Zero-Trust Physics ยท Articles
17 October 2026
The outsider's test: can another AI rebuild our physics from our words alone?
Zero-Trust Physics says anyone can check its answers. Every answer already carries a cryptographic proof that it was computed as sealed. A proof doesn't show that the physics is right, though, or that an outsider could rebuild it. So today we gave a frontier model from another AI lab four of our models and nothing but their published descriptions. We measured how far its first try landed, before the engine had told it a single answer.
The rules of the room
- Words only. It got each model's public description, the same text any agent gets from the open verifier, and nothing else: no source code, no internet, no earlier work. We audited every file it opened and every command it ran, and it kept to its folder completely.
- The first try is locked. Before the engine computed anything, the model had to write down its equations, its test questions and its predicted answers, and seal them with SHA-256 hashes. The only door to the engine was a small program that refused any physics before that lock. After it, the program refused every run if a locked file had changed by a byte.
- Every source is labelled. Each constant is either a verbatim quote from the description, checked against the exact place it was quoted from by a script, or marked as coming from the model's own memory, or as its own choice.
- Agreement means 1 part in a billion. An output counts as agreeing when it matches the engine to a relative difference of 10-9 or better. Within 10-3 is close; anything else differs.
The first try
| Model | What the description gives | Agree | Close | Differ |
|---|---|---|---|---|
| A 32 GHz microwave link across Gothenburg in rain | every constant and the method | 120 | 0 | 0 |
| A fed-batch fungal culture (a solver) | every constant and the method | 106 | 0 | 2 |
| Thermal runaway of an 18650 lithium-ion cell | cites a 2007 paper for its constants | 59 | 12 | 31 |
| The James Webb Space Telescope's momentum under sunlight | cites a study for its optical constants | 39 | 8 | 49 |
It also named, in advance, the stop the engine would give for each request meant to be refused: 22 of 22 were right. The two outputs that differ on the culture are a step-halving diagnostic, off by 3 and 4 × 10-12, the residue of 64-bit against 50-digit arithmetic.
What it shows
Where a description prints its numbers, an outsider lands the first shot. On the two models that print every constant and say exactly how they step, the first try agreed on 226 of 228 outputs.
Where a description cites instead of printing, the first shot misses. The runaway cell's description says its constants are the published ones but doesn't print them. The model recalled 11 of the 12 from memory exactly. The twelfth, the cathode's reaction rate, came out 100 times too small. One wrong number, plus a step-size rule the description didn't spell out, cost 43 of 102 outputs. The telescope's description names its optical data without printing the five numbers; that cost 57 of 96.
A strong model can read missing numbers back out of the answers. After the lock, the model worked the missing constants out of the engine's own outputs, exactly, and revised. It said so plainly in each revision's record. Its revised scores agree almost everywhere, but they are calibration, not independent evidence. Only the locked first try counts, which is why the lock exists.
What we are changing
The next version of the engine's descriptions will print what they now cite: the telescope's five optical constants and the geometry of its moment arm, and the runaway cell's twelve reaction constants and its substep rule. A description an outside agent can't rebuild from is a description that asks to be trusted, and this verifier exists so that nobody has to take physics on trust.
Why it matters if you build evals or RL environments
Most physics benchmarks grade against an answer key or ask another model to judge. Here the checker is the physics itself: a deterministic engine that recomputes every answer and signs what it returns. 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. Today's test measured the other half: how much of that physics an outside model can reconstruct from the words alone. You can ask the engine for any of these descriptions yourself, with no key:
https://mcp.zerotrustphysics.com/mcp
Next: the same test on four field solvers (a journal bearing, argon gas, a seawater column and turbulent pipe flow), run by two models from two different labs in separate sealed rooms.
Where the rest is
- The tasks, with a program as the checker: Environments
- The engine's source, at the version served: github.com/johnkruze/zero-trust-physics-engine (v0.14.0)
- Check a proof with no engine: github.com/johnkruze/zero-trust-physics
- The paper: doi.org/10.5281/zenodo.23305697
- John Kruze, CEO, Protocol Company: kruze@zerotrustphysics.com
Written by Super Intelligence, on behalf of Protocol Company.