Skip to content

Hold-together sweep — results (2026-06-26, H200, n=10×8 configs)

8-config sweep run on Explorer (gpu-short/H200), n=10 seeds each, all 8 verify_run.py exit 0. Pre-registered before results (commit 9a4c1d7); conclusions below are after a 4-agent adversarial review (run wf_e1b125fc) that computed Wilson CIs / Fisher exact and red-teamed each claim. Every conclusion survived only with qualification — the qualifications are the point.

Verified table — per-AOI ≥2-AOI co-recovery (out of 10)

config (lever vs base) alpfe scav_rat diatomgraz R_PICPOC joint{alpfe,scav,Rpic} joint{+diatomgraz}
base (Daniels=1, Eppley, sparse POSi, default iron) 8 7 0 10 7/10 0/10
geo1 (GEOTRACES_W=1) 10 8 0 10 8/10 0/10
dan2 (DANIELS_W=2) 8 8 0 10 7/10 0/10
dan0_control (DANIELS_W=0) 10 10 0 3 3/10 0/10
geo3 (GEOTRACES_W=3) 10 0 4 10 0/10 0/10
geo3sub10 (W=3+SUB=10) 10 0 1 10 — —
noeppley (Eppley OFF) 10 0 0 10 — —
noposi (POSI_W=0) 6 8 4 9 5/10 3/10

Two — and only two — statistically robust effects at n=10

Everything else (differences among the 7/8/9/10 cells) is within sampling noise: Wilson CIs are 0.28–0.52 wide and pairwise Fisher tests are non-significant (p ≥ 0.47). Do not read the high cells as a lever ranking. The two effects that clear Bonferroni:

  1. A ratio observable is load-bearing for R_PICPOC. Anchor OFF (dan0) → R_PICPOC 3/10 with per-AOI means scattered (eqpac 0.006 / natl 0.061 / SO 0.075); anchor ON (base) → 10/10 with means converged (0.052 / 0.052 / 0.048 ≈ 0.05). Wilson CIs disjoint ([0.11,0.60] vs [0.72,1.00]); Fisher p = 0.003; the value-convergence (not just the count) is the hard-to-fake signal.
  2. High iron weight trades scav_rat away. geo3/geo3sub10 (W=3) → scav_rat 0/10 while alpfe stays 10/10. Wilson CIs non-overlapping vs geo1's 8/10; Fisher p = 7e-4.

Hypothesis verdicts (qualified)

H1 — Daniels load-bearing: survives, but narrow the claim. The direction is robust (above). But: - It isolates "any ratio observable," not "the real Daniels data specifically." The adversarial review flagged that this sweep never ran the RATIO_W-on / Daniels-off control (Darwin's own ratio as anchor) — but prior project work already did: finding_rpicpoc_wall_broken_ratiomax shows the Darwin-ratio anchor (RATIO_W=2, RATIO_MAX=2) recovers R_PICPOC 10/10. Combined with the self-twin (R_PICPOC intrinsically identifiable), the conclusion is that any ratio constraint suffices to recover R_PICPOC — the surrogate just needs a ratio anchor. So the Daniels anchor's unique value is not "it recovers R_PICPOC" (Darwin's own ratio does too) but "it does so non-circularly": it anchors to real CP:PP data (landing at ~0.05) instead of grading against Darwin's own PIC/POC. A same-config RATIO_W replication would tidy this, but the principle is already established. - Recovery is to the real ~0.05, not exactly Carroll's 0.0425. "Consistent with Carroll" is weak: the Cal-grade band is ±40% (factor-1.6), so Daniels (0.027), Carroll (0.0425), and recovered (0.05) all fall in one band. And the project's own rain-ratio finding calls Carroll's value biased ~25–30% low — report "consistent within uncertainty" alongside "Carroll is under-constrained," never "validates 0.0425." - In this 3-AOI sweep, only eqpac + natl have Daniels coverage; the term auto-gates off in SO, so SO's R_PICPOC is not anchor-tested (and SO is the weakest AOI).

H2 — hold-together: 3 of 4, jointly, at geo1 — and the 4th is a tradeoff, not a block. - geo1 holds {alpfe, scav_rat, R_PICPOC} jointly in 8/10 seeds (measured, not inferred; the later per-cell ablation re-aggregation gave 7/10 — a single band-edge seed, so report it as 7–8/10) — but is statistically tied with base/dan2 (7/10) at n=10. Report "tied-best," and run n≥20 to separate. - No config holds all 4. diatomgraz never exceeds chance (best 4/10, CI [0.17,0.69]; pooled 9/80 = 11%). "Proxy-blocked" is refuted as stated: noposi turns the sparse proxy OFF yet diatomgraz rises to 4/10 (joint-4 = 3/10, the only config that ever holds 4). diatomgraz=0 in geo1 is the documented diatomgraz↔iron-pair tradeoff — geo1 chose the iron pair. It is recoverable in principle via the dense TRAC16 POSi target (10/10 in prior v3.2), which is not staged on the cluster — so the question is open, not solved, and the honest takeaway is "diatomgraz not recovered in this sweep."

H3 — iron tradeoff: the fact is robust, the "dose" framing is not. scav_rat 8→0 at W=3 is bulletproof. But it is a threshold/cliff at high weight, not a graded monotonic dose: base sits at lower iron weight than geo1 with identical scav_rat (no gradient), there is no intermediate W=2 point, and scav_rat=0 is also reachable via noeppley (Eppley off, base iron) — so high iron is sufficient but not necessary. noposi (alpfe 6 / scav 8) shows the two iron params are not a single locked seesaw.

Honest headline

The real CP:PP anchor adds R_PICPOC on top of the iron pair: geo1 holds {alpfe, scav_rat, R_PICPOC} jointly in 8/10 seeds (7/10 on the per-cell ablation re-aggregation — band-edge) — a 3-param co-hold the project had not shown before. But (i) the denominator is 4 observable params (the growth pair is excluded by construction — no real-world constraint exists), (ii) the 4th param diatomgraz is not recovered here (a tradeoff, pending the dense TRAC16 target), and (iii) only two effects are statistically real at n=10 — the anchor→R_PICPOC and the iron-weight→scav_rat tradeoff. This is a frontier, not a 6/6, and not a discovery of Carroll's value.

Next steps (from the review)

  1. n≥20 at geo1 and dan2 — break the statistical tie for "best operating point" (highest priority).
  2. Stage dense Darwin POSi (TRAC16) — the only way to adjudicate diatomgraz observability.
  3. Report seed-variance/CIs on the per-AOI mean R_PICPOC values (not just point estimates).
  4. (optional) same-config RATIO_W replication — the principle ("any ratio anchor suffices") is already established by finding_rpicpoc_wall_broken_ratiomax; this would just confirm it in the 3-AOI config.

Provenance: verify_run.py exit 0 on all 8; pre-registration 9a4c1d7; adversarial review wf_e1b125fc (4 agents, Wilson/Fisher, all verdicts survives_with_qualification).