The Nine Failures Stay On The List

The Wiro face-guard rejection is the #1 render failure and the failure is not about sex. The signature on the wire is first frame real human detected, content policy, task fails with no outputs. The root cause is a deepfake face-guard on the first-frame still, and any readable human face trips it. Explicit oral at shot 39 rendered clean. A kiss at shot 33 failed because 2 faces were readable inside the frame. The fix is to frame the face out - silhouette, from behind, neck-down, deep shadow - and never name facial features in the prompt. The identity block carries body plus hair plus skin only, plus face not visible. A diagnostic runs in seconds and the diagnostic is the cheap gate: submit a body-only still (succeeds) against a face still (fails) and the boundary confirms in 2 renders. Balance code 97 is insufficient balance and it is a hard stop that is not retryable. The whole remaining batch dies until funds are topped up. The operator rule is to fire 1 shot and read the code back before any multi-shot run, because the read is the only signal that the wallet is live. The Sal y Miel production paused at 20 of 32 shots mid-run on a 97 because the run had not verified the balance up front. Balance code 96 is the concurrent task limit and the fix is to wait about 30 seconds and retry. The cheapest gate is 1 Wiro task at a time, because parallel submission is the only path to a 96 and a serial pipeline never trips it. A driver that submits in parallel pays the wait 4 to 6 times across a 32 shot run, and the wait is what loses the hour. The 2 codes look the same to the human eye and they are not the same at all. 97 is fatal, 96 is patient. Reading the code back is what separates the 2 cases, and the read is the cheap gate that ships every multi-shot run.

Ethnicity drift bit the Salvador shoot when the prompt named a barefoot girl and the renderer defaulted to the training-data majority. The reviewer wrote the child was clearly Asian and the render was unusable for a Brazilian-set story. The root cause is that the renderer does not remember between generations, so ethnicity has to be named verbatim in every character prompt or the model drifts back to its training mean. The fix is ethnicity plus skin tone plus hair texture in every character prompt, verbatim, no exceptions. Afro-Brazilian, deep brown skin, honeyed sheen, long dark wavy hair is the locked spec for the Salvador character. The lock carries across still and video, and the lock is the cheap gate that holds the character's identity across 32 shots. The super-AI glossy look is failure 4 and the look is the polished digital-painting look that the eye reads as obviously synthetic. The root cause is photorealistic plus blanket cinematic film still plus beautiful plus perfect plus flawless on a realism-biased model. The 5 words stack into a glossy composite that has nothing to do with documentary and everything to do with stock imagery. The fix is to drop photorealistic, drop blanket cinematic, and replace them with documentary language plus negative fill plus Tier 1 negatives. Documentary film, 35 mm, natural skin texture, visible pores, unretouched, natural light, negative fill. The Tier 1 negative prompt lists glossy, airbrushed, plastic, waxy, CGI, 3D render, beauty filter, uncanny, morphing face, identity drift, extra fingers, deformed hands, watermark, text. The composite holds the look in documentary register and the register is what reads as real. The 2 failures share 1 underlying cause. The renderer drifts toward the training-data mean whenever a name is left unspecified, and the drift shows up as a glossy composite because the mean is glossy. The fix for both is the same verb: name. Name the ethnicity. Name the medium. Name the Kelvin. Naming is the only counter to drift.

Face drift on held shots happens because the model re-invents the face every frame and the model has no memory of what it generated a moment ago. A long static face hold dissolves in front of the viewer and the dissolve is what reads as 2 different actors in the same scene. The dissolve is most visible at eye level under steady light, and the steady light is what catches the drift in profile. The fix is to keep face shots short, 1 to 3 seconds, or moving, and to hold medium-to-wide framing rather than extreme close-up. Partial shadow or hair across the face hides micro-drift, and the hide is what keeps a long take on screen without the eye flagging the dissolve. The rule is to limit face close-ups to the emotional peaks and to hold wide for everything else, and the holding is what ships a watchable film. Split-screen duplicate or mirrored stills is failure 6 and the failure mode is 2 flavors. The first is the diptych: a still prompt that names a temporal sequence (the light flickers twice then goes out) made the still model render a before-and-after image side-by-side. The second is the pair: a still model that draws 2 identical fans or 2 identical chairs in the frame with no divider. The fix is to treat stills as 1 frozen moment. Sequence and motion live in the video prompt only, never in the still prompt. The cheap gate is deterministic duplicate detection before the paid vision call: left-versus-right half correlation, duplicate at r greater than 0.72, mirror at r greater than 0.78. The detection runs in milliseconds and the rejection runs before the credit, and the credit is what the gate protects. A still is a freeze frame. A still is not a storyboard panel. The renderer treats sequence as 2 panels and the 2 panels are the failure. The fix is the prompt and the fix is the gate, and the 2 together ship a single frozen moment per render.

