# Honest limitations and troubleshooting Phase 182, `182.C.39.02`. Ten conditions the item names: profile, occlusion, fast speech, singing, overlapping dialogue, tiny faces, rapid movement, weak references, unsupported languages, and model drift. Every number here is a measurement this repo holds, with the module that holds it. Where nothing was measured, the entry says so instead of estimating. ## Start here: the three that fail silently A limitation you can see is a limitation you can report. These three produce output that looks correct to every check we run, which makes them the ones to suspect when a take is wrong and nothing flagged it. | Condition | What the metrics say | What is actually wrong | | -------------------- | ------------------------------------------------- | ------------------------------------------------------ | | Overlapping dialogue | Pronunciation score exactly 0.0 — clean | Timing is 513.8 ms out, four times the reject boundary | | A motionless cast | Speaker attribution at maximum margin — confident | Nobody in the scene is reacting | | Frames nobody scored | Aggregate sync high — passing | The score is only over the still, frontal frames | **Overlapping speech breaks the aligner silently** (`182.C.18.08`, `human-video-delivery-conditions.ts`). A second voice starting half a second in gives a 513.8 ms onset error while the goodness-of-pronunciation stays at exactly 0.0. Forced alignment is given the words and asked where they are; it will always find them somewhere. `assertAlignmentMeasurable` refuses a timing claim on overlapping speech rather than discounting it, because a discounted wrong number is still a wrong number. **A frozen supporting cast buys a perfect attribution** (`182.C.20.03`, `human-video-speaker-attribution.ts`). Nothing else in frame moves, so the speaker is unambiguous: a reacting listener cuts the margin to 0.65 where a still one leaves it at 1.00. The confident-looking number is _caused by_ the defect, so both are reported together. **The measurability filter is a sampling bias** (`182.C.20.02`, `human-video-motion-sampling.ts`). Excluding windows where the face is turned or occluded is correct, and the excluded frames are the moving ones — so the aggregate is "sync, given the face was frontal and still". On a clip that is half movement, the same band scores read 0.908 over what was scored and 0.750 under the clip's own motion mix. The figure is an upper bound, not an estimate. ## The ten conditions ### Profile Sync is scorable on-camera only. Past 60° of head yaw the mouth geometry is foreshortened; past 120° there is no visible speech at all (`human-video-offscreen-dialogue.ts`, boundaries named provisional against a 1.55° pose-measurement noise floor). A back turn is reported **unevaluable**, never as a pass — a shot of the back of a head passes every check there is. _What to do:_ plan the window's visibility. A deliberate profile needs a stated reason, and it is then excluded from the sync arithmetic rather than counted against the take. ### Occlusion A hand, a prop or a mask removes the mouth from measurement. `human-video-measurability.ts` gives every window one of three verdicts and never lets "no findings" read as "no problems". Intermittent occlusion — wind, a passing hand — deliberately stays **scorable**: treating it as unevaluable would excuse a real failure (`human-video-challenge-cases.ts`). ### Fast and slow speech `182.C.18.08` measured it: onset error rises from 32.9 ms to 68.4 ms at 1.6×, still inside the ITU boundary — the aligner survives. What breaks is the **pronunciation** measure, which falls to −4.99 at 1.25× and −10.20 at 0.80× on audio whose words and speaker are identical to the clean clip, against a floor of −3.0. A phase vocoder's smearing looks like bad pronunciation to a frame-level acoustic model, and slow is worse than fast. _What to do:_ nothing, on the creator's side. `assertGopMeasurable` refuses the pronunciation score on time-stretched audio instead of reporting a mispronunciation that did not happen. ### Singing **Not measured, and not fakeable.** Singing is one of six delivery conditions with no fixture (`UNMEASURED_CONDITIONS`), because the conditions that _were_ built are transformations of a signal — a tempo change, a mix, an offset copy — and singing is different phonation. Filtering a normal voice into "a whisper" produces a quiet normal voice, and testing a detector against that proves the detector reads the filter. No claim is made about sung dialogue in either direction. ### Overlapping dialogue See the silent-failure section. Beyond the aligner, attribution cannot work at all: two people speaking at once is one envelope and two mouths moving with it, so the correct answer is **not evaluable by this method**, not "ambiguous" (`human-video-conversation-dynamics.ts`). Filing overlap as a defect makes the ambiguous rate rise with how naturally people talk. Separating them needs per-speaker audio, which no route supplies today. ### Tiny faces Below 7% of frame height the landmarker declined entirely on the measured frame; below 10% identity margin is under a quarter of its clean value (`WARNING_THRESHOLDS` in `human-video-pre-generation-warnings.ts`). These are warnings raised **before** generation, because the alternative is finding out after paying for the render. _What to do:_ change the shot size. The expected face height per shot size is tabulated in `human-video-measurability.ts`. ### Rapid movement At 21 px of motion blur only 4% of identity margin survives. At 31 px a face scored **0.251 against its own clean self**, under the 0.282 two-different-people baseline (`182.C.19.08`). Past that point identity is not low, it is **undecidable**, and the system declines rather than reporting a failure — a take is not rejected for failing a test that could not be run on it. ### Weak references The pack check refuses a pairwise cosine under 0.4. It has a **known false negative** and the code says so: `182.C.07.03` measured two different subjects at 0.261 and one subject against a degraded copy of itself at 0.739 and 0.819, so 0.4 sits between those — but an earlier probe put two strangers at 0.422, above the line. A pack that mixes two people who look alike passes this check. It catches a gross mismatch before money is spent; **it is not an identity gate.** ### Unsupported languages A locale whose normalizer has no rules is a refusal, not a pass-through (`localeNormalizable` in `human-video-pre-generation-warnings.ts`, `hv.capability.locale_unsupported`). Silently sending unnormalised text produces a take with the wrong words in it and no record of why. ### Model drift A provider can change a model under a stable name. Re-runs are triggered by change rather than by calendar, and the demotion tolerance is **derived from measured noise** rather than chosen: a single window of eight scores spanned 0.0400 while three steady windows moved 0.0025 between them, so a tolerance of 0.02 would fire on the spread and demote a route for nothing (`human-video-auto-demotion.ts`). Demotion is automatic because it is the safe direction; promotion is not, because a bounded re-run cannot re-establish what a promotion asserted. ## Troubleshooting: from symptom to cause | Symptom | Look at | | --------------------------------------------------- | ----------------------------------------------------------------------------------------------------- | | "The words are right but the timing is off" | Overlapping speech in the source audio — the aligner will not say so | | "It says mispronounced and it sounds fine" | Time-stretched audio; the pronunciation score is refused, not scored | | "It says nothing about sync at all" | The window was unevaluable: profile, occlusion, or a back turn | | "Identity failed on an action shot" | Check the blur band; past 31 px the answer should be _undecided_ | | "The scene looks dead" | A frozen non-speaking cast, which scores _better_ on attribution | | "The pass rate dropped with no release change" | Model drift; the route's re-run is triggered by change, not by date | | "Two references of the same person were rejected" | The 0.4 cosine floor is provisional and has a known false negative | | "A low-SNR warning on clean audio with music under" | Music under speech is not noise; the percentile SNR reads 16.8 dB at a 5 dB bed, against a gate at 20 | ## What this document does not cover Everything here is measured on fixtures and committed frames. **No real generated clip has been through this pipeline**, so none of these limits has been confirmed against production output, and the pass rates a creator would want — how often a take is good — are `182.C.30.03`'s, waiting on the corpus in `182.C.33`.