A physics verifier for Super Intelligence (SI): agents ask it about real hardware, and every answer comes with a cryptographic proof any agent can recompute. A model stops at a mathematical or physical limit and names it.
A deterministic physics runtime with a proof on every answer, for agents and RL rollouts, with calibrated Harbor and HUD task packs for RL environment vendors and post-training teams. It runs inside --network none rollout containers, with zero internet egress from the sandbox; your host harness calls the model. An agent asks about a real physical thing over the Model Context Protocol, the engine computes the answer with deterministic physics, holds a conservation law at every step, and returns it with a SHA-384 proof, so a reward is computed again and not judged.
run_trajectory calls (about 5,500 steps each; the shorter half 4.5 ms), on one Apple silicon CPU core.--network none: 0 differ.https://mcp.zerotrustphysics.com/mcpOpen to any agent, with no key, 1 physics task at a time per address: a tasting tap, not a rollout path. Engine 0.15.3 serves 36 cards, 25 field solvers and 18 tools: 61 models and solvers, one for each real object; 46 of them are compared with measurement, 3 held to a reference calculation and 12 are company specimens. The reference source on GitHub is frozen at v0.15.2; models added since are served, not published. The paper and the sealed datasets below are free. The free sample of Group 1 is published; 5 sample tasks of Groups 2 to 5 are on request as each group is built: kruze@zerotrustphysics.com.
Try the tasks: the free Group 1 sample.
The Domain Pack, $9,500, non-exclusive: 30 withheld multi-step physics tasks and 100,000 sealed runs from one domain group, in Harbor, HUD, Inspect and native MCP. Group 1 (Batteries, Thermal Runaway & Grid Frequency) is on self-serve checkout and ships with the compiled runtime for its 9 models. Groups 2 to 5 are by contact, each air-gapped once its own binary passes. All-Five Bundle $39,500, Saturation Refresh $9,500/month, Exclusive $38,000 or $152,000. Keyed multi-permit hosted HTTP on request.
One call, from any shell. It asks the field solver of the Princeton Superpipe for its worked question, turbulent air at Reynolds number one million, and the answer comes back with its SHA-384 solution_seal and the operator's ML-DSA-87 signature:
curl -sS https://mcp.zerotrustphysics.com/mcp -H 'Content-Type: application/json' -H 'Accept: application/json, text/event-stream' -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"solve","arguments":{"solver":"pipe_flow_superpipe","example":"example_solve"}}}'Send the same call again and the seal is the same. An agent gives its MCP client the address and starts with find_model, or list_cards and list_solvers.
key_id ca06027a…eb1be38.UndervoltageCutoff at the 36.0 V cutoff, and the run is sealed to the last step that held.RetiredModel, with the ID that serves the object now: reactor_spert3_twin is served as reactor_spert3.| What it holds | |
|---|---|
| Free sample, Group 1 | 5 calibrated multi-step tasks from Batteries, Thermal Runaway & Grid Frequency, with their pass@1, pass@8 and cheat-suite results. Evaluation only. |
| Environments | Version 0.5 (tag v0.5.0): 170 tasks, each scored by a program that sends the agent's own request to the verifier again, with no judge; the reference agent scores 170 of 170 on engine 0.15.2. For evals and RL environments. The version 0.3 scores (153 tasks, engine 0.14.0) and their traces are on main: Gemini 4 Argon 152 of 153, Gemini 3.8 Flash 136, GPT-OSS 120B 36. |
| SPERT III E-core | reactor_spert3 on the 70 tests of IDO-17281 (1969) and a sweep around them: 233 sealed rows, engine 0.15.2, each row signed by the open address. Compared with each test left out of the fit in turn, at the median test: peak power within 6.8 percent, energy to the peak within 8.6. |
| SPERT III C-core | reactor_spert3_ccore on the 111 transients of IDO-16918 and a sweep: 196 sealed rows, engine 0.15.2, each row signed by the open address. Peak power within 10.4 percent RMS on each tenth left out of the fit in turn. |
| MSRE fuel pump | reactor_msre_pump, the Molten-Salt Reactor Experiment's pump start-up and coast-down of ORNL-4233 and a sweep: 45 sealed rows, engine 0.15.2, each row signed by the open address. Within 6.5 pcm RMS on the 73 coast-down rod readings not in the fit. |
| TRIGA element 6753 | fuel_rod_triga_jsi, the instrumented fuel element of the JSI TRIGA Mark II reactor through a pulse: 141 sealed solves, 16 of them chained from a sealed run of the core, engine 0.15.2, each row signed by the open address. |
| Lithium-ion cells | Three models of real cells, 252 sealed runs: every published discharge and drive cycle of three Tesla Model 3 2170 cells from -20 to 40 degC (234), eleven 1C discharges of a Panasonic NCR18650PF cell, and an 18650 LiCoO2 cell heated in an oven to runaway, beside the laboratories' measurements. |
| GB grid frequency | The 86 single importing-interconnector trips NESO's one-second frequency recorded from May 2017 to June 2026, each computed again for its first 60 s: the nadir, the rate of change of frequency and where it settles, beside the measured frequency. |
| Entry and heat shield | The Stardust and InSight entries (15 and 21 sealed rows) and the Stardust PICA shield's wear chained from the entry, beside what the flights and the recovered shield recorded. |
| Orbits and pointing | LAGEOS-1 and LAGEOS-2 seven-day orbit predictions against the ILRS precise orbits, and the James Webb Space Telescope's momentum build-up and fine-guidance pointing against its own telemetry. |
| Agent test | Fifteen tasks, every exchange of 645 agent runs with their scores, and a small program that puts your own agent to the same tasks and scores it with no model in the loop. Does an agent ask the verifier, or answer from memory? When the model stops at its limit, does the agent stop too? |
| Edges | 936 rows: for every model served at the time, the last request that stands and the nearest one past it, each with its proof. |
| Companions | Beside each open dataset a model was compared with: its predictions on the runs not in the fit, each with its proof. |
Try the verifier in a browser: the Space. Read what it measured: the articles, dated, by Protocol Company. Everything is in one place in the collection.
v0.15.2. Build it and get that release's proofs. The open address serves 0.15.3, whose source is private.verify_signature.py and dilithium-py 1.5.1.com.zerotrustphysics/verifier, version 0.15.3.Every offer on one page: https://zerotrustphysics.com/offers/
Protocol Company · John Kruze, CEO · kruze@zerotrustphysics.com
Protocol Company.