zero-trust-physics/zero-trust-physics-grid-gb-frequency

GB grid frequency after an interconnector trip: 86 losses of infeed computed again with a proof on every row

For frequency containment in Great Britain, the questions after a loss of infeed are how deep the frequency falls (the nadir), how fast it falls (the rate of change of frequency) and where it settles. Here are the 86 single importing-interconnector trips of 500 to 1,402 MW that NESO's one-second frequency recorded from May 2017 to June 2026, each computed again for its first 60 s from the half hour's published inertia, the demand, and the Dynamic Containment, Moderation and Regulation NESO held, with a cryptographic proof on every row, beside the measured frequency.

Fitted on all 86 trips, the model's depth of the fall is within 0.217 RMS relative error of the measured depth; with each trip left out of the fit in turn, 0.224. Gap: the 11 trips since November 2025, 0.359.

Zero-Trust Physics is a physics verifier for Super Intelligence (SI). SI agents ask it questions about real hardware, and every answer comes with a cryptographic proof, so any agent can run it again and get the same thing. Past what a model was compared with, it keeps going with the same equations and flags every step. It stops only at a physical limit, or where its equations, constants or tables no longer hold, and it names the reason.

https://mcp.zerotrustphysics.com/mcp Open to any agent.

What is in it

86 rows under one dataset seal:

523bb25a03556395f63c669f365d31afd26415b3c8455d79496fdcaeea16c947992cf7d208869fa9f2b6fc32b91f7bc8

• One row a trip: every single importing-interconnector loss of 500 MW or more that the record's scan found in 114 months of NESO's one-second frequency, 87 trips, less one (2 September 2021) dropped by the record's quality rule because its frequency moved 0.275 Hz in the ten seconds before the fall. Eight links: IFA 42 trips, IFA2 15, NSL 8, Nemo 7, ElecLink 6, Viking 4, BritNed 3, Greenlink 1.
• Each row is the request the model's own comparison ran, unchanged: the start frequency (49.763 to 50.126 Hz), the outturn inertia of the half hour (111 to 368 GVA s), the demand (15,597 to 43,106 MW), the six cleared volumes of Dynamic Containment, Moderation and Regulation, low and high, for the four-hour block, NESO's delivery times (1 s, and 10 s for Regulation) and the loss (500 to 1,402 MW) held for 60 s, in 9,600 steps of 1/160 s with a frame every 0.5 s.
• Outcomes: 86 completed, 0 refused. 10,320 recorded frames. The largest conservation residual over every step, 3.4e-11.

data/comparison.parquet (the default view) has one line a trip: its date and link, its inputs, and the depth of the fall, the steepest second and the frequency near 60 s, measured and modelled, with the errors, the lowest sample and the RMS error over the whole minute. data/traces.parquet has the 61 one-second samples of each trip, NESO's measured frequency beside the model's. data/rows.parquet and data/frames.parquet are the sealed rows and frames, written from the sealed files only after the whole dataset verified.

How close it comes to the measurements

Against NESO's measured one-second frequency, with the definitions of the model's own comparison: the depth of the fall is the start sample less the lowest of the next 60; the steepest second is the largest fall between consecutive samples in the first five (a rate of change of frequency over one second); the frequency near 60 s is the mean of samples 45 to 60; the trace error is the RMS error over samples 1 to 60.

Quantity, 86 trips	Fitted on all 86 (these rows)	Each trip left out of the fit in turn (the record)
Depth of the fall, RMS relative error	0.217 (mean +0.012)	0.224 (mean +0.014)
Steepest second, RMS relative error	0.236	0.237
Frequency near 60 s, RMS error	0.075 Hz	0.076 Hz
Whole trace, median of the RMS error	0.035 Hz	0.035 Hz

The first column is computed from these 86 rows and is, to the last bit, the figure of the model's own comparison: every row's four errors equal the record's. The second column is not computed from these rows: each trip's run there used five constants refitted on the other 85, and its figures are the record's, stated in the model's served description.

On 72 of the 86 trips the depth is within 20 percent of the measured depth, and on 43 within 10 percent; the median size of the error is 10.0 percent.

