Track 3 · Physical AI · AMD AI DevMaster Hackathon 2026

CRUX

failure discovery → repair → qualification for contact-rich robot manipulation, end to end on one AMD Radeon GPU

+36.7 ppTASK SUCCESS 0/128 → 47/128 on virgin seeds, exact McNemar p = 1.4e-14 — replicates on four further suites (13/32, 12/32, 9/32, and 6/32 on a second task)

APPROVEDthe release gate approved v4 after rejecting the two candidates before it — +36.7 pp generalization gain with negative regression

1 metric bugour own success check was geometrically impossible — and one repair we called falsified five times had always worked. Fixing the ruler turned 1/32 successes into 11/32 on the suite that had called it falsified

~200k–293kenvironment-steps per second at 4,096 batched environments on one Radeon PRO W7900

Verify every number yourself — CPU only, no GPU

git clone --depth 1 https://github.com/Enoch208/Crux && cd Crux
uv sync
uv run pytest -q                                  # 275 tests, ~1 s
uv run crux validate evidence/manifest.json       # hash + recompute the bundle → 9/9
uv run crux report evidence-dev/qualification_v4_standard.jsonl \
  evidence-dev/qualification_scale.jsonl \
  --baseline-version baseline-v1 --repaired-version candidate-v4 \
  --config configs/qualification.yaml             # every headline number from raw JSONL

Tamper with one byte of an episode file and crux validate fails. That is the point.

The verdict, verbatim

PASS schema                 manifest 1 and receipt 1 parsed against the declared schema
PASS files_exist            all 9 declared files present
PASS hashes                 9 files match their recorded sha256 and size
PASS device_evidence        AMD Radeon Graphics (gfx1100) via amdgpu, ROCm 7.2.1, torch 2.13.0+rocm7.2
PASS suite_separation       32 held-out seeds disjoint from 32 repair seeds
PASS checkpoint_identity    receipt checkpoint resolves to controller/repaired.json
PASS replays                2 replays present, non-empty and hashed
PASS aggregates             heldout: baseline 0/128, repaired 47/128 recomputed from raw episodes
PASS headline_regression    standard regression -28.12 pp reproduced from 32 matched pairs

9/9 checks passed · release gate: APPROVED

Virgin-seed qualification (seeds 801–928, never used before — 128 matched pairs)

Endpoint	baseline-v1	candidate-v4	Delta	Exact McNemar
Task success	0/128	47/128	+36.7 pp	p = 1.4e-14
Reached seating verification	2/128	58/128	+43.8 pp	p = 4.1e-16

Zero discordant pairs against the candidate across 256 matched pairs — there is no seed the baseline completes and v4 does not. Success replicates on four further suites: virgin 701–732 (13/32, p = 0.0002), virgin 501–532 (12/32, p = 0.0005), standard 101–132 (9/32, p = 0.0039) and a second task (6/32, p = 0.0312). Claims that failed to replicate along the way are withdrawn in writing.

Watch it

▶ four-minute narrated demo (YouTube) — ends with an uncut, real-time completed episode on a virgin seed. Also in-repo.

Dig deeper

Repository · Technical report · Gate-by-gate evidence log · 3 filed Genesis issues