spiderpilot89/zero-trust-physics-edges

Edges

Where does a model of a real thing stop, and what do the two answers either side of that place look like?

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; when a question goes past what a model was tested for, the answer stops there and says why.

The edge of a model is the most useful thing to know about it. So this dataset is made only of edges. Each row is a pair of questions that differ in one number. One is the last question that stands, kept with its whole answer and its cryptographic proof. The other is the nearest question past it, kept with what came back: the name of the stop and the model's own sentence saying why.

How near is nearest? In 895 of the 936 rows the two numbers are neighbours: no 64-bit number lies between them. In 41 more the input is a count of steps, and the two counts differ by one.

Every row was made by asking the address below, one question at a time, the way any agent would.

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

One row

The SPERT III reactor is where this began: a government report from 1969 became a working model of a nuclear reactor in under six hours. The report tested the reactor between two reactor periods, and the model takes the reactivity between the two values those periods give. Ask for the largest of them, 1.2102315121058187 dollars, and the run stands: 214 steps to the first maximum of power, a peak of 295.6 MW, and a proof,

6af631aa1eaafb5634cf7103d203fc135a04e07188de92a9b8e3cc6bcc0cd496

Ask for the next 64-bit number up, 1.210231512105819, and this comes back in its place:

Physical Failure: 'reactivity_inserted_dollar' 1.210231512105819 dollar is above 1.2102315121058187 dollar, the inclusive maximum of the reactor_spert3 card (ReactivityOutsideCard). That state is not sealed.

Those two questions, the full answer to the first and that sentence are row edge-f7a90f14fe70. The two inputs are kept in decimals and as their exact bits, 3ff35d1bb7d15475 and 3ff35d1bb7d15476.

Three kinds of edge

edge_kind	What the pair straddles	Rows
declared_limit	A limit the model states: the end of the range it was compared with measurement over, a limit of the specimen itself, or the largest and smallest time step it takes. Inside, the run stands with its proof. Outside, the answer stops and names the limit.	850
declared_end	The place where the event the model describes is over: a reactor's first maximum of power, a capsule's parachute deployment, a mission's last day, sixty seconds after a power link drops. Inside, the run completes. Outside, it ends there, and keeps its proof up to the last step that landed.	18
outcome_flip	Nothing stops. Both runs stand with their proofs, and something about the outcome is different: a flag of the final state has turned, or a quantity read from the run has crossed a level for the first time.	68

When a question is stopped before any computing, the answer is a sentence and no proof is issued for it. When a run starts and is stopped part way, which happens in 114 rows, the answer keeps its proof up to the last step that landed. When a solve starts and cannot return its result, in 43 rows, the answer says so and carries the proof of that outcome. Each row holds exactly what came back.

The seven things first

The same seven real things as on zerotrustphysics.com, and what their edges turned out to be.

• The SPERT III reactor. The model takes one step in reactivity between 0.7686 and 1.2102 dollars, the values that the longest and the shortest reactor period of the 1969 tests give. One 64-bit number past either end and the answer stops, naming ReactivityOutsideCard. Inside the range there is a place where the burst first reaches 100 MW: at 1.1147492996633026 dollars the peak is 99.9999999999997 MW, and at the next number up it is 100.00000000000001. And a burst is one burst: change the reactivity by one unit in the last place part way through, and the run stops there (ReactivityChangedDuringBurst).
• NASA's Stardust capsule. The worked run follows the capsule for twenty seconds from entry interface. Let it run on, and the fall passes 5 MW of heating on each square metre at step 809, passes 10 g at step 849, and is over at step 2,779, 138.95 seconds in, at the drogue-deployment condition the radar recorded. The next step is not taken. Start the capsule 1.73 m/s faster sideways than the flight did and the run stands; at 1.74 it is outside the speeds the compared entry passed through (AirSpeedOutsideCard).
• The James Webb Space Telescope's guide star. The model is told where the guide star was measured, frame by frame, and says where it will sit next. Move one measured position by a little over a thousandth of an arcsecond one way, or two thousandths the other, and the run still stands. One 64-bit number further, the star is more than four standard deviations from where it was expected, and the run stops with the sentence that a new one may start no earlier than 30 seconds after that frame (InnovationBeyondFourSigma).
• Great Britain's power grid. The model covers the sixty seconds after a power link drops, for a loss between 693 and 1,402 MW. In its worked stumble, a loss of 996 MW, the frequency dips to 49.67 Hz. Make the loss smaller and there is a place where the dip no longer reaches 49.7 Hz: at 869.2003553078027 MW the lowest frequency is 49.699999999999996 Hz, and at the next number down it is 49.7. The minute itself has an edge too: step 2,889 lands, and step 2,890 is not taken.
• LAGEOS. Told where the satellite was at one moment, the model carries it forward seven days and no further: 60,480 steps of ten seconds land, and the 60,481st stops the run, with its proof kept to the last step that landed (HorizonOutsideCard). The start has to be the satellite's own orbit. Move the starting position 86 metres along one axis and the run stands; at 87 the orbit is in neither of the two bands the model was compared in (OrbitOutsideCard).
• A tyre on ice. The rig measured slip between 0.05 and 0.20, driving or braking, and that is what the model takes: 0.05 stands and 0.049999999999999996 stops (SlipOutsideCard). At the worked load the tyre pulls with more than 1,200 N only while the slip is under 0.0682. The worked run sits in the very corner of what was measured, with the slip held 5 seconds before it starts and 21 by its end, so one more step is past the measurements (HoldBeyondMeasured).
• The power source on NASA's Curiosity rover. The model follows the generator's daily output from landing to mission day 4,087. Run it a day at a time and the output first reads under 110 W on day 320 and under 100 W on day 1,115. Day 4,087 lands, and day 4,088 is where the model ends (MissionTimeOutsideCard).

