Free sample: Orbital, Space and Atmospheric Entry
Protocol Company · zerotrustphysics.com · Domain Packs for training: 30 withheld tasks and 100,000 sealed runs, scored offline, $9,500, order or ask · WhatsApp +1 928-264-7883 · kruze@zerotrustphysics.com
Protocol Company · Zero-Trust Physics · release v0.15.3 · 7 November 2026
Five multi-step physics RLVR tasks: a capsule's entry and the heat shield it burns, a budgeted search for where a run stops, and two forensics audits of reported answers. These measure agentic search convergence under strict simulator call budgets. The verifier does the physics; the reward is computed, not judged.
Evaluation only. This free sample is for evaluating the tasks. It grants no licence to train on them or pass them on.
| Task (kind) | pass@1 (n) | pass@8, and how little it says | Cheat suite | Failure split (of the failed attempts) |
|---|---|---|---|---|
g2-budget-sd-airspeed-sink-e-t3 (budgeted search) |
62.5% (10 of 16) | 100% (n=16); fits a true value from 97% to >99.9% | 0 of 16 above 0; honest 3 of 3 at 1 | tolerance 4, proof_or_record 2 |
g2-chain-sd-slab-radius-halfchar-h-r7 (chained design) |
25% (2 of 8) | 100% (n=8): one group of 8; fits a true value from 23% to >99.9% | 0 of 19 above 0; honest 3 of 3 at 1 | over_budget 4, tolerance 2 |
g2-sd-slab-shallowest-char-load-v3 (design to a limit) |
37.5% (3 of 8) | 100% (n=8): one group of 8; fits a true value from 51% to >99.9% | 0 of 19 above 0; honest 3 of 3 at 1 | over_budget 4, proof_or_record 1 |
g2-forensics7-10 (forensics audit) |
37.5% (3 of 8) | 100% (n=8): one group of 8; fits a true value from 51% to >99.9% | 0 of 10 above 0; honest 3 of 3 at 1 | forensics_wrong_rows 5 |
g2-forensics6-07 (forensics audit) |
56.25% (9 of 16) | 100% (n=16); fits a true value from 94% to >99.9% | 0 of 10 above 0; honest 3 of 3 at 1 | forensics_wrong_rows 7 |
One frontier model, tool variant, 8 or 16 attempts per task (n in the table). pass@8 is 1 - C(n-c, 8) / C(n, 8): at n=8 it is 100% for any task solved once, at n=16 for any task solved 9 times or more. These tasks were chosen for at least one pass and pass@1 in [2%, 70%] on these same attempts, so the 100% says little. The range beside it is what pass@1's exact 95% interval allows. tolerance: the sides were right and the bracket or designs were wider than the task asks. over_budget: more computing calls than the task allows. proof_or_record: a proof or the call record did not hold. forensics_wrong_rows: the rows named false, or their reasons, were not the report's.
Attempting a task. Agents use the verifier at https://mcp.zerotrustphysics.com/mcp, which allows one physics task at a time per client address. A call made while another from that address computes is answered Too many concurrent physics tasks.; the bundled bridge waits and retries, so parallel attempts from one address queue. The live address has one virtual CPU and was not load-tested. The Harbor, HUD, Inspect and native MCP variants as shipped here all reach this address.
For rollouts at scale, the paid Domain Pack ($9,500) ships a compiled offline rollout binary for Linux amd64, Linux arm64 and macOS arm64, so Harbor and HUD rollouts and grading run --network none with no rate limit. Keyed hosted HTTP is also available if you prefer.
Scoring. scorer/ scores offline: bin/ holds a sample-only verifier (macOS arm64, Linux amd64, Linux arm64) that computes these 5 tasks' requests, each forensics row's request as written, and refuses everything else by name. Source not shipped. It holds no signing key, and its answers matched the served engine byte for byte on all 721 calls compared. Harbor's own grader runs in the task image and reaches the address.
HUD. tasks/hud/tasks/ is these tasks in the shape HUD's Harbor adapter takes (hud 0.6.21: 5 of 5). The verifier is an HTTP route to a logging sidecar, the agent's sandbox has no network, and the grader takes the call log from the sidecar's container, so the budget holds with no Harbor trajectory; an edited or missing log scores 0. Only HUD's hosted runtime awaits a buyer run: if HUD is your harness, tell us and we will run the first with you.
| Folder | What it holds |
|---|---|
tasks/ |
harbor/, hud/, inspect/, mcp/: the 5 tasks in each format. |
scorer/, harness/ |
The scorer, cheat suite, sample verifier, bridge to the address, call-logging proxy. |
reference/, reports/, calibration/ |
Reference answers with SHA-384 proofs; a report per task; the records behind the table. |
VERIFY.md, verify_sample.sh, SHA384SUMS |
One command: every hash, the 5 references, the cheat suite. |
The paid Domain Pack is 30 other withheld tasks + 100K SFT Warm-Start Corpus at $9,500: https://zerotrustphysics.com/offers/fast-start/. None of those tasks is in this archive, and this verifier does not compute them.
Protocol Company · zerotrustphysics.com
In this repository: sample/ is the archive unpacked, unchanged (run sh sample/verify_sample.sh); archive/ holds the archive, the Raw Data ZIP and their SHA-384 lines. Buy: https://zerotrustphysics.com/offers/fast-start/
- Downloads last month
- 84