---
source_hash: 5b42e9251b8ef3f32db2d6bdca39fa4f85818f327856c21f376679c87cfd9284
post: /posts/workflows/minimax-h3-lock-the-still
---

# Context Pack: Lock the still, move one action

Reference for Ari's August 2026 independent production account, route /documents/workflows/minimax-h3-lock-the-still. Model identities/workflow are reported process; retained files establish metadata but do not authenticate generation provenance.

## Verbatim excerpts

> Lock the still. Move one action. Fix one variable.

> That confirms the file properties. It does not prove native generation rather than upscaling, or authenticate a model name.

> A thumbnail cannot establish whether the movement holds.

## Chain and locks

Codex session Terra conducts GPTImage2 references → front/rear product plates → short MiniMaxH3 motion → dailies/QC/correction. Conduct owns the sequence. Five craft locks: product identity; world composition/light; one verb plus one camera/subject move; named correction while other requirements stay visible; ultrawide delivery.

Product is a black performance SUV, originally described as Defender-class. Front/rear plates specify identity/spare; references do not guarantee preservation. First/middle/end checks plus movement inspection catch spare/paint/geometry problems invisible in a good middle frame. Findings describe this run, not a general benchmark.

Splat failure combined camera jump/body loss. Revised brief locks camera, narrows rear dissolve, keeps front metal. One-variable fix is a working discipline, not proof that the pair changed only one experimental parameter. Exit pair compares missing-spare failure to selected rear/spare result. Keep all pairs and thumbnails; stills cannot prove temporal success.

Five raw selects: 01ice_wall cold-light readability; 02glacial_river water/body solidity; 03splat_controlled solid front/dissolving rear; 04basalt_arches world scale; 05reality_exit rear/spare. The public assets include seven distinct MP4s, including fail_splat_multi and fail_exit_no_spare; 13 image occurrences, 9 video occurrences and 6 S7 frames remain in the body.

## Delivery evidence and limits

Read-only ffprobe confirms every retained MP4: H.264, 2944×1248, 24 fps, duration 4.458344 seconds. Ratio 92:39 ≈ 2.359:1; 21:9 is nominal. No inference of native rendering/no-upscale/model identity follows from metadata. Ari's original direct-from-motion-model/no-upscale account remains attributed. A 16:9/phone crop changes this composition and needs review, not a blanket claim every crop is wrong.

The private 30-second corrected conform was reference only, too large for the public asset cap, not embedded or newly verified. Music/voiceover remained unselected; no finished mix. Independent production study, not manufacturer work/approval, not one prompt producing a complete commercial. Do not invent a finishing pipeline or infer controlled causality from selected samples.

Assets: apps/web/public/posts/minimax-h3-lock-the-still. Preserve the conduct diagram, still stack, motion card, illustrative correction dial, splat pair and delivery ruler. Companion essays: /documents/workflows/fable-made-me-start-using-codex and /documents/workflows/every-stage-its-own-folder.