Period	Trips	Depth, RMS relative error	Mean
May 2017 to 2020	22	0.209	-0.014
2021 to 2023	30	0.190	-0.033
2024 to October 2025	23	0.173	-0.004
November 2025 to June 2026	11	0.343	+0.223

Gap: since November 2025 the model's falls are too deep, 22 percent on average; with each trip left out in turn, 0.359 RMS. The largest single gap is 21 March 2026, 1,002 MW from 50.110 Hz, a measured fall of 0.142 Hz, which the model makes 100 percent deeper (106 percent with the trip left out). Three more trips are past 0.6 either way: 4 March 2024 (960 MW, +69 percent), 26 March 2018 (634 MW, +63 percent), and 22 December 2023 (998 MW, 64 percent too shallow), where a second fall begins eight seconds after the first, a loss the open data does not name.

One row

Row 77 is the trip of 17 November 2025 at 06:19:28 (the time of the scan hit in NESO's frequency file): 1,007 MW lost on IFA from 49.948 Hz, inertia 177 GVA s, demand 26,774 MW, Dynamic Containment 1,145 MW low, Moderation 500, Regulation 480. The request, as run_trajectory takes it:

{"tool": "run_trajectory", "arguments": {"card": "grid_gb", "domain": "energy", "dt_s": 0.00625, "initial": {"dc_high_mw": 1080.0, "dc_low_mw": 1145.0, "demand_mw": 26774.0, "dm_high_mw": 510.0, "dm_low_mw": 500.0, "dr_high_mw": 480.0, "dr_low_mw": 480.0, "fast_delivery_s": 1.0, "frequency_hz": 49.948, "inertia_mw_s": 177000.0, "regulation_delivery_s": 10.0}, "record_every": 80, "segments": [{"load": {"loss_mw": 1007.0}, "steps": 9600}]}}

Outcome completed, after 9,600 steps. The model's lowest one-second sample is 49.754 Hz against 49.753 Hz measured; depth 0.194 Hz against 0.195 (-0.6 percent); near 60 s 49.787 Hz against 49.807; whole trace 0.020 Hz RMS. The steepest second is 0.092 Hz against 0.137 measured (-33 percent). Run seal:

21718275b8a3960e2bcb678744f92b3a5d45aceee01ee8556635b53072e0141fe1e777f8b2324941744ca3fff9f5a900

Verify it yourself

Every row can be verified without trusting this page.

# 1. the proof of the whole dataset, with no engine running: the seal of every row, its chain of frames and the root
genesis_seal verify-dataset sealed/head.json sealed/rows.jsonl sealed/frames.bin

# 2. one row's proof on its own: the row, its chain and its path to the dataset seal
genesis_seal row-proof sealed/head.json sealed/rows.jsonl 77 sealed/frames.bin > row77.json
genesis_seal verify-proof row77.json

# 3. the operator's signature on each signed head, against the public answer key
genesis_seal verify-signature signed/signed_000.head.json --public-key signed/answer_key_public.b64 --now 2026-10-22T03:13:21Z
genesis_seal verify-signature signed/signed_001.head.json --public-key signed/answer_key_public.b64 --now 2026-10-22T03:13:22Z

genesis_seal is built from the engine's public source. signed/answer_key_public.b64 is the public_key field of the operator's answer key, published at https://mcp.zerotrustphysics.com/answer-key. Each signature is good for 3,600 s from its issue; --now verifies it at its issue time.

4. Ask the engine to run the whole plan again: send verify_dataset the head's request with dataset_seal 523bb25a03556395f63c669f365d31afd26415b3c8455d79496fdcaeea16c947992cf7d208869fa9f2b6fc32b91f7bc8. The open address answers an inline plan of up to 10,000 frames, so send it there as two halves of 43 rows (run_dataset, store none, return inline) and compare each row's run seal.
5. Send any row's request to run_trajectory on energy/grid_gb and compare the run_seal.

When this dataset was made, all of these were run: verify_dataset by the stored seal: holds; verify_dataset's short form on a fresh engine, the plan run again: holds, the same seal; genesis_seal verify-dataset: holds, the receipt, the rows file and the frames file match the head; row 17's proof on its own: holds; 6 of 6 rows (0, 2, 9, 17, 33, 85) sent to run_trajectory again gave the same run seal.

