Findings¶
Per-version technical writeups behind the headline results. Project Status is the canonical live summary; these pages are the underlying experimental record.
Current frame: 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.R_PICPOCis recoverable given a real calcite anchor (Daniels/MODIS) — the "6/6 wall / needs the Darwin port" framing in the older writeups below is superseded. The v2.x/v3.x pages are point-in-time records.
Version milestones¶
- v2.1 — GLODAP DIC/ALK hybrid — first real-observation target;
R_PICPOC360% → 74% off Carroll. - v2.2 — 5-PFT box — full 5-PFT box matching Darwin v05; project-first 4/6 calibration-grade.
- v2.7 — 2-layer box scoping — subsurface DFe anchoring alone does not unblock
scav_rat. - v2.8 — Darwin ICs + POC loss — project-first reproducible
scav_ratrecovery; bimodal degeneracy. - v2 Track 1 closeout — carbonate-cycle arc summary.
- v3.0 — multi-AOI scoping — joint multi-AOI training; the 5/6 plateau (historical 'parameter conservation' framing — superseded; see lead note).
- v3.1 — closeout — 3-AOI Basin C; 38/40 iron-pair recovery; two complementary 5/6 paths.
- v3.2 — dense POSi + Eppley — first reproducible 5/6 at 3-AOI via forward-model fidelity (valid prior result against Darwin's own silica field; the
/6framing is superseded by the 4-observable frame, anddiatomgrazdoes not recover in the real-data sweep — it is an iron-pair tradeoff).
R_PICPOC recovery (real calcite anchor)¶
R_PICPOC is recoverable given a real calcite observation (Daniels CP:PP / MODIS PIC). Any ratio anchor recovers it (Darwin's own too, via RATIO_MAX=2) — the real anchor supplies non-circularity, not recoverability. The campaign below is point-in-time.
- Ratio loss (structural) — the PIC:POC ratio observable recovers
R_PICPOCin eqpac without the iron-pair mutex. SUPERSEDED: the apparent ≥2-AOI "wall" was a contaminated Southern-Ocean ratio target, not forward-model fidelity — fixed byRATIO_MAX=2(R_PICPOC10/10) and recoverable from a real calcite anchor. - ALK anchor mutex test — pre-registered null; an apparent co-recovery falsified as a cell-weighted averaging artifact.
Scaling & verification¶
- Compute-time scaling
- Memory scaling — eager (RTX 5090)
- Memory scaling — compiled (Explorer H200) — measured 82.9 B/(cell·step), 4.3× below eager
- Pre-scaleup verification
- Stage-1 per-AOI gating (2-AOI)