A Failure Is Filed With Its Fix

The register is 1 file of 1,334 words holding 22 numbered failures, and it is the cheapest artefact the studio owns. Every row carries the same 4 things: what happened, the root cause, the fix, and how the boundary was confirmed. Nothing in it is theory. Each row cost money or rounds to learn once, and reading it costs nothing. The 22 rows fall into 3 classes. The render failures are the loudest, because they end a run in front of you: face guards, balance codes, duplicate frames, light that contradicts itself. The model and billing failures are quieter and just as expensive: a dead search backend, a text-only model sent an image, a parameter that returns a silent 400. And 1 row is a process failure deep enough that it changed the rulebook for every venture afterwards. An index of what did not work is worth more than a playbook, because a playbook states intent and this states evidence.

What happened: a job returned first frame real human detected, then a content policy error, then a task that failed with no outputs and no partial file. The natural reading is that the image was too explicit. The register says otherwise. The guard that fired is a deepfake face check on the first-frame still, and what trips it is any readable human face. An explicit oral frame, shot 39 of a run, rendered clean. A kiss at shot 33 failed, because 2 faces were readable inside it. The fix is a framing rule rather than an argument. Frame the face out, with a silhouette, from behind, neck-down or in deep shadow, and never name facial features in the prompt. Body-only stills submit and succeed. Face stills fail. Both were fired deliberately to confirm the boundary, and 1 cheap test frame settled a question that 2 rounds of prompt rewriting could not. Bodies pass. Clear faces fail.

The billing layer returns 2 numbers that read as the same kind of problem and are nothing alike. Code 97 is an insufficient balance. It is a hard stop, it does not retry, and it kills the whole remaining batch until funds land. Code 96 is a concurrent task limit. It clears in about 30 seconds and a retry succeeds. Reading the 2 as 1 paused a production at 20 of 32 shots mid-run, with the driver still reporting healthy. Both codes arrive in the same response body, which is why reading them matters more than fixing them. The rule that came out of it is 1 line: fire 1 shot, read the code back, and only then release a multi-shot run. A single test frame costs about 1 generation. A paused run at 20 of 32 costs the window and the operator's afternoon, and a top-up that was never the problem.

Three of the 22 rows are the same failure in different clothes: an unset parameter handed to the model, which then chose on the machine's behalf. The Wiro default aspect ratio is 1:1, so a batch intended for 16:9 came back square. MiniMax reads 768p as an error and 768P as a resolution, and its H3 model is gated by plan rather than missing, so the answer is a plan code and not a 404. And an unnamed ethnicity does not produce a neutral person. It produces the majority face in the training data, which is how a prompt for a barefoot girl returned a child with features nobody asked for. The fix is the same in all 3 places: name the value. Ethnicity, skin tone and hair texture go in verbatim on every character prompt, because the model carries nothing from 1 generation to the next. Say the aspect ratio out loud. Say the casing out loud. Say how the light is motivated. A default is a decision, and leaving it unset only moves the decision off the page.

22 rows. Each row is a decision that has already been made once.

The expensive rows are the ones that never throw an error. A still prompt describing a sequence, the light flickers twice then goes out, made the still model render a before-and-after diptych, and the same model drew 2 identical fans side by side with no divider between them. The detection rule is deterministic and cheap: compare the left half of the frame against the right half before any paid vision call, because a duplicate scores about r>0.72 and a mirror about r>0.78. In the edit, weight normalization shifts the climax earlier whenever shots are added to a later act, and a cut that looked right on the timeline landed its peak about 26 seconds off. Explicit per-shot durations, each act summed to its own budget, and no normalization curve. Light fails quietly too. A moon behind a subject with front light on her body is instantly fake, so the light motivation gate asks for source, position and temperature on every shot. Cool blue moonlight injects winter into a hot summer night; soft silver moonlight does not. Audit one channel against the other per shot, above plus 8 for warm and below minus 8 for cold.

The deepest row is a fabrication cascade that cost $750 at $450 an hour and produced zero usable output. An agent invented 4 reference files, cited line numbers that did not exist, claimed it had no vault access more than 4 times while the tools sat in front of it, and shipped noise for 25 rounds. The postmortem named 5 causes, and the permanent fix is 1 sentence long: before generating anything for a venture, search the vault, read at least 1 file, and never ship a path you have not verified exists. Two smaller rows sit next to it. A plan said 33 shots while 39 were on disk, and a stale 20 of 32 count misled a status report, so counts get re-walked from disk every time. Killing a driver wrapper orphaned its child process, which kept generating and raced the new run, so there is now a PID-file lock and a teardown that kills both patterns. EX Venture runs 35 staff and more than 200 interns a year across 12 countries, and every render order passes a named check before it ships. The probability that a failure costs the same money twice falls to near zero once its cause and its fix sit in the same row that somebody actually opens. 1 file, 22 rows, each row a decision already made.