Skip to content

R_PICPOC wall broken by a contaminated-ratio-target fix; robust 6/6 hits a structural ceiling

⚠ SUPERSEDED FRAMING (2026-06-27). Point-in-time record; data stands, framing corrected by STATUS.md. The project is a surrogate-to-model identifiability study over 4 observable params; the growth pair is unobservable by construction. R_PICPOC is recoverable with a real calcite anchor (the '6/6 wall / 5/6 ceiling / needs the Darwin port' framing is refuted). The dimensional surrogate gap (box homogenizes) — not calcite physics — is the real limit.

Date: 2026-06-24 (H200) · Status: COMPLETE — all numbers gated through scripts/verify_run.py (recompute-from-raw, exit 0). Code behind NEW default-OFF flags (RATIO_MAX, RATIO_AOI_W_*, RATIO_SCHED_START, USE_ENV_RAIN_RATIO); legacy reproduces bitwise when unset. Uncommitted; runs in /projects/schultz/qi.zim/runs/rpicpoc_* on NU Explorer.

TL;DR

  1. R_PICPOC — the project's sole documented 6/6 wall — is broken. With one fix it recovers 10/10 joint Cal, genuine per-AOI (no straddle), R_PICPOC≈0.030–0.046 vs Carroll 0.0425, robust across the ratio cap (2→5) and weight (1→2).
  2. The cause was a data-pipeline bug, not missing physics. The Southern-Ocean PIC:POC ratio-loss target was contaminated by near-zero-POC cells (per-cell pic/poc up to ~1e8, target mean 4.7e7). That inflated the loss scale normalizer (mean(target²)) and silently collapsed the SO ratio term to ~0, leaving R_PICPOC unconstrained there. Capping the target at a physical rain ratio (RATIO_MAX=2) fixes it. This explains the prior "ratio loss recovers R_PICPOC only in eqpac" finding (rpicpoc_ratio_structural.md).
  3. The deep-research's prescribed fix (environmental rain-ratio gating, ADR/option 3) is REFUTED — it produces fewer recoveries (R_PICPOC 3/10) and a straddle artifact. The simple data fix beats the elaborate mechanism.
  4. An all-six event appears (3/10 seeds, verified) — but it is a tuned, init-anchored result on the looser JOINT metric, NOT a robust recovery and not a headline (iron pair init-anchored in the full loss; the real grounding for the iron pair is the separate real-GEOTRACES validation, not this joint-metric event — see the 2026-06-25 profile/FIM diagnostic). Across 8 lever classes the binding constraint is a structural identifiability tension: you can robustly recover {R_PICPOC, iron pair, diatomgraz} OR the growth pair {Smallgrow, Biggrow}, not all six together. Longer training does not help (3000 epochs made Biggrow worse) — it is a loss-landscape tradeoff, not undertraining.

Setup

  • Config: v3.2 Eppley 3-AOI operating point (AOIS=eqpac,natlsubpolar,southernoceanpac, USE_EPPLEY_T=1, POSI_W=1, AOI_W_NATL=AOI_W_SO=2, CHL1_W_EXTRA=3, POSI_DARWIN_W=0.5, DARWIN_IC=1), 1° resolution, compiled, n=10 seeds, 1500 epochs.
  • Recovery: RATIO_W>0 PIC:POC ratio loss (identifies R_PICPOC orthogonally to the iron pair — mort_total cancels at steady state), COCCOLITH_ONLY=0.
  • Targets (corrected, ratio-of-means): eqpac 0.033 / natl 0.68 / SO 0.0067.
  • Forward-model feasibility probe first (scripts/probe_env_rain_ratio.py): an env rain ratio g(T,Ω_c) CAN reproduce the spread (log-MSE 0.000) — but the n=10 fit then showed the gating is unnecessary and harmful (see below).

The fix: RATIO_MAX (cap the per-cell ratio target)

scripts/run_v3.0_joint_multi_aoi.py: cells with Darwin pic/poc > RATIO_MAX are dropped from the ratio mask (default inf = legacy). Excludes the low-POC artifact cells (87 in eqpac, 41 in natl, the SO blow-up) without touching genuine cells.

Verified results (n=10 each; verify_run.py exit 0)

arm 6/6 (joint) 6/6 (per-AOI ≥2) mean cal R_PICPOC iron pair note
baseline RATIO_W=2, cap=2 3/10 0/10 4.4 10/10 clean 5/10 best joint 6/6
baseline-v2 (cache-swap control) 3/10 0/10 4.3 10/10 5/10 reproduces ✓
cap=1 0/10 — 3.0 4/10 straddle 2/10 cap too tight
cap=5 2/10 — 4.5 10/10 (7 Exc) 7/10 alpfe 9/10
env-ON (gating) 0/10 — 2.9 3/10 straddle 4/10 REFUTED
RATIO_W=0.5 0/10 — 4.3 9/10 straddle 8/10 iron best, RPP weak
RATIO_W=1 1/10 1/10 4.6 10/10 straddle 7/10
RATIO_W=4 0/10 — 3.9 10/10 2/10 scav_rat collapse
natl ratio=0 (per-AOI wt) 2/10 0/10 4.4 9/10 straddle 6/10 no help
natl ratio=0.5 1/10 — 4.3 10/10 straddle 6/10
PRIMPROD_W=1 0/10 — 4.1 10/10 straddle 5/10 Biggrow worse (2/10)
PRIMPROD_W=3 1/10 — 3.4 10/10 1/10 alpfe collapse
POC_ABS_W=0.1 0/10 0/10 2.4 straddle 0/10 iron pair WIPED (magnitude mutex)
scheduled (warmup 0.5) 2/10 1/10 4.7 10/10 clean 7/10 best operating point
3000 epochs 0/10 0/10 4.3 10/10 straddle 6/10 longer ≠ better (Biggrow 1/10)

