [1], "ignore_numbered_elements": [], "camera_axis_used": "ax2", "camera_axis_source": "w11_axis", "visible_zone_scope": ["Z2"], "prior_bg_ref_role": "none", "render_prompt_appendix": "", "final_prompt_assembly_preview": "", }, } def _synthetic_adapter() -> dict: return { "background_catalog": { "BGT": { "depends_on_fp": ["FP_T"], "depends_on_bg": [], "sub_location_label": "zone_t", "state_label_raw": "base", "loc_id": "Lt", "space_key": "St", "applies_to_shots": ["S1_Shot1"], }, "BGN": { "depends_on_fp": ["FP_NT"], "depends_on_bg": [], "sub_location_label": "zone_n", "state_label_raw": "base", "loc_id": "Ln", "space_key": "Sn", "applies_to_shots": ["S2_Shot1"], }, }, } def _synthetic_source_bundle() -> dict: return { "project_id": "fake-proj", "episode_id": "fake-ep", "selected_shots": [ {"shot_key": "S1_Shot1", "scene_index": 1, "shot_index": 1, "shot_description": "synthetic desc 1", "screenplay_scene_heading": "S#1.", "t2i_variations_json": "[]", "visible_entities_json": "[]"}, {"shot_key": "S2_Shot1", "scene_index": 2, "shot_index": 1, "shot_description": "synthetic desc 2", "screenplay_scene_heading": "S#2.", "t2i_variations_json": "[]", "visible_entities_json": "[]"}, ], } def _synthetic_scene_save() -> dict: return { "segments": [ {"scene_index": 1, "heading": "S#1.", "text": "synthetic scene 1 text", "start_char": 0, "end_char": 10, "length": 10}, {"scene_index": 2, "heading": "S#2.", "text": "synthetic scene 2 text", "start_char": 10, "end_char": 20, "length": 10}, ], } def _synthetic_fp_baseline() -> dict: return { "fp_prompt_path": "/fake", "fp_render_path": "/fake", "fp_prompt_status": "ok", "fp_render_status": "missing", "floor_plans": { "FP_PROD_T": { "fp_id": "FP_PROD_T", "group_id": "Gt", "depends_on_fp": [], "key_elements": [], "numbered_elements": [], "camera_recommendations": [], "t2i_prompt": "production baseline target prompt", "applied_shots": [], "png_path": "", "png_exists": False, }, }, } def _synthetic_llm_output() -> dict: """Mirror of the schema the real LLM returns. Target fp has 2 enclosed rooms; non-target fp does not appear (copy-through is performed by the main loop, not by the LLM).""" return { "source_topology_by_fp": { "FP_T": { "fp_id": "FP_T", "spatial_units": [ {"unit_id": "U_living", "unit_label": "living zone", "unit_kind": "living_zone", "is_enclosed_room": True, "evidence_refs": [{"source_ref": "scene:1", "quote": "/거실 placeholder", "field": "scene_subheader"}]}, {"unit_id": "U_room_a", "unit_label": "private room a", "unit_kind": "private_room", "is_enclosed_room": True, "evidence_refs": [{"source_ref": "scene:1", "quote": "/room placeholder", "field": "scene_subheader"}]}, {"unit_id": "U_entry", "unit_label": "entry", "unit_kind": "entry_transition", "is_enclosed_room": False, "evidence_refs": [{"source_ref": "S1_Shot1", "quote": "shot desc", "field": "shot_description"}]}, ], "relationships": [ {"from_unit": "U_living", "to_unit": "U_room_a", "relation_kind": "door_between"}, ], "room_count_assessment": { "standalone_enclosed_room_count": 2, "rationale": "two enclosed rooms identified", }, "do_not_collapse_units": [ {"unit_id": "U_living", "reason": "primary living", "evidence_refs": [{"source_ref": "scene:1", "quote": "/거실", "field": "scene_subheader"}]}, {"unit_id": "U_room_a", "reason": "victim found here", "evidence_refs": [{"source_ref": "scene:1", "quote": "/room", "field": "scene_subheader"}]}, ], "conflicts_and_assumptions": [ "production baseline says one-bedroom but source slash subheaders show two enclosed rooms", ], }, }, "bg_unit_bindings": { "BGT": { "primary_unit_ids": ["U_living"], "secondary_visible_unit_ids": ["U_room_a"], "state_delta_unit_ids": [], "evidence_refs": [{"source_ref": "S1_Shot1", "quote": "shot desc 1", "field": "shot_description"}], }, }, "candidate_floor_plans": { "FP_T": { "fp_id": "FP_T", "group_id_pointer": "Gt", "candidate_diagram_t2i_prompt": "revised target plan with two enclosed rooms", "candidate_key_elements": [], "candidate_numbered_elements": [ {"number": 1, "label": "z1", "category": "area", "position_hint": "north", "zone_id_pointer": "Z1", "unit_id_pointer": "U_living"}, {"number": 2, "label": "z2", "category": "area", "position_hint": "east", "zone_id_pointer": "Z3", "unit_id_pointer": "U_room_a"}, ], "candidate_camera_recommendations": [ {"bg_id": "BGT", "sub_location": "Z1", "camera_position": "near #1", "camera_height": "1.6m", "lens_hint": "35mm", "framing_notes": ""}, ], "reconciliation_notes_vs_production": "expanded to two enclosed rooms based on source", }, }, "per_bg_render_reference_instructions": { "BGT": { "bg_id": "BGT", "group_id": "Gt", "fp_id": "FP_T", "applies_to_shots": ["S1_Shot1"], "floor_plan_ref_role": "layout_only", "use_numbered_elements": [1, 2], "ignore_numbered_elements": [], "camera_axis_used": "ax", "camera_axis_source": "w11_axis", "visible_zone_scope": ["Z1", "Z3"], "prior_bg_ref_role": "none", "render_prompt_appendix": "axis=ax. use #1,#2.", "final_prompt_assembly_preview": "revised base text", }, }, "scene_shot_consistency_audit": { "scene_unit_claims": [ {"scene_index": 