Skip to content

R_PICPOC observability/mechanism — overnight deep-research synthesis

⚠ SUPERSEDED FRAMING (2026-06-27).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.

Date: 2026-06-24 (overnight) · Status: LITERATURE SYNTHESIS, decision-ready (no experiment run). Feeds issue #143 ("spine D: R_PICPOC observational/mechanistic retry — calcite port refuted") and the forward-model-fidelity-roadmap thread.

How this was produced. A deep-research workflow fanned out 5 search angles → 20 primary sources → 90 extracted claims → adversarial 3-vote verification on the top 25. It hit the Claude.ai monthly spend limit mid-verification, which crashed most verifier votes and the automatic synthesis step. I (main agent) wrote the synthesis by hand and reconciled it against our own internal findings (docs/findings/rpicpoc_ratio_structural.md, docs/findings/alk_anchor_rpicpoc_mutex.md). Read the confidence tags literally: - ✅ VERIFIED — survived adversarial verification (3-0 or 2-1). - ◻️ EXTRACTED — direct quote from a fetched primary source but the verifier crashed on the spend limit before voting (0-0, 3 abstain). Not refuted — just unconfirmed. Treat as a strong lead, not a settled fact. - ❌ REFUTED — failed verification (0-3); excluded from the argument.

TL;DR — the verdict

The problem is primarily MISSING-MECHANISM, not missing-observation — and specifically it's a missing environmental gating of the rain ratio, not a missing calcifier PFT.

We already proved internally that the estimator/observation side is solvable: the PIC:POC ratio loss identifies R_PICPOC orthogonally to the iron pair (eqpac 10/10, no iron collapse). What blocks ≥2-AOI recovery is that the box's calcite is a rigid readout — PIC_1/POC_1 = R_PICPOC everywhere — while the real/Darwin realized rain ratio varies regionally by a large factor. No new observable fixes a forward model that cannot represent the spread in the first place. So the fix is to give the forward model the spread, by the cheapest route the BGC-modeling field already uses: make the rain ratio a smooth function of environment.

Ranked by (identifiability gain × tractability in a differentiable box)

Rank Option Identifiability gain Tractability Net
1 (3) Environment-dependent rain ratio R_PICPOC = f(T, Ω_c/CO₂, PO₄, light) High — directly creates the regional spread the box lacks; standard practice in CESM/PISCES/CMIP; pairs with our existing ratio loss as the orthogonal observable High — a smooth differentiable function of state variables the box already carries; no new PFT, no new tracer Best
2 (2) TA* / excess-alkalinity loss Medium in principle — real-ocean CaCO₃ observable — but in OUR box collapses to the same R_PICPOC × mort_total product as PIC (see internal ALK null), so it's gated on the same forward-model fix Medium — TA* recipe is differentiable-friendly, but absolute-ALK anchor already shown IC-dominated & degenerate Complement to #1, not a standalone fix
3 (4) Minimal coccolithophore bloom mechanism Medium-High — would create real regional spread via a calcifier PFT Low — needs a new PFT + temperature/grazing niche dynamics; regional dominance is top-down (grazing) controlled, harder to fit; and CMIP shows explicit calcifiers are not required to get the spread Reserve; only if #1 underfits
4 (1) Satellite PIC observable (MODIS/PACE) Low where it matters — ✅ confirmed unreliable in the Southern Ocean / high latitudes (>25× overestimate; diatom-silica artifact), and the alkalinity rain-ratio pattern points at the wrong regions High (tooling already built: modis_pic_loader.py, pace_loader.py) Deprioritize

This inverts the planned priority. next.md ranked satellite PIC #1 ("highest EV; tooling already built"). The verified evidence says satellite PIC is the weakest option in exactly the regions (SO/high-lat) where our target is lowest and recovery is hardest. Tooling-built ≠ scientifically informative.

Reconciliation with our own prior findings (the key value-add)

The external literature does not contradict our internal work — it explains why our internal results came out the way they did, and it points past the wall.

