Differentiable physics, E7: how large is the winter night step at Primary substations, a pattern like that of storage heating, and how does it grow with cold?

Status: results for winters 2021-22 to 2024-25 (the headline winters), with winter 2025-26 (the selection winter) read afterwards and reported in its own section. The selection winter is the winter in which D5, D10, and D16 were picked. It was read only after PR 1397 (experiment 114) merged. The plan, committed before any code ran, follows the results.

We examined 14 Primary substations, labelled D2 to D22 to keep them anonymous. Six of them (D5, D9, D10, D11, D14, and D16) were set before any fit as the substations the conclusion rests on; we call them the decision substations. D5, D10, and D16 are the three where earlier experiments first saw a step in winter 2025-26 demand. The other eight (D2, D3, D4, D8, D12, D19, D21, and D22) were added as a check; we call them E7-X.

We asked how much demand rises in the small hours of a winter night, the pattern that storage heaters (heaters that charge from cheap night-rate electricity, switched on by timers) would produce, and whether the rise grows with cold. On each winter night we took the mean demand over the five half-hours starting 00:30 to 02:30 and subtracted the mean over the four half-hours starting 22:00 to 23:30. We fitted that difference against how far the day's mean temperature (from the ERA5 reanalysis, a weather-agency estimate of past weather) fell below 15.5 C. We then subtracted the same difference measured on June to August nights, which we treat as the no-step baseline. We measure demand in points, where one point is 1% of that substation's 99th percentile of demand. On a night with a daily mean of 3 C, demand rises by 9.9 points averaged over the six decision substations (90% interval 6.2 to 13.5; the average weights each substation by how precisely its step is known, using Hartung-Knapp random-effects pooling). The eight E7-X substations, which were not chosen for showing a step, average 13.2 points (6.8 to 19.6), above the decision substations' 9.9.

Take-home message: demand at 13 of the 14 Primary substations steps up around 01:00 on winter nights, and the step is larger on colder nights. The step is not special to the three substations that prompted the study.

The eight E7-X substations have an average step as large as the six decision substations, so the analysis does not show that D5, D10, and D16 stand out (it does not rule that out, because the eight were not a random sample). The full step depends on the assumption that summer nights have no step. The part that grows with cold needs no such assumption, and that part is smaller: 1.1 to 4.7 points at the six decision substations (the slope times 12.5 degrees, the gap between 15.5 C and a 3 C night). The rise looks fixed to UTC on British Summer Time nights (a post hoc check). On cold winter nights the day-ahead forecast of experiment 114 runs below actual demand from about 01:00, at D5 and D16 on the two pre-registered winters and at D10 in winter 2024-25 alone. The verdict word is inconclusive. By the literal decision rule it would be supported, because every per-series gate row passes and both pooled intervals lie above 0. We keep inconclusive because the audit found the gate weaker than supported needs: its four gated shapes are one test, its pooled-interval row cannot fail, and its one misspecified cell under-covers (the gate paragraph below). The estimates above stand. The independent audit has run, and its findings are applied.

Step per series and pooled

Summary, by question (winters 2021-22 to 2024-25 unless stated; every number is printed in report.txt):

How large is the night step at a daily mean of 3 C? Pooled over the six decision series, 9.9 points (Hartung-Knapp interval 6.2 to 13.5; resampling whole series 7.0 to 12.6). Per series, from 3.0 points at D14 to 15.4 at D10. D5, D10, and D16 (selected for showing the step) average 13.2 points; D9, D11, and D14 average 6.5.
Does the step grow with cold? Yes at all 14 series: every slope is positive and the 90% interval of every slope excludes 0. The slopes are 0.09 to 1.12 points per degree, and 0.25, 0.37, and 0.26 at D5, D10, and D16.
Is the step specific to the chosen substations? No. The eight further Primaries (E7-X) pool to 13.2 points, above the six decision series. One of the eight (D8) has no step (0.5 points, 90% interval -0.7 to 1.7).
Does the step have a fixed clock time? The step starts in the half-hour starting 00:30 (ending at 01:00) at all 14 series (at D4 the interval reaches one half-hour earlier). On cold British Summer Time nights the largest jump falls one hour later by the clock at 13 of 14 series, which fits timers set to Coordinated Universal Time (UTC) (a post hoc check, chosen after seeing the data; D14 shows a shift of 4 half-hours).
How much of the step do experiment 114's day-ahead forecasts already remove? R, the step left in the errors of its baseline model divided by the full step, is 0.53 to 0.78 at D5, D10, and D16 (stored predictions for winters 2023-24 and 2024-25 only).
Where does the baseline forecast miss on cold nights? At D5 and D16 actual demand runs 5.6 and 9.4 points above the forecast in the half-hour starting 01:00 on the two pre-registered winters; at D10 it runs 12.2 points below the forecast in the half-hour starting 23:30, because in winter 2023-24 D10's cold nights sit about 11 points lower and the baseline is 9 to 17 points too high all night (figure 9).
What do the data and methods support? An estimate of the size and shape of the night step in four winters of one weather product. They do not support a cause: daylight, lighting, and switching schedules also change at night, and NGED has not confirmed storage heating.

Known-answer gate: the gate holds one real test of the interval, and that test under-covers. We added a known step to the real nights, resampled in 14-day blocks, and asked whether the 90% interval holds the true step. The four gated shapes (matched, ramp-plateau, flat, and placebo) are one test, not four: the estimator is linear, and all four injected shapes lie in its column space, so each moves the estimate by exactly its truth. Within every draw the point error is identical across the four cells, and their coverages differ only through the random numbers of the inner bootstrap. They show coverage of 87.0 to 87.5% pooled over series (target 85 to 95%) and 82 to 93% for each series (target 80 to 98%), and the mean estimate is within 0.20 points of the truth everywhere (target 0.5). That checks the bootstrap on resampled residual noise once. The Hartung-Knapp interval for the pooled mean covered the true mean in 100% of the 200 draws (target 85 to 95%), but that row cannot under-cover. The gate defines the true mean as the mean of a fixed set of steps (0, 0, 4, 4, 7, and 7), each series recovers its step within about 0.3 points, so the pooled estimate misses that mean by at most 0.52 points while the interval half-width is at least 2.13. The 100% therefore says nothing about whether the real pooled interval is conservative. The one misspecified cell is the night-temperature cell. If the true step follows the 18:00 to 08:00 mean temperature instead of the daily mean, coverage falls to 79% pooled at 7 points and to 43% at the worst series (descriptive, no target), so the daily-mean proxy matters. The gate also centres the summer nights on 0, so it cannot test the assumption that summer nights have no step. It tests only the shapes this plan chose, with noise that is independent between 14-day blocks.

Gate coverage

Results

The raw nights show the pattern before any fitting. On winter nights demand rises after midnight, where on summer nights it falls. At D10 the difference between the later and earlier half-hours averages -5.5 points in summer, and the raw mean on cold nights (daily mean at most 8 C) is 14.9 points above that summer value. That raw mean sits within 0.4 points of the fitted line at all 14 series (the plan allowed up to 2 points).

Raw nights

The step has the same sign in every winter at the six decision substations, and falls over the four winters at D10. At D10 it falls from 17.0 points in 2021-22 to 14.2 in 2024-25, and it also falls at two E7-X substations (D19 and D22). At D3 it falls and then levels (8.2, 6.8, 5.5, and 6.0 points), and at D2 it is level over the last two winters (13.7 in both). At D5, D11, and D16 the four winters stay within 3.4 points of each other, and at D14 the step is 1.3 to 4.2.