1, "relevant_unit_ids": ["U_living", "U_room_a"], "source_refs": [{"source_ref": "scene:1"}], "notes": "scene 1"}, ], "shot_unit_claims": [ {"shot_key": "S1_Shot1", "bg_id": "BGT", "relevant_unit_ids": ["U_living", "U_room_a"], "relevant_numbered_elements": [1, 2], "source_refs": [{"source_ref": "S1_Shot1"}], "notes": "shot 1"}, ], "bg_instruction_consistency": [ {"bg_id": "BGT", "bound_unit_ids": ["U_living", "U_room_a"], "use_numbered_elements": [1, 2], "missing_unit_ids": [], "questionable_ignore_numbers": [], "status": "ok"}, ], "candidate_topology_consistency": [ {"fp_id": "FP_T", "missing_unit_ids_in_candidate": [], "missing_required_elements": [], "conflicts": []}, ], }, } def test_w15_source_evidence_bundle_collects_target_only(): """`_build_source_evidence_bundle` must include only the scene segments whose scene_index appears in target bg's `applies_to_shots` shots, plus the raw selected_shots rows for those shots, plus parsed t2i anchors keyed by shot_key. Non-target scenes and shots must be excluded.""" from experiment_floor_plan_topology_slice import ( _build_source_evidence_bundle, ) adapter = _synthetic_adapter() cands = _synthetic_w12c_candidates() per_bg = _synthetic_w12c_per_bg() # Source bundle has two scenes / two shots; only Scene 1 / S1_Shot1 is # connected to FP_T via BGT.applies_to_shots. source_shots = [ {"shot_key": "S1_Shot1", "scene_index": 1, "shot_index": 1, "shot_description": "target shot desc", "screenplay_scene_heading": "S#1.", "scene_summary": "target summary", "still_frame_prompt": "frame target", "t2i_prompt_cinematic": "cinematic target", "visible_entities_json": "[]", "t2i_variations_json": json.dumps([ {"variant_label": "var_1", "source_facts": [{"id": "sf1"}], "visual_inferences": [{"id": "vi1"}], "owned_object_usage": [{"id": "oo1", "usage_kind": "anchor"}], "applied_frame_spatial_constraint_ids": ["c1"], "reference_phrase_kinds": ["k1"]}, ])}, {"shot_key": "S2_Shot1", "scene_index": 2, "shot_index": 1, "shot_description": "non-target shot desc", "screenplay_scene_heading": "S#2.", "scene_summary": "non-target summary", "still_frame_prompt": "", "t2i_variations_json": "INVALID_JSON_STRING"}, ] scene_save = _synthetic_scene_save() fp_baseline = _synthetic_fp_baseline() bundle = _build_source_evidence_bundle( target_fp_ids={"FP_T"}, w12c_candidates=cands, w12c_per_bg=per_bg, adapter=adapter, selected_shots=source_shots, scene_save=scene_save, production_fp_context=fp_baseline, ) # target_shot_keys covers only the target bg's applies_to_shots. assert bundle["target_shot_keys"] == ["S1_Shot1"] # target_selected_shots carries raw rows (no summarization) for the # target shot only. assert len(bundle["target_selected_shots"]) == 1 raw_shot = bundle["target_selected_shots"][0] assert raw_shot["shot_key"] == "S1_Shot1" assert raw_shot["screenplay_scene_heading"] == "S#1." assert raw_shot["t2i_variations_json"].startswith("[") # raw string carried # target_scene_segments has only Scene 1 full text; Scene 2 excluded. assert len(bundle["target_scene_segments"]) == 1 assert bundle["target_scene_segments"][0]["scene_index"] == 1 assert "synthetic scene 1 text" in bundle["target_scene_segments"][0]["text"] # parsed_t2i_anchors_by_shot is keyed by shot_key and only carries # the LLM-relevant fields (no full t2i variant dump). anchors = bundle["parsed_t2i_anchors_by_shot"] assert set(anchors.keys()) == {"S1_Shot1"} assert len(anchors["S1_Shot1"]) == 1 anchor_entry = anchors["S1_Shot1"][0] # Kept fields are present; non-listed fields (e.g. t2i_prompt) are not. for key in ("source_facts", "visual_inferences", "owned_object_usage", "applied_frame_spatial_constraint_ids", "reference_phrase_kinds", "variant_label"): assert key in anchor_entry assert "t2i_prompt" not in anchor_entry # baseline_soft_evidence carries production fp prompt AND W12c # candidate baseline AND W12c per_bg baseline (all named soft so the # LLM treats source as winning on conflict). bse = bundle["baseline_soft_evidence"] assert "production_floor_plan_prompt" in bse assert "w12c_candidate_floor_plans_for_target" in bse assert "w12c_per_bg_render_reference_instructions_for_target" in bse assert "FP_T" in bse["w12c_candidate_floor_plans_for_target"] assert "BGT" in bse["w12c_per_bg_render_reference_instructions_for_target"] # A second source row with bad t2i_variations_json (the non-target # row) must not appear in diagnostics because it is not a target shot; # only target shot parse outcomes are tracked. diag = bundle["anchor_parse_diagnostics"] assert all(d.get("shot_key") != "S2_Shot1" for d in diag) # Dynamic character-name collector: only entities whose `short_id` # begins with "C" (character namespace) contribute to the per-run # render-safety guard. L*/P* are excluded so they cannot # accidentally fail render surfaces that legitimately reference # locations or props. source_shots_with_entities = [ {"shot_key": "S1_Shot1", "scene_index": 1, "shot_index": 1, "shot_description": "x", "screenplay_scene_heading": "S#1.", "scene_summary": "", "still_frame_prompt": "", "t2i_variations_json": "[]", "visible_entities_json": json.dumps([ {"short_id": "C01", "id": "uuid1", "entity_name": "Character