Written 2026-09-09 for 182.C.03. Every mode, threshold, and limitation below
is bounded by what the audited code and the dated provider catalog can actually
do (capability-audit.md, and the OpenRouter routing snapshot in
TODOS/phase-182.md). Where a capability has never been probed, this document
says so rather than promising it.
Two rules govern everything here:
- A mode is defined by what the creator is promised, not by which model runs it. Routing may change model within a mode; it may never change mode.
- No claim in this document may outrun its measurement. A threshold with no
evaluator behind it is written as a target awaiting
182.C.17–182.C.20, never as a delivered guarantee.
182.C.03.01 Supported modes#
Seven modes. Each names the promise, the inputs that are mandatory for it, the lanes that can carry it today, and the reason it is a separate mode rather than a parameter of another one.
1. text-to-scene — generated cast#
Promise: a scene containing people the model invents, matching a described situation. Mandatory inputs: prompt. Forbidden inputs: identity references, any named real person.
Lanes: any video lane. Identity rule: because no cast member has an
authorization record, output from this mode may never be labelled, titled, or
described as depicting a specific real person, and it never enters the identity
gates — there is no baseline to preserve. A request that names a real person in
the prompt is a reference-to-video request without consent, and is refused
under 182.C.03.03, not silently rendered.
2. reference-to-video — authorized likeness, directed performance#
Promise: the person in the supplied authorized references appears, moving and acting as directed. Mandatory inputs: at least one typed identity reference plus its authorization record. Not promised: exact words — this mode's dialogue, if any, is model-invented.
Lanes: alibaba/wan-2.7 (frame_images first/last plus images
passthrough), bytedance/seedance-2.5 (image-only input_references, 4–30 s),
alibaba/wan-2.6 (first frame plus input_references only). Reference count
limits are not in the catalog and must come from the live probe at 182.C.11.
3. exact-dialogue-performance#
Promise: the person says exactly the approved script, and the visible speech
matches the approved audio. Mandatory inputs: identity references with
authorization, the canonical script (182.C.03.04), and creator-approved speech
audio or deterministic TTS derived from that script.
Lanes: only models exposing an audio passthrough — today
alibaba/wan-2.7 and alibaba/wan-2.6. This is the hard boundary of the phase:
a model that generates dialogue from the prompt cannot be routed here at any
quality tier, because it may paraphrase, drop words, mispronounce names, or
re-time delivery. generate_audio must be explicitly set per adapter in this
mode (F-08) — an unrequested model-invented audio track is a correctness defect
here, not a cost annoyance.
4. multi-character-scene#
Promise: two or more authorized cast members in one shot, each speaking their own attributed lines. Mandatory inputs: everything mode 3 requires, per speaker, plus a speaker-labelled script.
Separate mode because the failure is different: the defect that matters is
not "bad sync" but speech assigned to the wrong mouth, which the single-person
sync metrics cannot see. It stays gated on the speaker-attribution evaluator in
182.C.19/182.C.20, and until that exists no multi-character take is
releasable, however good its sync score.
5. continuation-extension#
Promise: a new take that continues an existing approved take, preserving identity, wardrobe, scene, and camera. Mandatory inputs: the prior take's artifact reference plus its continuity record.
Lane: direct only. OpenRouter exposes no extend operation on any lane;
the in-repo path is @isis/seedance-provider's video-extension.ts
(buildVideoSplicePlan, assertSupportedVideoExtensionRequest). Per F-13 the
per-shot fan-out belongs to the workflow layer, not to this mode.
6. video-edit#
Promise: a bounded change to an existing approved take — replace a region, a prop, or a background — leaving everything else byte-comparable in intent. Mandatory inputs: the source artifact, the edit region, and the same authorization that covered the original.
Lane: direct only, @isis/seedance-provider's video-inpainting.ts and
character-replacement.ts. minimax/hailuo-3 (instruction-guided editing) and
runway/aleph-2 (in-context editing with keyframes) are probe candidates,
not lanes: neither has a committed fixture, so neither may carry a creator
promise yet.