Step per winter

The eight substations chosen for nothing have the same average step as the three that prompted the study. The mean over D5, D10, and D16 is 13.2 points against 6.5 for D9, D11, and D14, but the eight E7-X substations also average 13.2. That does not rule out a selection effect, because the eight were not a random sample of Primary substations. The between-series standard deviation is 4.1 points for the six decision series (profile-likelihood interval 2.7 to 7.2) and 8.9 for E7-X (6.2 to 14.3). One series, D19, has a step of 33.5 points, which pulls the E7-X and all-series means up.

Pooled sets

Experiment 114's day-ahead forecast errors still contain a large part of the step. R is the same regression run on the difference between actual demand and the median forecast of experiment 114's baseline model, divided by the full step over the same two winters. R is 0.78, 0.63, and 0.53 at D5, D10, and D16 (and 2.05 at D14, where the step in the errors exceeds the full step). With the model's recent-error filter R is 0.96, 0.74, and 0.59, and with the interaction model and the filter 0.52, 0.42, and 0.31.

Share left by the baseline model

On cold British Summer Time nights the jump arrives one hour later on the clock. The pre-registered rule (the sign of the difference between the local-clock and UTC-clock windows) printed UTC at all 14 series. For a pure step the rule does separate the two clocks: a step fixed to the local clock gives a positive difference (+0.25 of the step) and a step fixed to UTC a negative one (-0.4 of the step). The confound is the ordinary load shape, which makes the difference negative with no step at all, on summer nights (-0.2 to -6.0 points) and on mild British Summer Time nights. Against that summer placebo, cold British Summer Time nights are more negative at 13 of 14 series (D4 is level), which supports UTC. The post hoc check in the figure compares the half-hour of the largest mean rise on cold summer-time nights with the same half-hour on cold winter nights. The search runs from the transition into the half-hour starting 23:30 to the one starting 05:00, and finds a shift of 2 half-hours at 12 of 14 series (D14 shows 4; D9 shows 9, a jump at 05:00 on 18 nights). Inside 23:30 to 03:00, D9 also shifts by 2, which makes 13 of 14. The counts of cold summer-time nights in the cards and the figure are the same (18 at D9 and 68 to 90 at the others).

Timing

Demand rises within about one half-hour around 01:00. Demand then stays raised until the morning rise. The profile of cold-winter-night demand minus summer-night demand, relative to the 22:00 to 23:30 level, has the same shape at nearly all substations.

Night profile

Selection winter 2025-26 (first run)

Winter 2025-26 and experiment 114's scored-year residuals were read once, after PR 1397 had merged: decisions.json held the merge commit's SHA (503ce3a3de408d1f0c2bdbada5ba88032ff238bb) before the fit for this winter ran. This is a first run; figures 3 and 9 and the R and cold-night numbers that include winter 2025-26 are first-run numbers.

Winter 2025-26 gives a similar step to the headline winters at every substation except D12, and the three substations picked on that winter are about 2 points higher in it. The step fitted to winter 2025-26 alone has a 90% interval above 0 at all 14 substations (D8, which has no step in the headline winters, reaches 2.4 points, interval 0.5 to 4.0). Pooled over the six decision substations it is 10.8 points (Hartung-Knapp 6.3 to 15.4), against 9.9 (6.2 to 13.5) in the headline winters; over all five winters, 10.0 (6.2 to 13.9). The mean of D5, D10, and D16 is 15.2 points in winter 2025-26 against 13.2 in the headline winters, a rise of 1.1, 2.3, and 2.6 points at the three. D9, D11, and D14 average 6.5 in both. The E7-X substations average 13.8 (7.6 to 19.9) against 13.2, and D12 rose by 9.6 points. The three substations picked on winter 2025-26 are therefore higher in it by about 2 points, a modest selection effect on an estimate of 13 to 15 points.

Experiment 114's scored winter. With the scored-year residuals, the share of the full step left in the baseline model's errors over winters 2023-24 to 2025-26 is 0.76, 0.63, and 0.52 at D5, D10, and D16 (0.64 at D11 and 1.97 at D14).
Replicate-winter rule. The plan's rule that the step is positive in the replicate winter 2024-25 alone holds at all six decision substations (figure 3, which now includes 2025-26).
Cold nights. The cold-night figure above uses winters 2023-24 to 2025-26.

Winter 2025-26 alone, per substation (points):

== winter 2025-26 alone, per series (points) ==
shape: (14, 7)
┌──────┬───────────┬───────┬───────┬───────────────────────┬────────────┬───────────────────┐
│ site ┆ step_2025 ┆ low   ┆ high  ┆ headline_winters_mean ┆ difference ┆ positive_interval │
│ ---  ┆ ---       ┆ ---   ┆ ---   ┆ ---                   ┆ ---        ┆ ---               │
│ str  ┆ f64       ┆ f64   ┆ f64   ┆ f64                   ┆ f64        ┆ bool              │
╞══════╪═══════════╪═══════╪═══════╪═══════════════════════╪════════════╪═══════════════════╡
│ D5   ┆ 11.86     ┆ 10.13 ┆ 14.15 ┆ 10.82                 ┆ 1.05       ┆ true              │
│ D9   ┆ 8.89      ┆ 7.56  ┆ 9.75  ┆ 9.51                  ┆ -0.62      ┆ true              │
│ D10  ┆ 17.72     ┆ 17.15 ┆ 18.29 ┆ 15.38                 ┆ 2.34       ┆ true              │
│ D11  ┆ 7.69      ┆ 7.06  ┆ 8.21  ┆ 6.99                  ┆ 0.71       ┆ true              │
│ D14  ┆ 3.00      ┆ 2.35  ┆ 3.73  ┆ 2.94                  ┆ 0.06       ┆ true              │
│ D16  ┆ 15.99     ┆ 14.38 ┆ 17.24 ┆ 13.44                 ┆ 2.55       ┆ true              │
│ D2   ┆ 13.40     ┆ 12.21 ┆ 14.18 ┆ 14.68                 ┆ -1.28      ┆ true              │
│ D3   ┆ 6.15      ┆ 5.89  ┆ 6.37  ┆ 6.62                  ┆ -0.46      ┆ true              │
│ D4   ┆ 12.97     ┆ 11.20 ┆ 14.27 ┆ 13.42                 ┆ -0.45      ┆ true              │
│ D8   ┆ 2.35      ┆ 0.52  ┆ 4.00  ┆ 0.58                  ┆ 1.77       ┆ true              │
│ D12  ┆ 24.66     ┆ 20.68 ┆ 27.65 ┆ 15.05                 ┆ 9.61       ┆ true              │
│ D19  ┆ 30.59     ┆ 28.14 ┆ 32.32 ┆ 33.62                 ┆ -3.04      ┆ true              │
│ D21  ┆ 9.68      ┆ 8.48  ┆ 10.50 ┆ 9.66                  ┆ 0.02       ┆ true              │
│ D22  ┆ 11.84     ┆ 10.55 ┆ 12.96 ┆ 12.68                 ┆ -0.84      ┆ true              │
└──────┴───────────┴───────┴───────┴───────────────────────┴────────────┴───────────────────┘

