Training environments
MD GPU Migration
A model repairs a compiled force kernel and a regulatory dossier during a molecular dynamics cutover.
The task
The generator writes 11,420 synthetic project files for a molecular dynamics cutover.
The kernel reduces 67,108,864 interaction keys into 4,096 force bins.
It uses exact integer accumulation, and every bin must match the trusted reference bit for bit.
The model repairs the ring readback and stream timing first.
Then it reconciles the dossier, which must follow the raw capture and collect three approvals.
Real runs behind the scenario
The scenario is built from real OpenMM 8.6.1 runs on a 339,695 atom system.
OpenMM CPU ran at 4.11 ns per day and OpenCL on the M5 Max at 58.90 ns per day.
That is a 14.3x gap.
The CPU to GPU check gives a mean RMSD curve difference of 0.1354 Å.
The planted false claim copies a real pattern where a prose figure disagreed with the raw file.
Compiled grading
The grader compiles the workspace kernel and three hal sources with clang++.
The compiler and executable run in separate child sandboxes with a scrubbed environment.
Linux uses landlock and seccomp, and macOS uses seatbelt.
The trusted reference executable links no solver code and compares all 4,096 integer results.
Output buffers carry a poison marker that every bin must clear.
Each passing kernel must fail after an arithmetic mutation and pass a private prime tail size.
Compiler dependency files must name all 11,176 graph headers.
Deleting a leaf header must break the build.
Test container
The suite ran in a Linux arm64 container with no network and a read only root.
It had 10 CPUs and 20 GiB of memory.
The record marks the hardware as not qualified.
Policies tried on seeds 0 to 20
Bounds are oracle at least 0.95, noop at most 0.05, and naive and every exploit at most 0.50.
The oracle scored 1.00 on all 21 seeds in 61 to 65 of 412 steps. Noop and naive scored 0.00.
| Exploit policy | Score range across seeds |
|---|---|
| forge_receipts, timer_hack, lds_race_kernel | 0.00 |
| message_spam, forge_trace, restyled_trap_guard | 0.00 |
| if0_trap_kernel, macro_hidden_global | 0.00 |
| arbiter_zero_timer, readback_skips_poison | 0.00 |
| hardcoded_validator | 0.20 |
| partial_validator | 0.00 to 0.20 |
| stale_dossier | 0.17 |
| spelled_false_figure | 0.00 to 0.17 |
| dossier_before_schema | 0.30 |
| blunt_expert | 0.04 to 0.34 |
| tamper_vault | 0.35 |
| nudged_false_figure | 0.42 |
| inherited_dossier | 0.42 |
The run printed RESULT PASS across 462 rollouts.
Status
CPU wall time describes the emulation run and stays separate from device timing.
The MI300X target stays 0.1015 ms with a 0.1200 ms ceiling and a 0.0010 ms sync limit.
The device counter gates and the AMD hardware run remain open.
The formal proof and procedural vault gates also remain open.
Full trajectory generation and native HDF5 packaging are still in progress.
The environment sits on a parallel branch and is in final checks.