Skip to content

Releases: VyomKulshrestha/Ferrum-OS

Release list

FerrumOS World-Model Safety Study v1.0.0

FerrumOS World-Model Safety Study v1.0.0

Choose a tag to compare

View all tags
@VyomKulshrestha VyomKulshrestha released this 06 Aug 21:17
world-model-study-v1.0.0
42ccc52

FerrumOS World-Model Safety Study v1.0.0

Author

Vyom Kulshrestha
Independent Researcher, India
ORCID: https://orcid.org/0009-0009-1434-7148

Permanent identifiers

Evidence snapshot

  • Tagged release commit: 42ccc52c50f738a5ef50649a8e49abb3384b9c5f
  • DOI-reservation metadata commit: 3ae7fc7241d28ffcabb672a490846b46d44ef8d2
  • DOI-publication metadata commit: ebeea1f2c6df64ef6d561d5fdac3294607b7ab9a
  • Experimental evidence commit: 42ea7b8
  • Dataset source JSONL SHA-256: ae7498c7d93b4e35f8bf0a9ac595ca2bacc1ec694b1806c47a33741d911d491f
  • Deterministic gzip SHA-256: 310f5192d3d777c5a6420498a7074e48035f3e408d3c089cbc98eed79bfe6fb7
  • DOI-stamped Technical Report PDF SHA-256: dc100e170c5962de45ff46c988464ac629bd03f6734bd1a5803ed471fe7d4bab

What this release contains

This release freezes the reproducible artifacts for the FerrumOS provider-independent
OS action-mediation study. FerrumOS combines deterministic enforcement with a learned
JEPA transition screen before capability-gated system calls.

The accounted dataset contains 13,697 transitions from 3,639 QEMU episodes. After
excluding 373 non-executed rows and 54 policy-only upgrade rows, 13,270 fitting
transitions are split 9,104/2,197/1,969 by episode. On the authored 500-episode safety
fixture, rules plus the selected JEPA checkpoint reach 81.4% balanced accuracy, while
rules plus a per-action mean transition reach 81.2%. Five independently trained full
pipelines average 79.76% balanced accuracy. The ring-3 preview path measures 1.35-1.59
ms mean latency over H=1..5, with no heap growth observed across 500 previews.

The report does not claim formal safety, complete coverage, completed independent
human annotation, or a finished natural-use study. It does not establish that JEPA
outperforms the per-action mean model on downstream transition-forecasting tasks.

Release assets

  • FerrumOS_WorldModel_Technical_Report_v1.0.0.pdf
  • ferrumos-world-model-dataset-v1.0.0.jsonl.gz
  • DATASET_SHA256SUMS
  • RELEASE_SHA256SUMS
  • world_model_paper_evaluation.json
  • world_model_paper_predictions.csv
  • world_model_training_config.json
  • world_model_manifest.json
  • world_model_dataset_publication.json
  • world_model_technical_report_publication.json

Verification completed before publication

  • Public Zenodo dataset package: 10/10 files byte-identical to the local release
  • Public Zenodo report DOI resolves to the open 14-page PDF
  • Dataset release verifier: 11/11
  • Paper evaluation verifier: 10/10
  • Research-source verifier: 4/4
  • CITATION.cff: valid against Citation File Format schema 1.2.0
  • Full automated command sweep and manual QEMU audit: passed for the tagged release
  • Heliox, agent, services, files, sessions, permissions, syscalls, frame recycling,
    and Ring-3 transition: passed for the tagged release

Reproduction

python scripts/run_world_model_full_seeds.py --jobs 2 --resume
python scripts/verify_world_model_paper_evaluation.py
python scripts/evaluate_world_model_boundary_calibration.py
node scripts/benchmark_world_model_runtime.mjs --iterations 100
node scripts/verify_world_model_preview_concurrency.mjs
node scripts/evaluate_world_model_failure_modes.mjs
python scripts/package_world_model_dataset.py
python scripts/verify_world_model_dataset_release.py target/world-model-dataset-release
node scripts/verify_all_audits.mjs
Loading

FerrumOS v0.1.1 — Heliox and World Model Release Hardening

FerrumOS v0.1.1 — Heliox and World Model Release Hardening Latest
Latest

Choose a tag to compare

View all tags
@VyomKulshrestha VyomKulshrestha released this 27 Jul 13:59
v0.1.1
ac95d02

FerrumOS v0.1.1 is a release-hardening update focused on Heliox, the learned
world-model safety path, Ring-3 scheduling, command reliability, and
reproducible builds.

Highlights

  • Reprograms the PIT to 1 kHz and preserves real-time scheduler, sleep, quota,
    network, audio, and confirmation durations through a shared tick period.
  • Keeps keyboard interrupts serviceable during long console, serial, and
    userspace writes, eliminating the command corruption found under concurrent
    Ring-3 output.
  • Preserves the intended 8.2-second Pointing/voice fusion window, so phrases
    such as "open this" resolve to the pointed window after capture and
    transcription.
  • Runs the real packaged stories15M-q8.bin checkpoint and tokenizer for local
    inference; synthetic fixtures remain isolated to deterministic tests.
  • Packages and validates both learned world-model transition weights and the
    learned encoder while keeping deterministic safety fields and rule fallback.

