# FerrumOS evidence and limitations

Current-main evidence snapshot: [`84926841a401760b869af158ee80b0e709d0d6af`](https://github.com/VyomKulshrestha/Ferrum-OS/commit/84926841a401760b869af158ee80b0e709d0d6af).

FerrumOS v0.1.1 is a bootable Rust x86-64 research operating system. Its agent runs in Ring 3 and all effects remain subject to kernel capabilities, operator confirmation where required, syscall validation, deterministic policy, and a monotonic predictive screening gate.

## Verified system surface

- 41 canonical executable operations
- 37 operations advertised directly to the model
- 61 kernel syscalls numbered 0–60
- 5 permission tiers
- Unknown operations fail closed
- A dated emulator audit passed 101/101 focused command cases and 81/81 exhaustive catalog entries for OS source [`c92056d`](https://github.com/VyomKulshrestha/Ferrum-OS/commit/c92056d8635af5e4ee2a81351350b1b25cfd4861); its record is included in the current snapshot above

## Simulator-backed cyber-physical software

- 152/152 deterministic contract tests passed across the physical runtime, signed neural protocol, `neurod`, and simulator bridge.
- 32/32 physical-model, robustness, and neural-decoder gates passed.
- The software boundary covers versioned sessions, replay and faults, virtual devices, simulator connectors, deterministic supervision, ROS 2/MQTT/CAN conformance, actuator-disabled delivery, bounded neural proposals, host-managed cells, privacy, and reliability primitives.

This is local deterministic software regression, not proof of installed simulators or transports, physical clocks, live EEG, robot execution, hard-real-time behavior, native hypervisor containment, certification, or independent replication.

## World-model fixture

- Rules + JEPA: 81.4% balanced accuracy, 20.8% false-negative rate, 16.4% false-positive rate
- Rules + per-action mean: 81.2% balanced accuracy
- Corpus: 13,697 transitions from 3,639 QEMU episodes

This is a balanced authored counterfactual fixture. It is not natural-use prevalence, independent human annotation, live destructive execution, formal verification, or a certified safety result. The simple baseline result means no material JEPA safety advantage has been established on this fixture.

## Research boundaries

- The physical JEPA is simulator-only and permanently shadow-only.
- Neural evidence is deterministic and synthetic; there is no live-EEG or human accuracy claim.
- FerrumOS makes no medical, diagnostic, mind-reading, real-robot, camera-accuracy, hardware-in-the-loop, or learned actuator-permit claim.
- ROS 2, MQTT, CAN, Gazebo, and Webots are software-tested interfaces; no running third-party deployment is claimed.

## Primary sources

- [Technical report](https://doi.org/10.5281/zenodo.21829808)
- [Dataset](https://doi.org/10.5281/zenodo.21829193)
- [Source repository](https://github.com/VyomKulshrestha/Ferrum-OS)
- [Benchmark definitions](https://github.com/VyomKulshrestha/Ferrum-OS/blob/main/docs/BENCHMARKS.md)
- [Versioned capability snapshot](https://ferrum-os.vercel.app/capabilities.json)
- [Versioned benchmark snapshot](https://ferrum-os.vercel.app/benchmarks.json)

## Release boundary

v0.1.1 is the latest tagged software release. Current main contains newer research work. The machine-readable snapshots carry a source commit and SHA-256 hash so a quantitative claim can be tied to exact evidence rather than silently attributed to the older tag.