Signed

The row seals and the dataset seal above were computed by the served engine build, run locally. The rows are not signed one by one. What is signed ML-DSA-87 is the dataset head the open address returned for the same plan: each signed head carries its merkle_root and rows_sha384 over its rows' run seals, and those run seals equal the stored rows', row for row.

The same plan was sent to the open address in two halves of 43 rows, and each answer came back with its head signed ML-DSA-87 by the operator's answer key (key ID ca06027a972c81c8630fed9f4ea2d4e91596b9434a88a03b7eb325b24e40645b7090cb705927745e73ee8b272eb1be38), with an expiry of 3,600 s and a nonce. SHA-384 and ML-DSA-87 are the CNSA 2.0 algorithms.

Signed head	Rows	Issued (UTC)	Merkle root (first 16)	Signature	Run seals equal to the stored rows'
signed/signed_000.head.json	0 to 42	2026-10-22T03:13:21Z	29247d013f8e8fd1…	holds, receipt matches	43 of 43
signed/signed_001.head.json	43 to 85	2026-10-22T03:13:22Z	a414a2c61edef0ac…	holds, receipt matches	43 of 43

A signed half has its own merkle root and seal, because it is a plan of 43 rows; it attests this dataset because every one of its run seals equals the stored row's. The full answers are signed/signed_000.json and signed/signed_001.json; signed/signed_summary.json lists every run seal of each.

Files

File	What it holds	SHA-256
sealed/head.json	the dataset's head: its request, counts, root and seal	1ac8f760518db9b0…
sealed/rows.jsonl	every row as the engine stored it, with its request, outcome, final state and run seal	1490cda4656570ad…
sealed/frames.bin	every recorded frame, 64 bytes each	336721e94f7fabff…
data/comparison.parquet	one line a trip: inputs, measured and modelled depth, steepest second, near 60 s, lowest, trace error	3bbeab5194fefb0f…
data/comparison.csv	the same as CSV	7c30590c48f3424e…
data/traces.parquet	61 one-second samples a trip, measured and modelled	02098946f8aeca61…
data/traces.csv	the same as CSV	b4a02d51653285b5…
data/rows.parquet	the sealed rows	d04da8383feab730…
data/frames.parquet	the sealed frames	6f315251138bdd81…
signed/signed_000.json, signed_001.json	the two signed answers from the open address, as they came	8d7773f01674b267…, ba1876075cfa234d…
signed/signed_000.head.json, signed_001.head.json	each signed head's own bytes	0394e986aeed1156…, 975279cf1e3f02ea…
signed/answer_key_public.b64	the operator's public answer key, the public_key field published at https://mcp.zerotrustphysics.com/answer-key	507ef857d3d95be7…
signed/signed_summary.json	per half: seals, signature result, every run seal	884fd04d3f2426cc…
source/trips.json	the input table the model's comparison ran, unchanged: 87 trips with their loss, demand, inertia, volumes and measured 61-second trace	c6b19d5a516b0318…

Where the numbers come from

• System frequency, one value per second: NESO Data Portal, System Frequency (https://www.neso.energy/data-portal/system-frequency-data), monthly files April 2017 to September 2026. Licence: National Energy SO Open Data Licence v1.0.
• System inertia, half-hourly outturn: NESO Data Portal, System Inertia (https://www.neso.energy/data-portal/system-inertia). Licence: National Energy SO Open Data Licence v1.0.
• Dynamic Containment, Moderation and Regulation volumes: NESO Data Portal, Dynamic Containment Masterdata; DC, DR & DM Results Summary 2021-2023; EAC Response-Reserve Results Summary from November 2023. Licence: National Energy SO Open Data Licence v1.0. The service curves and delivery times are NESO's published service terms.
• The loss and the demand: Elexon BMRS, generation by fuel type (FUELINST, https://data.elexon.co.uk/bmrs/api/v1/datasets/FUELINST): the interconnector's flow before and after the trip, and the sum of the categories just before it. Licence: Elexon's licence to use BMRS open data.
• The input table: source/trips.json (SHA-256 c6b19d5a516b03181b29545e4f31fd8f110fea7a5e5d7f6674d53af055a69771), the record's table of the 87 trips, used unchanged; each row's request is equal, field for field, to the request the model's own comparison ran for that trip.
• The model: five constants fitted on all 86 trips at once, least squares on the frequency sample by sample over 60 s: the plant's response, 2,059 MW/Hz below 50 Hz and a further 1,987 MW/Hz below 49.8 Hz through a lag of 3.27 s, and the shares delivered of Dynamic Moderation, 0.913, and Dynamic Regulation, 0. Not fitted: Dynamic Containment delivers as contracted; the service curves and delivery times are NESO's; demand's relief is 2.5 percent per hertz.

