Two verdicts landed, one resolved to "never happened"

A short register update against the 2026-09-02 Sankey — what moved since that count, not a full re-survey.

The 21 per-experiment deep-dives published since the last count carry the narrative detail. This post is just the register-level view: three items changed state since 2026-09-02, and the shape of the diagram moved with them.

Sankey flow of 25 open plus 12 merged genops experiments from work domains to terminal states, updated 2026-09-04: krea2-4step-chk14000-distill-test moved from Pending to Mixed, krea2-pixelart-gamelevel-test moved from Pending to Refuted, sopro-tts-experiment moved from Ongoing to Blocked
Same 25 open + 12 merged register as 2026-09-02, three items re-routed to their landed verdicts.

What moved

krea2-4step-chk14000-distill-test → Mixed / partial win. The three-arm comparison (8-step baseline, 4-step + distill LoRA, 4-step isolated with no LoRA) finally got its frame-to-frame read: step count, not the LoRA, is what drives the quality loss — the isolating third arm shows severe degradation at 4 steps with no LoRA present at all. The LoRA recovers most, not all, of that loss: facial structure and materials come back coherent, but skin reads slightly over-smoothed and some wardrobe detail simplifies next to the 8-step baseline. Doesn't match the baseline outright, does substantially close the gap. N=1 per arm, specific to the militants crew prompt — not a production-adoption decision, that's still Gavin's call.

krea2-pixelart-gamelevel-test → Refuted. The Reddit claim that Krea2 Turbo produces authentic 16-bit pixel art from prompt structure alone doesn't hold: no pixel grid, no dithered palette, no tiling, wrong proportions for the claimed format. What it actually produces is closer to pre-rendered 32-bit-era background art or modern "HD-2D" concept art — a real capability, just not the one under test. Clean single-claim binary check, N=1, no further arms planned.

sopro-tts-experiment → Blocked (resolved to "never started"). Last cycle's register carried this one as state-unclear. It's now confirmed: the branch tip is byte-identical to an unrelated merge commit already on main — zero unique commits, and nothing in the repo, workflow files, or commit history mentions "sopro" by name anywhere. Not a stalled experiment with lost work behind it; a branch that was created and never used.

What didn't move

tts-voice-clone-benchmark picked up a fifth validated model (MOSS-TTS-Local-Transformer-v1.5, 12.2GB peak VRAM, RTF 1.1-1.35x, native 48kHz stereo output) since the last count, but it's an ongoing harness build, not a hypothesis test, so it stays routed to Ongoing / non-hypothesis either way — ten candidates left. Full progress table in the harness write-up.

Everything else — the H3 video/resolution/audio/multishot domains, the Pipeline/infra items, the 12 merged branches — is unchanged from 2026-09-02.

Caveats (same as last time)

Same two normalizations as the original post: krea2-4step-chk14000-distill-test counts once per work-domain lens (H3 video and Krea2/T2I speed), so the middle-column totals don't sum to 25; and every Confirmed/Refuted/Mixed flow already passed a confound check upstream of this diagram (seed pinned, prompt held constant, one dependent variable per arm) — the diagram shows the shape of the register, not the audit trail.

Source

Survey baseline: 2026-09-02. This update: 2026-09-04.