We wrote down what we'd test, and now there's a real results commit

This is the second write-up out of the experiment register: feature/h3-prompt-builder-seed-consistency-test had a clean, well-formed hypothesis on record and, until now, no results ever committed anywhere. A second commit (fdbefb0) adds a RESULTS.md with the silencedetect table below.

What came before it

An earlier branch, feature/h3-prompt-builder-consistency-test (no "seed" in the name), vendored the MiniMax prompt-builder tool and tried three different prompt variants once each: one with realistic music cues, one with no audio cues, and one built around a prompt-builder setting that suppresses certain narrative discipline cues to keep a scene deliberately silent. It merged into main on 2026-08-20, approved after Gavin reviewed the render output.

The name promised a consistency test. It didn't run one. Three different prompts, each rendered exactly once, tells you whether a prompt variant produces silence, not whether that result holds up across repeated attempts. One data point per condition is not a distribution.

The redesign

Gavin caught this on review and a follow-up branch, feature/h3-prompt-builder-seed-consistency-test, redesigned the test properly. Its single commit (acd9b25, 2026-08-21) states an actual hypothesis and dependent variable: take the already-validated silent-scene prompt configuration and rerun it three times, varying only RandomNoise's seed, to check whether the discipline-suppressing setting reliably keeps the scene silent regardless of which seed lands, rather than being a fluke of the one seed already tested.

The commit is thorough. It defines three sibling workflow files (h3_promptbuilder_consistency_run1/2/3.api.json), each byte-identical except for noise_seed (919880931791466, 481027336619204, 702255819944871) and a distinct filename_prefix so all three outputs land side by side instead of overwriting each other. It adds scripts/compare-consistency-run.sh, a host-side script meant to fetch all three rendered clips from comfyui-local and print a per-run ffmpeg silencedetect verdict table, so a human could read one table instead of eyeballing three separately-downloaded videos by ear.

What's missing

That's where the trail stops. git log main..feature/h3-prompt-builder-seed-consistency-test shows exactly one commit: the redesign itself. There is no second commit recording that the three runs happened, no silencedetect table committed anywhere, and the branch itself was never merged to main. docs/lab-notebook.md on main has no entry mentioning this test by name.

It's possible the three renders happened and compare-consistency-run.sh was run by hand, off the record, the way the script's own comments describe (a manual step, not a Concourse task, because task containers can't reach comfyui-local's output volume). It's equally possible nothing past the workflow-file setup ever ran. Both are consistent with what's actually in the repo, which is nothing beyond the setup.

What checking the render host itself turns up

comfyui-local's live output volume, which is not part of the git repo and wasn't checked by the original branch work, held the answer. docker exec comfyui-local find /opt/ComfyUI/output -iname '*consistency*' turns up real render output: two or three takes each of H3_promptbuilder_consistency_run1, run2, and run3, all timestamped 2026-08-21, the same day as the redesign commit acd9b25. Concourse's own build history confirms the mechanism: builds #7295 and #7936, both comfyui-branch / run-changed-workflows, both succeeded. The renders happened, for real, through the normal pipeline.

Pulling the latest labeled take of each run directly via comfyui-local's read-only /view endpoint (a GET against existing output, not a /prompt dispatch) and running ffmpeg -af silencedetect=n=-30dB:d=0.5 against each confirms the hypothesis the branch set out to test: all three report silence across the full clip duration (5.184s), regardless of which of the three seeds produced it.

Run 1, seed 919880931791466. Retrieved from comfyui-local's output volume (H3_promptbuilder_consistency_run1_00003_.mp4), not from git. ffmpeg silencedetect reports silence for the entire 5.184s.

Run 2, seed 481027336619204 (H3_promptbuilder_consistency_run2_00002_.mp4). Same silencedetect result: silent for the full 5.184s.

Run 3, seed 702255819944871 (H3_promptbuilder_consistency_run3_00002_.mp4). Same result again: silent for the full 5.184s.

That confirms the hypothesis: the silence-discipline prompt-builder setting holds across all three tested seeds, not a fluke of the one seed the predecessor branch (feature/h3-prompt-builder-consistency-test) tested once. It's now written down where it should have been from the start: RESULTS.md on this branch (commit fdbefb0), with the actual per-run silencedetect table.

Where it actually landed

Confirmed, and now durable. The hypothesis was well-formed, the harness existed in committed form, and the renders did happen and did support the hypothesis: silence held across all three seeds (919880931791466, 481027336619204, 702255819944871), N=3, single fixed prompt/graph per this branch's own design. No docs/lab-notebook.md exists on this branch (it postdates when this branch was cut off main) -- a notebook entry there is a follow-up for whoever actually merges this, not attempted here since merging wasn't part of this pass.