Executable operations
41The catalog, declared count, and dispatch implementation independently agree. Thirty-seven are exposed directly to the model.
Evidence center / current mainSource 84926841a401760b869af158ee80b0e709d0d6af
Proof before promise.
FerrumOS publishes what was measured, how it was measured, and what the result does not prove. Learned forecasts can add caution. They cannot remove deterministic warnings or grant execution authority.
System surface
The authority boundary is countable.
Kernel syscalls
61The source enum spans syscall numbers 0–60. Agent effects still cross this deterministic ABI.
Permission tiers
5Observe, safe, network, modify, and destructive classes remain distinct from model prediction.
Unknown actions
ClosedUnrecognized catalog entries are rejected rather than silently mapped to a permissive fallback.
QEMU command audit
101 / 101A dated emulator audit passed all focused cases and 81/81 catalog entries for OS source c92056d; its evidence record is included in the current snapshot above.
Cyber-physical contracts
152 / 152Deterministic physical-runtime, neural-protocol, neurod, and simulator-bridge tests passed on the named source snapshot.
Model & decoder gates
32 / 32Physical-model, robustness, and neural-simulator gates passed without promoting learned output into actuator authority.
World-model study
A measured gain, not a mythology.
| Gate | Balanced accuracy | False negative rate | Boundary |
|---|---|---|---|
| Rules + JEPA | 81.4% | 20.8% | Authored 500-episode counterfactual fixture |
| Rules + action mean | 81.2% | — | No material JEPA advantage established over this simple baseline |
| Five complete pipelines | 79.76% mean | — | Independent training runs on the same authored fixture |
Claim boundary — do not overread this resultThe fixture is balanced and authored. It is not natural-use prevalence, independent human annotation, 500 live destructive executions, formal verification, or a certified safety result.
Runtime
Preview latency stays bounded in the measured profile.
H=1 preview
1.29–1.40 msRing-3 benchmark run means under the documented QEMU/WHPX profile.
H=5 preview
1.43–1.57 msBounded lookahead in the same environment.
Model loading
26–30 msWith zero measured heap-growth bytes in all three benchmark runs.
Queue response
96 / 96Every serialized preview response returned in each run; this is not parallel inference throughput.
Cyber-physical software
The reference vertical is wired—inside a strict evidence boundary.
Current main joins versioned provenance, deterministic replay and faults, virtual devices, Gazebo/ROS 2 and Webots bridge boundaries, watchdog and recovery logic, ROS 2/MQTT/CAN conformance, actuator-disabled delivery, bounded neural proposals, host-managed cells, privacy, and reliability primitives.
Claim boundaryThis local deterministic regression does not prove installed third-party infrastructure, physical clocks, live EEG, robot execution, hard-real-time behavior, native hypervisor containment, certification, or independent replication.
New interfaces
Strong simulator results retain narrow claims.
| Interface | Measured result | What it proves | What it does not |
|---|---|---|---|
| Physical JEPA | 99.44% balanced accuracy; 1 FN, 16 FP | Deterministic simulator screening; permanently shadow-only | No robot, camera, hardware-in-the-loop, or actuator-permit claim |
| Neural decoder | 600/600 signals; 400/400 artifact abstentions; 0/10,000 idle candidates | Deterministic synthetic-evidence contract | No human participant, live-EEG accuracy, medical, or mind-reading claim |
Primary artifacts
Reproduce the evidence yourself.
- Technical report on Zenodo (DOI 10.5281/zenodo.21829808) — 14-page report, CC BY 4.0.
- World-model dataset v1.0.0 — 13,697 transitions from 3,639 QEMU episodes.
- Benchmark definitions and limitations.
- Paper evaluation protocol.
- Machine-readable benchmark snapshot.