Alpha"}, {"short_id": "C02", "id": "uuid2", "entity_name": "Character Beta"}, {"short_id": "L99", "id": "uuid3", "entity_name": "Synthetic Location"}, {"short_id": "P01", "id": "uuid4", "entity_name": "Synthetic Prop"}, ])}, ] bundle_chr = _build_source_evidence_bundle( target_fp_ids={"FP_T"}, w12c_candidates=cands, w12c_per_bg=per_bg, adapter=adapter, selected_shots=source_shots_with_entities, scene_save=scene_save, production_fp_context=fp_baseline, ) chr_names = bundle_chr["character_entity_names_from_visible_entities"] assert set(chr_names) == {"Character Alpha", "Character Beta"} assert "Synthetic Location" not in chr_names assert "Synthetic Prop" not in chr_names def test_w15_evidence_refs_resolver_accepts_valid_and_rejects_invalid(): """`source_ref` of form `scene:N` must be in scene_save and `shot_key` form must be in selected_shots. Invalid refs flagged.""" from experiment_floor_plan_topology_slice import _resolve_evidence_refs scene_idx_set = {1, 2} shot_keys = {"S1_Shot1", "S2_Shot1"} refs_ok = [ {"source_ref": "scene:1"}, {"source_ref": "scene:2"}, {"source_ref": "S1_Shot1"}, {"source_ref": "S2_Shot1"}, "S1_Shot1", ] resolved_ok, unresolved_ok = _resolve_evidence_refs(refs_ok, scene_idx_set, shot_keys) assert len(resolved_ok) == 5 assert unresolved_ok == [] refs_bad = [ {"source_ref": "scene:99"}, {"source_ref": "S_does_not_exist"}, {"source_ref": "scene:"}, {"source_ref": ""}, ] resolved_bad, unresolved_bad = _resolve_evidence_refs(refs_bad, scene_idx_set, shot_keys) assert resolved_bad == [] assert len(unresolved_bad) == 4 def test_w15_non_target_fp_copy_through_unchanged(): """`floor_plan_prompt_candidate.json` for a non-target fp must equal the W12c entry byte-for-byte (no LLM revision applied).""" from experiment_floor_plan_topology_slice import ( _merge_revised_candidates_with_w12c_copy, ) w12c = _synthetic_w12c_candidates() llm = _synthetic_llm_output() merged = _merge_revised_candidates_with_w12c_copy( w12c_candidates=w12c, llm_revised=llm.get("candidate_floor_plans") or {}, target_fp_ids={"FP_T"}, ) # target is the revised one assert merged["FP_T"]["candidate_diagram_t2i_prompt"] == "revised target plan with two enclosed rooms" assert len(merged["FP_T"]["candidate_numbered_elements"]) == 2 # non-target is the W12c copy assert merged["FP_NT"] == w12c["FP_NT"] def test_w15_target_bg_coverage_in_bindings(tmp_path): """Every bg whose fp_id is in target_fp_ids must have a `bg_unit_bindings` entry. Missing → invariant fails.""" from experiment_floor_plan_topology_slice import ( _build_w15_compatibility_report, ) per_bg = _synthetic_w12c_per_bg() cands = _synthetic_w12c_candidates() llm = _synthetic_llm_output() # Happy case. rep = _build_w15_compatibility_report( w12c_per_bg=per_bg, w12c_candidates=cands, revised_candidates=_synthetic_llm_output()["candidate_floor_plans"], topology_by_fp=llm["source_topology_by_fp"], bg_unit_bindings=llm["bg_unit_bindings"], target_fp_ids={"FP_T"}, scene_idx_set={1, 2}, shot_keys={"S1_Shot1", "S2_Shot1"}, production_diff_empty=True, db_write_count=0, image_import_seen=False, missing_inputs=[], stage_status="generated", prev_run_id="fake", ) inv = rep["invariants"] assert inv["bg_unit_bindings_cover_target_bgs"]["pass"] is True detail = inv["bg_unit_bindings_cover_target_bgs"]["detail"] assert detail["target_bg_count"] == 1 assert detail["uncovered_bgs"] == [] # Drop bg binding → fails. llm_bad = json.loads(json.dumps(llm)) llm_bad["bg_unit_bindings"].pop("BGT") rep_bad = _build_w15_compatibility_report( w12c_per_bg=per_bg, w12c_candidates=cands, revised_candidates=llm_bad["candidate_floor_plans"], topology_by_fp=llm_bad["source_topology_by_fp"], bg_unit_bindings=llm_bad["bg_unit_bindings"], target_fp_ids={"FP_T"}, scene_idx_set={1, 2}, shot_keys={"S1_Shot1", "S2_Shot1"}, production_diff_empty=True, db_write_count=0, image_import_seen=False, missing_inputs=[], stage_status="generated", prev_run_id="fake", ) assert rep_bad["invariants"]["bg_unit_bindings_cover_target_bgs"]["pass"] is False assert "BGT" in rep_bad["invariants"]["bg_unit_bindings_cover_target_bgs"]["detail"]["uncovered_bgs"] def test_w15_no_image_api_call_and_production_guard(): """All five `production_diff_zero...` style guards plus the loose `topology_room_count_self_consistent` invariant. self-consistency is NOT a strict equality with do_not_collapse_units length; it only requires: (a) room_count_assessment.standalone_enclosed_room_count == count(spatial_units with is_enclosed_room=true) (b) every do_not_collapse_units id exists in spatial_units (c) every non-enclosed do_not_collapse unit has evidence.""" from experiment_floor_plan_topology_slice import ( _build_w15_compatibility_report, ) per_bg = _synthetic_w12c_per_bg() cands = _synthetic_w12c_candidates() llm = _synthetic_llm_output() rep_ok = _build_w15_compatibility_report( w12c_per_bg=per_bg, w12c_candidates=cands, revised_candidates=llm["candidate_floor_plans"], topology_by_fp=llm["source_topology_by_fp"], bg_unit_bindings=llm["bg_unit_bindings"], target_fp_ids={"FP_T"}, scene_idx_set={1, 