HELIOGRAPH

What it returns

From an active region's observed radial field, HELIOGRAPH reads a single state variable — the loaded "parasitic energy" of the dominant evolving field mode — against two fixed thresholds, and issues a deterministic two-tier alarm with a lead time and a magnitude estimate.

Output per region/time: alarm tier · hazard-window lead time · expected field-reorganization magnitude. The firing minute inside an open window is stochastic and is not forecast (see Scope).

And it is quick. A complete active-region assessment — 46 magnetograms, the full mode decomposition and all — computes in ~2.4 s on a single laptop core: no GPU, no cluster, no training, no field extrapolation. The whole active-region catalog can be re-swept continuously on commodity hardware.

Validated — blind, apples-to-apples

Larger events announce themselves earlier. The activation-crossing lead time scales with loaded field energy at r = +0.96 — a size-scaled lead the one prior timing method in the literature reports it could not find.

Blind, held-out set of ten solar-cycle-25 X-class flares, each active region compared against its own non-flaring window, identical pipeline on every window, no per-event tuning. Full method, protocol, and event table: 10.5281/zenodo.20696671.

View a sample hazard report — the 2024-10-03 X9.0 →

Rigor

Every number above survived a blind protocol built to break it. We self-certify the validation discipline, not just the result:

• Blind, apples-to-apples. Ten distinct cycle-25 X-flare regions, each scored against its own non-flaring window — so the signal can only come from the loading state, never from the fact that flaring regions are larger. Identical pipeline on every window, no per-event tuning: the only calibrated quantities are two threshold levels and one regression slope, each a median or a least-squares line, each stated openly.
• Self-falsified. Before the activation read-out survived, a series of candidate precursors did not, and were killed against the same blind controls — phase-locking, a marginal-stability regime, the torus-instability decay index, transverse-field current-helicity coherence. Each looked plausible; each turned out to track region size or a saturation artifact, not the flare. What remains is what survived being attacked.
• Scars reported, not buried. About one loaded quiet window in four crosses the threshold without a flare; the smallest events fire below the median threshold and are caught only by the broad watch. Both are stated on this page (see Scope) and in the preprint — not smoothed away.
• Deterministic and reproducible. No training, no fitted classifier, no randomness at inference. The same region at the same time returns the same assessment, reproducible from the published definitions applied to public data.

The instrument's own boundary was earned the same way: repeated blind tests for a deterministic trigger in the surface field all returned null — which is exactly why this page promises a hazard window and a magnitude, and never a firing minute.

Scope — stated plainly

• The window and the magnitude, not the minute. The firing time inside an open window is stochastic by physics and is not predicted.
• Field-energy, not GOES class directly. The activation variable is a magnetic energy; its link to soft-X-ray class is real but noisy.
• False alarms are real — about one in four loaded windows clears without a flare in the test interval: the price of a necessary-not-sufficient threshold over a stochastic trigger.
• The contribution is kind, not horizon. At ~12–14 h it does not out-lead the precursor literature; it is deterministic, data-direct, mode-decomposition-based, and size-scaled where the operational standard is probabilistic and trained.
• A small, deliberately blind sample. Ten regions, each its own control — a conservative test built to resist overfitting, not a population study. Large-catalog validation is the next step.