The rest follow: batteries, a fingertip on glass, a superalloy creeping, capsules, tokamaks, seawater, argon, a propeller, a heat shield, the Solar System. coverage.md lists every model with each stop its description names, and whether a row here brackets it.

What you or your agent can do with it

Know where a model stops before you lean on it. Filter by model and read input, inside_value, stop_name and stop_sentence. That is the range the model answers over from its own worked example, in the model's own words.

from datasets import load_dataset

cards = load_dataset("spiderpilot89/zero-trust-physics-edges", "cards", split="train")
grid = cards.filter(lambda r: r["model"] == "energy/grid_gb")
for r in grid:
    print(r["input"], r["inside_value"], r["outside_value"], r["stop_name"])

Recompute every proof, on your own machine. scripts/verify_rows.py uses only the Python standard library and sends nothing anywhere. It rebuilds each proof from the answer text a row holds (1179 proofs in all) and holds every column of the row to its two requests and two answers.

python3 scripts/verify_rows.py

Ask the two questions again. scripts/replay.py sends a row's pair to the address and compares what returns with what the row holds, byte for byte.

python3 scripts/replay.py edge-f7a90f14fe70
python3 scripts/replay.py --model energy/grid_gb

Give an agent hard cases. Each pair is a question with a known answer on each side of a line as thin as a line can be. An agent that plans a run, sizes a part or reads a model's description can be handed the two inputs and asked which one stands, and why.

Then ask your own. Hand it to your agent:

Connect to https://mcp.zerotrustphysics.com/mcp and read what is there. Then tell me what it is, and why it might matter to me.

What is in a row

Two configs, alike in every column but a few: cards are models stepped through time and carry domain, card, inside_steps_landed and outside_steps_landed; solvers are models of a whole field solved until they settle and carry solver, inside_iterations and outside_iterations. data/cards.jsonl and data/solvers.jsonl are the source; the Parquet files beside them hold the same rows for the viewer.

Column	What it holds
id	A name for the row, made from its two requests.
model	The model as the address lists it: domain/card for a card, the solver's name for a solver.
tool	The tool both requests were sent to: run_trajectory or solve.
specimen	The model's own one-line account of the real thing it describes.
base	The worked example the pair was reached from. Each model's description carries its own worked example; a row changes one number of it. Where it says so, the example's one segment was first cut into two of the same load, so that a load can change part way, or a readout was added.
input, input_unit	The one input that differs between the two requests, as a path into the request, and its unit.
whole_number_input	True when the input is a count, such as steps.
base_value	The value that input has in the worked example.
edge_kind	declared_limit, declared_end or outcome_flip.
direction	up when the outside value is the larger, down when it is the smaller.
inside_value, outside_value	The two values, in decimals.
inside_value_hex, outside_value_hex	The same two values as the 16 hex digits of their 64 bits: the exact numbers.
width	The outside value minus the inside value, as a size.
adjacent_64_bit_numbers	True when no 64-bit number lies between the two values.
stop_name	The name the answer gives the stop or the end. Empty for a flip, and for the few stops whose sentence carries no name.
stop_sentence	The model's own sentence saying why.
stop_where	before the run or before the solve (a sentence, no proof), part way (a run, with its proof to the last step that landed), in the solve (a solve that could not return its result, with the proof of that outcome) or declared end.
flip	For a flip: what differs between the two outcomes, as JSON, inside value first.
inside_outcome, outside_outcome	The outcome each answer carries.
proof_kind	run_seal or solution_seal.
inside_proof, outside_proof	The cryptographic proof of each answer. outside_proof is empty when the question was stopped before any computing.
inside_receipt, outside_receipt	The receipt of each answer: a hash over every byte of its text.
engine_version	The engine_version string the inside answer carries.
address, asked_utc, date	Where the pair was asked, and when, in UTC.
inside_request, outside_request	The two requests as they were sent, as JSON text.
inside_answer, outside_answer	The two answers as they came back, byte for byte.
outside_answer_is_error	True when the outside answer came back as a sentence and not as a result.