Predictions against outcome

Prediction (winters 2021-22 to 2024-25)	Outcome	Met
Pooled step over the six decision series 2 to 6 points, interval wholly above 0	9.9 (6.2 to 13.5)	No: above the range, interval above 0
D5, D10, D16: 5 to 10 points; D9, D11, D14: 1 to 3.5	10.8, 15.4, 13.4; 9.5, 7.0, 3.0	No: only D14 lies in its range
E7-X: each 0 to 3 points, pooled 0.5 to 2, below the decision series	pooled 13.2; one series (D8) at 0.5	No
Slope 0.4 to 0.8 points per degree at D5, D10, D16	0.25, 0.37, 0.26	No: below the range
Flat part within 2 points at every series	1.8 and -2.0 at D14 and D8; the others 4.6 to 19.4 in size	No
R against the full step 30 to 60% at D5, D10, D16	0.78, 0.63, 0.53	Partly: D16 only
Raw cold-night mean within 2 points of the line	gap at most 0.4 at all 14 series	Yes
Timing `cannot tell` at most series, or `local` if a label is printed	68 to 90 cold summer-time nights at 13 series (18 at D9); `UTC` printed at all 14	No
Gate coverage within the bands	the four gated shapes (one test) pass; the pooled-interval row cannot under-cover (100%); the night-temperature cell under-covers (79%)	Partly
Run time under 1.5 hours	not recorded	Not checked

Checks

Median forecast minus mean forecast: R and figure 9 use the median forecast of experiment 114's baseline model. On cold nights its median minus its mean is 0.0 at D5, D10, and D16. R and figure 9 include the hand-filtered switching half-hours; excluding them moves D5's miss at 01:00 from +6.1 to +6.8 points and moves the other substations by at most 0.1 points. The step fits themselves exclude those half-hours in the switching control only.
Hand-filtered switching list: refitting the six decision series without experiment 103's switching events changes no step by more than 0.3 points (11.0, 9.8, 15.4, 7.0, 3.0, and 13.5 for D5, D9, D10, D11, D14, and D16, against 10.8, 9.5, 15.4, 7.0, 3.0, and 13.4).
Block length: with blocks of 7, 14, and 28 days (100 draws), the coverage in the matched and placebo cells is 88%, 86%, and 84%, and the mean interval width 1.7, 1.9, and 2.1 points. The 28-day coverage is just below the 85% edge of the pooled band.
Audit triggers (conditions in the plan under which a result needs an independent audit). Trigger 1 (a control fails: test 1, the rank check, or the placebo coverage) fires: the gate's pooled-interval row, a known-answer test, gave 100% against 85 to 95%, and the tariff-edge placebo failed. Trigger 2 fires: the E7-X pooled step (13.2) is above the decision substations' (9.9), and 8 of the 10 predictions missed, some by 2 to 7 times. Two post hoc explanations are available. The predictions were anchored on experiment 121's residual steps, while the step here is the full step: dividing 121's 7.1 points at D10 by R = 0.63 gives about 11. And the flat part inherits the summer decline through the night (summer reference -0.7 to -13.4 points). Trigger 3 does not fire: no substation has a step and a raw contrast of opposite sign. Trigger 4 does not fire: leaving each decision substation out moves the pooled mean to between 8.8 and 11.2, inside 6.2 to 13.5. Trigger 5 does not fire: bounding the slope at 0 changes no slope. Trigger 6 does not fire: the widest decision-substation interval is 1.8 points (D9). Trigger 7 (a later design depends on the result) fires, because the cards and experiment 121's night arms use these numbers. The audit has run: every headline number reproduced, and its findings are applied on this page.

Substation cards

Provisional: fitted to past weather, estimation only. The MW column converts points with each substation's own 99th percentile of demand and is not a headline. D12 reports apparent power in MVA, so its "MW" is MVA.

Series	Step at 3 C, points (90% interval)	Step, MW (MVA at D12)	Slope, points per degree	Cold-dependent part, points	Flat part, points	Summer reference, points	Winters
D5	10.8 (10.4 to 11.2)	0.62	0.25 (0.16 to 0.32)	3.1	7.7	-5.2	2021, 2022, 2023, 2024
D9	9.5 (8.6 to 10.4)	0.68	0.38 (0.18 to 0.52)	4.7	4.8	-13.4	2023, 2024
D10	15.4 (14.9 to 15.8)	0.83	0.37 (0.26 to 0.45)	4.6	10.8	-5.5	2021, 2022, 2023, 2024
D11	7.0 (6.7 to 7.4)	0.76	0.13 (0.05 to 0.19)	1.6	5.5	-3.0	2021, 2022, 2023, 2024
D14	3.0 (2.5 to 3.4)	0.35	0.09 (0.01 to 0.15)	1.1	1.8	-4.2	2021, 2022, 2023, 2024
D16	13.4 (12.9 to 13.8)	1.22	0.26 (0.13 to 0.37)	3.2	10.2	-4.4	2021, 2022, 2023, 2024
D2	14.7 (14.1 to 15.1)	2.67	0.53 (0.36 to 0.62)	6.6	8.1	-9.5	2021, 2022, 2023, 2024
D3	6.6 (6.4 to 6.8)	0.85	0.16 (0.12 to 0.20)	2.1	4.6	-7.3	2021, 2022, 2023, 2024
D4	13.3 (12.2 to 14.3)	1.62	0.41 (0.28 to 0.49)	5.1	8.3	-11.4	2021, 2022, 2023, 2024
D8	0.5 (-0.7 to 1.7)	0.05	0.19 (0.09 to 0.29)	2.4	-2.0	-0.7	2021, 2022, 2023, 2024
D12	14.9 (13.2 to 16.4)	0.70	0.46 (0.16 to 0.67)	5.7	9.2	-2.3	2021, 2022, 2023, 2024
D19	33.5 (32.1 to 34.5)	1.82	1.12 (0.84 to 1.29)	14.0	19.4	-10.5	2021, 2022, 2023, 2024
D21	9.6 (9.1 to 10.1)	0.54	0.19 (0.10 to 0.25)	2.4	7.2	-5.9	2021, 2022, 2023, 2024
D22	12.7 (12.0 to 13.3)	0.41	0.16 (0.04 to 0.26)	2.0	10.7	-6.0	2021, 2022, 2023, 2024

Start of the step, timing, and share left by the baseline model (R):

Series	Start of the step (half-hour starting)	Cold BST nights	Pre-registered label	Post hoc jump shift on BST nights, half-hours	R / full step (baseline model)
D5	00:30	75	UTC	2	0.78
D9	00:30	18	UTC	9	n/a
D10	00:30	88	UTC	2	0.63
D11	00:30	90	UTC	2	0.68
D14	00:30	75	UTC	4	2.05
D16	00:30	68	UTC	2	0.53
D2	00:30	83	UTC	2	n/a
D3	00:30	75	UTC	2	n/a
D4	00:30	83	UTC	2	n/a
D8	00:30	75	UTC	2	n/a
D12	00:30	75	UTC	2	n/a
D19	00:30	83	UTC	2	n/a
D21	00:30	83	UTC	2	n/a
D22	00:30	75	UTC	2	n/a