1. Our ratio loss already solved the estimator side (eqpac 10/10). docs/findings/rpicpoc_ratio_structural.md: at steady state PIC_1/POC_1 = R_PICPOC·(W_SINK/W_SINK_PIC), so mort_total cancels and the ratio observable pins R_PICPOC orthogonally to the iron pair. It recovers R_PICPOC only in eqpac, because the box's calcite is a single-mort_total rigid readout while Darwin's realized ratio varies regionally. → The literature confirms the regional spread is real and large (see Option 3), so the residual wall is genuinely forward-model fidelity, exactly as we concluded. The estimator is not the bottleneck.

2. Our absolute-ALK anchor was a NULL — and this directly down-weights Option 2. docs/findings/alk_anchor_rpicpoc_mutex.md: the box's surface ALK is calcite-only (dALK_1 = −2·R_PICPOC·mort_total), so it shares PIC's exact factorization and is informationally equivalent to PIC for the R_PICPOC/mort_total split; it's also IC-dominated (calcite moves ALK ~0.17%), making the absolute-ALK signal ~60,000× weaker than PIC. The apparent co-recovery was a cell-weighted straddling artifact. → Critical caveat for Option 2: the TA literature is about disentangling real-ocean alkalinity (preformed + remineralization + calcite). Our box's ALK has no preformed/ remin/riverine structure — it is already a pure calcite signal. So computing TA on the box adds nothing; a TA/ALK target still constrains the same R_PICPOC × mort_total product. Option 2 is gated on the same forward-model fix as everything else* — it only becomes independently informative once the forward model carries a richer (environment- or PFT-resolved) calcite budget. Until then it reproduces the ALK null.

3. The target was re-scaled this session. The 2026-06-15 finding used the old SO ratio (~1.4); the corrected Darwin targets are eqpac 0.033 / natl 0.68 / SO 0.0067. The North Atlantic is the high-rain-ratio (coccolithophore-bloom) region; the Southern Ocean is low. This matters because it aligns with the alkalinity-derived rain-ratio literature (Option 3) and against the satellite-bloom pattern (Option 1) — see below.

Option 1 — Satellite PIC (MODIS / PACE): DEPRIORITIZE

The strongest verified result of the whole run is that satellite PIC fails precisely where we need it.

  • ✅ VERIFIED (3-0): MODIS PIC catastrophically overestimates true (in-situ) PIC south of the Antarctic Polar Front by >25×, where sample PIC averaged only ~15 nM. (agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2024EA004070)
  • ✅ VERIFIED (3-0): the SO overestimation is a non-carbonate artifact — heavily silicified colony-forming diatoms (Fragilariopsis, Pseudo-nitzschia) produce anomalous backscatter/birefringence that the PIC algorithm misreads as calcite. (same source)
  • ✅ VERIFIED (3-0): MODIS PIC and coccolith-based PIC are decoupled south of the Polar Front — satellite shows high PIC where no coccolithophores were observed. (bg.copernicus.org/articles/22/3143/2025)
  • ✅ VERIFIED (3-0): satellite PIC is overestimated in Antarctic waters because high reflectance also comes from suspended sediment and opal/diatom-frustule fragments. (same)
  • ✅ VERIFIED (2-1): authors conclude satellite PIC is unreliable south of the Polar Front; in-situ measurements are needed. (same)
  • ◻️ EXTRACTED: the alkalinity-derived rain-ratio pattern contradicts the satellite coccolithophore-bloom pattern — blooms cluster at high latitudes, yet the highest export ratios are at low latitudes (low-latitude non-bloom calcifiers, i.e. foraminifera/pteropods, dominate global calcification). (gfdl.noaa.gov/bibliography/related_files/jls0201.pdf — inferred Sarmiento et al. 2002, GBC)

Read: even setting aside our target re-scaling, satellite PIC is contaminated by diatom silica in the SO and structurally points at the wrong regions for rain-ratio constraint. The built loaders (reference_satellite_pic_loaders.md) are not wasted — they remain useful for low-latitude bloom regions — but satellite PIC is not the lever that unblocks the regional spread. Keep shelved.

