Pre-scale-up verification — before NU AICR (B200) / native resolution¶

⚠ SUPERSEDED FRAMING (2026-06-27). Point-in-time record; data/plans stand, framing corrected by STATUS.md. The project is a surrogate-to-model identifiability study over 4 OBSERVABLE params {alpfe, scav_rat, diatomgraz, R_PICPOC} — not a 6/6 chase or a '5/6 ceiling / parameter-conservation' result. The growth pair {Smallgrow, Biggrow} is unobservable by construction. R_PICPOC is recoverable at 1° with a real calcite anchor (Daniels/MODIS) + RATIO_MAX=2 — the differentiable Darwin calcite port and native resolution were tested and did not help, so R_PICPOC is not cluster-gated. geo1 holds {alpfe, scav_rat, R_PICPOC} jointly 7/10; diatomgraz is an open iron-pair tradeoff. The surrogate gap is dimensional (the 0-D box homogenizes spatial structure, CV→1e-15), so box-vs-Darwin pattern-matching is not a fidelity metric.

Companion to STATUS. Records what was verified (and fixed) on the laptop and the free Explorer H200 before spending paid AICR B200 time, on the principle maximize cheap verification before expensive cluster runs.

Method. A 5-dimension adversarial audit (parallel subagents grounded in the code, each finding independently re-verified) on 2026-06-18, then direct laptop execution of the cheap items. Hardware: RTX 5090 Laptop, 23.9 GB (not 32 GB).

Verified good — no action¶

Assumption	Result	Evidence
Batched-seeds ≡ sequential	exact, `0.0` diff	element-wise step; `test_seasonal_integrate`
`steps_per_month=122` converged	inorganic <0.03%; phyto diff on a near-zero value	spm 122 vs 976
torch Blackwell-capable (B200 = sm_100)	`sm_100` + `sm_90` in `get_arch_list()`	torch 2.11.0+cu128, all-wheel lock
Native LLC270 load + bin-to-1°	works end-to-end (ran live)	`llc270_loader`
Required tracers present	PIC/POC native; Chl1/CO2_flux via bin_average	disk audit
env builds without a compiler	uv.lock is all wheels, no sdist	login-node `uv sync` is safe

Fixed (laptop)¶

Stale editable install. Bare python resolved darwindiff to a stale worktree (.claude/worktrees/…), so the documented command ImportError'd and a fresh pip install -e . could run old code. Re-pointed the editable install to main src; runner docstring → uv run. (The project .venv was always correct.)
DIC/ALK runaway (commit 4b7f147). At the upper PARAM_BOUNDS the surface carbonate budget (dDIC_1/dALK_1 have no analytic floor, forward-Euler has no clamp) ran away monotonically negative under multi-cycle seasonal spin-up — DIC_1 → −2.3e4 by 8 cycles — while staying torch.isfinite, so the one-step finiteness test missed it; a latent NaN-at-native-resolution. Added a >= 0 floor in carroll6_5pft_2layer_step. Provably result-neutral: an eqpac time-mean recovery (250 ep) is bit-identical pre/post; all 21 2-layer + seasonal tests pass; new test_param_bound_extremes_stay_nonnegative_under_spinup fails pre-fix.

Verified (laptop measurement)¶

Memory scaling holds at 5×. Measured 21.24 GB at 64M cell·step vs 21.25 GB predicted by the 356 B/(cell·step) fit — the linear law does not bend at 5× beyond the prior max measured point. Shrinks the LLC270 time-mean extrapolation gap 14.8× → 2.96×. (The harness + fit live on the unmerged PR #102 branch.) NB torch.compile cannot run on this Windows laptop (Inductor needs MSVC; silently falls back to eager), so the compiled memory constant and the seasonal speed-up remain Explorer/AICR checks.

Forward-model probes (laptop) — can the B200 ask be shrunk?¶

Adversarial point from the audit: R_PICPOC may be blocked by forward-model physics, not resolution — one global W_SINK serves both PIC and POC (collapsing surface PIC/POC to exactly R_PICPOC) and coccolithophore-only calcite is OFF by default, so one R_PICPOC scalar must fit the 23× eqpac↔natl PIC/POC spread.

Coccolith-only calcite + light PIC anchor (COCCOLITH_ONLY=1, PIC_ABS_W=0.02, 3-AOI, n=10): NEGATIVE — the mutex holds. Mean 0.30/6 (max 1/6). The iron pair is wiped (alpfe 0/10, recovered 0.10–0.39 vs Carroll 0.93; scav_rat 0/10) and R_PICPOC is not recovered (1/10 — a single fluke at 0.041; the rest drift high to 0.13–0.24 vs Carroll 0.042). Even the lightest PIC anchor the mutex was probed at (0.02) still triggers the binary PIC-anchor mutex with coccolith-only calcite — an 8th independent box-scale exclusion of R_PICPOC (joining loss weight, AOI mix, architecture, IC, Eppley physics, PIC anchor, ALK anchor). This does not shrink the B200 ask; it strengthens the native-resolution case. Untested laptop lever remaining: a PIC:POC ratio loss (vs the absolute PIC anchor) — the "estimator-side is laptop-fixable" claim.
Box-vs-Darwin forward fidelity (scripts/box_vs_darwin_fidelity.py; box at CARROLL_VALUES, 3-AOI, 200 steps): the box pins surface PIC:POC to a single scalar. At Carroll's R_PICPOC=0.042 the box's steady-state PIC:POC is 0.0424 / 0.0425 / 0.0424 (eqpac / natl / SO) — essentially constant — while Darwin's actual ratio spans ~100×: 0.033 / 0.676 / 0.0067. The box matches eqpac (1.28×) but is 16× too low in natl and 6× too high in SO. So R_PICPOC is structurally unidentifiable jointly — not for lack of resolution, but because the box's PIC production (R_PICPOC × mortality) cannot represent Darwin's spatially-varying calcite:organic ratio. (The box relaxed off the Darwin-field IC to its own R_PICPOC steady state, so this is robust; per-field z-scored r is confounded by that IC and not reported.)

Synthesis — the two probes complete the B200 argument¶

The 1° box cannot hold Darwin's 100× PIC:POC spread with one R_PICPOC scalar (fidelity), and coccolith-only calcite at 1° fails the mutex (coccolith probe). Together these say the fix is coccolithophore-resolved calcite at native resolution — where Chl2 (coccolithophore) bloom structure supplies the spatial PIC variation a single 1° scalar cannot. That is precisely the mechanism the native-res B200 run targets, now empirically grounded by two laptop experiments rather than asserted.

Cluster-gated — do on the FREE Explorer H200 before paid B200¶

Env build / uv sync on a login node (confirm download-r2.pytorch.org reachable).
compile ≡ eager recovery-grade equivalence (Linux has the compiler).
GPU-util + wall-clock profile at native cell counts — the "~1 h/run, ~0% util" is assumed, never measured, yet it underpins both the GPU-hour ask and the AICR idle-GPU-reaper risk (which would kill the Python-loop-bound seasonal runner).
Native-resolution single-AOI seasonal recovery on H200.

B200-only (unavoidable paid checks)¶

sm_100 runtime driver satisfies the cu128 runtime (the sbatch matmul line).
Global LLC270 seasonal fits 192 GB without checkpointing.

Takeaway¶

Most of the AICR premise is testable without AICR, and most "native-resolution" testing runs on the free, granted Explorer H200 — the genuinely B200-only checks are narrow. Run the fast forward-model probes first: if they recover R_PICPOC on the laptop, the wall is physics not resolution and the B200 ask shrinks accordingly.