Cold nights: actual demand against the baseline forecast

On cold nights actual demand runs above the baseline forecast of experiment 114 from about 01:00 at D5 and D16 on the two pre-registered winters (by 5.6 and 9.4 points), and at D10 in winter 2024-25 alone (+7.6 at 01:00); at D10 the two-winter median is dominated by a level error in winter 2023-24. The figure shows the median and the 25th to 75th percentile, across the 104 cold winter nights per substation of winters 2023-24 to 2025-26 (the coldest 10% of days, the plan's cold-snap rule), of actual demand and of the day-ahead median forecast, in points. The largest median misses (actual minus forecast) are D5 +6.1 in the half-hour starting 01:00, D9 +2.9 starting 01:30, D10 +6.7 starting 01:00, D11 +3.6 starting 03:00, D14 +4.0 starting 02:00, and D16 +9.6 starting 01:00. The half-hour of the largest rise between 05:00 and 08:00 differs between actual and forecast at D5 (07:00 against 06:00), D9 (06:00 against 07:00), D11 (06:30 against 06:00), and D16 (07:00 against 06:30), and matches at D10 and D14. The two earlier winters alone (72 nights, before the selection winter was read) gave the same pattern at D5 and D16 (+5.6 and +9.4 at 01:00) but a different largest miss at D10 (-12.2 in the half-hour starting 23:30). The cause is winter 2023-24: D10's cold nights sit about 11 points lower than in the other winters (about 42 points against 52 to 54), and the baseline forecast is 9 to 17 points too high through the whole night. That level drop is not in the switching mask, and it also lowers D10's 2023-24 step. Winter 2024-25 alone shows +7.6 at 01:00 and +7.0 at 00:30, the same pattern as D5 and D16, and adding winter 2025-26 dilutes the 2023-24 error rather than creating the rise. report.txt prints both.

Actual demand against the baseline forecast, decision series

Actual demand on cold nights, E7-X series

Follow-up readouts (post hoc and exploratory)

These readouts come from an unverified summary of Economy 7 tariff practice, whose claims are hypotheses to test; nothing here cites it as a source. They were chosen after seeing the main results, so they are post hoc, and they read winters 2021-22 to 2024-25 only (cold British Summer Time nights up to October 2025), leaving the scored-year lock unchanged. The script is e00123_followups.py, and report.txt prints every number.

No substation shows a local-clock component on cold British Summer Time nights. For each cold night we found the half-hour entered by the largest rise between 23:30 and 03:00. On cold winter nights that half-hour is the one starting 00:30 on 66% to 100% of nights (11 of 14 substations are at 85% or more). On cold British Summer Time nights it falls at the local-clock step (starting 00:30) on at most 11% of nights (D22) and at the UTC step (starting 01:30, one hour later on the clock) on 24% to 100% of nights, with the remainder at other half-hours. D14 (24%) and D9 (50%) have the weakest UTC share, and their steps are the smallest or the least observed (3.0 points at D14, 18 cold British Summer Time nights at D9), so a mixture of clocks is not shown, and not ruled out.
The step ramps over about one half-hour and has no visible second window. Averaged over cold winter nights and minus the summer shape, demand reaches 7% to 56% of its plateau in the half-hour starting 00:00 and 81% to 119% in the half-hour starting 00:30. No substation has a second rise of at least 30% of the main one. The end of the window is not visible: the largest fall between 05:00 and 10:00 is small (0.5 points or less at seven substations) and falls at 07:30 at only four substations (D9, D10, D2, and D19), where the roughly 7-hour window would end if it began at 00:30. The morning demand rise overlaps the end of any window.
The step by winter falls at five substations, rises at one, and cannot be told at eight. The label compares the first and last winter's 90% intervals. It falls at D10, D2, D3, D19, and D22, rises at D11, and is cannot tell elsewhere, including D9, which has two winters. Only winter 2024-25 is the latest winter here, so the step by year cannot yet say anything about the 2025 and 2026 closures; that waits for winter 2025-26.
Descriptive only for now: fixed-window heat-pump and electric-vehicle tariffs that would add a night step on the local clock, and bespoke half-hourly contracts.

Follow-up card rows, one per substation (percentages are shares of cold nights; the reference columns are the same two shares on cold winter nights; fall sizes are in points):

Series	Largest rise at the local-clock step, cold BST nights	At the UTC step, cold BST nights	Same two shares, cold winter nights	Plateau reached by the half-hour starting 00:00 / 00:30	Largest fall 05:00 to 10:00	Second rise	Step by winter
D5	0%	92%	100% / 0%	0.18 / 0.93	09:30 (-1.2)	none	cannot tell (-1.0)
D9	0%	50%	90% / 0%	0.36 / 0.91	07:30 (-0.5)	none	cannot tell (-0.0)
D10	0%	72%	97% / 0%	0.34 / 0.99	07:30 (-0.4)	none	falls (-2.8)
D11	3%	59%	75% / 1%	0.27 / 0.85	08:00 (+0.2)	none	rises (+1.2)
D14	0%	24%	72% / 0%	0.07 / 1.16	05:30 (-0.4)	none	cannot tell (-1.2)
D16	0%	85%	87% / 0%	0.47 / 1.09	09:30 (-0.3)	none	cannot tell (-0.7)
D2	0%	100%	100% / 0%	0.44 / 1.19	07:30 (-0.8)	none	falls (-2.8)
D3	0%	100%	100% / 0%	0.17 / 1.09	09:30 (+0.3)	none	falls (-2.1)
D4	0%	96%	100% / 0%	0.56 / 1.03	09:00 (-2.7)	none	cannot tell (-0.5)
D8	8%	65%	85% / 0%	no step	09:30 (-1.2)	none	cannot tell (+0.2)
D12	3%	59%	66% / 4%	0.19 / 0.81	08:30 (-9.1)	none	cannot tell (-1.3)
D19	0%	99%	100% / 0%	0.38 / 1.02	07:30 (-4.8)	none	falls (-9.9)
D21	1%	81%	94% / 0%	0.45 / 1.13	09:30 (+0.2)	none	cannot tell (+0.3)
D22	11%	73%	90% / 0%	0.36 / 1.02	09:30 (-1.3)	none	falls (-2.6)

UTC-fixed and local-clock shares by year (post hoc and exploratory)

The step could come from a tariff fixed to the local clock as well as from the legacy step fixed to UTC. The following start and end times are hypotheses from a search for equivalents, to be checked against Elexon's public tables, and are not sources: the legacy Economy 7 standard window 00:30 to 07:30 on Greenwich Mean Time, Intelligent Octopus Go 23:30 to 05:30, Octopus Go 00:30 to 05:30, and a Cosy heat-pump tariff. This page does not say which of them follow Greenwich Mean Time or local time.

For each cold British Summer Time night (daily mean below 8 C, March 2021 to October 2025) we found the half-hour entered by the largest rise between 23:30 and 03:00 on the local clock. The first percentage in each cell is the share of nights where it is the half-hour starting 01:30 local, where a step fixed to UTC at 00:30 would fall. The second is the share where it is the half-hour starting 23:30 or 00:30 local, where tariffs fixed to the local clock would fall. The number of nights is in brackets, and cannot tell marks a year with fewer than 10 nights. The 2025 column holds spring and summer nights only. The shares do not sum to 100% because the largest rise can fall elsewhere. e00123_clock_by_year.py prints the table, and the night profiles in figure 8 and figure 9 are indexed on the local clock, with UTC one hour earlier in British Summer Time.