7. talking-head-avatar#
Promise: a framed presenter delivering a script to camera. Mandatory inputs: the script and a voice binding.
Lane: heygen/avatar-iv (script as prompt, plus voice_id,
voice_settings, motion_prompt, expressiveness, fit, background).
Two rules that make this mode safe. It is selectable only by explicit
creator choice — the router may never fall back to it from an active-scene
request (182.C.03.07), because a talking head is a different product, not a
lower-quality version of the same one. And its voice_id must be bound to a
voice-consent record before submission; OpenRouter does not accept supplied
audio on this lane, so exact-audio mode is unavailable here and the mode is
labelled accordingly.
Mode–lane matrix#
| Mode | Needs identity auth | Needs approved audio | Lanes today | Blocked on |
|---|---|---|---|---|
text-to-scene |
no | no | any | — |
reference-to-video |
yes | no | wan-2.7, seedance-2.5, wan-2.6 | reference-count probe 182.C.11 |
exact-dialogue-performance |
yes | yes | wan-2.7, wan-2.6 only | audio-passthrough probe, ASR |
multi-character-scene |
yes (per cast) | yes | wan-2.7, wan-2.6 only | speaker attribution 182.C.19 |
continuation-extension |
inherited | inherited | direct @isis/seedance-provider |
direct-lane parity |
video-edit |
inherited | inherited | direct @isis/seedance-provider |
direct-lane parity |
talking-head-avatar |
yes | n/a (not accepted) | heygen/avatar-iv |
voice-consent binding |
Mode is never inferred. The request carries the mode; the router picks a
model within it. If no healthy lane satisfies the mode's mandatory inputs, the
request fails with the deterministic message defined in 182.C.03.07 — it is
never quietly re-planned as a different mode.
182.C.03.02 Active-person benchmark scenarios#
The point of this set is that a front-facing talking head is the easiest case and the least representative one. Each scenario below exists because it breaks a specific evaluator or a specific model assumption; a scenario that stresses nothing new does not belong in the set.
Every scenario is run in exact-dialogue mode with the same approved script and approved audio, so the only variable is what the body and camera are doing.
| # | Scenario | What it stresses that a talking head does not | Primary failure to detect | Evaluator that must catch it |
|---|---|---|---|---|
| S1 | Walking toward / across camera while talking | Continuous scale and pose change; face pixels shrink and grow | Sync drift as face size changes; identity drift at range | 182.C.17, 182.C.19 |
| S2 | Cooking (hands at chest height, looking down) | Downward gaze, chin occlusion, hands entering the mouth region | Mouth occluded and scored as "closed" rather than hidden | 182.C.17 visible-mouth |
| S3 | Presenting to a room (three-quarter profile) | Off-axis mouth geometry | Sync metric degrading purely from viewing angle | 182.C.17, 182.C.20 |
| S4 | Driving, stationary or safely staged | Fixed frame, seatbelt/interior clutter, side-lit face, hands off-frame | Identity drift from harsh side lighting | 182.C.19 |
| S5 | Dancing | Whole-body motion, motion blur, frequent full-face loss | Blur read as artifact; sync scored on unusable frames | 182.C.17, 182.C.21 |
| S6 | Sports / exertion | Breath-broken delivery, open-mouth non-speech, rapid head motion | Non-speech mouth movement counted as speech | 182.C.18, 182.C.20 |
| S7 | Working at a desk / screen | Long static shots, small in-frame face, glasses and screen glare | Identity drift in a low-face-pixel regime | 182.C.19 |
| S8 | Gesturing emphatically | Hands crossing the face; rapid hand-face overlap | Occlusion misattributed to the mouth model | 182.C.17 visible-mouth |
| S9 | Interacting with props (holding, handing) | Object permanence, hand-object-face overlap, wardrobe motion | Wardrobe/prop discontinuity across a take | 182.C.19 wardrobe |
| S10 | Moving through camera angles while speaking | Cut or move between angles mid-utterance | Sync re-anchoring per angle; speaker re-identification | 182.C.17, 182.C.19 |
Rules that make the set honest#
- Same script, same audio, across all ten. Differences in the result are then attributable to motion and framing, not to content.