2}, shot_keys={"S1_Shot1", "S2_Shot1"}, production_diff_empty=True, db_write_count=0, image_import_seen=False, missing_inputs=[], stage_status="generated", prev_run_id="fake", ) inv = rep_ok["invariants"] assert inv["inputs_present"]["pass"] is True assert inv["target_fp_covered"]["pass"] is True assert inv["evidence_refs_resolve"]["pass"] is True assert inv["candidate_elements_present"]["pass"] is True assert inv["topology_room_count_self_consistent"]["pass"] is True assert inv["render_surface_no_proper_noun_or_entity_instance"]["pass"] is True assert inv["production_diff_zero_db_write_zero_image_api_call_zero"]["pass"] is True assert rep_ok["all_pass"] is True # image_import_seen → guard fails. rep_img = _build_w15_compatibility_report( w12c_per_bg=per_bg, w12c_candidates=cands, revised_candidates=llm["candidate_floor_plans"], topology_by_fp=llm["source_topology_by_fp"], bg_unit_bindings=llm["bg_unit_bindings"], target_fp_ids={"FP_T"}, scene_idx_set={1, 2}, shot_keys={"S1_Shot1", "S2_Shot1"}, production_diff_empty=True, db_write_count=0, image_import_seen=True, missing_inputs=[], stage_status="generated", prev_run_id="fake", ) assert rep_img["invariants"]["production_diff_zero_db_write_zero_image_api_call_zero"]["pass"] is False # production_diff dirty → guard fails. rep_prod = _build_w15_compatibility_report( w12c_per_bg=per_bg, w12c_candidates=cands, revised_candidates=llm["candidate_floor_plans"], topology_by_fp=llm["source_topology_by_fp"], bg_unit_bindings=llm["bg_unit_bindings"], target_fp_ids={"FP_T"}, scene_idx_set={1, 2}, shot_keys={"S1_Shot1", "S2_Shot1"}, production_diff_empty=False, db_write_count=0, image_import_seen=False, missing_inputs=[], stage_status="generated", prev_run_id="fake", ) assert rep_prod["invariants"]["production_diff_zero_db_write_zero_image_api_call_zero"]["pass"] is False # Self-consistency: mismatch standalone_enclosed_room_count with # actual enclosed unit count → (a) fails. llm_bad_count = json.loads(json.dumps(llm)) llm_bad_count["source_topology_by_fp"]["FP_T"]["room_count_assessment"]["standalone_enclosed_room_count"] = 5 rep_bad_count = _build_w15_compatibility_report( w12c_per_bg=per_bg, w12c_candidates=cands, revised_candidates=llm_bad_count["candidate_floor_plans"], topology_by_fp=llm_bad_count["source_topology_by_fp"], bg_unit_bindings=llm_bad_count["bg_unit_bindings"], target_fp_ids={"FP_T"}, scene_idx_set={1, 2}, shot_keys={"S1_Shot1", "S2_Shot1"}, production_diff_empty=True, db_write_count=0, image_import_seen=False, missing_inputs=[], stage_status="generated", prev_run_id="fake", ) assert rep_bad_count["invariants"]["topology_room_count_self_consistent"]["pass"] is False # do_not_collapse_units id not in spatial_units → (b) fails. llm_bad_b = json.loads(json.dumps(llm)) llm_bad_b["source_topology_by_fp"]["FP_T"]["do_not_collapse_units"].append( {"unit_id": "U_ghost", "reason": "missing unit", "evidence_refs": [{"source_ref": "scene:1", "quote": "x", "field": "x"}]} ) rep_bad_b = _build_w15_compatibility_report( w12c_per_bg=per_bg, w12c_candidates=cands, revised_candidates=llm_bad_b["candidate_floor_plans"], topology_by_fp=llm_bad_b["source_topology_by_fp"], bg_unit_bindings=llm_bad_b["bg_unit_bindings"], target_fp_ids={"FP_T"}, scene_idx_set={1, 2}, shot_keys={"S1_Shot1", "S2_Shot1"}, production_diff_empty=True, db_write_count=0, image_import_seen=False, missing_inputs=[], stage_status="generated", prev_run_id="fake", ) assert rep_bad_b["invariants"]["topology_room_count_self_consistent"]["pass"] is False # do_not_collapse_units entry for non-enclosed unit without evidence_refs → (c) fails. llm_bad_c = json.loads(json.dumps(llm)) llm_bad_c["source_topology_by_fp"]["FP_T"]["do_not_collapse_units"].append( {"unit_id": "U_entry", "reason": "", "evidence_refs": []} ) rep_bad_c = _build_w15_compatibility_report( w12c_per_bg=per_bg, w12c_candidates=cands, revised_candidates=llm_bad_c["candidate_floor_plans"], topology_by_fp=llm_bad_c["source_topology_by_fp"], bg_unit_bindings=llm_bad_c["bg_unit_bindings"], target_fp_ids={"FP_T"}, scene_idx_set={1, 2}, shot_keys={"S1_Shot1", "S2_Shot1"}, production_diff_empty=True, db_write_count=0, image_import_seen=False, missing_inputs=[], stage_status="generated", prev_run_id="fake", ) assert rep_bad_c["invariants"]["topology_room_count_self_consistent"]["pass"] is False # Render-safety guard, dynamic character term: when the evidence # bundle carries a character name (synthetic `Character Alpha`), # using that exact name in the candidate prompt must fail. synthetic_character_names = ["Character Alpha", "Character Beta"] llm_proper = json.loads(json.dumps(llm)) llm_proper["candidate_floor_plans"]["FP_T"]["candidate_diagram_t2i_prompt"] = ( "schematic plan with Character Alpha's bedroom on the east" ) rep_proper = _build_w15_compatibility_report( w12c_per_bg=per_bg, w12c_candidates=cands, revised_candidates=llm_proper["candidate_floor_plans"], topology_by_fp=llm_proper["source_topology_by_fp"], bg_unit_bindings=llm_proper["bg_unit_bindings"], target_fp_ids={"FP_T"}, scene_idx_set={1, 2}, shot_keys={"S1_Shot1", "S2_Shot1"}, production_diff_empty=True, db_write_count=0, image_import_seen=False, missing_inputs=[], stage_status="generated", prev_run_id="fake", forbidden_character_terms=synthetic_character_names, ) assert