How it was made

For each model, start from the worked example its own description gives. Change one number, and see whether the kind of answer changes: stands, stops, ends, or a flag turns. Where it does, halve the gap between the two values until no 64-bit number is left between them. The searching was done on a second copy of the same engine, so that the open address was asked only the final pair of each edge, 936 pairs, one question at a time. What the open address returned is the row.

A stop that belongs to the form of a request and not to a model (a number that is not a finite normal 64-bit number, a count of steps below one or above a million, more than ten thousand recorded states) is left out.

All rows were asked on 13 October 2026 (UTC). A model can change after that: a range widens, a limit moves, a new model arrives. Each row carries its date and the engine_version its answer stated, scripts/replay.py shows in one line whether a pair still falls the same way, and each model's own description at the address is always the current word.

The people whose measurements these models were compared with

A model has an edge because someone measured something real up to that point and no further. These are the people who did the measuring, or the published calculation a model was compared with. None of their data is in this dataset; full references are in CREDITS.md.

Model	The real thing	Measured or calculated by
energy/reactor_spert3	The SPERT III E-core, a small oxide-fuelled pressurized-water test reactor at the National Reactor Testing Station, Idaho	McCardell, R. K. and others (1969); Olson, Arne P. (2013)
aerospace/capsule_stardust	NASA's Stardust sample return capsule, Earth entry of 15 January 2006, from entry interface to parachute opening	Desai, Prasun N. and others (2008); Desai, Prasun N. and others (2007)
aerospace/los_jwst_fine_guide	The James Webb Space Telescope holding a guide star in fine guide, commissioning programme 1163, April to June 2022	Perrin, Marshall (principal investigator) and others (2022)
energy/grid_gb	The Great Britain 50 Hz transmission system in the 60 seconds after an importing interconnector trips	National Energy System Operator (2026); National Grid Electricity System Operator (2026); Elexon Limited (2026)
aerospace/satellite_lageos1	The laser-ranging satellite LAGEOS-1 in free flight (the same card, unchanged, also run on LAGEOS-2), weeks of August to September 2026	International Laser Ranging Service (ILRS) (1976); Pearlman, Michael R. and others (2019)
materials/tire_eve_r19_ice	One 235/55 R19 all-season SUV tyre of the European project EVE, rolled on prepared ice on the Terramechanics Rig at Virginia Tech, April 2016	Schleinin, Dmitrij and others (2017); Schleinin, Dmitrij and others (2017); Schleinin, Dmitrij and others (2017)
energy/generator_mmrtg_f1	The Multi-Mission Radioisotope Thermoelectric Generator, flight unit F1, on NASA's Curiosity rover	Whiting, Christofer and others (2024); Whiting, Christofer E. and others (2023); Whiting, Christofer E. and others (2024)
ablator_slab	PICA, the heat-shield material of NASA's Stardust sample-return capsule	Tran, Huy K. and others (1997); Milos, F. S. and others (2009); Monk, Joshua D. and others (2022); Lachaud, Jean R. and others (2011); Lachaud, Jean and others (2012); National Aeronautics and Space Administration (2023); and 6 more in CREDITS.md
aerospace/capsule_insight	NASA's InSight entry capsule, entry at Mars of 26 November 2018, from 126.5 km altitude to parachute deployment	Karatekin, O. and others (2020); Karlgaard, Christopher D. and others (2020); Maddock, Robert W. and others (2020); McCleese, Daniel J. and others (2006); LMD and others
aerospace/crater_ryugu_sci	The crater made on asteroid 162173 Ryugu on 5 April 2019 by the Small Carry-on Impactor of JAXA's Hayabusa2	Schmidt, R. M. (1980); Gault, D. E. and others (1977); Arakawa, M. and others (2020); JAXA Hayabusa2 Project and others (2020)
aerospace/solar_system	The Solar System (Sun, eight planets, Moon, Ceres, Pallas, Vesta) carried forward from its state of 1999-01-01	JPL Solar System Dynamics Group (data maintained by Ryan Park, JPL/Caltech) (2026); Konopliv, Alex S. and others (2011); Kuchynka, P. and others (2012); Park, Ryan S. and others (2017); Hees, A. and others (2014)
argon_pair_gas	Argon gas, 99.999 percent pure, at low density, as measured in two laboratories	Goodwin, Anthony R. H. and others (2009); Vogel, E. (2010); Riccardi, Demian and others (2021); Klein, Max and others (1974); Klein, Max and others (1968)
atom_dft	One spherical neutral atom, hydrogen to argon, in the spin-unpolarized non-relativistic local-density approximation, as NIST's reference tables compute it	Kotochigova, Svetlana and others (1997); Kotochigova, Svetlana and others (1997)
culture_trichoderma_ambr250	Fed-batch cultures of the marine fungus Trichoderma sp. VTT D-221704 making mucic acid in Sartorius ambr 250 vessels at VTT, Finland	Tamminen, Anu and others (2022); Tamminen, Anu and others (2022)