- Every scenario carries a multi-character variant for the speaker attribution gate; those variants are reported separately and never averaged into the single-person pass rate.
- Scenarios are not a pass/fail list, they are a cohort key. Per
182.C.34.07and the existing Beta–Thompson router (which Phase 182 extends to per-model, per-cohort keys rather than replacing), each scenario is its own cohort with its own measured release rate. A model that passes S1 and fails S5 is routed accordingly — it is not called "passing". - Occlusion is a first-class outcome, not a failure. S2, S8, and S9 exist precisely to make the evaluator distinguish mouth hidden from mouth wrong. A take where the mouth is legitimately out of view for a justified duration is releasable; one where the evaluator could not tell is not.
- Duration bounds bite here. Every scenario must fit the lane's real limits
— 2–10 s on wan-2.7, 5–10 s on wan-2.6 — so a scenario needing longer coverage
is authored as a multi-shot plan (
182.C.09) and evaluated per shot plus across the seam, never as one impossible request. - Sample size and confidence intervals come from
182.C.34, not from this document. Nothing here is a threshold; this is the list of situations that must be measured before any active-scene claim is made.
182.C.03.03 Out of scope at launch#
"Out of scope" here has two distinct meanings, and conflating them is how a safety boundary quietly becomes a roadmap item. Each entry below says which it is:
- Refused — the system declines and says why. Enforcement already exists in
libs/oshun/trust-safety/src/policy-taxonomy/categories.ts; this phase binds the video path to it rather than inventing a parallel policy. - Unsupported — the system cannot do it well, says so, and does not pretend otherwise by substituting something else.
| Boundary | Kind | Enforcement point |
|---|---|---|
| Live / real-time impersonation | Refused | No real-time lane exists and none is built. synthetic-media:identity-impersonation |
| Unauthorized public figures | Refused | synthetic-media:nonconsented-likeness is a hard-block key; no authorization record → no provider submission |
| Any likeness without an authorization record | Refused | 182.C.26.01–.02 bind an authorization to the reference digests before upload |
| Voice cloning without voice consent | Refused | synthetic-media:voice-clone-abuse, hard-block; heygen voice_id binding |
| Removing or defeating the watermark | Refused | synthetic-media:watermark-removal, hard-block |
| Deceptive realism (passing synthetic as real) | Refused | synthetic-media:deceptive-realism plus the mandatory disclosure in synthesis-disclosure.ts |
| Unsafe stunts | Refused | Depiction that models a real physical risk to imitate; routed through content:graphic-violence review rather than auto-denied |
| Off-screen dialogue shown as visible sync | Unsupported | See below — this is a measurement rule, not a policy rule |
| Crowd dialogue (background speakers) | Unsupported | No evaluator attributes speech to an unenrolled background face |
The off-screen dialogue rule#
This one deserves its own statement because it is the boundary most easily
crossed by accident. If the approved audio contains speech while the speaker's
mouth is not visible, the system may deliver the take only as narration or
off-screen dialogue, explicitly labelled. It may never present that segment as
lip-synced performance, and the sync evaluator may never score those frames as
"passing sync" — an unmeasurable segment is unknown, and unknown is not
pass. A take whose mouth-visible coverage falls below the threshold in
182.C.03.05 for an unjustified duration is rejected, not relabelled.
The crowd rule#
Background people may appear. Background people may not be given attributed dialogue, because nothing in the pipeline can verify who said what for a face that has no enrolled identity and no consent record. A script that assigns lines to unnamed background characters is refused at validation with the reason stated, not silently rendered with the lines mixed into the bed.
What "refused" must look like#
A refusal names the boundary it hit and never degrades into a nearby product: no
substituting a generated look-alike for a refused likeness, no dropping to
text-to-scene when consent is missing, no swapping to talking-head-avatar
when an active scene is refused. That is the same rule as 182.C.03.07, applied
to policy rather than to capacity.
