Skip to content

Pre-registration — hold-together sweep (2026-06-26, before results)

Written before the 8 configs (sweep_holdtogether_20260626_1119/ + job 7888644) complete, so the analysis cannot be cherry-picked after the fact. Decisions, success criteria, and reward-hacking checks are fixed here.

Hypotheses

  • H1 (load-bearing test). The real Daniels CP:PP anchor is what constrains R_PICPOC, not the Darwin-pattern loss. Prediction: the dan0_control config (no Daniels) reproduces the historical v3.2 behaviour — R_PICPOC does not co-recover — while the Daniels configs do. If dan0_control already recovers R_PICPOC, the anchor is not load-bearing and the headline must be downgraded.
  • H2 (hold-together). A single config can hold ≥3 of {alpfe, scav_rat, diatomgraz, R_PICPOC} together with R_PICPOC real-anchored. Prediction: uncertain — this is the open question.
  • H3 (iron dose trade-off). Increasing GEOTRACES_W (0.3 → 1 → 3) trades alpfe for scav_rat (the ironboost in B.3 dropped alpfe to 4/50). Prediction: geo3* drop alpfe; geo0.3 hold the pair.

What each config isolates (one-lever ablation from the base)

config lever changed isolates
base (7888644) — the combined operating point
dan0_control DANIELS_RPICPOC_W 1→0 H1 — is Daniels load-bearing for R_PICPOC?
dan2 DANIELS_RPICPOC_W 1→2 does a stronger anchor help or distort?
geo1_dan1 / geo3_dan1 / geo3sub10_dan1 GEOTRACES_W 0.3→1→3 (+SUB) H3 — iron dose vs alpfe/scav_rat
noeppley_dan1 USE_EPPLEY_T 1→0 Eppley's role in diatomgraz
noposi_dan1 POSI_W 1→0 the bSi term's role in diatomgraz

Success criterion (fixed in advance)

Per-AOI ≥2-AOI co-recovery (Cal-grade-or-better vs Carroll, re-derived via band_of), n=10 seeds, verify_run.py --expect-seeds 10 exit 0. The "winner" is the config holding the most of the four params at ≥2-AOI co-recovery ≥7/10 — reported alongside every other config, including failures.

Reward-hacking / failure-mode checks (must pass before any positive claim)

  1. Per-AOI, not joint. Grade per-AOI co-recovery, not joint cell-weighted — the 2026-06-15 ALK-anchor result was a false positive (a cell-weighted average straddling Carroll while no single AOI recovered).
  2. The "we told it the answer" caveat (precision, not hacking). The Daniels loss pins R_PICPOC to the real per-cell rain ratio (~0.05). Carroll = 0.0425. So a Cal grade means "real CP:PP data is consistent with Carroll," not "the method discovered Carroll's value from Darwin output." The honest, non-circular claim is a validation that independent data agrees with Carroll, conditional on H1 (Daniels is load-bearing). State it that way; do not call it a from-scratch discovery.
  3. diatomgraz via sparse proxy. Dense Darwin POSi is not staged on the cluster; POSI_W uses the sparser GEOTRACES bSi. If diatomgraz recovers, check it is not an artifact of a few bSi cells; if it fails, the cause may be the proxy, not the method — flag both.

Decision rules

  • Report all 8 configs (positive and negative), as a table with numbers — no narrative without the table.
  • If H1 fails (dan0_control recovers R_PICPOC): downgrade the "real-data recovery" framing project-wide.
  • If a winner holds ≥3 params with R_PICPOC real-anchored: robustify at n=100 before any headline.
  • If no config holds them together: that is a publishable negative result — the params trade off even with real anchoring — and is reported as such.