Option 2 — TA / excess alkalinity from GLODAP: complement, not a standalone fix*

  • ✅ VERIFIED (3-0): TA is a constructed tracer isolating the CaCO₃-dissolution imprint on alkalinity: TA = TA0 + TAr + TA*, with TA0 = a0 + a1·S + a2·T + a3·PO (PO = O₂ + 170·PO₄) and TAr = 1.26·(16/170)·AOU. A concrete, differentiable-friendly recipe. (bg.copernicus.org/preprints/12/20223/2015 — inferred Battaglia et al. 2016, GBC)*
  • ✅ VERIFIED (weak, 1-1): CaCO₃ dissolution parameters are hardly constrained by TA or current flux compilations — models with and without saturation-dependent dissolution both achieve skill. (same) → alkalinity is a weak* constraint on dissolution mechanism.
  • ◻️ EXTRACTED: real-ocean CaCO₃ signal in surface alkalinity is not regionally flat — near-constant at low latitudes, strong poleward increase (~−110 µmol/kg SO / subarctic N Pacific). (sciencedirect S0304420315001073)
  • ◻️ EXTRACTED: a salinity-normalized alkalinity (sAlk) vertical-gradient emergent constraint links the alkalinity field to PIC export across ESMs (44–55 Tmol/yr at 100 m). (bg.copernicus.org/articles/20/1195/2023)
  • ◻️ EXTRACTED: the CaCO₃:Corg export ratio can be estimated from vertical gradients of potential alkalinity + sAlk-nitrate alone, without absolute transport magnitudes. (gfdl jls0201.pdf)
  • ❌ REFUTED (0-3): a specific "0.82 GtPIC/yr global pelagic export" figure — excluded.

Read: TA is a legitimate real-ocean observable and its vertical-gradient form (sAlk) carries genuine export information — better than the absolute ALK anchor we already nulled. But (per internal finding #2) our box's ALK is calcite-only and IC-dominated, so any ALK/TA target still constrains R_PICPOC × mort_total and reproduces the null until the forward model is enriched. Sequence it after Option 3: once the rain ratio varies with environment, a TA-gradient loss becomes an independent cross-check rather than a degenerate restatement of PIC. The dissolution-constraint weakness (1-1) further argues TA informs export/dissolution, not surface production rain ratio — which is what we're after.

This is the missing mechanism, and it is standard practice in production ocean BGC models.

  • ◻️ EXTRACTED: in CESM, the coccolithophore calcification-to-organic ratio (realized PIC:POC) is an explicit function of aqueous CO₂, phosphate limitation, and temperature — not a uniform mean rain ratio. (agupubs 2018MS001483 — inferred Krumhardt et al., JAMES)
  • ◻️ EXTRACTED: making the rain ratio environment-dependent produces large regional spread in calcification (N Atlantic, W Pacific, parts of SO increase under rising CO₂ via relief of carbon limitation). (same) → directly the regional spread our box lacks.
  • ◻️ EXTRACTED: realized PIC:POC parameterized as a saturating (Michaelis-Menten) function of calcite saturation Ω_c: (PIC/POC)_max = 0.8, half-saturation K = 0.4, fit to E. huxleyi data with r² = 0.92. (bg.copernicus.org/articles/4/505/2007 — inferred Ridgwell et al. 2007) → a concrete, drop-in differentiable form.
  • ◻️ EXTRACTED: across CMIP5/CMIP6 ESMs, implicit pelagic calcification (the rain ratio) is a function of nutrients (PO₄/NO₃/Fe/Si), temperature, light, depth, and Ω_c. (bg 20/1195/2023)
  • ◻️ EXTRACTED: no CMIP5/6 ESM explicitly represents a calcifying PFT — all treat calcification implicitly, yet still reproduce regional PIC:POC spread. (same) → strong argument that Option 3 ≫ Option 4 on cost/benefit: you do not need a coccolithophore PFT to get the spread.
  • ◻️ EXTRACTED: the rain ratio varies strongly and systematically by region — equatorial max, smaller in subtropical/subpolar/high-lat (Atlantic eq 0.084 vs subpolar 0.023 vs SO 0.020; Pacific eq 0.087 vs subtropics 0.045). (gfdl jls0201.pdf) → observed spread is real and large-amplitude.

