I have the exact source. Now writing the assessment. Note one precision point for Part 1: the production runner uses the 2-layer model (carroll6_5pft_2layer), but carroll6.py (read above) is the canonical reference for the equations, bounding, and CARROLL_VALUES. I'll be precise about this.

DarwinDiff Scientific Assessment: The Neural Network Mechanism and Utilization of ECCO-Darwin v05 / Carroll¶

⚠ SUPERSEDED FRAMING (2026-06-27). This is a point-in-time record; its data and negative results stand, but its framing is corrected by STATUS.md and docs/research_notes/2026-06-27_box_homogenization_DEFINITIVE.md. Specifically: R_PICPOC is NOT a "6/6 wall" and is NOT cluster-gated — it recovers at 1° box scale given a real calcite anchor (Daniels CP:PP / MODIS PIC) plus the RATIO_MAX=2 fix for the contaminated Southern-Ocean ratio target; the differentiable Darwin calcite port and native resolution were tested and did not help. The project is reframed as a surrogate-to-model identifiability study over 4 OBSERVABLE params {alpfe, scav_rat, diatomgraz, R_PICPOC}; the growth pair {Smallgrow, Biggrow} is unobservable by construction (excluded, not failed). The surrogate gap is dimensional (the 0-D box homogenizes spatial structure, tracer CV→1e-15), so box-vs-Darwin spatial-pattern correlations are not fidelity metrics — identifiability comes from real absolute anchors.

Prepared for the DarwinDiff team and Jonathan Lauderdale. Date: 2026-06-11. Branch 2imi9/stage1-per-aoi-gating, HEAD 1704a44. Code-grounded against src/darwindiff/networks.py, carroll6.py, carroll6_5pft_2layer.py, scripts/run_v3.0_joint_multi_aoi.py, and scripts/build_darwin_ic_cache.py.

Part 1 — What the neural network is, and how it enables Carroll-6 recovery¶

1.1 The DINN is a per-cell function from environment to the six parameters¶

The learnable object is not the Carroll-6 parameters themselves. It is a small neural network DINN (networks.py:78-133) whose weights W define a function

θ(x) = DINN_W( env(x) )            # env(x) ∈ R^{C}, θ(x) ∈ R^{6}, at every grid cell x

Architecturally DINN is a stack of 1×1 convolutions (networks.py:118-128): Conv2d(C, 16, k=1) → Tanh → Conv2d(16,16,k=1) → Tanh → Conv2d(16,6,k=1). A 1×1 kernel has no spatial receptive field, so the network is mathematically identical to applying one tiny MLP independently at each cell — cell x's six outputs depend only on cell x's own covariate vector env(x) (docstring networks.py:9-16, 81-83). This is a deliberate design constraint, not an accident: it is the formal content of the "per-cell" identifiability argument in §1.3.

In the production v3.0 runner the input is deliberately thin: n_input_channels = 1 + USE_AOI_ID_CHANNEL + USE_MLD_CHANNEL (run_v3.0_joint_multi_aoi.py:831), both optional channels default OFF, so a default run feeds SST z-score only (one channel). The richer 4-channel form (SST+MLD+wind+lat) lives in DINNDeep and the nb20-series builders but is not the production path.

1.2 Sigmoid bounding into physical Carroll-6 ranges¶

The DINN output is unbounded (networks.py:104). It is mapped into physical ranges by bounded_params (carroll6.py:307-354):

θ_physical = lo + (hi − lo) · sigmoid(θ_raw)

with per-parameter [lo, hi] from PARAM_BOUNDS (carroll6.py:82-94): alpfe ∈ [0.05, 1.0], scav_rat ∈ [3e-8, 3e-6] s⁻¹ (a 100× window around Carroll's 6.0e-7), Smallgrow/Biggrow ∈ [0.10, 2.0] d⁻¹, diatomgraz ∈ [0.05, 1.0], R_PICPOC ∈ [0.005, 1.5]. Two consequences matter scientifically:

The optimizer never leaves the physical box, and gradients are well-behaved (sigmoid is smooth, saturating at the bounds). The R_PICPOC bound was deliberately raised from 0.20 to 1.5 (carroll6.py:88-93) after a diagnostic showed the binding constraint is inter-AOI tension (eqpac PIC/POC median 0.031 vs natl 0.722, a 23× spread), not the bound — a finding that directly motivates several Part-2 levers.
bounded_params broadcasts over [6], [6,H,W], [B,6,H,W] via param_axis (carroll6.py:347-354), so the identical bounding applies whether we fit one scalar, a per-cell map, or a batched ensemble of seeds.