182.C.03.04 Exact-dialogue semantics#
"Exact" is a claim about a specific artifact, so it needs a definition precise enough to fail. Seven parts:
1. The canonical script#
One immutable text is the authority for a job: the canonical script, stored with a SHA-256 digest and a language tag, carrying per-speaker turns. Everything downstream compares against that digest — the prompt compiler, the TTS input, the ASR alignment, and the release proof. If the creator edits a word, that is a new canonical script with a new digest, and every prior candidate is invalidated rather than re-scored. A take may not be released against a script digest that differs from the one its audio was produced from.
2. Approved pronunciation#
Names, acronyms, numbers, and domain terms are the part a recognizer gets wrong
and a creator notices immediately. The canonical script may carry a
pronunciation lexicon: per-token IPA or a phoneme spelling plus an optional
audio exemplar. Lexicon entries are approval, not suggestion — a delivered
take whose pronunciation of a lexicon token does not match is a defect even when
the word is transcribed correctly, and it is reported per token rather than
folded into an aggregate score (182.C.18.03).
3. Allowed disfluencies#
Default: none. The delivered speech must contain the script's words and no
inserted "um", "uh", false start, or repetition. A creator may opt in per job to
disfluencies: permitted, which moves inserted non-lexical fillers from
"insertion error" to "ignored token" — it never permits inserted words.
4. Expected language#
Each turn declares its language tag, and code-switched spans are marked inside
the turn. This matters twice: the recognizer is selected per declared language,
and a delivered take in the wrong language is a hard failure rather than a low
score. Where a single recognizer is not validated for a declared language,
182.C.18.07 requires an ensemble or human review — the take does not silently
inherit a weaker measurement.
5. Word-error target#
Two different measurements, because averaging them hides the failure that matters:
- Content-word error rate (CWER) — target 0. After normalization (case, punctuation, whitespace, number-to-word, contraction expansion), every content word in the canonical script must be present, in order, with no substitutions and no inserted words. Any content-word deletion, substitution, or insertion blocks release in exact-dialogue mode.
- Function-word / filler WER — reported, not gated by default. Articles and particles that the recognizer commonly mangles are measured and shown, but do not by themselves block, because they are the region where ASR error and model error are least separable.
ASR error is not model error. Because a recognizer has its own error rate on
clean speech, a CWER > 0 result is a finding, not a verdict: it routes to a
second independent recognition or human review (182.C.18.07, needs_review in
the existing job enum) before the take is rejected as unfaithful. What is never
allowed is the opposite — treating an unverified transcription as proof of
exactness.
6. Timing tolerance#
Two clocks, two tolerances:
- Audio-to-video sync, against the published broadcast thresholds: ITU-R
BT.1359-1 puts the detectability of a sound/vision offset at roughly +45
ms (audio early) and −125 ms (audio late), with acceptability at +90 /
−185 ms; ATSC IS-191 tightens the production target to +15 / −45 ms. Phase
182 adopts the ATSC production window as the release target and the ITU
detectability window as the hard reject boundary. A take outside
detectability is never released; a take between the two is
needs_review. - Utterance timing against the approved audio. In exact-audio mode the audio
is supplied, so the delivered audio must match the approved audio's duration
and word onsets within a tolerance, not merely "sound similar". Stretching,
truncation, or re-synthesis of the approved audio is a defect and is detected
by
182.C.18.02— a model returning its own re-rendered voice track instead of the supplied one is the specific failure this catches.
7. Nonverbal vocalizations#
Laughter, breath, sighs, and audible reactions are permitted in the approved audio and are then part of the ground truth like any other segment: they must appear where approved and must be lip-consistent. They are not permitted as model inventions — a laugh the model adds is an insertion, scored under CWER's insertion rule even though it is not a word. Non-speech mouth movement in an exertion scenario (S6) is exactly why this is stated: the evaluator must not count an open mouth as speech, and must not count approved breath as an error.