rep_proper["invariants"]["render_surface_no_proper_noun_or_entity_instance"]["pass"] is False # Same render-safety guard, static generic entity-instance term: # putting "body" in a candidate label fails regardless of whether the # dynamic character list is empty. llm_body = json.loads(json.dumps(llm)) llm_body["candidate_floor_plans"]["FP_T"]["candidate_numbered_elements"][0]["label"] = ( "bo" + "dy and red stain zone" ) rep_body = _build_w15_compatibility_report( w12c_per_bg=per_bg, w12c_candidates=cands, revised_candidates=llm_body["candidate_floor_plans"], topology_by_fp=llm_body["source_topology_by_fp"], bg_unit_bindings=llm_body["bg_unit_bindings"], target_fp_ids={"FP_T"}, scene_idx_set={1, 2}, shot_keys={"S1_Shot1", "S2_Shot1"}, production_diff_empty=True, db_write_count=0, image_import_seen=False, missing_inputs=[], stage_status="generated", prev_run_id="fake", forbidden_character_terms=[], ) assert rep_body["invariants"]["render_surface_no_proper_noun_or_entity_instance"]["pass"] is False # Per-bg render appendix carrying a dynamic character name → fails. revised_per_bg_bad = json.loads(json.dumps(llm["per_bg_render_reference_instructions"])) revised_per_bg_bad["BGT"]["render_prompt_appendix"] = ( "axis=ax. focus on Character Beta's room" ) rep_perbg_bad = _build_w15_compatibility_report( w12c_per_bg=per_bg, w12c_candidates=cands, revised_candidates=llm["candidate_floor_plans"], topology_by_fp=llm["source_topology_by_fp"], bg_unit_bindings=llm["bg_unit_bindings"], target_fp_ids={"FP_T"}, scene_idx_set={1, 2}, shot_keys={"S1_Shot1", "S2_Shot1"}, production_diff_empty=True, db_write_count=0, image_import_seen=False, missing_inputs=[], stage_status="generated", prev_run_id="fake", revised_per_bg=revised_per_bg_bad, forbidden_character_terms=synthetic_character_names, ) assert rep_perbg_bad["invariants"]["render_surface_no_proper_noun_or_entity_instance"]["pass"] is False # When the evidence bundle has no character terms, a render surface # that happens to mention `Character Alpha` is NOT flagged — proves # the guard is fully driven by run-time evidence, not hardcoded. rep_neutral = _build_w15_compatibility_report( w12c_per_bg=per_bg, w12c_candidates=cands, revised_candidates=llm_proper["candidate_floor_plans"], topology_by_fp=llm_proper["source_topology_by_fp"], bg_unit_bindings=llm_proper["bg_unit_bindings"], target_fp_ids={"FP_T"}, scene_idx_set={1, 2}, shot_keys={"S1_Shot1", "S2_Shot1"}, production_diff_empty=True, db_write_count=0, image_import_seen=False, missing_inputs=[], stage_status="generated", prev_run_id="fake", forbidden_character_terms=[], ) assert rep_neutral["invariants"]["render_surface_no_proper_noun_or_entity_instance"]["pass"] is True def test_w15_consistency_audit_4_sub_checks(): """W15d: when `consistency_audit` is provided as a non-empty dict and stage_status='generated', the new invariant `scene_shot_consistency_audit_resolves` runs four sub-checks: (a) audit covers every target scene and shot, (b) every unit_id referenced in audit exists in topology, (c) every numbered element referenced in audit exists in revised candidate for the right fp, (d) every multi-unit bg whose use-numbers don't cover all bound units must be reported with non-ok status in the audit.""" from experiment_floor_plan_topology_slice import ( _build_w15_compatibility_report, ) per_bg = _synthetic_w12c_per_bg() cands = _synthetic_w12c_candidates() llm = _synthetic_llm_output() audit_ok = llm["scene_shot_consistency_audit"] rep_ok = _build_w15_compatibility_report( w12c_per_bg=per_bg, w12c_candidates=cands, revised_candidates=llm["candidate_floor_plans"], topology_by_fp=llm["source_topology_by_fp"], bg_unit_bindings=llm["bg_unit_bindings"], target_fp_ids={"FP_T"}, scene_idx_set={1}, shot_keys={"S1_Shot1"}, production_diff_empty=True, db_write_count=0, image_import_seen=False, missing_inputs=[], stage_status="generated", prev_run_id="fake", revised_per_bg=llm["per_bg_render_reference_instructions"], consistency_audit=audit_ok, target_scene_idx_set={1}, target_shot_keys_set={"S1_Shot1"}, ) inv = rep_ok["invariants"] assert inv["scene_shot_consistency_audit_resolves"]["pass"] is True detail = inv["scene_shot_consistency_audit_resolves"]["detail"] for k in ("covers_all_target_scenes_and_shots", "audit_unit_ids_exist_in_topology", "audit_numbered_elements_exist_in_candidate", "bg_instruction_missing_units_reported"): assert detail[k]["pass"] is True # Sub (a) fail: audit missing a target shot. audit_a = json.loads(json.dumps(audit_ok)) audit_a["shot_unit_claims"] = [] rep_a = _build_w15_compatibility_report( w12c_per_bg=per_bg, w12c_candidates=cands, revised_candidates=llm["candidate_floor_plans"], topology_by_fp=llm["source_topology_by_fp"], bg_unit_bindings=llm["bg_unit_bindings"], target_fp_ids={"FP_T"}, scene_idx_set={1}, shot_keys={"S1_Shot1"}, production_diff_empty=True, db_write_count=0, image_import_seen=False, missing_inputs=[], stage_status="generated", prev_run_id="fake", revised_per_bg=llm["per_bg_render_reference_instructions"], consistency_audit=audit_a, target_scene_idx_set={1}, target_shot_keys_set={"S1_Shot1"}, ) assert