Program G1 — the gold-standard solar pipeline (charter)

Status: charter only — no experiment below has run. Written in response to the independent review (docs/REVIEW-SR-ASTROPHYSICS-2026-07.md, M7 and plan items 10–11): the existing solar legs (P2-solar, R3, S2, S5, T3) are software demonstrations on convenience products; no astrophysical discovery claim is made from them, and none will be made until every item in this charter is satisfied. Pre-registered in the series’ self-registered sense: verdict rules below are committed before any run.

Why a rebuild rather than a patch

The review’s diagnosis, accepted in full: (1) 45-second line-of-sight magnetograms were used as the normal boundary field without radial conversion or uncertainty propagation; (2) mean-flux subtraction + periodic FFT continuation on cropped windows manufactures boundary conditions, and the repository’s own window tests show low nulls are window tenants; (3) Newton seeding is not a census — completeness and recall are unknowable from it; (4) the AR11158 sequence is too sparse to track null identity; (5) all quoted uncertainties are numerical, while the dominant error is the boundary data and model.

G1a — data products

• Primary: HMI 720-s definitive vector series, SHARP CEA remapped patches (hmi.sharp_cea_720s), using the radial component $B_r$ directly; disambiguation as distributed; per-pixel uncertainty maps kept.
• Secondary (cross-check): the 720-s LOS series with explicit μ-correction, to quantify what the earlier LOS-as-radial approximation did.
• Every figure states: NOAA AR number, HARP number, exact T_REC, series name, projection, pixel scale, disk position, and the uncertainty product used.

G1b — extrapolation without manufactured boundaries

• Potential extrapolation with non-periodic boundary treatment: zero-padded Green’s-function or apodized-window comparison against the FFT continuation; the difference field is the boundary-systematic estimate.
• Window ensembles as standard: ±16/±32 px shifts, two window sizes, two resolutions (full and 2× down), noise ensemble (add per-pixel HMI uncertainty realizations, n ≥ 20). A null is reported only with its survival fraction across the ensemble.

G1c — the census, done right

• External baseline: the trilinear method of Haynes & Parnell (2007) — cell-by-cell sign test + trilinear sub-grid positioning + Poincaré-index verification — as the completeness reference. Newton polishing only after the cell census.
• Report per region: census count, Newton-confirmed count, disagreements, and the weakest null’s survival fraction. No “N nulls” claim without the census.

G1d — temporal topology (the fold hunt, redone)

• AR11158-class emergence at hourly or better cadence, co-moving CEA window locked to the HARP centroid, null identity tracked by nearest-neighbour in (position, sign) with explicit boundary-crossing accounting: a count change is attributed to a fold only if the pair is interior, opposite-signed, and the collision is bracketed in time by the tracker.
• The S5c spectral-certification machinery applies unchanged once a candidate interval is found; the certificate vocabulary (“fold in a boundary-interpolation family”) stays until a tracked temporal fold passes.

G1e — verdict rules (pre-registered)

1. A null is reported iff survival fraction ≥ 0.8 across the G1b ensemble AND found by the G1c census AND its |B|-minimum exceeds the local noise floor propagated from the HMI uncertainty map.
2. The SR growth-vector reading remains labelled by the three-level taxonomy (blind / independent confirmation / tangent-cone check) — G1 does not upgrade it.
3. A fold event is claimed only under G1d’s tracked-interior-bracketed rule.
4. Negative outcomes are published with the same prominence as positives.

Cost estimate

G1a data pull ~1 h (JSOC exports); G1b/G1c pipeline work ~2–3 sessions; G1d needs ~50–100 frames per region (hourly, 2–4 days) — storage trivial, compute ~hours. Blocking dependency: none — drms/sunpy already in the venv.

Relation to existing results

Until G1 runs, the solar results stand as: methods demonstrations on convenience products with documented window sensitivity (R3 gallery: strongest-2-per-region, not a census; S5c: certified fold in a boundary-interpolation family; T3: sparse census, no fold caught). The review’s grade for that layer (astro grounding 2.4–2.6/5) is accepted as the current state.