Weight-normalization shifting the climax is the silent cut. The implementation is per equals song divided by sum of weights and the normalization looks harmless. The normalization shifts the climax about 26 seconds earlier whenever shots are added to a later act, and the shift is invisible to the editor until the cut lands a peak at the wrong beat. The fix is to write explicit per-shot durations, sum each act to its budget, and never weight-normalize. A 32 shot film with 4 acts and explicit durations holds the climax on the right beat because the math is the math. A 32 shot film with weight-normalized durations drifts earlier on every add, and the drift is what loses the climax. The cardinal light error is failure 8 and the error is moon behind the subject with front light on her body. The physics does not exist. A moon at the back of the frame means the subject is backlit, rim light, cool front in shadow. Front light needs a motivated source - a candle, a lamp, a streetlight - and the source carries its own temperature. A render that ships moon behind plus front light is rejected by the viewer in 1 frame because the rejection is pre-linguistic. The fix is the Light Motivation Gate and the gate is 3 fields written into every prompt: source plus position plus temperature out loud per shot. Moon behind means rim light at 5600 K, front in shadow at 3200 K ambient. Front light means a candle at 1900 K, off-axis at 30 degrees, hard shadow on the wall. The gate catches the inconsistency before the inconsistency ships, and the catch is what saves the frame. The 2 failures look unrelated and they are not. Weight-normalization and the cardinal light error are both silent cuts that pass the editor and fail the viewer. The fix for both is to write the gate explicitly. The math is the math for time. The physics is the physics for light. The gate holds both.

The 9 failures stayed on the list because every 1 of them cost money or rounds to learn, and the cheapest read is the one that names the root cause and the fix together.

Warm-cool temperature drift flips the look of a hot summer night into a winter night inside the same scene. The drift happens because words like cool blue moonlight silently inject cold, and the cold pulls the white point down to 6500 K and the kelvin down to 4500 K. The viewer reads the flip as a season change and the season change is the failure. The fix is silver instead of blue. Soft silver moonlight carries the same emotional register without the cold injection. Silver is a metal and metals read as warm or neutral depending on the key. Blue is a temperature and temperatures carry a number, and the number is what flips the scene. The audit is R-B channel per shot. R minus B greater than plus 8 is warm. R minus B less than minus 8 is cold and the cold is flagged. The flag runs on every still before the still is approved, and the flag is what catches the silent drift before the drift ships. A scene that holds within plus or minus 8 across all shots reads as 1 scene. A scene that drifts 30 in 1 cut reads as 2 seasons, and 2 seasons is the failure. The 3-layer rule is the fix. Layer 1 is the named Kelvin in every prompt (2000 K match flare, 2700 K candle, 5600 K moon, dawn amber). Layer 2 is the held Kelvin across the scene (no swaps mid-act). Layer 3 is the R-B audit per shot at plus or minus 8. The 3 layers hold the temperature, and the temperature is what holds the scene. A scene that reads as 1 climate reads as 1 film. A scene that reads as 2 climates reads as 2 films stitched together, and the stitch is the failure. The Kelvin is the lock and the lock is what ships the film.

The 9 failures stayed on the list because every 1 of them cost money or rounds to learn, and the cheapest read is the one that names the root cause and the fix together. The list is not a museum. The list is the operating manual for the next shoot, and the next shoot is paid for by the gates the list holds. 3 patterns cut across the 9 failures. Pattern 1 is unspecified defaults. Ethnicity, Kelvin, aspect ratio, lens, and motion are all fields the renderer leaves to the training mean when the prompt leaves them blank. The fix is to name every field the eye reads, and the reading is what catches the drift. Pattern 2 is silent cuts. Weight-normalization, temperature drift, face drift, and light motivation all pass the editor because the editor is looking for the obvious error. The fix is the pre-render gate, and the gate is the audit that catches the silent cut. Pattern 3 is shared infrastructure. The face-guard frame-out, the negative prompt, the deterministic duplicate detection, the R-B audit, the Light Motivation Gate - the gates are reusable across the next 100 shoots and the reuse is what holds the cost down. The probability that a future shoot hits a fresh failure is not zero. The probability that the fresh failure lands on the list and the list holds the root cause and the fix is high, and the list is what makes the next shoot cheaper than this one. The operating rule is 1 sentence. Before firing any shot, read the list. Before approving any still, run the gates. Before shipping any film, walk the 9 failures one more time and confirm none of them is in the cut. The walk is the last gate, and the gate is what the next shoot inherits. Reading the list is cheaper than re-learning the failure. The list is the operating manual, the manual is the edge, and the edge is what the next shoot ships on.