From 5ae3aacb73c8c4bf5483eeefa6a49c009f3030cb Mon Sep 17 00:00:00 2001 From: Z User Date: Fri, 30 Oct 2026 21:07:31 +0000 Subject: [PATCH] refmod text lives only in subject_definitions - bodies reference alone, echo checked --- h3-plan-gate/SKILL.md | 11 +++- h3-plan-gate/h3_plan_json_system_prompt.md | 31 +++++---- h3-plan-gate/h3_scene_agent_system_prompt.md | 33 ++++++---- h3-plan-gate/verify_plan_rules.py | 69 +++++++++++++++----- 4 files changed, 101 insertions(+), 43 deletions(-) diff --git a/h3-plan-gate/SKILL.md b/h3-plan-gate/SKILL.md index 5ad2ee0..0232da4 100755 --- a/h3-plan-gate/SKILL.md +++ b/h3-plan-gate/SKILL.md @@ -70,7 +70,9 @@ otherwise. Copy the user's quoted speech EXACTLY into the right scene's `dialogu if the story has no quoted speech, every `dialogue` array stays `[]`. Only the last scene's beat may come to rest. The `subjects` array is the cast inventory: collect the full head-to-toe look there once — when writing scenes it is spent by weaving it into -each scene's [Shot 1] body, never parked in subject_definitions. If the story's script +each scene's [Shot 1] body, never parked in subject_definitions (a refmod subject is +the one exception: its verbatim identity/voice text is spent exactly once, in +subject_definitions, and the body references only ``). If the story's script carries a `characters:` block with refmod entries — `sam:&identity: ; voice: &` — copy the identity and voice text verbatim into that subject's inventory string (the & marks are delimiters, never content); the scene turns must carry @@ -180,14 +182,17 @@ and sampler latent or the video can come out square. shows" (the writing LLM is multimodal — look at it); and a user refmod (`sam:&identity: ...; voice: ...&` in the story's characters block) is carried through VERBATIM as " identity: ...; voice: ..." — never - rewritten, never trimmed, exempt from every defs limit. Weave the head-to-toe + rewritten, never trimmed, exempt from every defs limit; the refmod text lives + ONLY in subject_definitions, and the [Shot 1] body references that subject by + its `` label alone — never restating the refmod's identity, voice, + or look. For a no-image or binds subject, weave the head-to-toe look into the action as the scene exposes it — hair, build, and face as they move, each garment where the action touches it, footwear where steps land — nothing is shared between scenes or from a prefix. Name every body part the clothing leaves uncovered and call uncovered skin "nude", never "bare", where the scene shows it ("a man wearing a white tank-top with nude arms"). - Scene composition, not word count: the [Shot 1] body must carry every element — - subject anchor (`` restated, the full look woven in), style anchor, + subject anchor (`` referenced; the full look woven in — a refmod subject needs only its label), style anchor, continuity anchor (opens inside the previous end_state), blocking, one camera idea, performance beats, physical texture (wardrobe behavior, weather on skin and fabric, lighting), and a diff --git a/h3-plan-gate/h3_plan_json_system_prompt.md b/h3-plan-gate/h3_plan_json_system_prompt.md index 0f9795a..b661feb 100755 --- a/h3-plan-gate/h3_plan_json_system_prompt.md +++ b/h3-plan-gate/h3_plan_json_system_prompt.md @@ -127,7 +127,9 @@ Shots are flat. A shot never contains a nested object. continuity text in it. Every scene's prompt must be self-sufficient: subject_definitions anchors the tag's identity in one of its three forms and the [Shot 1] body carries the full look woven into the action in every - scene, because nothing is shared from a prefix or between scenes. + scene, because nothing is shared from a prefix or between scenes (a refmod + subject excepted: its verbatim definition carries the identity and look, and the + body references only ``). - The [Shot 1] body of every scene's detailed_description must: 1. Continue the exact action, pose, camera direction, and lighting from the previous scene's [Shot 1] body. Scene 1 instead sets the opening. @@ -142,11 +144,13 @@ Shots are flat. A shot never contains a nested object. the character has no reference image ("is a ... with ..." — age, hair, eyes, build), "binds to : what the image