Background Semantic Extractor — generic experiment, minimal 2-stage brief (sample fixture: l05_rooftop_interior)

W1 dry-run (no LLM call) run_id=20260524_1620_b843c3 / plan_version=bsx_w1 / model=model_for_future_generate / 2-stage flow: world brief → minimal spatial brief

§0. Why minimal + why 2-stage + why an Evidence Filter

Minimal means we extract the smallest evidence-backed brief that keeps T2I/I2I background generation honest. Over-extraction of door/window/furniture inventory tends to hurt downstream shot generation, so the schema avoids object counting and rigid inventory.

2-stage means Stage A compresses the world grounding (era, geography, technology, social/economic tone, genre mood) and explicitly drops plot/lore that does not change visual background. Stage B uses only the short world hints from Stage A, never the full Stage A dict — enforced by the structural input-contract check.

Full SourceBundle was read only by the Evidence Filter. Downstream background briefs use only filtered evidence. The Filter is the single gateway over the raw SourceBundle; the downstream world / spatial prompt builders only receive the filtered BackgroundEvidencePack, never the bundle.

§1. SourceBundle summary

total sources = 79 / char_budget_estimate = 52919 (fail-closed at W2 if exceeds model context limit)

§1.5. Evidence Filter — BackgroundEvidencePack (gateway over the SourceBundle)

top-level: schema_version, source_bundle_ref, evidence_items, coverage_notes, rejected_or_irrelevant_summary

evidence_item use_for enum: world_context, place_identity, spatial_scale_or_layout_feel, state_variation, visual_default_or_avoid

Caps: evidence_items ≤ 40; why_background_relevant ≤ 180 chars; pass_down_hint ≤ 160 chars; coverage_notes ≤ 8; rejected_or_irrelevant_summary ≤ 8. Over-cap is FAIL; original pack preserved verbatim.

placeholder pack Filter LLM not called in dry-run; downstream prompts use this placeholder.

§2. Stage A — BackgroundWorldBrief schema

top-level: schema_version, bundle_id, world_background_brief, confidence_band

world_background_brief subfields: era_and_time_period, geographic_cultural_grounding, technology_and_material_baseline, social_economic_visual_tone, genre_mood_constraints, background_relevant_do_not_assume, irrelevant_or_do_not_pass_down, evidence_refs

§3. Stage B — MinimalSpatialBrief schema

top-level: schema_version, bundle_id, world_brief_ref, world_hints_for_background, place_identity, continuity_groups, essential_spatial_relations, visual_anchors, state_variations, generation_notes, confidence_band

visual_anchor.base_plate_role enum: identity_anchor, optional, avoid (identity_anchor means base plate should be informed by this anchor, not that the anchor is rendered as an inventory item)

§4. Prompt preview

evidence_filter_system.txt (first 600 chars)

You are a background generation evidence filter.

Read the provided SourceBundle in full. Extract only the evidence items that a production designer would actually need to generate a background image — identity of place, scale and layout feel, world grounding, state variations, and visual defaults to avoid. Drop plot facts, character relationships, mythology lore, event chronology, and anything that does not change the visual background. The downstream world brief and spatial brief stages will receive ONLY this pack — never the raw SourceBundle.

Output a single BackgroundEvidencePack JSON obj

world_system.txt (first 600 chars)

You are a background world grounding assistant.

Read the provided SourceBundle and produce a single BackgroundWorldBrief JSON object that captures only the world grounding needed to keep T2I background generation from defaulting to luxury, Western, futuristic, or fantasy aesthetics.

Extract only background-generation-relevant world grounding. Do not pass plot lore unless it changes visual background assumptions. Put plot facts, character relationships, mythology lore, and event chronology that do NOT change visual background into `irrelevant_or_do_not_pass_down` so the downstream stage skips

spatial_system.txt (first 600 chars)

You are a minimal spatial brief assistant for T2I/I2I background generation.

Read the provided SourceBundle plus the short world hints from Stage A. Produce a single MinimalSpatialBrief JSON object.

Use only the compact world hints from Stage A, not the full source world analysis. Do not re-extract plot or lore that Stage A marked as irrelevant.

Keep the brief compact. Avoid over specification. Do not inventory objects or count doors / windows / furniture. Do not assume the location is a house, room, set, building, or interior. Extract spatial decomposition that works for any place type — c

§5. Validation

planned only checker not executed (--generate off).

• 1. stage_a_schema_required_fields (4 top-level + 8 wbb subfields)
• 2. stage_b_schema_required_fields (11 top-level + per-item required)
• 3. stage_a_evidence_refs_non_empty (world_background_brief)
• 4. stage_b_evidence_refs_non_empty (place_identity + 4 list groups)
• 5. source_ref_resolvable (both stages, matches SourceBundle)
• 6. quote_exact_containment (both stages, source.text contains exact quote)
• 7. confidence_enum (both stages, trusted/plausible/weak/unknown)
• 8. no_forbidden_fields (topology fields + 'count' JSON key + Stage B embedded world dict)
• 9. minimality_caps (array length + string length caps; equal-to-cap PASS, over-cap FAIL — NO auto-truncate/slice/drop/chunk/auto-shorten; original brief content preserved verbatim)
• 10. stage_b_input_contract (structural: world keys exactly {world_brief_ref, world_hints_for_background}; no embedded full Stage A dict; no irrelevant_or_do_not_pass_down)

checker does NOT classify semantic meaning, does NOT scan brief content for forbidden words, and does NOT apply any lexicon/regex/boundary heuristic. checker does NOT truncate, slice, drop, chunk, or auto-shorten any brief field or source text — cap violations are recorded as fail entries; original values are preserved verbatim. Semantic checks (LLM passing plot/lore into hints despite Stage A irrelevant list) are deferred to W2 LLM/human review.