Per-param at the best (scheduled) arm — per-AOI co-recovery (≥2 AOI): R_PICPOC 10/10, alpfe 10/10, scav_rat 10/10, diatomgraz 10/10, Smallgrow 1/10, Biggrow 3/10. Four of six are robustly per-AOI; the growth pair {Smallgrow, Biggrow} is the residual wall.

Interpretation

  • R_PICPOC recovery is genuine, not a cell-weighted straddle — verify_run.py raises RPICPOC_STRADDLE for cap=1 and env-ON but NOT for cap=2/cap=5/scheduled.
  • It recovers Carroll's scalar (0.0425), which is the project's 6/6 criterion. It does NOT reproduce Darwin's 100× spatial PIC:POC spread — the flat-ratio box can't, and the env-gating that could is refuted as a recovery tool. Carroll himself used a single global R_PICPOC, so recovering ≈0.0425 is the right target.
  • Structural ceiling, not a tuning miss: 8 lever classes (env-gating, RATIO_MAX cap, scalar ratio weight, per-AOI ratio weight, PRIMPROD, POC anchor, schedule, epochs) all show the same R_PICPOC/iron ↔ growth-pair trade. The ratio loss that wins R_PICPOC perturbs the growth params (the documented 5/6 had no ratio loss); pinning growth/POC directly wipes the iron pair (magnitude mutex). No scalar or per-AOI weighting, and no extra epochs, satisfies all six robustly.
  • The scheduled ratio weight (ratio loss held at 0 for the first half of training, then ramped) is the cleanest operating point: it lets the iron/growth params settle before R_PICPOC engages, recovering R_PICPOC + iron pair + diatomgraz per-AOI 10/10 and the first strict per-AOI 6/6 — but the growth pair still caps robust 6/6.

Conclusion

The headline is verified and durable: R_PICPOC was never fundamentally unidentifiable — the box-scale wall was a contaminated observation in the loss. Fixing it removes the contaminated-target blocker and lets R_PICPOC recover ≈Carroll's scalar (consistent with real MODIS-Aqua calcite, optimum ~0.028 — Carroll's 0.042 in the acceptable low-ratio basin). The occasional all-six event is a tuned, init-anchored joint-metric artifact, not a robust 6/6 recovery and not a headline. Robust 6/6 is blocked by a genuine growth-pair identifiability tension at this box + observable set — a clean, characterized ceiling, and a stronger result than the env-gating mechanism (now refuted) would have been.

Reproduce

# best operating point (scheduled ratio weight), n=10:
RPP_ENV=0 RATIO_MAX=2.0 RATIO_W=2 RATIO_SCHED_START=0.5 \
  NB23_SEEDS=0,1,2,3,4,5,6,7,8,9 NB23_N_EPOCHS=1500 \
  OUTPUT_DIR=$RUNS/rpicpoc_sched sbatch --export=ALL --partition=sharing \
  --time=01:00:00 scripts/slurm/run_rpicpoc_env_fit.sbatch
# gate + aggregate:
python scripts/verify_run.py $RUNS/rpicpoc_sched --expect-seeds 10
python scripts/aggregate_rpicpoc_env.py $RUNS/rpicpoc_sched
Artifacts: scripts/{probe_env_rain_ratio,aggregate_rpicpoc_env,build_primprod_cache}.py, scripts/slurm/run_rpicpoc_env_fit.sbatch, src/darwindiff/carbonate.py (calcite_saturation), src/darwindiff/carroll6_5pft_2layer.py (env rain ratio).

DRAFT comment for issue #143 (NOT posted — for review)

R_PICPOC wall broken (verified) — it was a contaminated loss target, not missing physics. Full note: docs/findings/2026-06-24_rpicpoc_ratio_target_fix.md.

The SO PIC:POC ratio-loss target was contaminated by near-zero-POC cells (target mean 4.7e7), collapsing the SO ratio term via the scale normalizer and leaving R_PICPOC unconstrained — the real cause of "ratio loss recovers R_PICPOC only in eqpac." Capping the per-cell target (RATIO_MAX=2) → R_PICPOC 10/10 joint Cal, genuine per-AOI (no straddle), ≈Carroll, robust across cap 2–5. First-ever 6/6 (3/10 seeds, verified).

The deep-research's environmental rain-ratio gating (option 3) is refuted: fewer recoveries + straddle artifact. The simple data fix wins.

Robust 6/6 not achieved — 8 lever classes (incl. scheduled ratio weight, POC anchor, 3000 epochs) all hit the same structural trade: {R_PICPOC, iron, diatomgraz} vs the growth pair {Smallgrow, Biggrow}. Best operating point = scheduled ratio weight (4 params per-AOI 10/10; first strict per-AOI 6/6). This is a characterized identifiability ceiling, not a tuning miss. Recommend closing the "R_PICPOC is the wall" framing and reframing spine D around the growth-pair ceiling.