What this section does not yet have#
Every gate above needs the exact-speech evaluator, which is blocked on the ASR
vendor decision at 182.C.02.12. Until that decision is funded, exact-dialogue
mode has a complete definition and no measurement, and therefore cannot be
released — which is the honest state, recorded rather than papered over.
182.C.03.05 "Perfect delivered sync" — the definition that can fail#
The product promise is no known bad-sync take is released — not "every take
is perfect". So "perfect delivered sync" is defined as a conjunction of six
conditions measured on the delivered bytes. All six must hold. Any one
unmeasurable makes the result unknown, and unknown never releases.
| # | Condition | Metric | Release threshold |
|---|---|---|---|
| C1 | Audiovisual alignment | SyncNet-class LSE-C (confidence, higher better) and LSE-D (distance, lower better) | LSE-C ≥ 6 and LSE-D ≤ 8, plus offset inside the ATSC window (182.C.03.04) |
| C2 | Exact speech | Content-word error rate against the canonical script digest | CWER = 0 in exact-dialogue mode |
| C3 | Speaker attribution | Per-utterance active-speaker assignment | 100% of utterances attributed to the enrolled speaker who was scripted to say them |
| C4 | Identity preservation | ArcFace-class cosine against the authorized reference embeddings | cosine ≥ 0.35, and no per-shot drift beyond the identity-drift critical band |
| C5 | Visible-mouth coverage | Fraction of speech-active frames where the mouth region is measurable | ≥ 85% of speech-active frames, and every uncovered span justified (182.C.03.03) |
| C6 | No critical artifacts | Temporal-consistency and blank/black/frozen-frame detection; FVD-class distance | FVD ≤ 120, head-pose MAE ≤ 8°, zero blank/frozen/duplicate-frame findings |
Where these numbers come from, and what they are not#
The C1, C4, and C6 numbers are inherited, not invented: they are the
defaults already encoded in
libs/nous/portrait-animation/src/portrait-animation-quality-metrics-pipeline.ts
(lseC 6, lseD 8, arcFaceCosine 0.35, fvd 120, head-pose MAE 8°), and
they sit in the right regime for SyncNet-class scoring, where real footage
typically lands near LSE-D ≈ 6–7 / LSE-C ≈ 7.
But that pipeline ingests these scores; it never computes them (F-09). So
today every one of C1, C4, and C6 is a threshold with no producer. They become
real at 182.C.17 (sync), 182.C.19 (identity/temporal), and 182.C.20
(speaker attribution), and each must be recalibrated on our own labelled
corpus at 182.C.34 before it gates anything — a threshold carried over from
another codebase's defaults is a starting point, not a calibration.
The rules that stop this definition from being gamed#
- Conjunction, not average. There is no weighted score in which a great identity result compensates for a failed speech result. Six booleans, all required.
- Measured from the released bytes. Per
182.C.01.05, every value is computed from the exact artifact that would ship — not a provider preview, not the pre-watermark master, not metadata. unknown≠pass. An evaluator that could not run, a face it could not find, a segment whose mouth was hidden — allunknown. The take is blocked or routed toneeds_review, never released on the strength of a missing measurement.- Per-shot, then per-delivery. A multi-shot delivery passes only if every shot passes and the seams pass; a single failing shot fails the delivery rather than being averaged away.
- The pass rate is published. The creator-facing promise is the gate, and
the measured pass rate per cohort (
182.C.03.02) is shown alongside it. "Perfect delivered sync" describes what ships, never the model's first-sample hit rate.