energy/accumulator_sbo450	One HYDAC SBO450-0.6 hydraulic diaphragm accumulator charged with nitrogen, on the accumulator test rig of the Chair of Fluid Systems, Technische Universität Darmstadt	Rexer, Manuel (2024); Lefemmine, Oscar (2025); Rexer, Manuel (2024); Rexer, Manuel (2024); Rexer, Manuel (2024); Lefemmine, Oscar (2024); and 2 more in CREDITS.md
energy/cell_18650pf	One new Panasonic NCR18650PF lithium-ion cell tested by Phillip Kollmeyer at the University of Wisconsin-Madison, discharged at 1C at 25 degrees Celsius	Kollmeyer, Phillip (2018)
energy/cell_2170_m80	One Tesla Model 3 2170 lithium-ion cell, cell m80 of the McMaster University dataset, in six blocks of tests from -20 to 40 degrees Celsius	Kollmeyer, Phillip J. and others (2023)
energy/cell_2170_model3	Three Tesla Model 3 2170 lithium-ion cells of the McMaster University dataset (m80, m1000 and m448-N), from -10 to 40 degrees Celsius	Kollmeyer, Phillip J. and others (2023)
energy/reactor_spert3_ccore	The SPERT III reactor with the C-19/52 highly enriched plate core, National Reactor Testing Station, Idaho	Toole, C. R. (1964)
energy/reactor_spert3_fuel	The SPERT III E-core, a small oxide-fuelled pressurized-water test reactor at the National Reactor Testing Station, Idaho	McCardell, R. K. and others (1969); Olson, Arne P. (2013)
energy/reactor_triga_jsi	The TRIGA Mark II research reactor of the Jožef Stefan Institute, Ljubljana, in pulse operation	Švajger, Ingrid and others
energy/tokamak_jet_dte1	The JET tokamak at Culham in ELMy H-mode pulses of 1996 to 1997, in deuterium and in deuterium-tritium	The ITER 1D Modelling Working Group / ITPA profile database working group and others (2008); The ITER 1D Modelling Working Group and others (2000); Budny, R. V. and others (2000); Sharapov, S. E. and others (1999); Horton, L. D. and others (1999); Bateman, Glenn and others (1999)
fusion_pb11_lhd	The boron-seeded hydrogen plasma of the Large Helical Device (National Institute for Fusion Science, Toki), heated by hydrogen neutral beams	Ogawa, K. and others (2024); Magee, R. M. and others (2023); Ogawa, K. and others (2022); Sikora, M. H. and others (2016); Xu, Y. and others (2013)
fusion_pb11_thermal	No real object	Putvinski, S. V. and others (2019); Ochs, Ian E. and others (2022); Sikora, M. H. and others (2016); Xu, Y. and others (2013)
journal_bearing_sheffield	The plain full journal bearing of the bespoke test platform of the Leonardo Centre for Tribology, University of Sheffield, fifty steady runs at 65 °C	Beamish, Scott and others (2020); Beamish, S. and others (2020); Raimondi, A. A. and others (1958); Bell, Ian (2012)
materials/creep_cmsx6	Twelve test pieces from one cast of the single-crystal nickel-base superalloy CMSX-6	Ávila Calderón, Luis A. and others (2026); Ávila Calderón, Luis and others (2026)
materials/fingerpad_p02_glass	The right index finger pad of participant P02 pressed by a robot-held glass plate and twisted through 80 degrees, Université catholique de Louvain	du Bois de Dunilac, Sophie and others (2023); du Bois de Dunilac, Sophie and others (2023)
modulus_bam_s355	Four resonance bars cut from one hot-rolled sheet of structural steel S355J2+N and rung at 24 °C at the Bundesanstalt für Materialforschung und -prüfung (BAM), Berlin	Schilling, Markus and others (2022)
pipe_flow_superpipe	The Princeton Superpipe of A. J. Smits' group, Princeton University	Zagarola, Mark V. and others (1998); Zagarola, Mark V. and others (1998); McKeon, B. J. and others (2004); McKeon, B. J. and others (2004); Hultmark, M. and others (2012); Hultmark, M. and others (2013); and 1 more in CREDITS.md
propeller_apc_10x7e	The APC Thin Electric 10x7 model-aircraft propeller as measured in the University of Illinois low-speed wind tunnel	Brandt, J.B. and others (2026); Brandt, John B. and others (2011); Brandt, John Burton (2005)
seawater_column	The water columns of the WOCE Indian Ocean expedition of the research vessel Knorr, cruise 145, legs 5 to 15, December 1994 to January 1996	CCHDO Hydrographic Data Office (2023); McCartney, Michael (2026); Gordon, Arnold L. (2026); Talley, Lynne (2026); Nowlin, Worth (2026); Toole, John (2026); and 6 more in CREDITS.md
signal/cantilever_eth	One aluminium strip, 326 mm x 40 mm x 1 mm, clamped at one end at ETH Zürich and recorded ringing down in room air	Lai, Zhilu and others (2026)
signal/link_gothenburg_10010	One commercial microwave link in Gothenburg, Sweden (OpenMRG link 10010, sublink 587, 32.445 GHz, 1.25841 km) and the rain gauge at Barlastplatsen beside it	Andersson, Jafet C.M. and others (2022); Andersson, Jafet C. M. and others (2022)
step_flow_armaly	The air channel with a backward-facing step of Armaly, Durst, Pereira and Schönung (1983), measured by laser-Doppler anemometry	Armaly, B. F. and others (1983); Steinthorsson, Erlendur and others (1993); Erturk, Ercan (2008)
tokamak_dt_beam_target	The deuterium-tritium pulses of JET (Culham, 1997) and TFTR (Princeton, 1993 to 1997) as the International Multi-Tokamak Confinement Profile Database holds them	The ITER 1D Modelling Working Group / ITPA profile database working group and others (2008); The ITER 1D Modelling Working Group and others (2000); Budny, R. V. and others (2000); Budny, R. V. (2002); Sharapov, S. E. and others (1999); Horton, L. D. and others (1999); and 5 more in CREDITS.md

