HEXA-OS Family — AI inference appliance OS · law-enforced · self-hosting · unikernel-first
Operating systems · unikernel · Firecracker · law kernel · syscall-free · p99-first · self-hosting · serving-only
Not a Linux replacement. law-enforced · self-hosting · unikernel-first. Serving p99 30–50% ↓ is the whole point.
user .hexa
→ hexa_compile → bytecode
→ hexa_vm_run (user proc, single address space)
→ syscall = hexa fn call (not int 0x80, ~300 ns → 20–80 ns)
→ kernel fn: law_check(caller_cap, target, op)
pass → real I/O
fail → EPERM + growth_bus record
Training gains are capped at ~5% (compute-bound) — serving only. Scoped as a cloud inference appliance. v0.1 target: 7B-model unikernel serving, cold-start < 500 ms.
YouTube · Email · 💖 Sponsor · 💳 PayPal · 🗺️ Atlas · 📄 Papers
🔭 NEXUS — Universal Discovery Engine. 216 lenses + OUROBOROS evolution + 5-phase singularity cycle.
🧠 Anima — Consciousness implementation. PureField repulsion-field engine + 1030 laws + Φ ratchet.
🏗️ CANON — Architecture from perfect number 6. 225 AI techniques + chip design + crypto/OS/display.
💎 HEXA-LANG — The Perfect Number Programming Language. Working compiler + REPL.
- v0.1 spec live (design stage) — 6 modules, SPEC_FIRST closure, 4/4 verify scripts PASS
- 6 modules:
boot·kernel·serving·deployment·ios·macos - Family substrate: this repo + sibling
hexa-ios/+hexa-macos/placeholders - Depends on
hexa-langP7–9 fixpoint (C runtime eliminated,@nostdfreestanding codegen) - Operational unikernel UNVERIFIED — design target, not a measurement; ~5% training cap is the falsifiable real-limit claim
- DOI 10.5281/zenodo.19703255 (concept release)
# 1. Install hexa-lang (gives you `hexa` + `hx` package manager)
/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/dancinlab/hexa-lang/main/install.sh)"
# 2. Install hexa-os
hx install hexa-os| 🎯 | Scoped — cloud inference appliance only. Not a Linux replacement |
| ⚡ | No syscalls — ~300 ns → fn call 20–80 ns (5–20×) |
| 🧊 | Unikernel-first — Firecracker guest · virtio-only drivers · single address space |
| ⚖️ | Law kernel — law_check(cap, target, op) + growth_bus log. raw.json is the capability source |
| 📈 | v0.1 target — 7B serving · cold-start < 500 ms · p99 30–50% ↓ |
| 🧬 | Self-hosting — boots 100% in .hexa · freestanding, no libc |
| Bottleneck | Linux | Hexa OS | Gain |
|---|---|---|---|
| syscall round-trip | ~300 ns | fn call 20–80 ns | 5–20× |
| copy_to_user | memcpy | pointer pass | bandwidth 1.5–2× |
| context switch | 1–3 μs | cooperative yield 100–300 ns | 10× |
| latency p99 | ~200 ms | 30–50% ↓ | real commercial value |
| QPS at SLA | baseline | 20–40% ↑ | downstream effect |
Training is compute-bound, so OS-layer gains cap at ~5%. Serving is dominated by syscall · copy · context-switch, so OS-layer wins land directly on p99.
Unikernel + Virtio-only + Firecracker guest
× Tensor primitive + KV-cache inode + Serving loop
× Single address space + Zero-copy
× Law kernel (raw.json)
# closure verification (the SPEC_FIRST entry point)
hexa run verify/run_all.hexa # → 4/4 scripts passed
# individual verify scripts
hexa run verify/spec_presence.hexa # 6/6 module spec docs at declared paths
hexa run verify/lattice_arithmetic.hexa # n=6 lattice identities (aux check)
hexa run verify/real_limits_anchor.hexa # H1/H5/H6/H7/H9/S1 real-limit anchors
hexa run verify/closure_consistency.hexa # scoreboard cross-check
# inspect spec docs
ls docs/ # PERF_P99 · HEXA_SERVE_V01 · DEPLOYMENT · BRAINSTORM
cat hexa-ios/hexa-ios.md # mobile target spec
cat hexa-macos/hexa-macos.md # desktop target specOperational boot / kernel / serving invocations are pre-impl (gated on hexa-lang P7–9 fixpoint).
self/os/
boot.hexa UEFI stub + kernel load
kernel.hexa scheduler / memory / IPC
fs.hexa law-tagged inode
drv/ virtio-only drivers
user.hexa user-space runtime
_freestanding_probe.hexa libc-less POC (first step)
docs/
ROADMAP.md 3 stages × prerequisites
BRAINSTORM.md 124 raw items
perf_model.md p99 gain mechanism
- Freestanding hexa compile POC —
@nostdELF, no libc - Syscall-free bench POC — measure p99 on a user-space mini-server
- Unikernel v0.1 — Firecracker guest + 7B serving
hexa-langP7–9 fixpoint (C runtime eliminated) — roadmap- Freestanding codegen +
@nostdsupport
hexa-os is a SPEC_FIRST substrate: 6 design modules, each backed by a
spec doc on disk. Closure is verified by 4 small .hexa scripts under
verify/:
# from repo root:
hexa run verify/run_all.hexa
# expected: 4/4 scripts passed| # | script | what it checks |
|---|---|---|
| 1 | verify/spec_presence.hexa |
6/6 module spec docs at declared paths (+ 2 crosslink sub-specs) |
| 2 | verify/lattice_arithmetic.hexa |
n=6 lattice identities (aux per LATTICE_POLICY.md §1.3 rule 1 — never sole verification) |
| 3 | verify/real_limits_anchor.hexa |
LIMIT_BREAKTHROUGH.md anchors (Turing H1 · Brewer H6 · Lamport H7 · NVMe H5 · speed of light H9 · Linux syscall S1 baseline) |
| 4 | verify/closure_consistency.hexa |
scoreboard cross-check (CLI registry · hexa.toml · README badge · AGENTS.md policy registration) |
The 6 modules:
| # | module | spec doc |
|---|---|---|
| 1 | boot |
ROADMAP.md — UEFI / Multiboot2 / Firecracker guest |
| 2 | kernel |
docs/PERF_P99.md — scheduler · memory · IPC · law_check |
| 3 | serving |
docs/HEXA_SERVE_V01.md — hexa-serve v0.1 (7B unikernel) |
| 4 | deployment |
docs/DEPLOYMENT.md — operational notes |
| 5 | ios |
hexa-ios/hexa-ios.md — mobile target spec |
| 6 | macos |
hexa-macos/hexa-macos.md — desktop target spec |
UNVERIFIED until the operational unikernel boots (gated on hexa-lang P7–9 fixpoint). It is a design target, not a measurement. The ~5% training cap is a falsifiable real-limit claim (GPU-FLOPs-dominated); it is preserved verbatim from the headline above. Actual OS vendors (Linux Foundation, Microsoft, Apple, Google) use their own benchmarks (LMBench / sysbench / fio / etc.) — no lattice-fit is asserted on any external system.
ROADMAP · Docs · LATTICE_POLICY · LIMIT_BREAKTHROUGH · Releases · Paper (hexa-lang · P-HEXA)
MIT — permissive open source.
🖥️ syscalls become fn calls. p99 becomes commercial value. · dancinlab