Read & concrete design. Replace the constant R_PICPOC with a smooth gating function of state variables the box already carries, e.g. a Ridgwell-style saturating form R_PICPOC(cell) = R_max · Ω_c/(Ω_c + K) · g(T) · h(nutrient), or the CESM f(CO₂, PO₄, T) form. The free parameters of f(env) become the new recovery targets, and our existing ratio loss (already orthogonal to the iron pair) supplies the identifying observable. This is the highest (gain × tractability): it injects the missing spread, needs no new PFT or tracer, is differentiable by construction, and slots into the lever we've already built.

Option 4 — Minimal coccolithophore bloom mechanism: reserve

  • ◻️ EXTRACTED: in ROMS-BEC (Southern Ocean), coccolithophore-vs-diatom dominance is set primarily by top-down (grazing) control, not growth-rate differences — specific grazing differences are 2× (40–50°S) to 8× (50–60°S) larger than growth-rate differences. (bg.copernicus.org/articles/15/6997/2018 — inferred Nissen et al. 2018)
  • ◻️ EXTRACTED: coccolithophores do both PIC and POC and thereby set the water-column rain ratio. (agupubs 2022JG007355)

Read: a faithful calcifier PFT would create real regional spread, but it's the most expensive path — it requires a new PFT plus temperature/grazing niche dynamics, and the regional signal is top-down controlled, which is exactly the hard-to-fit part. Combined with the CMIP result that explicit calcifiers are not required, Option 4 is overhead for the identifiability goal. Hold it in reserve for the differentiable-Darwin port endgame (differentiable-full-darwin-regional thread), not for unblocking R_PICPOC now.

  1. Implement Option 3 behind a default-OFF flag (RPICPOC_ENV or similar) in carroll6_5pft_2layer.py: make R_PICPOC a Ridgwell-style saturating function of Ω_c (and optionally T, PO₄), with Ω_c derived from the box's existing carbonate state.
  2. Pre-register a forward-model probe first (as we did for the ALK null): confirm the gated R_PICPOC(env) actually produces an AOI spread matching the corrected targets (eqpac 0.033 / natl 0.68 / SO 0.0067) before running the optimizer — i.e. test that the mechanism can make the spread.
  3. Fit with the existing ratio loss (RATIO_W>0, cocco off) across the 3 AOIs; score per-AOI (the straddling-artifact lesson). Success = R_PICPOC-controlling env params Cal-grade in ≥2 AOIs without iron-pair collapse.
  4. Only then add a TA*-gradient (sAlk) cross-check (Option 2) as independent confirmation.
  5. Treat Options 1 & 4 as out-of-scope for this retry.

Answer to the central question, in one line: missing-mechanism — specifically a missing environmental gating of the rain ratio; the estimator and a workable observable already exist, but the box's rigid single-ratio calcite cannot represent the regional spread that both Darwin and the real ocean exhibit.

Verification ledger

  • Pipeline: 5 angles → 20 primary sources → 90 claims → top 25 verified.
  • Outcome: 7 confirmed (✅), 1 genuinely refuted (❌, the 0.82 GtPIC/yr figure), ~17 crashed-to-abstain (◻️, unconfirmed — spend limit), synthesis step crashed.
  • Caveat: every ◻️ claim is a verbatim quote from a fetched primary source but was not adversarially checked. The four-option ranking rests on (a) the ✅ satellite-PIC failures, (b) the ✅ TA* recipe + weak-dissolution-constraint, and (c) our own internally-verified ratio-loss and ALK-null findings — the ◻️ Option-3 claims are corroborating, not load-bearing. Re-run the verification + synthesis once the spend limit resets to upgrade the ◻️ tier.