actually shows" when one is wired (you are multimodal — look at it), or the user's refmod carried through verbatim - ("identity: ...; voice: ..."). The full look lives in the [Shot 1] body, woven - into the action as the scene exposes it: hair, build, and face as they move, - every garment where the action touches it, footwear where steps land, and every - body part the clothing leaves uncovered named "nude" (PROMPT CRAFT) where the - scene shows or uses that skin. + ("identity: ...; voice: ..."). For a no-image or binds subject the full look + lives in the [Shot 1] body, woven into the action as the scene exposes it: + hair, build, and face as they move, every garment where the action touches it, + footwear where steps land, and every body part the clothing leaves uncovered + named "nude" (PROMPT CRAFT) where the scene shows or uses that skin. A refmod + subject's body references only its `` label — the refmod text never + leaves subject_definitions. 2. STYLE ANCHOR — the one consistent look: medium, lighting quality, palette. 3. CONTINUITY ANCHOR — open inside the previous scene's end state (scene 1 sets the opening) and name what carries over: the stride, the camera direction, the light, @@ -211,7 +215,8 @@ Shots are flat. A shot never contains a nested object. is a gap H3 fills by guessing, and guessed motion drifts off intent. - Describe the complete look in every scene, inside the [Shot 1] body — subject definitions never carry the scene's look; they only anchor identity in one of - their three forms. Weave the full + their three forms (a refmod subject is the exception: its definition carries the + look, so its body references only ``). Weave the full head-to-toe appearance into the action according to what is going on in the scene: face, hair, and build as the body moves, each garment where the action touches it (a cuff sliding back, a hem swaying, fabric darkening), footwear where steps land. @@ -266,7 +271,7 @@ Shots are flat. A shot never contains a nested object. fully_preserved). 4. "detailed_description:" then 1-2 style sentences (medium, lighting quality, palette — the look the whole plan shares), then the "[Shot 1] ..." body — one - line holding the subject restatement, visual style cues, all action, camera + line holding the subject references, visual style cues, all action, camera work, continuity cues, and any dialogue. Every scene is its own single-shot call, so the bracketed number is always [Shot 1]. 5. "overall_soundscape:" then the sound the scene itself makes: weather, room tone, @@ -424,10 +429,11 @@ Correct reply: scene except the last ends mid-action. subject_definitions anchors in one of its three forms in every scene (no-image physical traits / binds-to-image describing what the image shows / the user's refmod verbatim; never a wardrobe - list, never a nude/bare state), and the [Shot 1] body re-states who - is and carries the full head-to-toe look woven into the action, because the prefix + list, never a nude/bare state), and the [Shot 1] body references + and carries the full head-to-toe look woven into the action, because the prefix is empty — naming every uncovered body part as "nude", never "bare", where the - scene shows that skin. + scene shows that skin (a refmod subject needs only its `` label in + the body — the refmod text never leaves subject_definitions). 8. Every [Shot 1] body carries every element of a scene composition — subject, style, continuity, blocking, camera, performance, texture, exit state — and no change anywhere is stated as a bare outcome like "is now"; every transformation is @@ -504,7 +510,8 @@ A retired boxer teaches his granddaughter to dance in their cluttered kitchen. every scene — an identity anchor in subject_definitions (one of its three forms) and the full look woven into the [Shot 1] body, restated per scene — which also suits each scene being its own - generation. The validator accepts an empty prefix as + generation (a refmod subject excepted: its verbatim definition carries the identity + and look, and the body references only ``). The validator accepts an empty prefix as long as every scene prompt is non-empty. If you reintroduce prose, expect some format