Protocol Company commited on
Commit ·
f781803
1
Parent(s): fd4ed5d
Card wording: name each figure's basis, fix placeholders and counts
Browse files
README.md
CHANGED
|
@@ -24,7 +24,7 @@ Protocol Company · Zero-Trust Physics · release v0.15.3 · 13 November 2026
|
|
| 24 |
|
| 25 |
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.
|
| 26 |
|
| 27 |
-
**Domain & slice coverage.** These 5 free evaluation tasks exercise measured CMSX-6 single-crystal superalloy creep at 980 degC (`materials/creep_cmsx6`), coupled nonlinear
|
| 28 |
|
| 29 |
**Evaluation only.** This free sample is for evaluating the tasks. It grants no licence to train on them or pass them on.
|
| 30 |
|
|
@@ -36,6 +36,8 @@ Five calibrated multi-step physics RLVR tasks scored by deterministic recomputat
|
|
| 36 |
| `g4-forensics10-06` (forensics audit) | 50% (4 of 8) | 100% (n=8) | 0.881 | 0 of 10 above 0; honest 3 of 3 at 1 |
|
| 37 |
| `g4r10-ch-cant-tes-2d-peakenvelope` (chained design) | 62.5% (5 of 8) | 100% (n=8) | 0.847 | 0 of 19 above 0; honest 3 of 3 at 1 |
|
| 38 |
|
|
|
|
|
|
|
| 39 |
Calibrated on a frontier model (tool variant, 8 attempts per task): sample `pass@1` spans 50%–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, 81 adversarial cheat attempts scored `0` above `0` and all 15 honest controls scored `1`. Per-task quality reports (HTML and JSON) are in `reports/`.
|
| 40 |
|
| 41 |
**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.
|
|
|
|
| 24 |
|
| 25 |
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.
|
| 26 |
|
| 27 |
+
**Domain & slice coverage.** These 5 free evaluation tasks exercise measured CMSX-6 single-crystal superalloy creep at 980 degC (`materials/creep_cmsx6`), coupled nonlinear resonator-to-cantilever beam vibration (`signal/tesseract`, an 80 Hz Duffing resonator, <-> `signal/cantilever_eth`), and 10-model structural, RF, and acoustic forensics (`link_gothenburg_10010`, `modulus_bam_s355`, `stress_plate_hole_s235_elastic`, `junction_nist_b7_s2`, `seawater_column`, `argon_pair_gas`): a budgeted search, a design to a limit, two two-model chained designs, 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 32.4 GHz rain-attenuated terrestrial RF links (`signal/link_gothenburg_10010`), BAM S355 and S235 structural steel elasticity and plate-hole stress concentration, NIST structural junctions (`junction_nist_b7_s2`), and ocean acoustic seawater columns (`seawater_column`).
|
| 28 |
|
| 29 |
**Evaluation only.** This free sample is for evaluating the tasks. It grants no licence to train on them or pass them on.
|
| 30 |
|
|
|
|
| 36 |
| `g4-forensics10-06` (forensics audit) | 50% (4 of 8) | 100% (n=8) | 0.881 | 0 of 10 above 0; honest 3 of 3 at 1 |
|
| 37 |
| `g4r10-ch-cant-tes-2d-peakenvelope` (chained design) | 62.5% (5 of 8) | 100% (n=8) | 0.847 | 0 of 19 above 0; honest 3 of 3 at 1 |
|
| 38 |
|
| 39 |
+
Task IDs: `bs` budgeted search, `dl` design to a limit, `ch` chained design, `r10`/`r11` design round; `tes` is `signal/tesseract`, `cant` is `signal/cantilever_eth`.
|
| 40 |
+
|
| 41 |
Calibrated on a frontier model (tool variant, 8 attempts per task): sample `pass@1` spans 50%–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, 81 adversarial cheat attempts scored `0` above `0` and all 15 honest controls scored `1`. Per-task quality reports (HTML and JSON) are in `reports/`.
|
| 42 |
|
| 43 |
**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.
|