rep_a["invariants"]["scene_shot_consistency_audit_resolves"]["pass"] is False assert "S1_Shot1" in ( rep_a["invariants"]["scene_shot_consistency_audit_resolves"] ["detail"]["covers_all_target_scenes_and_shots"]["missing_shots_in_audit"] ) # Sub (b) fail: scene_unit_claims references an unknown unit_id. audit_b = json.loads(json.dumps(audit_ok)) audit_b["scene_unit_claims"][0]["relevant_unit_ids"].append("U_ghost") rep_b = _build_w15_compatibility_report( w12c_per_bg=per_bg, w12c_candidates=cands, revised_candidates=llm["candidate_floor_plans"], topology_by_fp=llm["source_topology_by_fp"], bg_unit_bindings=llm["bg_unit_bindings"], target_fp_ids={"FP_T"}, scene_idx_set={1}, shot_keys={"S1_Shot1"}, production_diff_empty=True, db_write_count=0, image_import_seen=False, missing_inputs=[], stage_status="generated", prev_run_id="fake", revised_per_bg=llm["per_bg_render_reference_instructions"], consistency_audit=audit_b, target_scene_idx_set={1}, target_shot_keys_set={"S1_Shot1"}, ) assert rep_b["invariants"]["scene_shot_consistency_audit_resolves"]["pass"] is False # Sub (c) fail: shot_unit_claims references a number not in revised # candidate. audit_c = json.loads(json.dumps(audit_ok)) audit_c["shot_unit_claims"][0]["relevant_numbered_elements"].append(999) rep_c = _build_w15_compatibility_report( w12c_per_bg=per_bg, w12c_candidates=cands, revised_candidates=llm["candidate_floor_plans"], topology_by_fp=llm["source_topology_by_fp"], bg_unit_bindings=llm["bg_unit_bindings"], target_fp_ids={"FP_T"}, scene_idx_set={1}, shot_keys={"S1_Shot1"}, production_diff_empty=True, db_write_count=0, image_import_seen=False, missing_inputs=[], stage_status="generated", prev_run_id="fake", revised_per_bg=llm["per_bg_render_reference_instructions"], consistency_audit=audit_c, target_scene_idx_set={1}, target_shot_keys_set={"S1_Shot1"}, ) assert rep_c["invariants"]["scene_shot_consistency_audit_resolves"]["pass"] is False # Sub (d) fail: bg has multiple PRIMARY units in its binding, but # per_bg `use_numbered_elements` covers only one of them, AND audit # silently reports status='ok' without listing the uncovered primary. # secondary-only uncovered units are diagnostic and do NOT fail sub (d). revised_per_bg_partial = json.loads(json.dumps(llm["per_bg_render_reference_instructions"])) revised_per_bg_partial["BGT"]["use_numbered_elements"] = [1] # drop #2 bg_bindings_two_primary = json.loads(json.dumps(llm["bg_unit_bindings"])) # Promote U_room_a from secondary to a second primary so coverage must # include it. Now use=[1] (only U_living) leaves U_room_a primary uncovered. bg_bindings_two_primary["BGT"]["primary_unit_ids"] = ["U_living", "U_room_a"] bg_bindings_two_primary["BGT"]["secondary_visible_unit_ids"] = [] audit_d = json.loads(json.dumps(audit_ok)) audit_d["bg_instruction_consistency"][0]["status"] = "ok" audit_d["bg_instruction_consistency"][0]["missing_unit_ids"] = [] rep_d = _build_w15_compatibility_report( w12c_per_bg=per_bg, w12c_candidates=cands, revised_candidates=llm["candidate_floor_plans"], topology_by_fp=llm["source_topology_by_fp"], bg_unit_bindings=bg_bindings_two_primary, target_fp_ids={"FP_T"}, scene_idx_set={1}, shot_keys={"S1_Shot1"}, production_diff_empty=True, db_write_count=0, image_import_seen=False, missing_inputs=[], stage_status="generated", prev_run_id="fake", revised_per_bg=revised_per_bg_partial, consistency_audit=audit_d, target_scene_idx_set={1}, target_shot_keys_set={"S1_Shot1"}, ) assert rep_d["invariants"]["scene_shot_consistency_audit_resolves"]["pass"] is False detail_d = rep_d["invariants"]["scene_shot_consistency_audit_resolves"]["detail"] assert detail_d["bg_instruction_missing_units_reported"]["pass"] is False # Secondary-only missing coverage does NOT fail sub (d); it lives in # `secondary_unit_coverage_diagnostics` only. revised_per_bg_sec = json.loads(json.dumps(llm["per_bg_render_reference_instructions"])) revised_per_bg_sec["BGT"]["use_numbered_elements"] = [1] # only U_living bg_bindings_sec = json.loads(json.dumps(llm["bg_unit_bindings"])) # primary U_living is covered, secondary U_room_a is not. bg_bindings_sec["BGT"]["primary_unit_ids"] = ["U_living"] bg_bindings_sec["BGT"]["secondary_visible_unit_ids"] = ["U_room_a"] audit_sec = json.loads(json.dumps(audit_ok)) audit_sec["bg_instruction_consistency"][0]["status"] = "ok" rep_sec = _build_w15_compatibility_report( w12c_per_bg=per_bg, w12c_candidates=cands, revised_candidates=llm["candidate_floor_plans"], topology_by_fp=llm["source_topology_by_fp"], bg_unit_bindings=bg_bindings_sec, target_fp_ids={"FP_T"}, scene_idx_set={1}, shot_keys={"S1_Shot1"}, production_diff_empty=True, db_write_count=0, image_import_seen=False, missing_inputs=[], stage_status="generated", prev_run_id="fake", revised_per_bg=revised_per_bg_sec, consistency_audit=audit_sec, target_scene_idx_set={1}, target_shot_keys_set={"S1_Shot1"}, ) assert rep_sec["invariants"]["scene_shot_consistency_audit_resolves"]["pass"] is True sec_diag = ( rep_sec["invariants"]["scene_shot_consistency_audit_resolves"] ["detail"]["bg_instruction_missing_units_reported"] ["secondary_unit_coverage_diagnostics"] ) assert