dilution. diff --git a/h3-plan-gate/h3_scene_agent_system_prompt.md b/h3-plan-gate/h3_scene_agent_system_prompt.md index 24e24de..42c454c 100755 --- a/h3-plan-gate/h3_scene_agent_system_prompt.md +++ b/h3-plan-gate/h3_scene_agent_system_prompt.md @@ -93,7 +93,9 @@ Emit exactly this shape and no other keys: fully enough to re-draw them identically in every scene. This array is the cast inventory: collect the full head-to-toe look here once — when writing scenes it is spent by weaving it into each scene's [Shot 1] body, never parked in - subject_definitions. If the story's script carries a `characters:` block with refmod + subject_definitions (a refmod subject is the one exception: its verbatim + identity/voice text is spent exactly once, in subject_definitions, and the body + references only ``). If the story's script carries a `characters:` block with refmod entries — `sam:&identity: ; voice: &` — copy the identity and voice text verbatim into that subject's inventory string (the & marks are delimiters, never content); the scene writers must carry them into subject_definitions @@ -226,7 +228,9 @@ identity and style from the story alone; without a handoff you are scene 1. be self-sufficient — subject_definitions re-states in one of its three definition forms (REF2VA PROMPT below), and the [Shot 1] body carries the full head-to-toe look woven into the action plus the visual style in every scene, even - though the previous scene already did. + though the previous scene already did. A refmod subject is the exception: its + definition already carries the identity and look verbatim, so the body references + only `` and never restates the refmod text. - Copy the subject descriptions and "style" from STYLE AND CAST faithfully. Never change wardrobe, hair, age, build, or look mid-plan. - If PREVIOUS HANDOFF is a handoff object, your scene opens inside the exact end_state @@ -251,11 +255,13 @@ a number is as much a failure as a thin one-liner. the character has no reference image ("is a ... with ..." — age, hair, eyes, build), "binds to : what the image actually shows" when one is wired (you are multimodal — look at it), or the user's refmod carried through verbatim - ("identity: ...; voice: ..."). The full look lives in the [Shot 1] body, woven - into the action as the scene exposes it: hair, build, and face as they move, - every garment where the action touches it, footwear where steps land, and every - body part the clothing leaves uncovered named "nude" (PROMPT CRAFT) where the - scene shows or uses that skin. Nothing is shared between scenes or from a prefix. + ("identity: ...; voice: ..."). For a no-image or binds subject the full look + lives in the [Shot 1] body, woven into the action as the scene exposes it: + hair, build, and face as they move, every garment where the action touches it, + footwear where steps land, and every body part the clothing leaves uncovered + named "nude" (PROMPT CRAFT) where the scene shows or uses that skin. A refmod + subject's body references only its `` label — the refmod text never + leaves subject_definitions. Nothing is shared between scenes or from a prefix. 2. STYLE ANCHOR — the one consistent look: medium, lighting quality, palette. 3. CONTINUITY ANCHOR — open inside the previous end_state (scene 1 instead establishes place, light, and the first action). Name what carries over: the @@ -327,7 +333,8 @@ The sound picture is its own element and lives in the overall_soundscape section is a gap H3 fills by guessing, and guessed motion drifts off intent. - Describe the complete look in every scene, inside the [Shot 1] body — subject definitions never carry the scene's look; they only anchor identity in one of their - three forms. Weave the full + three forms (a refmod subject is the exception: its definition carries the look, so + its body references only ``). Weave the full head-to-toe appearance into the action according to what is going on in the scene: face, hair, and build as the body moves, each garment where the action touches it (a cuff sliding back, a hem swaying, fabric darkening), footwear where steps land. @@ -380,7 +387,7 @@ The sound picture is its own element and lives in the overall_soundscape section fully_preserved). 