1.3 The loss backpropagates through the forward-integrated box model¶

This is the "Informed" in DINN (networks.py:86-90). The training loop is:

Predict a per-cell Carroll-6 map: θ(x) = bounded_params(DINN_W(env(x))).
Integrate the box model forward from a fixed initial state, per cell, holding the ~94 non-learned Darwin knobs at literature defaults (carroll6.py:52-64: M_LIN, M_QUAD, G0_GRAZE, W_SINK, K_FE, PHI_DUST, Q_FE, LIGHT, H_MLD). The reference equations are in carroll6.py:97-141 (5-tracer) and :178-256 (7-tracer + carbonate); production uses the 15-tracer 2-layer variant carroll6_5pft_2layer.carroll6_5pft_2layer_step, stepped ~200× at dt=0.25 d (~50 d, ≈5× the sinking timescale) to a near-steady state.
Compare the endpoint to targets with a z-scored spatial-pattern MSE per tracer (run_v3.0_joint_multi_aoi.py:919-930, 954-966) across ~11 channels (FeT, Chl1-5, POC, PIC, DIC, ALK, CO2_flux) plus optional absolute-units anchors (PIC_ABS_W, POC_ABS_W, POSI_W, F_CO2_ABS_W) and GEOTRACES surface/subsurface terms.
Backpropagate the scalar loss through the entire forward integration (autograd computes the box model's adjoint) and through the carbonate solver, all the way to W. One optimizer step updates the network weights; the recovered Carroll-6 is read out at convergence and compared to CARROLL_VALUES (carroll6.py:69-76) by relative-offset bands.

The key move: the box model is differentiable, so "what parameter field best reproduces Darwin's fields" becomes a gradient-descent problem on W, not a Green's-functions ensemble of forward runs.

1.4 The structural identifiability argument (per-cell vs a single global vector)¶

DarwinDiff's claim over Green's-functions calibration is not network capacity — the nets are tiny by design (networks.py:22-25). It is about what the parameters are allowed to be conditioned on:

Green's-functions / global-scalar class: one Carroll-6 vector for the whole domain (the benchmark DINNRegional, networks.py:37-75, region-level MLP, 166 weights). A single vector must satisfy every regime's misfit simultaneously.
DINN per-cell class: θ(x) is a continuous function of local environment. Because the truth structure is genuinely per-cell, the per-cell map can express regime contrasts (HNLC eqpac vs iron-replete subpolar) that a single vector cannot, and the multi-AOI joint loss lets each parameter's identifying signal come from the regime where it is observable.

This is also why the project's hard problems are structural, not capacity-limited: the deeper DINNDeep (networks.py:175-243, ~8.5K weights, residual + GELU, per-cell LayerNorm at :136-150 specifically so shared norm stats never couple cells) recovers no more Carroll-6 params than the baseline and does not extrapolate spatially (held-out r=0.301 on FeT). Capacity is not the bottleneck; observational identifiability and box-model fidelity are.

1.5 Why seeding from the true Darwin pickup IC sharpened recovery (the v2.8 scav_rat result)¶

The forward integration must start somewhere. Through v2.7 the start state used synthetic/literature ICs. v2.8 replaced the 5 inorganic anchors (DFe, POC, PIC, DIC, ALK) × 2 layers with depth-weighted means extracted from the real Darwin pickup_ptracers (build_darwin_ic_cache.py:59-65; L1 = levels 0-4, L2 = levels 5-28). This produced the project-first reproducible scav_rat recovery. Mechanistically:

scav_rat is identified by the subsurface iron balance dDFe_2 = fe_remin − scav_rat·DFe₂·POC₂ + diffusion (carroll6_5pft_2layer.py:351-355). The 2-layer box exists precisely to break the surface alpfe↔scav_rat degeneracy via the subsurface DFe gradient (carroll6_5pft_2layer.py:6-11).
The scavenging sink is the product scav_rat · DFe₂ · POC₂. If DFe₂ and POC₂ start far from Darwin's true subsurface values, the optimizer can fit the surface fields by mis-setting scav_rat to compensate the wrong DFe₂·POC₂. Seeding DFe₂, POC₂ (and DIC/ALK/PIC) at Darwin's actual subsurface state pins the other factors of the product, so the gradient on scav_rat points at the real value rather than a compensating one. The IC sets where in the iron/carbon budget the optimizer starts, and for a product-form sink that starting point is load-bearing.

1.6 What the NN does NOT buy (be honest)¶

The DINN reweights and conditions only the 6 learnable parameters. It cannot fix two classes of error:

Box-model proxy bias. The forward model is a 5-7 (production: 15) tracer 1-deg, 0-D-per-cell, time-mean proxy of the 39-tracer Darwin SMS. Any structural mismatch — no macronutrient limitation (growth is iron-only, LIGHT=1.0 constant, carroll6.py:60, 120-122), no temperature/Eppley scaling of biology, no explicit zooplankton (grazing is one linear diatom-only term g_diatom·G0_GRAZE·Pl, carroll6.py:126), a single global sinking rate for POC and PIC, a spatially-uniform PHI_DUST, no horizontal advection (so eqpac upwelled iron is mis-attributed to dust) — is absorbed into the fixed ~94% of knobs and biases the recovered 6. No architecture removes this.
Fundamental observational under-determination. Where a parameter's signal is absent from the loss, no network recovers it. R_PICPOC under pattern-only z-loss is degenerate because z-scoring cancels the magnitude that R_PICPOC·mort_total controls (run_v3.0_joint_multi_aoi.py:269-277); scav_rat is unrecoverable without the subsurface DFe anchor / Darwin IC. The fixes for these are loss terms (absolute anchors), ICs, and additional AOIs — not the net.

Part 2 — Have we fully utilized ECCO-Darwin v05 and the Carroll papers?¶

Verdict: No — the forward-model physics and the time axis are the two largest under-utilized resources, and they are exactly the levers most coupled to the two stuck parameters. What is well-utilized is the iron-pair anchoring machinery (subsurface GEOTRACES + Darwin ICs) and the per-cell NN architecture. What is barely tapped is (i) the seasonal cycle (everything is fit to a 23-yr time-mean), (ii) Darwin's own dense gridded diagnostics (primProd, PAR, POSi, c01-c07), and (iii) Carroll's actual cost function / obs set (we score distance-to-optima, never out-of-sample obs misfit).

2.1 Highest-value untapped items (ordered by recovery value)¶

#	Untapped item	Type	Carroll-6 param it bears on	Value	Cost
1	No temperature/Eppley scaling of biology (growth = μ·f_Fe·1.0·P; T enters only the carbonate solver)	physics/forward-model	Smallgrow, Biggrow	High	Laptop
2	Monthly climatology / seasonal cycle (all targets are 23-yr time-mean; `monthly_climatology()` exists, unused)	data axis	Smallgrow, Biggrow, diatomgraz	High	Laptop→cluster
3	Native monthly `primProd` (Darwin's diagnosed NPP — never loaded)	data product	Smallgrow, Biggrow	High	Laptop
4	Native monthly `PAR` + a real `f_light` term (`LIGHT=1.0` hardcoded)	data product + physics	Smallgrow, Biggrow	High	Laptop
5	Native monthly `POSi` (dense gridded biogenic silica — direct diatomgraz target)	data product	diatomgraz	High	Laptop
6	`SiO2` macronutrient + Si-limitation in growth (box has no nutrient limitation at all)	physics/forward-model	diatomgraz, Biggrow	High	Laptop→cluster
7	Separate PIC sinking rate `wPIC_sink` (single global `W_SINK` aliases R_PICPOC by w_POC/w_PIC)	physics/forward-model	R_PICPOC	High	Laptop
8	Spatially-varying dust (`ironFile`) vs uniform `PHI_DUST` (recovered alpfe = product alpfe·PHI_DUST)	physics/forward-model	alpfe	High	Laptop
9	Coccolithophore-only calcite flag exists but OFF by default (default scales PIC by total mortality)	physics/forward-model	R_PICPOC	Med-High	Laptop
10	FeT_L2 as a loss target (cache has it; used only as start-state)	IC content	scav_rat	High	Laptop
11	Carroll's cost function / out-of-sample obs validation (we score distance-to-optima, never obs misfit; GLODAP loader built but unused in v3.x)	Carroll machinery	all 6 (methodology)	High	Laptop
12	Native monthly `POFe` (particulate iron — direct scav_rat sink target)	data product	scav_rat	Med	Laptop+box ext
13	Per-cell MLD as the carbonate flux divisor (hardcoded `H_MLD`/H1 = 50 m)	data product / physics	R_PICPOC (indirect)	Low-Med	Laptop
14	Carroll→PFT mapping (`lumped_mapping` OFF: Smallgrow learned on Pro-HL only vs Carroll's group rate)	physics/definition	Smallgrow, Biggrow	Med-High	Laptop