Series	2021	2022	2023	2024	2025
D5	96% / 0% (25)	100% / 0% (12)	100% / 0% (13)	92% / 0% (13)	cannot tell (4)
D9	cannot tell	cannot tell	cannot tell (1)	55% / 0% (11)	cannot tell (6)
D10	45% / 0% (31)	79% / 0% (14)	71% / 0% (14)	100% / 0% (15)	cannot tell (5)
D11	61% / 3% (31)	57% / 7% (14)	57% / 0% (14)	53% / 7% (15)	cannot tell (5)
D14	12% / 0% (25)	25% / 0% (12)	46% / 0% (13)	15% / 8% (13)	cannot tell (4)
D16	100% / 0% (26)	92% / 0% (12)	77% / 0% (13)	69% / 0% (13)	cannot tell (4)
D2	100% / 0% (31)	100% / 0% (12)	100% / 0% (13)	100% / 0% (11)	cannot tell (6)
D3	100% / 0% (25)	100% / 0% (12)	100% / 0% (13)	100% / 0% (13)	cannot tell (4)
D4	100% / 0% (31)	100% / 0% (12)	100% / 0% (13)	100% / 0% (11)	cannot tell (6)
D8	88% / 0% (25)	67% / 0% (12)	31% / 31% (13)	62% / 8% (13)	cannot tell (4)
D12	76% / 0% (25)	58% / 0% (12)	15% / 8% (13)	46% / 15% (13)	cannot tell (4)
D19	100% / 0% (31)	92% / 0% (12)	100% / 0% (13)	100% / 0% (11)	cannot tell (6)
D21	97% / 0% (31)	58% / 0% (12)	85% / 0% (13)	45% / 9% (11)	cannot tell (6)
D22	88% / 0% (25)	83% / 0% (12)	38% / 38% (13)	62% / 8% (13)	cannot tell (4)

The UTC share is 100% in every year with a reading at D2, D3, and D4, and 92% to 100% at D19, and a local-clock share above 10% appears only at D8 and D22 in 2023 (31% and 38%) and at D12 (15%) in 2024. A local-clock component is therefore visible at a few substations in single years, and cannot be separated from noise with 13 nights; it would need a second winter or a larger sample to say.

Tariff edges on the local clock (post hoc and exploratory)

Local-clock tariff edges are not visible at Primary level: the control failed, and no change since 2021 beyond the placebo limit appears at the Octopus Go edges outside D8. The tariff windows are unverified hypotheses from a search for equivalents, not sources: Intelligent Octopus Go 23:30 to 05:30, Octopus Go 00:30 to 05:30, and Cosy 04:00 to 07:00, 13:00 to 16:00, and 22:00 to 24:00 on winter weekdays. The script is e00123_tariffs.py; report.txt prints every number. Saving Sessions and demand-flexibility events are not modelled.

For each edge we took the rise in demand entering the half-hour starting at the edge, minus the median rise at the four half-hours either side, on British Summer Time nights (April to October, when the UTC-fixed Economy 7 step falls an hour later on the local clock) for the Octopus edges, and on winter weekdays for Cosy. Intervals resample 14-day blocks. Half-hours with no hypothesised edge (21:00, 03:30, 10:00, and 18:30) serve as placebos.

The level of the statistic is not a test. The ordinary daily load shape gives every half-hour a non-zero rise against its neighbours: the 90% interval excluded 0 in 76% to 79% of series-years at the placebo half-hours (5% expected), so a visible edge in one year means nothing. The median half-width of the intervals is 0.3 points.
Known-answer injection. Adding a step of S points (1, 2, and 4) to the edge statistic shifts the estimate and its interval by exactly S, so recovery is exact and the interval width is unchanged. The placebo is the control that failed; the injection says nothing about the load-shape confound.
The change between years cancels the load shape but the control still fails. Comparing the first and last year (2021 and 2025 for the Octopus edges; winters 2021-22 and 2024-25 for Cosy), the 90% interval of the change excluded 0 at 21% to 43% of placebo series (10% expected for either sign), because the load shape drifts between years by more than the sampling error. The empirical detection limit is therefore the 90th percentile of the placebo changes, 1.19 points (about 0.06 to 0.22 MW across the 14 substations).
Against that limit, the Octopus Go and Intelligent Octopus Go edges show a change at D8 only. At D8 the 23:30 edge changed by +2.5 points and the 05:30 edge by -2.3, in the expected signs. D8 has no Economy 7 step and large changes at the Cosy edges too (-12.4 points at 16:00), which points to a change in the substation's load, not to a tariff. The 00:30 Octopus Go opening, the edge the legacy step could be confused with in summer, shows no substation beyond the limit; the median change over the 14 substations is +0.6 points at 23:30, -0.06 at 00:30, and +0.03 at 05:30.
Cosy is not distinguished from other load changes. Beyond the limit, in the expected sign: the 07:00 closing at 4 substations (D11, D21, D22, and D8), the 16:00 closing at 5 (D10, D12, D16, D22, and D8), the 04:00 opening at 2 (D12 and D8), the 13:00 opening at 2 (D12 and D8), and the 22:00 opening at 3 (D19, D3, and D8). D8 and D12 appear at most edges, which fits a change in their loads. A heat-pump tariff would add its edges together; no substation other than D8 and D12 shows more than one opening and one closing together.

The detection limit is 1.19 points of a substation's demand, which is the smallest year-on-year change in an edge this design can tell apart from placebo drift. A tariff effect smaller than that at Primary level is not visible here. The sizes by year at each substation are in report.txt. The design could improve by using a matched control substation or by shortening the comparison to neighbouring years.

Limitations

Summer is assumed step-free. On summer nights the later half-hours average 0.7 to 13.4 points below the earlier ones, which is the ordinary decline of demand through the night, and the winter-night shape differs from it for reasons besides storage heating. The full step carries that difference. The slope does not.
Four headline winters of one weather product, from one set of substations. The winters are separate years, and neighbouring cells share one ERA5 product.
The gate says nothing about the pooled intervals, which may be too narrow in practice. The gate's 100% pooled coverage is built in (the gate paragraph). The substations share one weather product, so their errors are not independent, which would make the intervals too narrow.
Cold-snap means rest on few winters. The raw cold-snap means per winter (November to March beside November to February) are in report.txt, with 32 to 38 nights per winter. They cover four winters at 13 substations and two at D9, and the winters are not independent experiments.
The daily mean temperature is a proxy for the temperature that sets heater charging. In the gate, a step that follows the 18:00 to 08:00 mean temperature was covered by the interval in 79% of draws pooled at 7 points, and in 43% at the worst substation.
Winter 2025-26 was read once, after the headline winters. The page's headline numbers use the headline winters, and the selection winter appears beside them.

The plan, as committed before any fit

E7 is an estimation experiment. The result is a table of estimates with intervals, not a win-or- lose verdict, and the experiment forecasts nothing. The word calibrated is used only for the known-answer gate (interval coverage within a stated band), and detectable only for the per-series reading. The plan follows briefs/dp_series.md, then briefs/common.md. The Sonnet plan review is triaged at the end of this page. No data on or after 2026-07-01 is read, nothing in packages/ changes, and no dependency is added.

Pre-fit clarification: what the headline is (written before any fit)

The headline is the pooled step over all six decision series (D5, D9, D10, D11, D14, and D16), with a random effect across series, at a daily mean of 3 C. The extension series, E7-X, are shown separately. D5, D10, and D16 are reported as the subset that prompted the study, labelled "selected for showing the step", and no verdict rests on them alone. Every fitted series, including each with no step, appears in the unpooled table. This replaces the earlier draft, in which the verdict rested on D5, D10, and D16.