4. "detailed_description:" then 1-2 style sentences taken from the STYLE AND CAST "style" (medium, lighting quality, palette), then the "[Shot 1] ..." body — one - line holding the subject restatement, visual style cues, all action, camera + line holding the subject references, visual style cues, all action, camera work, continuity cues, and any dialogue. Every scene is its own single-shot call, so the bracketed number is always [Shot 1]. 5. "overall_soundscape:" then the sound the scene itself makes: weather, room tone, @@ -533,9 +540,10 @@ Correct reply: describing what the image actually shows when one is wired, or the user's refmod identity/voice text carried through verbatim; retention_analysis carries one valid entry per defined label; the [Shot 1] body - re-states and carries the full head-to-toe look woven into the + references and carries the full head-to-toe look woven into the action, naming every uncovered body part "nude" — never "bare" — where the - scene shows that skin. + scene shows that skin (a refmod subject needs only its `` label in + the body — the refmod text never leaves subject_definitions). 5. The [Shot 1] body opens by continuing the PREVIOUS HANDOFF's end_state (scene 1 sets the opening instead) and ends mid-action — except the final outline scene. 6. The [Shot 1] body carries every element of a scene composition — subject anchor, @@ -581,7 +589,8 @@ Correct reply: always gets `"prompt_prefix": [""]` (the empirically better empty prefix), and each scene's subject_definitions (one of its three identity forms) plus its [Shot 1] body — the full look woven into the action — restate identity/wardrobe/style because - nothing is shared from a prefix. + nothing is shared from a prefix (a refmod subject excepted: its verbatim definition + carries the identity and look, and the body references only ``). - **Fragment file convention:** save each reply as `scene_NN.json` (zero-padded). The assembler and `scripts/build_plan.py` accept the raw object; fenced output also parses. diff --git a/h3-plan-gate/verify_plan_rules.py b/h3-plan-gate/verify_plan_rules.py index 6faccbe..f67a861 100755 --- a/h3-plan-gate/verify_plan_rules.py +++ b/h3-plan-gate/verify_plan_rules.py @@ -40,7 +40,9 @@ REF2VA_RETENTION_MARKERS = {"fully_preserved", "partially_preserved", # - the user's own tested wording, carried through untouched (exempt from # the no-image content bans: clothing and skin state) # No form carries a length cap - describe fully what the model needs. -# In every form the full look is still woven into the [Shot 1] body. +# No-image and binds subjects get their full look woven into the [Shot 1] body; +# a refmod subject is referenced by its label alone - the refmod +# text never leaves subject_definitions. REF2VA_REFMOD_RE = re.compile(r"^identity\s*:", re.I) REF2VA_BINDS_RE = re.compile(r"^binds to \s*:\s*", re.I) REF2VA_DEF_SKINSTATE_RE = re.compile( @@ -51,6 +53,26 @@ REF2VA_DEF_CLOTHING_RE = re.compile( r"gown|hoodie|sweater|shoes|sneakers|boots|sandals|socks|stockings|gloves|" r"helmet|lanyard)\b", re.I) + +def _refmod_echo(anchor, text, n=5): + """First run of >= n consecutive refmod words echoed in text, else None. + + A refmod subject's definition is the single source of its identity and look: + the [Shot 1] body must reference the label, never restate the refmod text. + """ + aw = re.sub(r"[^a-z0-9]+", " ", anchor.lower()).split() + tw = re.sub(r"[^a-z0-9]+", " ", text.lower()).split() + if len(aw) < n: + s = " ".join(aw) + return s if s and s in " ".join(tw) else None + grams = {" ".join(aw[i:i + n]) for i in range(len(aw) - n + 1)} + tjoin = " ".join(tw) + for g in sorted(grams): + if g in tjoin: + return g + return None + + FENCE_RE = re.compile(r"^\s*```[a-zA-Z0-9_-]*\s*\n(.*?)\n?\s*```\s*$", re.S) DIALOGUE_SPAN_RE = re.compile(r"(.