Calibration

Step 8 of the method maps the winning evidence to a probability with two numbers:

p = 1 / (1 + exp(−(a + b · z))),    z = sign(r*) · log(1 + |r*|)

The release numbers

constant	value	provenance
intercept a	−4.590	logistic regression on the reviewed "sure" days of all eight Beta stations (2,288 days, 470 with a wrong sign), reference run of 22 September 2026
slope b	2.162	same fit
floor φ	8.03 × 10⁻⁹ MW	smallest non-zero overnight demand step over the eight Beta stations

With these numbers a day is corrected automatically (p ≥ 0.7) once its evidence reaches about 11.4, and kept automatically (p ≤ 0.3) below about 4.6.

Refitting for a new population

About 27 randomly chosen reviewed days are enough to refit the intercept; the slope stays.

from pynrpf import RELEASE_CALIBRATION, fit_calibration, run

scored = run(frame)                                    # release calibration, to get the evidence
reviewed = scored.site_days.merge(labels, on=["site", "date"])   # label: 1 wrong sign, 0 not
pair = fit_calibration(reviewed["evidence"], reviewed["label"],
                       slope=RELEASE_CALIBRATION.slope, provenance="network X, 27 reviewed days, 2027")
result = run(frame, calibration=pair)

Until a population has its own intercept, the release pair is the safe default: it demands more evidence, never less.

Show details

How stable the release numbers are

The held-out fits of the evaluation (one per station, each fitted without that station) range from −5.19 to −4.27 for the intercept and 2.06 to 2.35 for the slope, so the release pair is not sensitive to which stations were seen.

Why the intercept belongs to the population

The intercept says how much evidence a population's ordinary days produce. On Beta, real substations with real errors, ordinary sunny days can produce evidence of five to ten without any error (demand rising with solar, cloudy-day solar estimates), so the rule waits for about eleven. On Alpha, correctly signed meters with the wrong sign imitated by taking the absolute value, ordinary days produce almost none, so a rule fitted on Alpha corrects at about 0.3 and, applied to Beta, flips 214 clean days. The Beta-fitted rule applied to Alpha stays precise (energy precision 0.935) and only skips the tiny planted errors no reviewer would call. The release calibration is therefore the conservative one.

Why the slope carries over

Freezing the slope from one population and refitting only the intercept on the other reproduces that population's own held-out result in both directions (Beta energy precision 0.938 against 0.941 with its own fit; Alpha 0.892 against 0.889). The slope is the shape of the evidence-to-probability curve; the intercept is its position.

How many reviewed days a refit needs

In the learning-curve study the intercept fitted on 27 random days met the 0.90 energy-precision gate in 95% of draws; nine days did not. Fitting both numbers on so few days is unstable (a handful of error days cannot pin the slope), so the slope stays.

What the numbers are not

They are not training in the machine-learning sense: no features, no model file, two numbers a reader can check. They do not change which window wins or how days rank; the sigmoid is monotone. p is the probability that the day carries a wrong sign, not that the window's edges are exactly right.