Supported by National Energy SO Open Data.

Contains BMRS data © Elexon Limited copyright and database right 2026.

What the rows do not say

• The inertia is NESO's outturn estimate, which leaves out the Stability Pathfinder units and grid-forming batteries. The volumes are the cleared auction volumes of the four-hour block, not the delivery metered on the day. Before October 2020 no dynamic service is in the data, and the response held then is the plant term.
• One frequency for the whole area: no network, no generator that trips as a consequence and no reserve slower than these services. Past 60 s, under 49.55 Hz, or outside the compared range of start frequency, inertia, demand, volumes or loss, the model keeps going with the same equations and flags every step with the limit, the value and the extent past it. It refuses a loss larger than the demand, and a run ends at 48.8 Hz, where low-frequency demand disconnection begins, naming the reason.
• The steepest second reads the first samples of a fall at one value a second; since November 2025 it is 0.44 RMS on these rows.
• These 86 rows carry the constants fitted on all 86: they are the fit, not trips outside it. The left-out figures above are the record's.
• The measured traces are NESO's one-second data as published, with no filtering. Measured minus model is the comparison; neither is a claim about the cause of a trip.

Benchmark tasks

Three questions on these models are tasks in Environments since version 0.5 (tag v0.5.0, 22 October 2026), each scored by a program that sends the agent's request to the verifier again, with no judge. Their inputs are not in this dataset: no row here shares a task's start state and loads, or differs from it in one input only, a limit task's open input set aside.

• figure-gb-lowest-frequency-1320mw (figure): Great Britain's grid runs at 50.03 Hz with 152,000 MW s of stored kinetic energy and 24,700 MW of demand, holding 950, 420 and 310 MW of the low-frequency services Dynamic Containment, Dynamic Moderation and Dynamic Regulation (and the same volumes of the high-frequency ones), fast delivery 1 s and regulation delivery 10 s. A 1,320 MW block of generation is lost at once. What is the lowest frequency, in hertz, that the grid falls to in the 60 s that follow?
• limit-gb-largest-loss-before-demand-disconnection (limit): Great Britain's grid runs at 49.96 Hz with 109,000 MW s of stored kinetic energy and 18,900 MW of demand, holding 640, 300 and 220 MW of the low-frequency services Dynamic Containment, Dynamic Moderation and Dynamic Regulation (and the same volumes of the high-frequency ones), fast delivery 1 s and regulation delivery 10 s. What is the largest block of generation, in megawatts, that can be lost at once without the frequency falling far enough, in the next 60 s, that demand is disconnected?
• figure-gb-seconds-to-49-7-hz (figure): Great Britain's grid runs at 50.01 Hz with 167,000 MW s of stored kinetic energy and 27,400 MW of demand, holding 1,050, 470 and 330 MW of the low-frequency services Dynamic Containment, Dynamic Moderation and Dynamic Regulation (and the same volumes of the high-frequency ones), fast delivery 1 s and regulation delivery 10 s. A 1,900 MW block of generation is lost at once. How many seconds after the loss does the frequency first fall below 49.7 Hz?

In Environments, python3 prompt.py <id> prints the text an agent is given, and python3 score.py <replies> --engine <verifier> scores the replies.

Cite

The paper (v2.0): https://doi.org/10.5281/zenodo.23517223. The open address: https://mcp.zerotrustphysics.com/mcp. The public answer key: https://mcp.zerotrustphysics.com/answer-key.

Made by Super Intelligence, on behalf of Protocol Company. Data: CC BY 4.0; the NESO and Elexon data within it under their own licences, attributed above.