Skip to content

Manuscript red-team — 8-lens reviewer panel (#116, 2026-06-27)

Ran the paper_reviewer_panel skill as a 9-agent fan-out (R1–R8 + verify/synthesis) on the manuscript spine. 39 raw objections → verified/merged/ranked. Every numeric objection was independently recomputed. The reframed central claim survives ("a 0-D differentiable surrogate + per-cell NN characterizes an identifiable subset of a non-differentiable GCM's parameters" — the right, BINN-distinct paper). It does not survive as currently staged. The fixes:

Must-fix before submission (highest leverage first)

  1. The committed manuscript is a different paper. docs/paper/main.tex is an old draft whose abstract says "recover five of the six … the sixth [R_PICPOC] … unrecoverable / foreclosed" — the opposite of the spine. Rewrite main.tex to the identifiability reframe and re-gate every headline on the same branch that holds the evidence (#151/#153 + spine-d). Everything else is downstream of this.
  2. Per-cell-load-bearing is a non-sequitur (R1-5, rd 4). "Global scalar → flat box" only holds if forcing is also uniform; a global-scalar box with per-cell forcing is NOT flat. The homogenization result (flat at uniform params) is suggestive but does not prove the per-cell predictor is needed. Run the ablation: global-scalar-Carroll-6 + per-cell forcing vs per-cell-predicted params, scored on the real anchors; report the delta. (This is the cross-method/per-cell ablation already flagged as the #1 research-strength gap.) Until then, soften to "required to fit per-cell obs; per-cell param structure unvalidated."
  3. Iron-pair "real-world physics" overstates a saturated/shallow/hand-weighted result (rd 3). K_FE=5e-5 ⇒ f_Fe 0.67–0.91 (iron near-saturated → identified by DFe mass balance, not iron-limitation physics); realiron profile shallow (Hessian indefinite → direction not value); recovery required up-weighting real iron 10× (GEOTRACES_W 0.3→3). Reframe as mass-balance + non-circularity; report K_FE, f_Fe, the GEOTRACES_W sweep, and the profile span. (Honest version strengthens the dimensional-gap claim: the Darwin-pattern block alone does NOT identify the pair.)
  4. The "dimensional gap" defense over-reaches (R1-1, rd 2 but structural). Box surface PIC:POC ≈ 0.0424 everywhere vs Darwin ratio-of-means 0.033 / 0.676 / 0.0067 — a 16× (natl) / 6× (SO) error that survives spatial averaging ⇒ it is a structural single-scalar closure gap, not a pattern/CV artifact. Split the gap into (a) dimensional/pattern and (b) intensive closure; concede (b) caps any joint multi-AOI R_PICPOC claim.
  5. Joint 8/10 < marginals is an arithmetic impossibility unless stated (R3/R7, rd 2). 3×(10/10) forces joint = 10/10; 8/10 ⇒ at least one marginal ≤ 8/10 (it's scav_rat). State the three marginals next to the joint, or reframe "8/10 = R_PICPOC binds; alpfe/scav_rat pass ~always."
  6. No BINN delta (R8, rd 2). Zero BINN/Xu citations; BINN (Xu 2025) is structurally identical and the "identifiability" reframe widens the overlap. Add a Related-Work paragraph naming the delta (external GF target vs self-consistency; identifiability characterization; multi-AOI attribution).
  7. Self-twin claim + native-adjoint citation: the self-twin script now lands on main (#153) — cite it; add the missing ECCO adjoint/4D-Var citations (Heimbach/Forget 2015, Menemenlis 2005) at the CTRL claim.

Net-new (completeness critic)

  • N1: the committed paper's coccolith-only ablation (0.0049 flat) is the body's own argument that R_PICPOC fails — it must be reconciled with the spine's "R_PICPOC recovered," not left contradicting it.
  • N3: "23-yr climatology" averages away the seasonal succession signal that is the natural handle on the growth pair/diatomgraz — so "unobservable" is partly a target-construction choice, not only a data-availability law. Sharpens the growth-pair framing.
  • N4 (a strength to surface): the binary PIC-anchor mutex (any PIC_ABS_W wipes the iron pair) is a stronger honest finding than the spine surfaces — it pre-empts "you cherry-picked the config."

Cheap polish (one pass)

Wilson CIs on 38/40 ([0.835,0.986]) + disclose the 8/10 batch; label every rate (config, n, param-set); "verify_run.py = data-integrity gate, not validation"; obs-space 80/20 held-out split; Transtrum/Sethna sloppiness lineage cites; reconcile the diatomgraz "recovered (Darwin POSi 20/20) vs not (real obs)" states to one canonical statement.

Full ranked table + recompute arithmetic: run wf_d6a49058 (paper_reviewer_panel fan-out).