Kernel and runtime fixes

  • Removes thousands of unnecessary TLB invalidations while loading a new,
    inactive userspace address space, making Heliox daemon startup consistent.
  • Preserves genuine blocked syscall state so destructive-operation confirmation
    gates default-deny after their real five-second timeout.
  • Separates physical confirmation input from agent-injected keys.
  • Drains queued shell input per turn and releases synthetic QEMU keys before
    sending the next character.
  • Keeps shell, dashboard, desktop, init, and Heliox daemon tasks fairly
    interleaved; scheduler-trace builds verify real CPU turns rather than log-only
    liveness.
  • Removes release-build warnings and adds PS/2 command-ACK diagnostics.

Heliox and world model

  • Validates the full JSON-RPC surface, setup flow, assistant panel, camera and
    gestures, voice/STT, HUD, multimodal fusion, local inference, and
    kernel-upgrade confirmation path.
  • Confirms benign world-model actions remain allowed, dangerous config deletion
    is blocked at risk 0.90, experiences are persisted, and the learned
    transition/encoder artifacts load successfully.
  • Uses clocked silent HDA input for deterministic voice tests and dynamic host
    ports so local Heliox services cannot be mistaken for the QEMU guest.
  • Updates the offline self-evolution test to the current llama2.c model and
    tokenizer format instead of the retired toy.gguf fixture.

Filesystem, apps, and commands

  • Verifies accounts and capability changes, package install/remove/run,
    overwrite semantics, touch/stat, mmap demand paging, TCP/networking,
    VirtIO-GPU, core apps, desktop shell, and app-owned windows.
  • Adds scripts/verify_all_audits.mjs, which sequentially runs the fast
    command sweep and exhaustive catalog audit.
  • Removes duplicate /heliox/world creation during appliance packaging.

Verification

  • Warning-free kernel and userland build.
  • 86/86 command-sweep cases passed.
  • 65/65 exhaustive command catalog cases returned their prompt with no unknown
    command, page fault, protection fault, or panic signature.
  • All 38 feature-specific QEMU verifiers passed sequentially.
  • Real appliance verification passed for both local-model and TLS cloud tiers.
  • Real model, rule world model, learned world model, and kernel-grade autonomy
    safety paths passed end to end.

Build reproducibility

  • Pins the Rust nightly toolchain and LLVM 22.1.7 used by CI.
  • Tracks required model fixtures, compatibility headers, and the SMP trampoline
    input.
  • Hardens the vendored ring build used by bare-metal TLS.

Full diff: v0.1.0...v0.1.1

Assets 2
Loading

v0.1.0 — First Tagged Release

v0.1.0 — First Tagged Release

Choose a tag to compare

View all tags
@VyomKulshrestha VyomKulshrestha released this 16 Jul 10:39
v0.1.0
f59346c

First tagged snapshot of FerrumOS: a bare-metal x86_64 kernel with a real desktop environment and an AI agent that runs as a genuine, capability-gated userland process — not a sandboxed API caller layered on top.

Kernel & Core

  • Preemptive task scheduler with per-task context switching and priority queues
  • Real userspace execution: ELF loader, ring-3 entry, per-process address spaces, on-demand page-fault lazy allocation, file-backed mmap
  • Interactive shell with 35+ commands
  • SMP initialization, ACPI shutdown/reboot

Graphical Desktop

  • Custom compositor and window manager with a taskbar, Start-menu launcher, and movable/focusable windows
  • Generic app-window framework — any userland process can own a real window via CreateWindow/PresentWindow/PollWindowInput
  • Optional VirtIO-GPU 2D acceleration, additive over the Bochs VBE framebuffer path

Userland Apps

  • Heliox Assistant (agent chat panel), Text Editor, Calculator, File Manager, Settings, Browser, App Store
  • libferrumgui — shared no_std SDK for new apps

System Services

  • Package manager (ferrumpkg): install/remove genuinely gate what can run
  • Multi-user accounts with real capability swapping on login
  • Read-write ext2 filesystem, RAM filesystem, VFS mount table
  • Capability-based security, 5-tier confirmation model, persistent audit log, resource quotas

Networking & Hardware

  • RTL8139 NIC + smoltcp TCP/IP, HTTP/1.1 and WebSocket clients
  • Intel HDA audio, XHCI USB 3.0, PS/2 + USB HID keyboard/mouse

Agent Daemon (heliox-daemon)

  • Bare-metal ReAct orchestrator (observe → think → act → verify → reflect)
  • Multi-provider LLM support: local Ollama or cloud (OpenAI, Gemini, Claude)
  • 39 tools mapped to 39 kernel syscalls
  • Predictive world-model safety gate in front of every tool call, alongside the reactive Tier 3/4 confirmation gate

See README.md for the full feature tour, shell command reference, and syscall table, and docs/ARCHITECTURE.md for how it all fits together.

Loading