Items not listed (SSS-as-DINN-feature, seaIceArea, area-weighting, PO4/O2/DOM tracers, daily product, native carbonate-diagnostic duplicates, satellite PIC/PACE, SOCAT/BGC-Argo/WOA/WOD/GLODAPv2.2023) were confirmed untapped but rate low for Carroll-6 — mostly because the box has no state variable for them, they duplicate fields already in hand, or they touch the fixed ~94% rather than the 6.

2.2 The 3–5 levers most likely to move the stuck params or break the 5/6 ceiling¶

For diatomgraz (~10-15% Cal-grade, collapses to 0% with the Southern Ocean in the loss): - Dense gridded POSi (item 5). The sparse GEOTRACES-bottle proxy already lifted diatomgraz Cal-rate ~13%→80% in nb33; a dense, model-consistent target is the cleanest single lever, and it is diatom-specific so it does not compete with the iron-pair budget. This is the top candidate for the 6th parameter. Caveat: POSi needs a TRAC_MAPPING entry built from the .meta fldList, and the Southern Ocean bSi already fights Carroll's 0.83 (it prefers ~0.33) — so the apples-to-oranges tension in §2.3 applies. - Si-limitation in growth (item 6). Today diatomgraz must absorb variance Darwin attributes to Si-limitation; adding an f_Si co-limitation factor removes that structural mis-attribution.

For R_PICPOC (~3%, the hardest param): - Separate PIC sinking rate (item 7) + coccolithophore-only calcite (item 9). The single W_SINK collapses the steady-state surface PIC/POC to exactly R_PICPOC, biasing it low by w_POC/w_PIC; and scaling PIC by total mortality (default) instead of coccolithophore mortality makes a single R_PICPOC scalar try to fit the 23× eqpac/natl spread — structurally impossible. Both are forward-model fixes the absolute-anchor approach cannot substitute for, and they attack the documented binary PIC-anchor mutex from the physics side rather than the loss side.

For breaking the 5/6 ceiling generally: - Seasonal cycle (item 2). Every loss is on the 23-yr mean, integrating away exactly the spring-diatom-bloom / summer-pico contrast that separates the growth/diatom tier. This is the single biggest unexploited axis and converts one static snapshot into ~12 phase constraints per parameter. Caveat: monthly-resolved fitting changes the integrator from steady-state to transient (cluster-gated at scale). - Eppley temperature scaling (item 1). Across AOIs at 26 °C (eqpac) vs 5-10 °C (subpolar), a single learned μ cannot match both unless the DINN papers over a missing physical term via SST conditioning. This is a top suspect for the multi-AOI Smallgrow/Biggrow non-recovery and is a few-line forward-model change.