Sources (all primary unless noted)

# URL Inferred ref Used for
1 agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2024EA004070 — (Earth & Space Sci, 2024) ✅ SO satellite PIC >25× overestimate; diatom-silica artifact
2 bg.copernicus.org/articles/22/3143/2025 — (Biogeosciences, 2025) ✅ MODIS/coccolith PIC decoupling S of PF
3 bg.copernicus.org/preprints/12/20223/2015 Battaglia et al. 2016 (GBC) ✅ TA* recipe; weak dissolution constraint; ❌ 0.82 GtPIC/yr
4 sciencedirect.com/.../S0304420315001073 — (Marine Chemistry, 2015) ◻️ Alk poleward gradient; Alk recipe
5 aslopubs.onlinelibrary.wiley.com/doi/full/10.1002/lol2.10457 — (L&O Letters) ◻️ potential-alk & Ca as SO CaCO₃ tracers
6 agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2018MS001483 Krumhardt et al. (JAMES) ◻️ CESM r-ratio = f(CO₂, PO₄, T)
7 bg.copernicus.org/articles/4/505/2007 Ridgwell et al. 2007 ◻️ M-M rain ratio on Ω_c (r²=0.92)
8 bg.copernicus.org/articles/20/1195/2023 — (Biogeosciences, 2023) ◻️ CMIP5/6 implicit calcification; sAlk emergent constraint; no explicit calcifier PFT
9 agupubs.onlinelibrary.wiley.com/doi/10.1029/2022JG007355 — (JGR Biogeosciences) ◻️ coccolithophores set rain ratio
10 gfdl.noaa.gov/bibliography/related_files/jls0201.pdf Sarmiento et al. 2002 (GBC) ◻️ rain ratio from PA+NO₃ gradients; regional values; bloom-pattern contradiction
11 bg.copernicus.org/articles/15/6997/2018 Nissen et al. 2018 ◻️ SO cocco/diatom top-down grazing control

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

Overnight deep-research synthesis (literature, decision-ready). Full note: docs/research_notes/2026-06-24_rpicpoc_observability_deep_research.md.

Verdict: primarily missing-MECHANISM — specifically a missing environmental gating of the rain ratio, not a missing calcifier PFT or a missing observation. The estimator side is already solved (ratio loss, orthogonal to the iron pair) and a workable observable exists; the wall is that the box's rigid PIC/POC = R_PICPOC calcite cannot represent the regional spread both Darwin and the real ocean show.

Ranked fixes (gain × tractability): 1. Environment-dependent rain ratio R_PICPOC = f(Ω_c, T, PO₄) — standard in CESM/PISCES/CMIP; concrete differentiable form exists (Ridgwell M-M on Ω_c, r²=0.92); no new PFT/tracer; pairs with our existing ratio loss. ← do this. 2. TA/sAlk-gradient loss — real observable, but in our calcite-only box collapses to the same R_PICPOC × mort_total product (cf. the ALK null); useful as a cross-check after #1. 3. Minimal coccolithophore PFT — creates real spread but most expensive (top-down/grazing controlled); CMIP shows explicit calcifiers aren't required. Reserve. 4. Satellite PIC — deprioritize.* ✅ verified unreliable in the SO/high-lat (>25× overestimate; diatom-silica artifact) and the alkalinity rain-ratio pattern points at the wrong regions. This inverts the earlier "satellite PIC is highest-EV" plan.

Proposed next step: implement #1 behind a default-OFF flag, pre-register a forward-model probe that the gated ratio reproduces the corrected targets (eqpac 0.033 / natl 0.68 / SO 0.0067) before fitting, then fit with the ratio loss scored per-AOI.

Caveat: the research run hit a spend limit mid-verification — 7 claims adversarially confirmed (incl. all the satellite-PIC failures + the TA* recipe), the rest are source-quoted but unverified; re-verify when the limit resets.