182.C.03.06 Quality tiers#
Four tiers. A tier sets how much work the system spends before it decides,
not how much quality it is allowed to skip: the six release conditions of
182.C.03.05 are identical in every tier. This is the same principle the
existing hero path already states — "hero is about model QUALITY, not gate
bypass … a better model is not auto-released"
(apps/oshun/bff/src/generation/video-hero-gate.ts:12-14) — generalized to four
tiers.
| Tier | Candidates per shot | Retry budget (after evaluation) | Model pool | Evaluation depth | Release-eligible |
|---|---|---|---|---|---|
draft |
1 | 0 | cheapest healthy lane for the mode | Probe + shape + safety only | No — preview only |
standard |
2 | 1 (one reroute or one reseed) | measured-best lane for the mode/cohort | Full six conditions | Yes |
production |
4 | 2, at most 1 of them a model change | top-2 lanes by cohort release rate | Full six + per-shot seam checks | Yes |
hero |
6 | 3, model change allowed each time | frontier pool (OSHUN_HERO_VIDEO_MODEL) |
Full six + seams + mandatory human review | Yes, after human sign-off |
The rules that keep tiers honest#
draftis not releasable, and says so. It exists so a creator can check blocking and framing cheaply. It is watermarked and disclosed like everything else, and the UI must not offer a download path that implies delivery. A draft is never promoted to a delivery by relabelling — promotion means re-running at a releasable tier.- Candidates are generated, then selected on measurement. A candidate count
is not a quality claim; it is how many samples the selector gets to choose
among. Selection uses the
182.C.03.05conjunction plus the cohort-level release rate from the existing Beta–Thompson router, never provider metadata or job-completion order. - The retry budget is spent on evidence, not on hope. A retry is permitted only when the evaluation names a specific failed condition and the retry changes something addressed to it — a reseed for an artifact failure, a reroute for a lane-specific sync failure, a re-render for a truncated download. Retrying an unchanged request against an unchanged model is not a retry; it is a second sample, and it is counted against the candidate budget.
- Budget and quota gates bind at admission, per tier.
herocosts roughly six generations plus retries; on a per-second-with-audio SKU like Veo 3.1 that is a materially different spend fromdraft. The existinggeneration-budget-gate.tsandgeneration-quota-gate.tsadmit or refuse the job at its tier before any provider call (182.C.15.06), so a tier the tenant cannot afford fails at admission with a stated reason rather than mid-flight. - Exhausting the budget is a failure, not a downgrade. If no candidate meets
the six conditions within the tier's candidate and retry budget, the job ends
blockedorneeds_reviewwith the failed conditions listed. It is never delivered as "the best of a bad set", and it is never silently re-planned into a cheaper mode. - Hero requires a human. The frontier pool buys a better first sample, not a waived gate; hero deliveries carry a recorded human sign-off in addition to the automated conjunction.
182.C.03.07 The unsatisfiable-request failure#
When no currently healthy lane can satisfy a request's mandatory inputs, the job fails deterministically. Determinism here means three things: the same unsatisfiable request always produces the same machine code, the message names the specific missing capability rather than a generic apology, and the outcome never varies with which lane happened to be polled first.
The contract#
status: failed (or blocked, when the reason is policy rather than capacity)
code: human_video_unsatisfiable
reason: <one of the enumerated reasons below>
required: [ capability, ... ] # what the mode demanded
available: [ lane: [capability, ...], ... ] # what each healthy lane offers
The code style matches what the video path already emits — snake-cased, stable,
thrown loudly (provider_not_configured, video_provider_missing_output_url,
openrouter_videos_empty_response) — so this is one more member of an existing
family, not a new error dialect.