2.3 The honesty checks — methodology gaps and what is genuinely maxed-out¶

Two methodological caveats that bound every "recovery" claim (these are not headroom; they are framing risks Jon should weigh): - We never pinned down Carroll's actual assimilated obs set or cost function. Carroll 2020's headline target was seasonal surface pCO2 / air-sea CO2 flux vs SOCAT-type products, minimized as an absolute weighted misfit. Our loss is z-scored spatial-pattern MSE over Chl/POC/PIC/FeT against Darwin's own v05 output. These are different cost functions on different observables. A param landing far from Carroll's value could fit our obs (or even real obs) better — and we have no out-of-sample obs-misfit evaluation to check, only distance-to-Carroll bands. Building a withheld-GLODAP/SOCAT misfit test (the GLODAP loader is already built, glodap_loader.py, just unused in v3.x) would convert the comparison from apples-to-oranges (params) to apples-to-apples (obs). - Carroll's R_PICPOC = 0.04245 is incompatible with satellite PIC. MODIS PIC is ~16-20× higher than Darwin v05 PIC in eqpac, so a satellite anchor implies R_PICPOC ~0.5-0.6, not 0.042. Per Jon's resolution (no satellite in v05 targets), MODIS/PACE are correctly shelved — but this means the satellite loaders are not latent R_PICPOC headroom within the Carroll-6 framing.

What is genuinely maxed-out (do not over-claim headroom here): - Network capacity. DINNDeep (8.5K weights) recovers no more params than the SST-only baseline; the ceiling is structural, not capacity. Adding a bigger net buys nothing. - The iron-pair anchoring stack. Subsurface GEOTRACES iron + Darwin pickup ICs already deliver reproducible scav_rat (v2.8) and the multi-AOI HNLC-vs-replete contrast for alpfe. The remaining alpfe/scav_rat improvements are second-order (Q_FE quota bias, DOM iron recycling, spatial dust pattern), not a missing primary anchor. - The PIC/POC absolute anchors as a path to 6/6. PIC_ABS_W / POC_ABS_W were swept down to 0.02, paired and solo, at n=10-20: any nonzero PIC anchor wipes the iron pair (binary basin mutex). They are informative and tested-to-exhaustion; they are not undiscovered headroom. R_PICPOC progress must come from the forward-model side (items 7, 9) or the seasonal axis, not from doubling down on absolute PIC magnitude in the z-loss. - Per-cell gating as currently configured. The Stage-1 2-AOI per-AOI gradient gating result (HEAD 1704a44) is negative (mean 2.95→1.65/6); the 3-AOI confound-check is still pending and requires building the southernoceanpac IC cache.

Cheap / laptop-feasible now: items 1, 3, 4, 5, 7, 8, 9, 10, 14, and the GLODAP out-of-sample obs-misfit eval (item 11). Most are forward-model edits (carroll6_5pft_2layer.py) or single loader additions (a TRAC_MAPPING entry + one line in the native-load list at run_v3.0_joint_multi_aoi.py:368).

Cluster-gated: full seasonal/transient fitting at scale (item 2), macronutrient-co-limited box re-validation (item 6), a regional differentiable port of the full darwin_plankton.F, and time-resolved alpfe/scav_rat disambiguation requiring multiple pickup months (only one pickup is on disk).

Bottom line for the team: the NN and the IC/GEOTRACES iron stack are working as designed and are near their useful limit. The unexploited headroom is overwhelmingly in forward-model fidelity (temperature, light/PAR, Si-limitation, separate PIC sinking, coccolith-only calcite, spatial dust) and the time axis (seasonal climatology) — and those are precisely the resources coupled to the two parameters (R_PICPOC, diatomgraz) blocking 6/6. The cheapest highest-value next step is dense gridded POSi for diatomgraz; the cheapest highest-value R_PICPOC step is a separate wPIC_sink plus enabling coccolithophore-only calcite. Both are laptop-feasible forward-model changes, not cluster runs.

Relevant files: C:\Users\Frank\OneDrive\Desktop\Github\ecco-darwindiff\src\darwindiff\networks.py, ...\src\darwindiff\carroll6.py, ...\src\darwindiff\carroll6_5pft_2layer.py, ...\src\darwindiff\silica.py, ...\scripts\run_v3.0_joint_multi_aoi.py, ...\scripts\build_darwin_ic_cache.py, ...\src\darwindiff\llc270_loader.py, ...\docs\findings\v3.1_closeout.md, ...\docs\findings\v2.8_darwin_ic_poc_sub.md.