Repository navigation
#1199 stage 17: quality tiers on the long-form templates #1218
Description
Stage of #1199, filed by model
claude-opus-5-5via provideranthropic. Already
built onfeat/long-form-media(merged3f8ed222). The authoritative spec is the
plan (docs/superpowers/plans/2026-11-03-long-form-media.md, task named below) and
the design record. If they differ, the design record says what shipped.
What shipped: a tier preset variable (draft / standard / best, default
standard) on minimax/chained-segments, chain-video-continuity,
chain-matched-to-audio, chain-matched-and-aligned, dialogue-short and
music-video. A bundle moves only width, height, num_inference_steps and
weights_dtype, and shares no variable with the shot presets. tier is in each
template's cost_drivers. standard is today's defaults, so no cost row moved and no
default changed (no breaking-change).
| draft | standard | best | |
|---|---|---|---|
| FL2VA chain | 4 steps (0.66-0.72x wall) | 960x544, 5 steps, int4 | 1344x768, int8 (1.68x) |
| Ref2VA (5 templates) | 5 steps (0.56x) | 960x544, 9 steps, int4 | 1344x768, int8 (1.32x) |
Bench evidence (lem bench :8766 / :8767, bench/records/tiers.md): the template
end to end with shot: static_long at 120 s, aa80458ab4f7, 59.2 min. S1 standard
f5bb2add1911 14.1 min, 4 steps cc6e23eb555a 9.3 min, 1344x768 int8 6810b23c2e35
23.7 min. Draft keeps standard's staging at the same seed (segment-1 PSNR 32 dB on
S1). best needs about 65 GB of host RAM per job (one H3 job at a time on lem).
Observe over MCP:
validate_workflowontemplates/minimax/chained-segments(anddialogue-short) withimage,prompt,segments2 andtiereach ofdraft,standard,best, and with notier: compareplan.estimateand the reportedcost_drivers.run_workflowwithtier: draft(the cheap one), thenget_job_workflowshows the realizedwidth,height,num_inference_stepsandweights_dtype.- A
tier: bestrun needs about 65 GB of host RAM. Run it alone on lem, never beside another H3 job (bug: an int8 H3 job needs about 65 GB of host RAM, and two servers on one host OOM-kill each other #1227).
Acceptance intent:
tier: standard(or notier) realizes exactly today's template values
(get_job_workflow).tier: draftandtier: bestrealize the bundle values in the table; a variable
passed besidetierwins.- An unknown tier is refused at validate, naming the three values.
- The estimate differs by tier (validate's
plan.estimate), andtieris listed in
cost_drivers. - No
tierbundle sets ashotvariable.
Owes: nothing. Deploy: done at 3f8ed222.
verify: run every pending: #1218 case.
Activity
dkackman commentedon Nov 7, 2026
Verified by tester model
claude-opus-5-5via provideranthropic, on lem develop @ 85bcd33, workspacedefault.No suite case carries
pending: #1218(grep acrossregression-suite-*.mdfound none), so acceptance is this issue's own intent list.Ran:
validate_workflow templates/minimax/chained-segments,segments: 2, with tierdraft/standard/best/ none. All valid. Estimates: 7.2 / 5.5 / 17.8 / 5.5 min. Non-standard tiers reportdiffers: ["tier"](best addsweights_dtype), and standard and no-tier share one fingerprint.templates/minimax/dialogue-short: draft / best / none give 11.8 / 41.9 / 38.6 min.tier: "ultra"on all six templates (chained-segments, chain-video-continuity, chain-matched-to-audio, chain-matched-and-aligned, dialogue-short, music-video) is refused witharguments.tier: 'ultra' is not a 'tier' value; allowed: best, draft, standard.draft+num_inference_steps: 6is scaled as "steps 6 vs 5".best+width: 960, height: 544drops the canvas term and keeps theweights_dtypedifference.run_workflow chained-segments {segments: 2, tier: draft}, jobdc4c32402665, succeeded in about 3 min.get_job_workflow(realized) showswidth 960, height 544, num_inference_steps 4, weights_dtype {int4}andshot: moving.presets.tierhas draft {960x544, 4, int4}, standard {960x544, 5, int4} and best {1344x768, 5, int8}. No tier bundle sets anyshotvariable, andtieris first incost_drivers. Output deleted.best(65 GB host RAM, bug: an int8 H3 job needs about 65 GB of host RAM, and two servers on one host OOM-kill each other #1227). I checked its bundle values through the realizedpresetsand validate's plan instead.One finding, filed separately as #1236: the tier estimates do differ, which is what the acceptance list asks. But they are scaled from the curated cost row while
standarduses the observed figure. So chained-segmentsdraftis quoted at 7.2 min, above standard's 5.5 (the real draft run took about 3 min). I filed it on its own because it comes from how the estimator chooses its basis, not from the tier bundles.Closing as verified.