Enumerated reasons#
| Reason | Raised when |
|---|---|
no_audio_input_lane |
Exact-dialogue mode, but no healthy lane exposes an audio passthrough |
no_reference_lane |
Reference mode, but no healthy lane accepts image references |
reference_count_exceeded |
More typed references than the probed lane limit accepts |
duration_out_of_range |
Requested duration outside every healthy lane's supported range for the mode |
resolution_unavailable |
Requested resolution unsupported by every healthy lane for the mode |
no_edit_lane / no_extend_lane |
Edit or continuation requested with no direct lane configured |
capability_unprobed |
A lane advertises the capability but has no committed probe fixture yet |
all_lanes_unhealthy |
Lanes exist and match, but all are failing health or circuit-broken |
zdr_tenant_ineligible |
Tenant enforces zero data retention — OpenRouter video routing is unavailable to it entirely |
tier_budget_refused |
The tier's cost estimate exceeds the tenant's admitted budget or quota |
What must never happen instead#
- No mode downgrade. An active-scene request never becomes a
talking-head-avatarrender. They are different products (182.C.03.01). - No dialogue downgrade. An exact-audio request never becomes prompt-native
dialogue. A lane without an
audiopassthrough cannot say the approved words; substituting it converts a verifiable promise into an unverifiable one. - No silent quality downgrade. A
productionrequest does not quietly becomestandardbecause the production pool is unhealthy; it fails withall_lanes_unhealthyand the creator chooses. - No fabricated success. A completed provider job that fails the six
conditions is
blockedorneeds_review— never delivered with the failure omitted.
capability_unprobed is deliberate#
The catalog advertising a field is permission to probe, not evidence it works.
Until 182.C.11 commits a dated request/response fixture and an artifact digest
for a capability, a request depending on it fails with capability_unprobed
rather than being attempted and reported as a provider fault. This keeps the
distinction between "we have not proved this" and "the provider broke" visible
in the job record instead of blurring it into a retry loop.
182.C.03.08 Creator-visible limitations#
These are shown in the product, not buried in documentation, because each one changes what the creator should expect before they spend a generation. The column that matters is the third: what the system does when the condition is present. In every case the answer is measure, disclose, or refuse — never deliver quietly and hope.
| Condition | Why it is hard | What the system does |
|---|---|---|
| Side profiles (> ~60° off-axis) | Mouth geometry is foreshortened; SyncNet-class scoring degrades from the angle alone | Scores the span, flags angle as the cause, routes to needs_review rather than failing it as bad sync |
| Occlusion (hands, props, masks) | The mouth is unmeasurable, not wrong | Marks the span unknown; counts it against visible-mouth coverage (C5); releases only if coverage ≥ 85% and the span is justified |
| Very fast speech | Phoneme rate outruns the model's mouth articulation and the recognizer's segmentation | Reports elevated CWER risk up front; recommends splitting the turn across shots |
| Singing | Sustained vowels and pitch make lip shapes diverge from spoken-phoneme priors | Unsupported for exact-dialogue. Available in creative mode, labelled, never sync-gated as speech |
| Shouting | Wide mouth aperture, distorted formants, higher ASR error | Measured normally, with the elevated error rate reported per cohort rather than hidden in an average |
| Whispering | Little or no voicing; sync and ASR both degrade sharply | Unsupported for exact-dialogue. Refused at validation with the reason, not attempted |
| Overlapping dialogue | Two speakers at once defeats per-utterance attribution (C3) | Refused in multi-character-scene; the script must serialize the turns |
| Extreme motion blur | Frames carry no usable mouth signal | Blurred frames are unknown, not pass; if they exceed the coverage budget the take is rejected |
| Masks / face coverings | Identity and mouth signal are both suppressed | Identity gate (C4) usually fails; where the creator intends it, the take must run in a mode that makes no sync claim |
| Tiny faces (small in frame) | Too few face pixels for embedding or sync scoring | Below the evaluator's minimum face size the span is unknown; the plan should place a closer shot |
How these are surfaced#
- At request time, as a pre-flight notice on the specific limitation the request triggers — a whispered line in the script, a scenario tagged S5, a reference set with only profile angles.
- In the delivered evidence, as the per-condition result: a take that passed with 87% mouth coverage says so, rather than reporting an unqualified "sync passed".
- In the published pass rates per cohort (
182.C.03.02), so a creator can see that S5 dancing passes less often than S3 presenting instead of discovering it one generation at a time.
The standing rule#
Every limitation here is a measured or refused case, never a silently degraded one. Two of them — singing and whispering — are refusals in exact-dialogue mode today, and they stay refusals until an evaluator can score them, not until a model claims it can render them.