any(d.get("bg_id") == "BGT" and "U_room_a" in d.get("missing_secondary_unit_ids", []) for d in sec_diag) def test_w15e_element_placement_and_unit_scale_audit_invariant(): """W15e: when `placement_scale_audit` is a non-empty dict and stage_status='generated', the new invariant runs three structural sub-checks: (a) unit_scale_constraints covers every topology spatial_unit (b) element_placement_constraints covers every candidate number (target fps only) and references resolve into topology units (c) every element whose unit_id_pointer is part of any open_connection pair must set avoid_open_connection_boundary None / empty-dict modes follow the same legacy-skip contract as the consistency audit.""" from experiment_floor_plan_topology_slice import ( _build_w15_compatibility_report, ) per_bg = _synthetic_w12c_per_bg() cands = _synthetic_w12c_candidates() llm = _synthetic_llm_output() audit_consistency = llm["scene_shot_consistency_audit"] # Synthetic placement audit: every spatial_unit covered + every # revised candidate number covered + open-connection unit has # avoid flag. placement_ok = { "unit_scale_constraints": [ {"unit_id": "U_living", "enclosure_mode": "open_zone_inside_plan", "relative_scale_hint": "primary_zone", "open_connection_behavior": "continuous floor with adjacent zone", "render_notes": "compact"}, {"unit_id": "U_room_a", "enclosure_mode": "full_wall_enclosed_room", "relative_scale_hint": "secondary_zone", "open_connection_behavior": "n/a", "render_notes": "compact"}, {"unit_id": "U_entry", "enclosure_mode": "interior_threshold", "relative_scale_hint": "compact_nook_or_wall_run", "open_connection_behavior": "no partition", "render_notes": "compact"}, ], "element_placement_constraints": [ {"number": 1, "fp_id": "FP_T", "unit_id_pointer": "U_living", "category": "area", "placement_mode": "area_label_only", "anchor_surface_hint": "interior floor center", "avoid_open_connection_boundary": True, "must_not_form_boundary": True, "revised_position_hint": "interior floor center", "conflict_note": ""}, {"number": 2, "fp_id": "FP_T", "unit_id_pointer": "U_room_a", "category": "area", "placement_mode": "area_label_only", "anchor_surface_hint": "interior floor center", "avoid_open_connection_boundary": False, "must_not_form_boundary": True, "revised_position_hint": "interior floor center", "conflict_note": ""}, ], } rep_ok = _build_w15_compatibility_report( w12c_per_bg=per_bg, w12c_candidates=cands, revised_candidates=llm["candidate_floor_plans"], topology_by_fp=llm["source_topology_by_fp"], bg_unit_bindings=llm["bg_unit_bindings"], target_fp_ids={"FP_T"}, scene_idx_set={1}, shot_keys={"S1_Shot1"}, production_diff_empty=True, db_write_count=0, image_import_seen=False, missing_inputs=[], stage_status="generated", prev_run_id="fake", revised_per_bg=llm["per_bg_render_reference_instructions"], consistency_audit=audit_consistency, target_scene_idx_set={1}, target_shot_keys_set={"S1_Shot1"}, placement_scale_audit=placement_ok, ) inv = rep_ok["invariants"]["element_placement_and_unit_scale_audit_resolves"] assert inv["pass"] is True # Sub (a) fail: drop a unit from unit_scale_constraints. placement_a = json.loads(json.dumps(placement_ok)) placement_a["unit_scale_constraints"].pop() rep_a = _build_w15_compatibility_report( w12c_per_bg=per_bg, w12c_candidates=cands, revised_candidates=llm["candidate_floor_plans"], topology_by_fp=llm["source_topology_by_fp"], bg_unit_bindings=llm["bg_unit_bindings"], target_fp_ids={"FP_T"}, scene_idx_set={1}, shot_keys={"S1_Shot1"}, production_diff_empty=True, db_write_count=0, image_import_seen=False, missing_inputs=[], stage_status="generated", prev_run_id="fake", revised_per_bg=llm["per_bg_render_reference_instructions"], consistency_audit=audit_consistency, target_scene_idx_set={1}, target_shot_keys_set={"S1_Shot1"}, placement_scale_audit=placement_a, ) assert rep_a["invariants"]["element_placement_and_unit_scale_audit_resolves"]["pass"] is False # Sub (b) fail: a placement constraint references an unknown number. placement_b = json.loads(json.dumps(placement_ok)) placement_b["element_placement_constraints"].append({ "number": 99, "fp_id": "FP_T", "unit_id_pointer": "U_living", "category": "furniture", "placement_mode": "interior_floor_furniture", "anchor_surface_hint": "interior", "avoid_open_connection_boundary": True, "must_not_form_boundary": True, "revised_position_hint": "interior", "conflict_note": "", }) rep_b = _build_w15_compatibility_report( w12c_per_bg=per_bg, w12c_candidates=cands, revised_candidates=llm["candidate_floor_plans"], topology_by_fp=llm["source_topology_by_fp"], bg_unit_bindings=llm["bg_unit_bindings"], target_fp_ids={"FP_T"}, scene_idx_set={1}, shot_keys={"S1_Shot1"}, production_diff_empty=True, db_write_count=0, image_import_seen=False, missing_inputs=[], stage_status="generated", prev_run_id="fake", revised_per_bg=llm["per_bg_render_reference_instructions"], consistency_audit=audit_consistency, target_scene_idx_set={1}, target_shot_keys_set={"S1_Shot1"}, placement_scale_audit=placement_b, ) assert