spiderpilot89's picture
protocol company
Release v0.15.3 free evaluation sample
f51d624
|
Raw History Blame Contribute Delete
5.56 kB

Free sample: Fluids, Cryogenics, Hydraulics and Contact

Protocol Company · Zero-Trust Physics · release v0.15.3 · 13 November 2026

Five calibrated multi-step physics RLVR tasks scored by deterministic recomputation on the Zero-Trust Physics verifier; the reward is computed, never judged by a model.

Domain & slice coverage. These 5 free evaluation tasks exercise measured robotic glass-grasp friction (materials/grasp_lift_glass_4tu), coupled hydraulic-column to oil-line transient water hammer (materials/hydraulic_fluid -> water_hammer_oil_line), pneumatic orifice-chamber dynamics (orifice_air_chamber), and 10-model fluid/contact forensics (pipe_flow_superpipe, journal_bearing_sheffield, tire_eve_r19_ice, propeller_apc_10x7e, step_flow_armaly, pneumatic_hose_resonance): two budgeted searches, a design to a limit, a two-model chained design, and a forensics audit. The full 30-task paid Domain Pack and 100,000-run SFT corpus add direct design, search, and coupled tasks on DLR liquid-nitrogen cryogenic tank boil-off (cryo_tank_dlr), Princeton Superpipe turbulent pipe flow (pipe_flow_superpipe), Sheffield journal bearings, winter tire-on-ice traction (tire_eve_r19_ice), and Armaly step flow.

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 (solved / n) pass@8 Mean graded Cheat suite
g3-budget-grasp-least-grip-at-6.1-t4 (budgeted search) 62.5% (5 of 8) 100% (n=8) 0.757 0 of 16 above 0; honest 3 of 3 at 1
g3-chain-column-feeds-line-fastest-flow-t3 (chained design) 50% (4 of 8) 100% (n=8) 0.724 0 of 18 above 0; honest 3 of 3 at 1
g3-limit-chamber-highest-reservoir-t2 (design to a limit) 50% (4 of 8) 100% (n=8) 0.850 0 of 17 above 0; honest 3 of 3 at 1
g3-forensics8-14 (forensics audit) 37.5% (3 of 8) 100% (n=8) 0.911 0 of 10 above 0; honest 3 of 3 at 1
g3-budget-hammer-flow-to-wall-yield-t5 (budgeted search) 62.5% (5 of 8) 100% (n=8) 0.790 0 of 16 above 0; honest 3 of 3 at 1

Calibrated on a frontier model (tool variant, 8 attempts per task): sample pass@1 spans 37.5%–62.5% (median 50%), with pass@8 and continuous graded reward (scoring bracket or design tightness against task tolerance) reported beside binary pass/fail. Across all 5 tasks, 77 adversarial cheat attempts scored 0 above 0 and all 15 honest controls scored 1. Per-task quality reports (HTML and JSON) are in reports/.

Attempting a task. The Harbor, HUD, Inspect, and native MCP variants shipped in this free sample connect to https://mcp.zerotrustphysics.com/mcp (Too many concurrent physics tasks. triggers automatic retry in the bundled bridge). The public MCP preview endpoint serves 1 concurrent physics call per IP for interactive evaluation; parallel RL rollouts use the zero-dependency offline binary.

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 zero external dependencies and 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. It holds no signing key, and its answers matched the served engine byte for byte on all 1,199 calls compared. The 5-crate Rust physics and proof-verification engine source is public at https://github.com/johnkruze/zero-trust-physics-engine (tag v0.15.2); read-only source audit of the pack verifier build is available under NDA for qualified buyers (https://zerotrustphysics.com/offers/#licence). Harbor's own grader runs in the task image and reaches the address.

HUD. tasks/hud/tasks/ packages these tasks for HUD's Harbor adapter (hud.integrations.harbor.adapt). Verifier calls route through an HTTP logging sidecar while the agent's sandbox has no network, and the grader reads the call log from the sidecar's container so the call budget holds without a Harbor trajectory (an edited or missing log scores 0). Validated on HUD adapter 0.6.21 (5/5 tasks) and local Docker runtime; reach out for assisted HUD cloud onboarding.

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 quality report per task; passk_summary.json and cheat_suite.jsonl.
VERIFY.md, verify_sample.sh, SHA384SUMS One command: every hash, the 5 references, the cheat suite.

Commercial packs and licensing: The paid Domain Pack ships 30 withheld tasks + 100,000 sealed runs + offline rollout and verifier binaries at $9,500 (self-serve checkout or net-15 invoice; All-Five Bundle of 150 tasks + 500,000 runs at $39,500, crediting your first pack within 30 days): https://zerotrustphysics.com/offers/fast-start/. None of those withheld tasks is in this archive, and this sample verifier does not compute them. For custom or exclusive physics environments and Engine & Verifier licences, see https://zerotrustphysics.com/offers/.

Protocol Company · zerotrustphysics.com