Why the manuscript benchmarks pin --floor none

Decision 2026-07-12. The manuscript’s fit harnesses are pinned to --floor none, while the shipped CLI default stays --floor auto. This page exists because that looks like an inconsistency and is not one, and because three places in the tree cite the reasoning: packages/luxar/src/luxar/tests/test_benchmark_floor_pin.py, scripts/benchmark_progressive_psnr.py, and changelog.d/1184.md.

Relocated and reformatted from TODO.md when that file’s release section was retired. The reasoning is unchanged.

The short version

Keep the legacy no-floor numbers for the paper, for pragmatic — not scientific — reasons, and document the choice. The initial call was a full re-run with floor=auto; a quick measurement changed the how, not the science.

Scientifically, floor suppression is the right thing

The constant background pedestal is not real signal, so removing it before fitting is principled. The correct way to score a floor-suppressed fit is against a floor-suppressed reference — floor-recon vs floor-original.

The measured −5.86 dB is a reference mismatch, not evidence floor is worse

kidney_dapi 31.77 → 25.91; blastocyst +0.01. That comparison scores a background-free reconstruction against the original, pedestal-bearing volume, penalising the fit for correctly dropping non-signal. Under the principled floor-suppressed reference, floor would be fair — and appropriate.

Why keep no-floor for the paper anyway

Adopting floor properly means switching the evaluation to background-relative PSNR: a protocol and narrative change, plus re-checking the blind-spot cross-validation story. That was not worth doing immediately before bioRxiv. The committed legacy numbers are the --floor none numbers (committed kidney 31.81 dB ≈ floor=none 31.77) and are internally consistent — original-referenced throughout — so they stand.

The tool default is confirmed correct and stays

For general CLI use the floor should always be removed by default, because background is not signal. gsplat fit and gsplat cal keep --floor auto.

Do not change the shipped default. The --floor none pin is a paper-only deviation for original-referenced comparability, not a statement about the tool.

Resolution, without a GPU re-run

  1. The numbers stand as-is.

  2. Pin the manuscript fit harnesses to --floor none — run_analysis, run_convergence, run_noise2self, progressive, loss_comparison — so a future re-run stays reproducible instead of silently inheriting floor=auto.

    Amended 2026-08 per #1184: run_noise_floor was listed here in error. In the luxar-paper repo it only calls estimate_noise_floor, which never fits, so there is nothing to pin.

  3. A Methods paragraph stating that the benchmarks use --floor none with original-referenced PSNR for comparability, while floor=auto — the shipped default — is the more principled fit for real use, and noting background-relative evaluation as appropriate future work.

The harness pin and Methods paragraph landed in luxar-paper PRs #8 and #7, respectively, on 2026-07-12; this decision is closed.

Deferred

Post-bioRxiv / journal / future work: the floor-suppressed-reference evaluation, and the small open check of whether floor improves the blind-spot CV / K* selection (a cal sweep on a few datasets, not run_all).