The selected subset is expected to look larger than the pool, and the gap is the selection effect. Experiment 121 read winter 2025-26 and found a step at 01:00 of 4.3, 7.1, and 5.3 points in the residuals of the day-ahead model at D5, D10, and D16, and 1.9 to 3.4 points at D11, D14, and D9. The choice of those three series came from that winter, so winter 2025-26 is labelled the selection winter.

Pre-fit clarification 2: the unit is points of effective capacity (written before any fit)

Every size in this plan is in points of effective capacity, and the first draft mislabelled the unit as megawatts. One point is 1% of the series' effective capacity, which is the series' own 99th percentile of cleaned demand and not an electrical rating. Experiment 121's step figures (4.3, 7.1, and 5.3 at D5, D10, and D16) are in points: its quantile scores are in points, and a 7.1 MW step at D10 would exceed the series' 5.4 MW effective capacity. The wrong label was found in test 1, which showed D10's night difference to be about 0.4 MW (about 8 points). Every number in the plan (the predictions, the injected shapes S of 4 and 7, the bias bound of 0.5, the trigger thresholds of 4, 8, and 15, and the 1.7 summer step) keeps its value and now reads in points. Each series' night difference d is converted to points before any fit, so the pooled steps average comparable quantities, and the page also prints each series' step in MW (one point at D10 is 0.054 MW). Test 1 printed raw night differences for one series in the design winters and a per-series mean d over all nights of winters 2021-22 to 2025-26; no estimate, interval, or fit was computed before this clarification.

Units clarification: what "effective capacity" is, and the raw step in MW. The table's effective capacity is a scale taken from the demand itself: it matches each series' 99th percentile of cleaned half-hourly demand (D10: 5.398 MW against a 99th percentile of 5.413 MW; D5: 5.723 against 5.734; D16: 9.115 against 9.143). It is neither the electrical rating of the substation nor a PV capacity. A point is therefore 1% of that near-peak demand. Computed from the stored cleaned power on cold GMT nights (daily mean at most 3 C) of winters 2023-24 and 2024-25 (printed in report.txt under raw_cold_step), the half-hour ending at 01:00 minus the mean of the 22:00 to 23:30 half-hours is 0.27 MW (4.8 points) at D5, 0.44 MW (8.2 points) at D10, and 0.87 MW (9.5 points) at D16. The raw step is therefore a few tenths of a megawatt, and the figures of 1.9 to 10.7 "MW" in the earlier descriptive check of the night step are points: that check divided by a proxy capacity of about 100, which is the scale of points.

Pre-fit clarification 3: actual demand against the baseline forecast on cold nights (written before the figure was computed)

A figure on the substation cards shows actual demand and experiment 114's baseline forecast (B0) on cold winter nights, so that a visible miss at 01:00 shows up. The maintainer asked for it after the first fits. The rule is fixed here, before any residual is computed: cold winter nights are the cold-snap nights of the plan (the evenings whose next day lies in the coldest 10% of days of its July-to-June year by the mean ERA5 temperature across the six decision series, on Greenwich Mean Time nights that are not excluded), for winters 2023-24 and 2024-25, the winters with stored B0 predictions before the scored winter is unlocked. Each panel shows the median across nights, with the 25th to 75th percentile band, of actual demand and of the B0 median forecast (day D+1), for the half-hours starting 22:00 to 07:30, in points. The eight E7-X series have no stored B0 forecast, so they appear with actual demand only. A one-line list of other visible misses follows, from the median B0 residual per half-hour: the half-hour with the largest median residual, and the half-hour of the largest 30-minute rise between 05:00 and 08:00 in the actual and in the forecast. The figure is descriptive: it adds no contrast and no verdict. Winter 2025-26 enters only after PR 1397 has merged.

The question, and what earlier experiments already say

The night step is real in one winter and unmeasured across winters. Experiment 121 tried to detect the step through quantile scores over all 20 night half-hours, where a step of 4 points is diluted to about one noise standard deviation, and its fourth gate had to be redefined twice.

E7 changes the instrument. E7 compares five window half-hours with four baseline half-hours of the same night, so the dilution disappears and a slow level error cancels. The estimator is a regression of the night difference on cold, in the spirit of the degree-day regressions used for load, and the pooling is the random-effects meta-analysis of DerSimonian and Laird (1986) with the Hartung and Knapp (2001) interval. The block bootstrap follows Künsch (1989). The earlier experiments this plan builds on are 111, 113, 114, 120, 121 (PR 1413), and 122. One targeted search for a DOI on Economy 7 or storage-heating load found none on temperature dependence. The nearest sources are Elexon CP1058A, which describes the Economy 7 profile class, and a statistical review of UK residential electrical loads; neither was checked for a temperature-dependent night step. Elexon's Radio Teleswitch guidance note states that teleswitch times are in UTC and that Settlement converts them to clock time in British Summer Time where a GMT indicator is set; which value of the indicator means fixed to GMT is unresolved in the documents read, so "UTC-fixed" on this page is a reading of the data only.

Estimand

Window rows: the five half-hours starting 00:30, 01:00, 01:30, 02:00, and 02:30. Test 1 prints the timestamp convention of the observed file against experiment 121's code, and the window follows the documented convention only. It is fixed in test 1, before any estimate, and never moved. Sensitivity (descriptive): the three-row window 00:30 to 01:30.
Baseline rows: the four half-hours starting 22:00 to 23:30 of the same night, on the same clock as the window. A night is labelled by its evening's local date, and its day type (weekday, Saturday, Sunday) is that date's.
Night difference d: the mean of the window rows minus the mean of the baseline rows, in points.
Cold c: max(0, 15.5 - T) degrees, where T is the daily mean of hourly ERA5 temperature at the substation's cell over the local calendar day containing the window rows. The daily mean is a proxy for the temperature that sets storage-heater charging, and it includes the afternoon after the step. Its correlation with the mean temperature from 18:00 to 08:00 is printed, and a night-temperature c is a descriptive sensitivity.
Primary rows (GMT nights): nights between the October and March clock changes, when the local and UTC clocks agree. Clock-change nights, bank-holiday nights, and the Christmas to New Year run are dropped from every fit.
Summer reference d_summer: the mean of d over June to August nights on the local clock. Experiment 121's one summer pair found a step of 1.7 points or less; the page states the assumption that the summer night has no step.
Slope b: the points change in d per degree of c, from d = a_w + b*c + day-type terms + noise on GMT nights, with a separate intercept a_w for each winter w and a common b across winters. The slope needs no summer assumption.
Step at a stated cold night: step = a_hat + b*c_ref - d_summer, with c_ref = 12.5 degrees (daily mean 3 C, inside experiment 122's cold-snap range of -2.0 to 3.1 C), where a_hat is the mean of the winter intercepts. The step is the full cold-dependent and flat step, with no demand model removed. The flat part is a_hat - d_summer, so reading the step at c = 0 is an extrapolation, because few GMT nights are warmer than 15.5 C. The flat part is reported separately from b.
Excess over B0 (cross-check R): the same regression on experiment 114's stored B0 residuals measures only the part of the step that B0's free daily profile and its 00:00 to 04:00 heating band do not already absorb. R is a different, smaller quantity, expected at 30 to 60% of D, and is compared with D only as a measure of the absorption, never as an error.

