Scoreboard

The scored row's reference was pinned before the epoch's first day and is not recomputed inside it (a reference that drifts under the forecast it scores is not a reference). The in-window observed base rate is in-sample — not a fair reference; the number to beat: it is the frequency of the very days it scores, so no one could have forecast it in advance, and a forecaster that fails to beat it has added nothing to counting the days. The sealed training constant is shown so the live strip can be read on the same scale as the backtest strip's headline.

Every interval is a percentile bootstrap at the 95% level, resampling whole days with replacement — two thousand resamples from a fixed seed (BSS_CI_SEED in service/scoring.ts), so the published page renders the same interval on every build and anyone can reproduce it. Against the two benchmarks that change day by day, each resample rescores the forecaster and the benchmark on the same drawn days.

• This epoch's scored reference is the cycle-conditioned climatology counted on the live flare scale, p = 0.359 — pinned, and public, before the epoch's first day, and fixed for the life of the epoch. Earlier epochs' days keep their own references forever (score epoch zero, p = 0.144; score epoch one, p = 0.298): no day is ever re-scored under another epoch's reference.
• Why the fixed reference flatters the model near cycle maximum. That p = 0.144 is the average of a held-out tail that spans a deep minimum as well as an active stretch. The Sun is not at that average now. A constant that under-forecasts the current rate is itself a badly calibrated forecast, so its reference Brier is large — and skill measured against a large reference Brier looks impressive without the forecaster doing anything. Conditioning the climatology on where the cycle actually is puts the reference at p = 0.298: the M-class-day frequency over the sealed fifty-year flare series, restricted to the months whose smoothed sunspot number sat in the same decile as today's. That is roughly twice the fixed constant, and it is a far harder thing to beat. It is not swapped in mid-epoch — changing the scored reference is exactly what starts a new epoch, which is why it is the reference of the epoch that began when the recalibrated model first issued, and of no earlier day.
• Why score epoch two's constant is higher. It is the same rule — the same sunspot deciles, the same reference month — counted on the scale the live flare list reports, the scale this board resolves on and the early-warning model was fitted on. On that scale the older satellites' readings are corrected upward, so more of the same months' days reach M-class: on the very days it counts, the older scale gives p = 0.300 and the live scale p = 0.359. The step is the scale, not a different point in the cycle. It cuts both ways: while the Sun runs quieter than those months did, a higher constant is the easier one to beat — which is why the two benchmarks below, which follow the Sun, stay printed beside it.
• Sunspot numbers: source WDC-SILSO, Royal Observatory of Belgium, Brussels (CC BY-NC 4.0), vendored and checksum-pinned; the cycle-conditioned constants are derived by tools/cycle_climatology.mjs and pinned in contracts/scoreboard.pins.json.
• Why two benchmarks that move with the Sun sit beside the scored reference. A fixed reference cannot follow the Sun. This epoch's constant was pinned to the Sun's activity in early 2026; over this epoch the Sun has been quieter than that, so the constant sits above the recent rate: a badly calibrated forecast with a large Brier, and skill measured against it is partly the reference's doing. The scored reference is never changed inside an epoch, and a rolling rate can never be one — it would drift under the forecasts it scores — so two display-only benchmarks are printed beside it instead: NOAA SWPC's published day-ahead forecast, the operational forecast this one has to beat, and the last 120 days' flare-day rate, a fair no-skill reference. Neither is scored, and the scored reference is unchanged.
• NOAA SWPC's published day-ahead forecast is the probability of an M-class flare that NOAA's Space Weather Prediction Center publishes for the next UTC day in its daily Report of Solar-Geophysical Activity, issued at 22:00 UTC. It is read exactly as published — from the live report and from NCEI's archive of every past one — and counted only when the report came out before the day began: a forecast issued after its day started has already seen part of it. A day without such a report has no NOAA value and is left out of that row, the forecaster's score included; it is never filled in from anything else. Days so far with no NOAA value, because NOAA's report for them came out after the day had begun: 2026-09-04. Each capture is kept in ledger/noaa_reference.jsonl, append-only and hash-chained like the ledger. NOAA's forecasts are public-domain US-government data, attributed here to NOAA SWPC and shown unmodified; their use on this page implies no endorsement by, or affiliation with, NOAA or the National Weather Service.
• The last 120 days' flare-day rate is, for each day, the share of the observed days among the 120 before it that had an M-class flare — the no-skill reference operational flare-forecast verification uses: Leka et al. (2019), doi:10.3847/1538-4365/ab2e12, choosing the window after Sharpe & Murray (2017), doi:10.1002/2017SW001683. It reads only days before the one it scores, from the same count history as the board's other baselines — the sealed flare series, continued by this ledger's resolved days — and a day nobody observed is left out of the count, never counted as quiet.