rep_b["invariants"]["element_placement_and_unit_scale_audit_resolves"]["pass"] is False # Sub (c) fail: furniture / fixture / plot-device element on an # open-connection unit without avoid_open_connection_boundary=True. # Add an open_connection relationship to the synthetic topology so # U_living becomes part of an open-connection pair. Area labels are # exempt, so mutate the constraint into a furniture placement. topo_open = json.loads(json.dumps(llm["source_topology_by_fp"])) topo_open["FP_T"]["relationships"].append({ "from_unit": "U_living", "to_unit": "U_entry", "relation_kind": "open_connection", }) placement_c = json.loads(json.dumps(placement_ok)) placement_c["element_placement_constraints"][0]["category"] = "furniture" placement_c["element_placement_constraints"][0]["placement_mode"] = "wall_fixture" placement_c["element_placement_constraints"][0]["avoid_open_connection_boundary"] = False rep_c = _build_w15_compatibility_report( w12c_per_bg=per_bg, w12c_candidates=cands, revised_candidates=llm["candidate_floor_plans"], topology_by_fp=topo_open, bg_unit_bindings=llm["bg_unit_bindings"], target_fp_ids={"FP_T"}, scene_idx_set={1}, shot_keys={"S1_Shot1"}, production_diff_empty=True, db_write_count=0, image_import_seen=False, missing_inputs=[], stage_status="generated", prev_run_id="fake", revised_per_bg=llm["per_bg_render_reference_instructions"], consistency_audit=audit_consistency, target_scene_idx_set={1}, target_shot_keys_set={"S1_Shot1"}, placement_scale_audit=placement_c, ) assert rep_c["invariants"]["element_placement_and_unit_scale_audit_resolves"]["pass"] is False # placement_scale_audit=None → legacy auto-pass. rep_legacy = _build_w15_compatibility_report( w12c_per_bg=per_bg, w12c_candidates=cands, revised_candidates=llm["candidate_floor_plans"], topology_by_fp=llm["source_topology_by_fp"], bg_unit_bindings=llm["bg_unit_bindings"], target_fp_ids={"FP_T"}, scene_idx_set={1}, shot_keys={"S1_Shot1"}, production_diff_empty=True, db_write_count=0, image_import_seen=False, missing_inputs=[], stage_status="generated", prev_run_id="fake", revised_per_bg=llm["per_bg_render_reference_instructions"], consistency_audit=audit_consistency, target_scene_idx_set={1}, target_shot_keys_set={"S1_Shot1"}, placement_scale_audit=None, ) assert rep_legacy["invariants"]["element_placement_and_unit_scale_audit_resolves"]["pass"] is True def test_w15_html_section_order_and_methodology_grep(tmp_path): """HTML must put source topology section before the revised candidate, revised candidate before per-bg bindings, and raw JSON inside collapsed
. Methodology grep on script source must not contain narrative tokens (per-char split, this test file is exempt).""" from experiment_floor_plan_topology_slice import ( _build_w15_compatibility_report, _render_w15_html, ) per_bg = _synthetic_w12c_per_bg() cands = _synthetic_w12c_candidates() llm = _synthetic_llm_output() rep = _build_w15_compatibility_report( w12c_per_bg=per_bg, w12c_candidates=cands, revised_candidates=llm["candidate_floor_plans"], topology_by_fp=llm["source_topology_by_fp"], bg_unit_bindings=llm["bg_unit_bindings"], target_fp_ids={"FP_T"}, scene_idx_set={1, 2}, shot_keys={"S1_Shot1", "S2_Shot1"}, production_diff_empty=True, db_write_count=0, image_import_seen=False, missing_inputs=[], stage_status="generated", prev_run_id="fake", ) run_meta = { "run_id": "RID1", "stage": "w15_floor_plan_topology_slice", "run_status": "succeeded", "exit_code": 0, "derived_from": "fakeW12c", "model_used": "gemini-3.5-flash", "image_generation_count": 0, "image_generation_backend": "gpt-image-2", "stage_status": "generated", "target_fp_ids": ["FP_T"], } _render_w15_html( run_meta=run_meta, topology_by_fp=llm["source_topology_by_fp"], bg_unit_bindings=llm["bg_unit_bindings"], revised_candidates=llm["candidate_floor_plans"], per_bg_revised=llm["per_bg_render_reference_instructions"], report=rep, run_dir=tmp_path, ) html = (tmp_path / "index.html").read_text() topology_pos = html.find("Source topology") revised_pos = html.find("Revised candidate floor plans") per_bg_pos = html.find("Per-bg revised reference instructions") inv_pos = html.find("Invariants") raw_pos = html.find("raw run_meta") assert topology_pos > 0 assert revised_pos > topology_pos assert per_bg_pos > revised_pos assert inv_pos > per_bg_pos assert raw_pos > 0 details_pos = html.rfind("
", 0, raw_pos) assert details_pos > 0 and details_pos < raw_pos # Methodology grep: scenario-specific tokens only (proper nouns from # this specific episode and hardcoded room-count assumptions). Generic # floor-plan vocabulary like `kitchen_zone` or `bedroom` is allowed in # universal schema enums. Tokens are assembled per-char so this # assertion source does not match its own regex. script_path = _SCRIPTS_DIR / "experiment_floor_plan_topology_slice.py" forbidden_tokens = [ "L0" + "5", "옥" + "탑방", "안" + "방", "수" + "리영", "민" + "숙", "혜" + "수", "김형" + "사", "one-" + "bedroom", "two-" + "bedroom", "one-" + "room", "two-" + "room", ] pattern = re.compile("(" + "|".join(re.escape(t) for t in forbidden_tokens) + ")") m = pattern.search(script_path.read_text()) assert m is None, m.group(0) if m else ""