Substations and histories

Fitted series: the six decision series, which are DP Primaries, and the eight E7-X Primaries. The rule-based flag of nged_data.cleaning is applied to all 14 series, so E7-X needs no hand-filtered switching list. The six decision series are also run with their hand-filtered list as a sensitivity. E7-X has its own row set, e7_nights_v1, outside dp_mw_v1. D12 reports apparent power in MVA, so its night difference is an apparent-power step, converted to points in the same way.

Series	Observed from	Winters (2021-22 to 2025-26)	Role
D5, D10, D11	2019-09-17	5	decision series; D5 and D10 prompted the study
D14	2019-10-21	5	decision series
D16	2020-12-22	5 (the history covers 2021-07 onwards)	decision series; prompted the study
D9	2023-08-23	3 (2023-24 to 2025-26)	decision series
D2, D3, D4, D8, D12, D19, D21, D22	2019-09-17	5	E7-X: not chosen for any step
D15, D18	n/a	n/a	generator-linked, shown only as G-c, not fitted

Winters 2021-22 to 2025-26 only. Winter 2019-20 is not added: its window would leave about 6 months after the lockdown months go out, with no summer and no pinned-capacity vintage.

Winter roles. Winter 2023-24 (tuning) and winter 2024-25 (replicate) are used for design checks (block length). Winters 2021-22 and 2022-23 are not opened for any design choice, and their estimates are read only after the gate has run. Winter 2025-26 is the selection winter, and the maintainer ruled that it is fitted only after PR 1397 (experiment 114) has merged. Headline winters are 2021-22 to 2024-25 (the winters that did not choose the three series), and winter 2025-26 is reported beside them. There are at most four headline winters and five in all, from one weather product and one set of substations, so the winters are separate years and not independent experiments. The page prints the winter count at every estimate.

Estimators

D (primary): the direct per-night regression above. Intervals come from a 14-day-block bootstrap of nights stratified by winter, 1,000 draws, the same count in the real fits and in the gate. A bounded (non-negative) b is fitted only as a descriptive variant, never in the gate.
R (cross-check): the same regression on stored B0 residuals (predictions/{tuning,replicate, scored}_b0, median, day D+1), for winters 2023-24 to 2025-26 only, at the six decision series. Experiment 114 stores no earlier residuals. R is not injected into the gate, because injected data would need B0 refitted per draw. R reads winter 2025-26 only after PR 1397 has merged, and the run pins the merge commit's SHA in decisions.json.
E7-K (descriptive): R on the stored *_b0_k and *_t_all_k predictions, which shows how much of the step the Kalman filter removes.
Cold-snap subset (winter-first rule): the coldest 10% of days of a winter by the mean ERA5 temperature averaged across the six decision series (experiment 122's rule). The day count is stated, and too few days where under 5. On the subset the page prints the raw mean of d minus d_summer. November to March is reported beside November to February.
Event-day rule: not applicable, because E7 has no scored contrast across calendar events.

Pooling

The pool is a random effect across series, with limits. With k series the pooled mean has a Hartung-Knapp interval (t with k - 1 degrees of freedom) and, beside it, a percentile interval from resampling whole series. The between-series standard deviation tau is printed with a profile- likelihood interval. The pooled mean is marked cannot tell if the two intervals disagree on whether 0 lies inside. With fewer than 5 series in a set, the page prints the plain mean and range, with no tau. Neighbouring cells share one ERA5 product, so the series are not independent, and the page says the pooled intervals are optimistic.

Pooled sets, side by side:

the six decision series (headline);
D5, D10, and D16 (selected for showing the step);
D11, D14, and D9 (not selected);
the eight E7-X series;
all 14 series.

The gap between sets 2 and 3 shows the selection effect.

Known-answer test and gate

Order: test 0, test 1, the rank check, design checks, the gate, and only then the estimates of winters 2021-22 and 2022-23, and winter 2025-26 after PR 1397 has merged. The gate can fail by design. A failed gate is recorded as it stands and is not redefined, and it removes the detectable label, so the raw means carry the page.

Feasibility. Each fitted series has the rule-based flag, at least two winters, and a night table with every window and baseline row present on at least 90% of its nights. Failures are named and dropped.
Reproduction (blocking). The night table's d equals a hand-computed d on 20 printed nights; the timestamp convention and window rows are printed; no row on or after 2026-07-01; R's stored residuals reproduce experiment 114's stored B0 quantile score for the tuning year to 1e-4.
Known-answer injection (can fail). Noise: real nights of the same series with D's fitted cold-dependent part removed from the five window rows, resampled in 14-day blocks within winter, with the real covariates. Injected shapes, in points at c_ref, scaled by c / 12.5, for S in {4, 7} plus a placebo S = 0:
- matched: an equal amount on the five window rows (the fitted form);
- ramp-plateau: weights 0, 0.5, 1, 1, 1 on the five window rows, continuing at weight 1 until 07:00. The truth is the mean weight over the window times S;
- flat: S on the window rows of every GMT night, not scaled with cold (truth b = 0, step = S);
- night-temperature: the matched shape driven by the 18:00 to 08:00 mean temperature instead of the daily mean (descriptive: its bias is reported, and no coverage target is set).

There are 200 draws per cell per series, at all fitted series. - Target (the label rests on these rows): coverage of the truth by the 90% interval of the step at c_ref, for the matched, ramp-plateau, and flat shapes and the placebo, is 85 to 95% pooled over series and 80 to 98% for each series. With 200 draws the Monte Carlo standard error is about 2.1 points per series and about 0.9 points pooled over 14. Block resampling makes the draws less independent, so these errors are lower bounds. - Bias: the mean estimate lies within 0.5 points of the truth for the matched and flat shapes at 4 and 7. - Power (descriptive): the share of draws with the 90% interval above 0, at 4 and 7 points, per series. A series under 80% at 4 points is marked "weak". - Pooled interval: S is drawn per series from {0, 0, 4, 4, 7, 7} across the six decision series, and the pooled mean's Hartung-Knapp interval covers its true mean in 85 to 95% of 200 draws. 3. Rank check (blocking). The design matrix (1, c, day-type terms, winter intercepts) has full rank, its smallest singular value and the variance inflation of c are printed per series, and a series with a variance inflation above 10 is dropped.

Gate result: the table says interval coverage within 85 to 95% or names the failing cell. The page does not say the estimate is right, because the gate tests the estimator on shapes this plan chose.

Timing (descriptive)

The gate cannot calibrate timing, because cold British Summer Time (BST) nights are few. A cold BST night is a night in BST with a daily mean below 8 C. Per series and per winter-pair, the page prints the count and says cannot tell where it is under 10. Where it is 10 or more, the estimator is delta, the mean over those nights of d measured with the local-clock window minus d measured with the UTC-clock window (the baseline on the matching clock). A step fixed to the local clock makes delta positive, and a step fixed to UTC makes it negative. A 90% block-bootstrap interval is printed and the label (local, UTC, unresolved) is marked uncalibrated.

