Skip to content

Config / Results Matrix

The single source of truth for what each DarwinDiff configuration tested and what it found. Project Status links here for the current best; the archive holds the per-version writeups behind each row.

Scoring (diagnostics.band_of): Excellent ≤ 5 % off Carroll · Cal-grade ≤ 40 %. Counts are seeds-recovered / seeds-run, every number gated by scripts/verify_run.py (exit 0 = re-derived from raw). The honest target is 4 observable params {alpfe, scav_rat, diatomgraz, R_PICPOC}; the growth pair {Smallgrow, Biggrow} is unobservable by construction (no real data constrains growth rates) and is excluded from the target, not counted as a miss.

A. Box-scale parameter-recovery configs (the spine)

Config Scale / AOI What it tested Verified result What's distinct
v2.0 1° box, Eq Pac 7-tracer carbonate joint loss (carbonate.py) iron pair → 1.1 % (alpfe) / 40 % (scav_rat) off Carroll First degeneracy break — the carbonate cycle constrains the iron pair
v2.6 1° box, Eq Pac GEOTRACES IDP2025 absolute-iron MSE 4/6 Cal-grade, reproducible n=10; diatomgraz 9/10 A real absolute iron anchor (not z-scored) is what moves alpfe
v2.8 1° box, Eq Pac Darwin v05 pickup ICs + L2 POC z-score scav_rat 7/10 Cal (4/10 Excellent) First reproducible scav_rat; reveals (alpfe,scav_rat) bimodal degeneracy selected by POC_SUB_W
v3.0 1° box, 2-AOI (eqpac + natl) Shared Carroll-6 joint multi-AOI training 5/6 plateau over 50+ seeds; mean_cal 3.93 Establishes multi-AOI training; per-AOI DINN falsified at 2-AOI
v3.1 1° box, 3-AOI (+ SO Pac) Basin C lever set (POSI + AOI_W + CHL1) iron pair 38/40 (95 %) at n=40; 5/6 ceiling 2/856 3rd AOI makes the iron pair rock-solid; only 2 single-seed 5/6 in 856
v3.1.1 1° box, 4 AOI mixes (n=200) AOI ablation AOI mix selects which params recover; eqp+natl best ≥4/6 rate 24 %; no mix → 6/6 Recoverability is an AOI-attribution tradeoff, not a single ladder
v3.2 1° box, 3-AOI Eppley f(T) + dense Darwin POSi (TRAC16) mean 3.85/6; reproducible 5/6 (4/20, n=20); iron pair + diatomgraz together First gain from forward-model physics; breaks the alpfe↔silica mutex. diatomgraz here is vs Darwin's own silica, not real data

B. Resolution & the R_PICPOC resolution

Config Scale / AOI What it tested Verified result What's distinct
native LLC270 Native res, Eq Pac Does native resolution change iron recovery? Resolution selects which iron param recovers (native → alpfe, 1° → scav_rat; 8/10 each, per PRs #122 / #123 — not archived) — it does not lift the recovery count (cluster_setup.md) Native resolution is not a 6/6 unlock — it reshuffles which iron param recovers
R_PICPOC fix (RATIO_MAX + Daniels) 1° box, 3-AOI Was R_PICPOC a "6/6 wall"? R_PICPOC 10/10, landing at real ~0.05 (consistent with Carroll's 0.0425 within the ±40 % Cal band — not a validation) once the SO ratio target is sanitized (RATIO_MAX=2) + a real calcite anchor (Daniels CP:PP / MODIS) is added Refutes the "wall": the gap was a contaminated target + a missing real observation, not calcite physics or resolution. Deeper spine-D result: Carroll's global R_PICPOC is itself under-constrained and mis-specified vs a regionally-variable rain ratio (Daniels eqpac ~0.039, ≈1.6× global)
geo1 hold-together (current best) 1° box, 3-AOI Joint hold of the observable trio {alpfe, scav_rat, R_PICPOC} jointly 7/10 (3-of-4-observable frontier; tied with base/dan2 at n=10) The current operating point — real, Darwin-independent anchors hold 3 of 4
per-cell vs global 1° box, 3-AOI, geo1 Is the per-cell DINN load-bearing? per-cell 7/10 vs global scalar 0/10 for the trio (scav_rat 8/0, R_PICPOC 9/0; Fisher p < 0.01) Identifiability rests on per-cell anchoring, not pattern-matching against Darwin

C. Track-2 hybrid — feasibility only (self-twin / synthetic, not real Darwin)

These are feasibility probes on a self-twin: the model fits data it generated itself. They show the differentiable-UDE machinery runs — not that Darwin has been made differentiable, that real biology was learned, or that environment-gated calcification is proven. Real-data results live in tables A/B.

Probe Scale What it tested Verified result Honest scope
0-D hybrid probe 0-D box Do the 4 UDE ingredients run on the real box equations? All 4 pass: closure MAE ~0.5 %, parameter to 2e-5 %, RK4 drift < 0.001 % Feasibility, self-twin synthetic
spatial UDE probe 9,216-cell field Can an env-modulated closure be learned over a field? 2-D closure recovered 0.44 % MAE; torch.compile ≈ 3× speedup Feasibility of the spatial Track-2 thesis; still a synthetic twin

Reading the matrix

  • The current operating point is geo1 (table B): real GEOTRACES iron + a real Daniels/MODIS calcite anchor hold {alpfe, scav_rat, R_PICPOC} jointly in 7/10 seeds. diatomgraz is an open iron-pair tradeoff (recoverable in principle via dense Darwin POSi/TRAC16, not yet staged).
  • "6/6" is the wrong frame. The growth pair is unobservable; R_PICPOC was never a wall (table B refutes it). The honest target is the 4 observable params. The deeper spine-D result is that Carroll's global R_PICPOC is itself under-constrained and should be regional (the real rain ratio varies by AOI) — the recovery is consistency within the wide Cal band, not a validation of 0.0425.
  • The surrogate gap is dimensional. At uniform parameters the 0-D box homogenizes to a near-uniform state (tracer CV → ~1e-15 vs Darwin's O(1)), so box-vs-Darwin pattern correlations are not fidelity metrics — identifiability comes from real absolute anchors, which is exactly what makes the per-cell DINN load-bearing (table B, per-cell vs global).

Provenance

Each row's underlying writeup is archived (verified research record, kept out of the onboarding path):