Commit Graph

13 Commits

Author SHA1 Message Date
Z User
0723a2dbd4 Fix refmod isolation false positive: 'like' joins REF2VA_STOPWORDS - any refmod phrased 'dresses like ...' banned the common function word from every scene body (live test run: 'cries like an infant' rejected with found: like). No behavior change for distinctive tokens; embedded suite unchanged at 7,712 B, example plan PLAN OK 2026-11-01 01:41:04 +00:00
Z User
f6be7ffcc5 Refmod verbatim is now machine-checked: outline_refmods() extracts the story's identity/voice block from the outline's subjects strings and diffs every scene's subject_definitions line against it (fragment --outline + the merge); rewrites, form swaps and invented refmods fail with actionable errors; embedded suite gains the gate fixtures and stays under Cline's 8k output cap (7,712 B) 2026-10-31 01:19:46 +00:00
Z User
7df79bcf09 Restore the strict refmod-isolation validator: token-echo check catches scattered re-descriptions the n-gram check missed 2026-10-30 21:54:01 +00:00
Z User
fb271a8399 Six-section Ref2VA setup: three definition forms, refmod isolation, no length caps 2026-10-30 21:53:22 +00:00
Z User
5ae3aacb73 refmod text lives only in subject_definitions - bodies reference <Subject N> alone, echo checked 2026-10-30 21:07:31 +00:00
Z User
de19ecee44 no length caps and no lean doctrine in subject_definitions - node-author compactness rules removed (different H3 model) 2026-10-30 20:41:57 +00:00
Z User
962c667418 No length cap on the binds-to-image definition form 2026-10-30 20:35:58 +00:00
Z User
c2559f0572 subject_definitions takes exactly three forms: no-image traits, binds-to-image, verbatim refmod 2026-10-30 20:33:24 +00:00
Z User
28f9b80c83 subject_definitions is a lean identity anchor; the look is woven into detailed_description
Rule 6: subject_definitions never carries the full head-to-toe look or a
nude/bare skin state - it is a lean identity anchor (who the label is, tied
to its reference image). The full look is woven into the [Shot 1] body of
detailed_description according to what is going on in the scene: face, hair,
build as the body moves, each garment where the action touches it, footwear
where steps land, uncovered skin named "nude" where the scene exposes it,
and wardrobe/skin changes stated as part of the action that causes them.

- validator: defs lines reject nude/bare-family skin states and anything
  over 25 words (lean-anchor cap); body re-state message updated; two new
  suite FAIL cases (nude-in-defs, look-dump-in-defs); fixtures rewritten
  to lean defs with the look woven into both [Shot 1] bodies
- all three LLM prompts + SKILL.md: defs role flipped everywhere (template,
  SUBJECT ANCHOR, PROMPT CRAFT, REF2VA spec, worked examples, self-checks);
  outline 'subjects' array now documented as the cast inventory the look is
  spent from
2026-10-30 19:42:00 +00:00
Z User
30cbaeb8f6 Switch plan.json output to the native H3 ref2va prompt format
Each scene prompt is now the six native reference sections —
subject_definitions / summary [reference generation] / retention_analysis /
detailed_description (1-2 style sentences then [Shot 1]) /
overall_soundscape / non_diegetic_music — one line per element, headers on
their own line. The full head-to-toe look (nude-not-bare) lives in
subject_definitions and is restated compressed in the [Shot 1] body; both
cast members are defined per scene against the wired subject references.
verify_plan_rules.py auto-detects the format per shot and validates the
ref2va contract (canonical order, directive, definitions, retention,
style prefix, thin floor), so legacy t2va plans keep passing. Both system
prompts, the assembler prompt and SKILL.md carry the new skeleton,
worked examples and self-checks; example_output/README.md marks the old
run as the t2va craft reference.
2026-10-30 19:28:36 +00:00
Z User
fa764ec933 Full look every scene; 'nude' for uncovered body parts, never 'bare'
H3 needs every aspect of the look restated per scene, and uncovered
skin is steered by the skin-state word 'nude' (tank-top with nude
arms), not 'bare'. Added to both prompts' subject anchor + craft
rules + self-checks and SKILL.md; the validator prints a non-blocking
note when a description line uses 'bare'.
2026-10-29 23:18:56 +00:00
Z User
f5575fa678 Scene composition is the gate, not word counts
The skill exists so the LLM writes/rewrites each scene one at a time
with full attention on making it a complete scene composition — not so
it can hit a word target.

Prompts: the word-budget craft bullets are replaced by an 8-element
composition checklist (subject anchor, style anchor, continuity anchor,
blocking, camera, performance, physical texture, exit state) in both the
scene-writer PROMPT B (new SCENE COMPOSITION section) and the single-pass
prompt. Chase the elements, never a number; padding is as much a failure
as a thin one-liner.

Validator: hard word bands (<10s: 150-200, >=10s: 200-300) removed from
plan and fragment checks. Replaced by a thin-scene floor (100/120 words)
that only trips lazy output, plus a new continuity-anchor check — chained
fragments must open inside the previous end_state. No target band, no
upper limit; output lines say so.

SKILL.md/README: goal, Step 2, craft rules, table rows, and example
outputs updated to match. Suite, --check, chained fragment checks all
green.
2026-10-29 19:13:27 +00:00
Z User
bd50b1ba51 Add H3 plan-JSON gate: agent workflow prompts, validator, assembler, examples
- h3_scene_agent_system_prompt.md: outline turn + one-scene-per-turn writer
  with a truthful _handoff chain (style_and_cast frozen at outline)
- h3_plan_assembler_system_prompt.md: verbatim-copy LLM merge step
  (deterministic alternative: build_plan.py)
- h3_plan_json_system_prompt.md: single-pass variant for short 2-4 scene
  stories; prompt_prefix fixed to [""], native H3 three-line shot template
- verify_plan_rules.py: node-replica validator (plan + fragment modes,
  handoff-chain checks, embedded pass/fail suite)
- build_plan.py: deterministic fragment merger with --check for existing plans
- h3_agent_workflow_examples/: worked 3-scene greenhouse courier run
  (outline -> 3 fragments -> validated plan.json, 1042 frames = 43.42s)
2026-10-29 17:18:28 +00:00