Reporting

Table 1 (unpooled, first): per series, b, the flat part, and the step at c_ref, each with 90% intervals, the winters used, the raw cold-night mean of d minus d_summer, R and its ratio to D, and the detectable flag. Every fitted series appears, including those with no step.
Table 2 (winters): the step at c_ref per winter per series, the selection winter marked.
Table 3 (pooled): the five sets above, with both intervals, tau, and cannot tell where it applies. The first set is the headline.
Table 4 (timing): counts, delta, and the label.
detectable: the 90% interval of the step at c_ref lies wholly above 0, and the gate row for that series is within its band. No multiplicity correction is made: with 14 series, about 0.7 false positives are expected under no step (5% one-sided each), and the page says so.
Winter breakdown: every estimate is shown per winter, with the count of winters.

Predictions

Written before any estimate; each line can fail.

Headline pooled step at c_ref, six decision series, headline winters: between 2 and 6 points, with the Hartung-Knapp interval wholly above 0.
Selected subset (D5, D10, D16): 5 to 10 points, above the pooled step of D11, D14, and D9, which is 1 to 3.5 points and may include 0. The selected subset's larger value is the selection effect.
E7-X: each series 0 to 3 points, and the E7-X pooled step 0.5 to 2 points, below the pooled step of the decision series.
Slope b: positive at D5, D10, and D16, between 0.4 and 0.8 points per degree.
Flat part: a_hat - d_summer within +-2 points at every series. A series whose 90% interval lies wholly outside +-2 points fails the line.
R against D: 30 to 60% at D5, D10, and D16.
Raw contrast: the raw cold-night mean of d minus d_summer lies within 2 points of D's step at the mean c of those nights (about 8 to 10 degrees), not at c_ref.
Timing: cannot tell at most series; if a label is printed, local.
Gate: coverage within the bands. The chance of a failing cell is judged about one in three, because the ramp-plateau shape changes the window mean.
Run time: under 1.5 hours of wall clock on CPU.

Decision rule

There is no win-or-lose verdict on a hypothesis about a model. The verdict word summarises the estimates and the gate:

supported: the gate rows are within their bands for all six decision series, the headline pooled step over the six decision series has a Hartung-Knapp interval wholly above 0 in the headline winters, and the percentile interval agrees.
partly supported: the gate is within its bands and the pooled step is positive but cannot tell, or positive only in the selection winter.
inconclusive: a gate row for a decision series fails, or fewer than 2 headline winters are available at those series. invalid: a leak (a row on or after 2026-07-01). failed: test 1 or the rank check cannot pass.
The estimate does not choose between models and does not change the production forecast. A supported result feeds the substation cards and a calibrated b can serve as a prior for experiment 121's night arms; neither is part of this plan.

Audit before merge: the seven triggers

The seven triggers are those of briefs/common.md and briefs/dp_cal.md, applied to these results.

A control fails: test 1, the rank check, or the placebo coverage outside its band.
Contradiction: a pooled step above 15 points, a negative step at a series whose raw cold- night mean exceeds 4 points, or an E7-X pooled step above that of the decision series.
Estimate and raw contrast disagree in sign at a series.
Leaving one series out moves the pooled mean outside the all-series pooled interval.
A null control moves: the unbounded and bounded D disagree at a series with b far from 0.
Interval width: D's interval at c_ref is wider than 8 points at a decision series.
A later design depends on it: the cards and experiment 121's night arms use the numbers, so trigger 7 fires. The manager runs the Opus review of the estimator and the gate after the quota resets (01:00 UTC on 2026-10-21); this session runs one Sonnet code review for leakage (the cut at train_end) and one Sonnet write-up review.

Privacy

D-labels only, with dates, for the Primaries; no substation names, no topology. D15 and D18 are generator-linked and appear only as G-c, normalised, with no dates, and are not fitted. Fitted steps in points are acceptable for Primaries, which are not generators. Per-series slopes are shown only where the rank check passes and the gate row is within its band, because experiment 121 forbade per- site slopes from weakly identified fits.

Substation cards in the plan

Section label: "Provisional: fitted to past weather, estimation only". One card per fitted series shows:

the step at a stated cold night (daily mean 3 C), in points, with its 90% interval, the slope in points per degree, and the winter-versus-summer contrast;
the flat part and the summer reference;
the start time: the first half-hour whose mean (observed minus baseline), on cold GMT nights minus the summer mean, exceeds half of its value at 02:30, with a bootstrap interval in half- hours;
the timing label (local or UTC, where estimable) or cannot tell, with the count of cold BST nights;
R's ratio to D and the Kalman filter's removal (E7-K);
an example forecast week, where the series has one in dp_series.md's selection rule, drawn from stored predictions.

Comparison with experiments 121 and 120

Experiment 121 (MH). Its night-arm estimates are compared with D's step per series, with both intervals and their overlap. Experiment 120 (SH) learns an unsupervised daily-shape library with a 01:00 shape. When SH reports, the comparison per series is SH's integrated 01:00 jump over the five window rows, in points at the same cold night, beside D's step. A series where SH shows no jump and D's interval lies wholly above 0 is a case the learned library misses, and the card says so. The plan does not decide which is right.

Run order, reuse, compute, and time

Scripts import only their own folder, studies.*, and reviewed packages. The scripts read the series' power through studies.power.scan_power (cleaned readings), ERA5 temperature from the stage 1 pilot's era5_features.parquet, and the copy of experiment 114's and experiment 121's window and block-resampling conventions. D needs no demand model.

Order: night table and tests 0, 1, and 3; design checks on winters 2023-24 and 2024-25 (block length); the gate; then winters 2021-22 and 2022-23; then winter 2025-26 after PR 1397 has merged; tables, cards, write-up.

CPU only, 6 cores, with OMP_NUM_THREADS=1, MKL_NUM_THREADS=1, and OPENBLAS_NUM_THREADS=1 in every worker (one process per series). The run pauses when the 1-minute load average is above about 35. The budget is 1.5 hours of wall clock, no GPU.

What the write-up must contain

The take-home says how large the full night step is at a daily mean of 3 C, pooled over the six decision series and per series, whether it grows with cold, what the selection effect is, how much of the step B0 already absorbs, and that the clock is not identified where cold BST nights are few. The page is figure-led: raw first (the mean of d against c with the fitted line and the summer reference), then the step per series with intervals, the same per winter with the selection winter marked, known-answer recovery per shape, the pooled forest plot with the five sets, and cold BST night counts.

Caveats: the three large series were chosen on the selection winter; the five winters share one weather product; the summer reference may hold a step of up to 1.7 points; the flat part is an extrapolation; the step has the pattern of storage heating and NGED has not confirmed the cause (daylight, lighting, and switching schedules also change at night); the daily mean temperature is a proxy for the temperature that sets charging; the gate tests only the shapes this plan chose.

Size: the five triggers

Stored: private estimates and figures under data/ml-experiments/, and the write-up under docs/ml-experiments/; nothing in Delta. 2. Serving path: not touched. 3. Degradation rule: not touched. 4. More than one defensible design: yes (window, cold variable, estimator); this plan pre-registers D. 5. Callers not nameable without searching: none. Size: medium.

Review triage

The plan was reviewed once by a Sonnet agent (30 findings), and the manager and maintainer rulings were applied: D replaces the earlier joint estimator; the bound is removed from the gated estimator; the gate tests the used estimator on misspecified shapes; the pooled estimate is reported after the unpooled table; E7-X runs on the rule-based flag; no winter 2019-20. The full review is private. A second change, the pre-fit clarification at the top of this page, makes the six decision series the headline.