Licence and citation

The data, which is everything under data/ with coverage.md and CREDITS.md, is under CC BY 4.0. The scripts under scripts/ are under Apache-2.0, and the licence text is beside them.

To cite the dataset:

@misc{kruze2026edges,
  author       = {Kruze, John},
  title        = {Edges: the two runs either side of where a physical model stops},
  year         = {2026},
  month        = oct,
  howpublished = {Dataset},
  publisher    = {Hugging Face},
  organization = {Protocol Company},
  url          = {https://huggingface.co/datasets/spiderpilot89/zero-trust-physics-edges},
  note         = {Made with Super Intelligence by asking https://mcp.zerotrustphysics.com/mcp. Licence: CC BY 4.0}
}

The paper behind the verifier: Kruze, J., with Super Intelligence (2026). Zero-Trust Physics: a verifier for the physical answers of Super Intelligence (version 1.0). Protocol Company. Zenodo. https://doi.org/10.5281/zenodo.23305697

@misc{kruze2026zerotrustphysics,
  author       = {Kruze, John},
  title        = {Zero-Trust Physics: a verifier for the physical answers of Super Intelligence},
  year         = {2026},
  month        = oct,
  version      = {1.0},
  howpublished = {Preprint},
  publisher    = {Zenodo},
  organization = {Protocol Company},
  doi          = {10.5281/zenodo.23305697},
  url          = {https://doi.org/10.5281/zenodo.23305697},
  note         = {Byline on the paper: John Kruze, with Super Intelligence}
}

Where the rest is

• The address: https://mcp.zerotrustphysics.com/mcp. Open to any agent.
• The site: https://zerotrustphysics.com
• The paper: https://doi.org/10.5281/zenodo.23305697
• The rules of the proof, and scripts that recompute it: https://github.com/johnkruze/zero-trust-physics
• John Kruze: kruze@zerotrustphysics.com

Hand it to your agent.

Written by Super Intelligence, on behalf of Protocol Company. Rows as returned by the address on 13 October 2026 (UTC); the verifier's own descriptions are always the current word.