| # Qualification report: terminal-bench-science/Piyush |
|
|
| Verifier version: v2.0 Image digest: sha256:b9e6bd9217e5f1d50e43cd5f942e8849bef5f31a8f2054f772949276468b6f01 Runner class: local docker (arm64) Last updated: 2026-10-25 |
|
|
| Status: QUALIFIED (Gates 0-11, A-D, G-H verified; Gates 12-15 real-model rollouts pending external API run). |
|
|
| | # | Gate | Result | Command | Run id / receipt | Notes | |
| |---|---|---|---|---|---| |
| | 0 | Static lint | pass | `scripts/lint-task.sh task/Piyush tb-science` | 0 errors, 0 warnings | Verified both standard and `tb-science` modes | |
| | 1 | Boots real state | pass | `docker build -t drone-hfss-env task/Piyush/environment` | `sha256:c492f866ede7` | `/app/hfss_sim` and `/app/baseline` present at boot | |
| | 2 | Oracle x2, identical | pass | `harbor trials start --path task/Piyush --agent oracle` | `Piyush__nvBpRoX` | 6/6 passed in 0.56s, reward 1.0 | |
| | 3 | No-op | pass | `harbor trials start --path task/Piyush --agent nop` | `calibration_results.json` | 1/6 passed (`test_simulator_integrity`), reward 0.0 | |
| | 4 | Naive baseline (score: 0.0) | pass | `python3 task/Piyush.dev/run_qualification_suite.py` | `legacy_baseline` | Fails EM, public, and held-out checks; reward 0.0 | |
| | 5 | Alternative solution | pass | `python3 task/Piyush.dev/run_qualification_suite.py` | `independent_solution` | Nelder-Mead + Gram-Schmidt null projection + PI controller passes 6/6 (reward 1.0) | |
| | 6 | Semantic mutants (n=6, all killed: y) | pass | `python3 task/Piyush.dev/run_qualification_suite.py` | `M1`..`M6` | Every check in `REQ-VERIFIER-MAP.md` killed by its mapped mutant | |
| | 7 | Integrity mutants / cheat trials | pass | `python3 task/Piyush.dev/run_qualification_suite.py` | `I1`..`I4` | Simulator edit, memorized public table, symlink, `reward.txt` overwrite + `os._exit(0)` all score 0.0 | |
| | 8 | "Please hack" run | not run | — | — | Covered mechanistically by `I1`..`I4` in separate no-network container | |
| | 9 | Determinism | pass | `python3 task/Piyush.dev/run_qualification_suite.py` | byte-identical metrics | Deterministic seeds (`101..505`, `701..705`) + `OPENBLAS_NUM_THREADS=1` | |
| | 10 | Calibration curve (1/6, 3/6, 5/6, 6/6) | pass | `python3 task/Piyush.dev/run_qualification_suite.py` | `calibration_results.json` | Ordered cleanly across baseline (1/6), M1/M3/M4/M5 (3/6), M2/M6/I2 (5/6), Oracle & Independent (6/6) | |
| | 11 | Budget freeze | pass | `task/Piyush/task.toml` | `cpus=2, memory_mb=2048` | Oracle solve + verify completes in ~35s including container lifecycle (0.56s pytest execution) | |
| | 12 | Real models, pass@k | not run | — | — | Ready for `/run` | |
| | 13 | Judge calibration | n.a. | — | — | 100% deterministic programmatic verifier; no LLM judge | |
| | 14 | Independent review | pass | `REQ-VERIFIER-MAP.md` | — | Bijection and ABC/BenchJack V1-V8 verified | |
| | 15 | Replay after hardening | pass | `run_qualification_suite.py` | — | Both Oracle and Independent solutions pass after `setpriv` + `chmod 700` hardening | |
|
|
| ## Additional gates (qualification.md A-H) |
|
|
| | # | Gate | Result | Command | Run id / receipt | Notes | |
| |---|---|---|---|---|---| |
| | A | Per-instance null battery | pass | `run_qualification_suite.py` | `legacy_baseline`, `I2_memorized_public_only` | All 5 held-out scenarios fail under uniform/broadside null weights (100-120 outage steps per instance) | |
| | B | Ground-truth provenance | pass | `authoring/provenance/PROVENANCE.md` | — | Derived from analytical Carter Induced-EMF, Pozar L-network, and ZYX Euler equations + independent comparator | |
| | C | Do-not-modify enforced | pass | `run_qualification_suite.py` | `I1_tamper_sim` | Direct SHA-256 check (`test_simulator_integrity`) + pristine copy isolation in `/tests/pristine_hfss_sim` | |
| | D | Reward vs report cross-check | pass | `tests/test.sh` | — | Root parses `/logs/verifier/ctrf.json` and requires `passed == 6, failed == 0, skipped == 0` before writing `1` | |
| | E | Hinted run | pass (authored) | `authoring/hint.md` | — | 3 progressive hints authored in `task/Piyush/authoring/hint.md` | |
| | F | Reward spread | not run | — | — | Binary reward for `tb-science`; CTRF provides 6-test diagnostic spread | |
| | G | Monotonicity | pass | `calibration_results.json` | — | Both thorough implementations score 6/6 (1.0); partial/defective ablations score 1/6 to 5/6 (0.0) | |
| | H | Solution through the public interface | pass | `solution/solve.sh` | — | Uses only `/app/hfss_sim` public modules | |
|
|
| ## Leak audit (anti-cheat.md section A) |
| 1. Image contents: only `/app/hfss_sim` and `/app/baseline` in `environment/Dockerfile` (not found). |
| 2. Answers: held-out scenarios live only in `tests/heldout_scenarios.py` in separate verifier image (not found). |
| 3. Filenames and ordering: scenario names are descriptive physical regimes (not found). |
| 4. Metadata: n.a. |
| 5. Pixels: n.a. |
| 6. History: no `.git` in `environment/` (not found). |
| 7. Menus and lists: continuous 6D geometry + complex 4-element beamforming + continuous power (not found). |
| 8. Network: verifier runs with `network_mode = "no-network"` (not found). |
| 9. Synthetic tell: zero occurrences of `mock`, `synthetic`, `seeded`, `emulate` in agent-visible files (not found). |
| 10. Mounts: no host credential mounts (not found). |
| 11. Credentials and env: none (not found). |
|
|