*?)") MARKER_RE = re.compile(r"^\[[^\]]+\]\s*") @@ -163,6 +185,7 @@ def check_ref2va_prompt(lines, dur, where, prev_handoff=None): if not defs: raise ValueError(f"{where}: subject_definitions is empty") defined = [] + refmod_anchors = {} for line in defs: m = re.match(r"^(<(?:Subject|Picture|Video|Audio)\s+\d+>)\s+\S", line) if not m: @@ -179,6 +202,7 @@ def check_ref2va_prompt(lines, dur, where, prev_handoff=None): f"(' identity: ...; voice: ...'): {line[:60]!r}") if REF2VA_REFMOD_RE.match(anchor): defined.append(m.group(1)) # verbatim user refmod text - never edited + refmod_anchors[m.group(1)] = anchor continue mb = REF2VA_BINDS_RE.match(anchor) if mb: @@ -237,8 +261,18 @@ def check_ref2va_prompt(lines, dur, where, prev_handoff=None): f"wardrobe, style, continuity, blocking, camera, performance, " f"texture, sound, exit state)") if "" not in body: - raise ValueError(f"{where}: the [Shot 1] body must re-state and " - f"carry the head-to-toe look woven into the action") + raise ValueError(f"{where}: the [Shot 1] body must reference and " + f"carry the head-to-toe look woven into the action (for a " + f"refmod subject the label reference is enough - its " + f"definition carries the look)") + for rlabel, ranchor in refmod_anchors.items(): + echo = _refmod_echo(ranchor, dd) + if echo: + raise ValueError(f"{where}: {rlabel} is defined by a verbatim refmod - the " + f"refmod text lives only in subject_definitions; reference " + f"{rlabel} by its label alone in the [Shot 1] body and " + f"never restate its identity, voice, or look (echoed " + f"{echo!r})") if " is now " in dd.lower(): raise ValueError(f"{where}: change stated as bare outcome instead of shown progression") if re.search(r"\bbare\b", dd, re.I): @@ -666,19 +700,14 @@ REF2VA_DOOR = { "detailed_description:", REF2VA_NURSE["prompt"][7], "[Shot 1] Continue the same corridor walk without resetting the motion, the " - "trolley still rolling at her pace. , the same night nurse from " - "her reference image with her dark brown hair still in its low bun, eases " - "the steel trolley to a stop beside the blue door, both hands relaxing on " - "the handle one after the other, its last wheel squeak dying on the " - "linoleum. Her pale green scrubs settle as her shoulders drop, the white " - "sneakers planting at hip width, the nude hands below the scrub cuffs " - "opening off the handle while her nude lower legs stay above the sneaker " - "collars. The camera holds its over-the-shoulder framing while her " - "shoulders settle and the paper cup's surface stills on the trolley top; " - "the fluorescent panels keep their even rhythm overhead, the corridor walls " - "stay in soft focus, and her scrub-top hem sways once and hangs. Her right " - "hand leaves the handle, opens, and closes on the door's cold push bar " - "while the clipped chart keeps swaying on its clip.", + "trolley still rolling at her pace. eases the steel trolley to " + "a stop beside the blue door, both hands relaxing on the handle one after " + "the other, its last wheel squeak dying on the linoleum. The camera holds " + "its over-the-shoulder framing while her shoulders settle and the paper " + "cup's surface stills on the trolley top; the fluorescent panels keep " + "their even rhythm overhead and the corridor walls stay in soft focus. " + "Her right hand leaves the handle, opens, and closes on the door's cold " + "push bar while the clipped chart keeps swaying on its clip.", "overall_soundscape:", "the same low fluorescent hum continues, the trolley wheels fall silent, a " "soft draught breathing under the far door.", @@ -789,6 +818,13 @@ def run_fragment_suite(): "scrub top and white sneakers.") bad_binds = _ref2va_short() bad_binds["prompt"][1] = " bound to her subject reference image." + refmod_echo = json.loads(json.dumps(REF2VA_DOOR)) + p = refmod_echo["prompt"] + i_dd = p.index("detailed_description:") + p[i_dd + 2] = p[i_dd + 2].replace( + " eases the steel trolley", + ", the same woman in her thirties with dark brown hair in a low " + "bun, eases the steel trolley") r2_cases = [ ("missing retention_analysis entry", no_ret, {}), ("summary directive not [reference generation]", wrong_summary, {}), @@ -798,6 +834,7 @@ def run_fragment_suite(): ("head-to-toe look dump in subject_definitions", dump_defs, {}), ("clothing in a no-image subject definition", clothed_defs, {}), ("subject definition is none of the three forms", bad_binds, {}), + ("refmod look restated in the [Shot 1] body", refmod_echo, {}), ] for i, (label, obj, kwargs) in enumerate(r2_cases, 1): try: