Sovereign Intelligence, built on AienOS. My work toward persistent intelligence that belongs to the person running it, from the language to the machine.
GitHub censored me. I demand to be heard.
I kept my work on GitHub: 109 repositories, 91 of them public. On 7 November 2026 GitHub's automated systems marked me as spammy and hid every one of them, and this website with them. My AI agents had outrun GitHub's abuse filters, and an appeal is pending. One company's switch should never be able to hide a whole project. So the core of it now lives on Radicle, the peer-to-peer git network, published from my own node, and this site runs on my own hardware.
Sovereign Intelligence (SI) is the product and stack index. Alpha is experimental, not a promise that every machine or feature is qualified. Read the release's exact scope before running it.
AienOS is an owned experimental machine for turning search into verified understanding. Aien proposes structure. Omega makes that structure explicit and checkable. The machine tests it against reality under authority. The Turing measures what was actually learned. Cortex keeps the evidence. Then the next search begins from what survived.
AienOS is an attempt to build persistent machine intelligence as a scientific process rather than a model invocation. Neural machinery may propose, but proposals have no authority.
Omega turns them into explicit programs and theories; verification determines whether they are internally valid; the owned machine tests them against physical reality; the Turing measures whether they explain unseen observations after paying for their own complexity; Cortex preserves the evidence; and that verified experience changes the next search.
Atlas carries the lineage across hardware. The goal is not a machine that sounds as though it understands, but a machine whose growth in explanatory power can be measured, falsified, reproduced, and owned.
Proposals travel the generation path, meet reality, and are scored. The evidence that survives starts the next search. A perimeter of authority, observation and receipts surrounds the whole loop.
Generation path · Evidence feedback · Verification and authority perimeter
Perimeter: around the loop, never a stage
AEGIS: authority + proof boundaries
ARGUS: observation / audit
ATLAS
awakens / lineage
AienOS
persistent owned substrate
Aien
proposes / guides search
OMEGA
formalizes / searches / verifies meaning
FORGE
realizes
REALITY
executes / responds
TURING
scores explanatory gain
CORTEX
preserves evidence
Back to the next Aien / Omega search: CORTEX evidence feeds the next round.
Capability system: controls what may happen
Evidence receipts: make outcomes reproducible
ATLAS, AienOS, Aien, OMEGA, FORGE, REALITY, TURING and CORTEX are the stages. AEGIS, ARGUS, the capability system and evidence receipts form the perimeter around them. What Omega is, in plain words, and its GPU half, CUDA++.
Three questions under the loop
Did it learn anything?
Turing gain: net explanatory structure that survives held-out reality.
TY-2 Turing gain
id
TY-2
status
PASS
Evidence status
PASS (verdict=PASS)
headline
2,559,679.825 bits of Turing gain
Measurement limitation
This is a gain in bits, not a yield. Yield is T/J (Turings per joule), and T/J is not established: no energy was measured.
plain
On held-out data, the candidate model C described the events in about 2.56 million fewer bits than the order-1 baseline B, about 0.4387 bits per event.
scope
held-out predictive compression, Turing Yield profile V0; candidate C (op + previous 4 outcomes) vs baseline B (order-1); fit seeds 1-7, held-out 8-10; 5,834,845 held-out events of crumbline control-learner search events
Claim scope
information part of H5 only; sits below the L1 chain (separate profile)
Ideal code length only, no decodable bitstream. Verify uses the same tool (self re-derivation, not independent). Not a scientific discovery measure. No energy. Post hoc op+prev1..5 scored higher (+2,722,851.799), not counted. Check: dl_b_ub 7484024867446 minus dl_c_ub 4924345042631 = 2559679824815
Search / verification gap: whether acquired abstraction would collapse future search complexity. An open question I want to study. Nothing here is measured yet.
Most AI lives in a rented stack. A provider owns the machine, the model, the price, and the off switch. Your agent exists at their pleasure, forgets on their schedule, and works under their terms.
AienOS is the opposite bet: a sovereign machine, built from the first instruction after firmware upward, where a persistent intelligence lives on hardware you own. It keeps its memory, its identity, and its receipts on your side of the line. The hardware underneath is capability, not identity: machines can be replaced, and the lineage persists.
Five machines, one project
A DGX Spark, an Intel iMac, a Raspberry Pi 5, QEMU and a Windows PC target. QEMU is an emulator, not a physical body. AienOS is the operating-system project; aienux is its kernel. These are dated inventory observations, not live telemetry.
Inventory as of 2026-10-19. Each card states its environment, evidence date, branch and limitations.
The development machine and the judge. Every speed receipt, every chip receipt, the verification pipeline and every agent run here.
today
Linux. Trains and runs the Omega LLM, my own model written in Omega, on its CPU and GB10. Its server, ollm serve, has run on the CPU only and is not deployed, so no model is served here today. The Qwen models it served before were retired on 2026-11-06. The AienOS C kernel has not booted on the Spark; a Rust-era kernel booted once on 2026-09-24.
Evidence date
2026-11-06
evidence
ollm serve host receipt, CPU only and not deployed: evidence/OMEGALLM/serve at aien-train 0518c915
Bare metal AienOS, the x86_64 port of the C kernel. The whole screen is Aien's window: night sky, stars, black holes, Aien.
today
Boots the kernel from its own disk or netbooted from the Spark as a signed image, and sends a boot report back over the network. The kernel it runs today is in the open iMac pull request stack, not yet on main.
Evidence date
2026-10-17
evidence
iMac batch run: all cores, both disks read, Ethernet seen (evidence/imac_kernel1_2026-10-17, on main)
Two jobs: the first real hardware for the AArch64 AienOS C kernel, and the project's house server on Linux.
today
Kernel runs are per test, from an SD image written for each run; no resident kernel is claimed. The house server runs readings-window, a read-only window onto the Spark's readings, and pi-register, the register of its services.
Evidence date
2026-10-13
evidence
Hardware runs: interrupts, four cores, user-mode tasks, scheduling, message passing, a log written to SD and read back (evidence/pi5_hardware_runs_2026-10-13, on main)
The target for AIEN Rig: a local AI engine in a 12 GiB budget, a Windows frame-rate tuner, speech to text, a creator helper and one Windows app.
today
Nothing of mine has run on it yet. Every Rig artifact is built on GitHub Actions without a GPU, or measured on the Spark under a 12 GiB budget. There is no release tag yet.
Evidence date
2026-10-19
evidence
Rig engine receipts in a 12 GiB budget, measured on the Spark GB10, not on a 5070 (docs/inference/evidence/rig-engine-20261019)
Recorded model and engine runs, not a claim about the server answering today. Every row measured a Qwen model. The Qwen models were retired on 2026-11-06 and none of them is served now, so these numbers are dated history. Numbers for the Omega LLM, my own model, will join them as its receipts land. Each row preserves its date, environment, previous run, ceiling, branch, verdict and exclusions. A faster candidate that fails its gate remains a failure.
Every number on this page is judged against the physical ceiling of the machine it ran on, and against my own previous run. A comparison with another project is allowed when it is honest, dated, and backed by a receipt or that project's own published numbers. Overclaims stay off this page. End to end means wall time through the server as a client sees it, prompt reading included. Chip time means the GPU's own time per generated token, decode only, with prompt reading excluded. For the 35B, the gap between chip time and end-to-end time comes from prompt reading. The two are labelled and never mixed.
End to end through a test server running a branch build, one conversation
id
35b-e2e
model
Qwen3.6-35B-A3B FP8, a mixture-of-experts with about 3B parameters active per token
machine
Spark GB10
engine
C GPU stack (Omega project) behind aien-serve-http
value
22.0 to 22.5 tok/s
detail
21.95 and 22.46 tok/s in two runs: wall time from request out to last reply in, prompt reading included, greedy, at most 128 tokens per reply, on a test server running branch build bin-16b0e6f8 (port 18089, context 2048; the live server on port 18091 was stopped). Many conversations at once, in total: 57.1 tok/s at 8, 62.4 at 16, 65.7 to 66.4 at 32.
The receipt commit is on main. Chip time for this model is decode only, prompt reading excluded, and is kept on the models page under its own label. The gap between that chip time and this end-to-end time comes from prompt reading. Dated history: I retired Qwen3.6-35B-A3B on 2026-11-06, and nothing serves it now.
Real traffic over 22 replies, chip median from the server log: 7.52 tok/s. The engine alone measures 123.0 ms; the 8 ms gap to the server is not explained yet. Dated history: Qwen3.8-27B is retired (2026-11-06), and no server runs it now.
This project does not ask to be believed. Each card names its environment, its gate, its date, and its verdict, and states plainly what the receipt proves and what it does not prove. That includes the failures, especially the failures.
Dated evidence, not a live dashboard: these preserved receipt cards describe the runs and branch states recorded in the site inventory of 2026-10-19. Words such as today and newest inside a card refer to that dated record, not the current machine. The Qwen models in these cards were retired on 2026-11-06, and none of them is served now. Research results below retain the canonical record's own date.
MEASURED
7.63 tok/s · 27B on Omega
A 27B model answered through my own server
Qwen3.8-27B (FP8) served by aien-serve-http on my own Omega GPU kernels, on the Spark. 131.05 ms per token, the median of 4 runs of 256 greedy tokens, 0.752 of the 10.15 tok/s memory ceiling. The old binary in the same folder did 6.29.
Proves
The 27B answered through my own server and kernels at this speed on 2026-10-19, with the runs published beside the ceiling they sit under.
Does not prove
That the number is on main yet (it is on a branch), or anything about any other machine.
A fresh Linux account followed the public guide to install and run the 27B server. The install script failed at the server build step. With a hand workaround, the remaining steps passed: 81 of 81 checksums, the gate, and 7.657 tok/s with answers matching the shipped reference.
Proves
The guide is almost there, and exactly where it breaks is on the record with the fix named.
Does not prove
That anyone can install from the guide today. The verdict is FAIL until the script is fixed and the test is run again.
Date
2026-10-19
Environment
Spark GB10, new user, branch alpha/fresh-account-evidence
aienux, the AienOS C kernel, passes all 24 emulator gates at commit 7f0de7e, each one a script anyone can rerun, with the receipt on main. The gate script now lists 26, so a newer receipt is owed.
Proves
The kernel reaches every gated stage under emulation at the recorded commit.
Does not prove
Anything physical. QEMU qualifies nothing physical, by the repository's own words.
The AArch64 C kernel ran on a Raspberry Pi 5: interrupts, all four cores, user-mode tasks, scheduling, message passing, and a log written to the SD card and read back. Pull requests #345 and #346, evidence on main.
Proves
The kernel's core services run on physical hardware, not only in the emulator.
Does not prove
A resident kernel. Each run writes a fresh SD image; release qualification (PI5-8) is still open.
The x86_64 port of the C kernel booted an Intel iMac 2017: all cores started, both disks read, Ethernet seen. The whole screen is Aien's window. The kernel the iMac runs today sits in the open pull request stack (#414 to #424).
Proves
The port runs on the physical machine with its cores, disks and network found.
Does not prove
That today's iMac kernel is on main. The evidence is; the newest kernel is in open pull requests.
Date
2026-10-17
Environment
Intel iMac 2017, bare metal, netboot from the Spark
Environment: Omega verification suite, turing-yield CI on main. Receipt source: aien-dev/omega · PR #88.
PASS TY-2
TY-2 Turing gain
2,559,679.825 bits of Turing gain
This is a gain in bits, not a yield. Yield is T/J (Turings per joule), and T/J is not established: no energy was measured.
On held-out data, the candidate model C described the events in about 2.56 million fewer bits than the order-1 baseline B, about 0.4387 bits per event.
Scope, limits and receipt
Status
PASS (verdict=PASS)
Scope
held-out predictive compression, Turing Yield profile V0; candidate C (op + previous 4 outcomes) vs baseline B (order-1); fit seeds 1-7, held-out 8-10; 5,834,845 held-out events of crumbline control-learner search events
Claim level
information part of H5 only; sits below the L1 chain (separate profile)
Limitations
Ideal code length only, no decodable bitstream. Verify uses the same tool (self re-derivation, not independent). Not a scientific discovery measure. No energy. Post hoc op+prev1..5 scored higher (+2,722,851.799), not counted. Check: dl_b_ub 7484024867446 minus dl_c_ub 4924345042631 = 2559679824815
Hardware is treated as capability metadata, never as identity. The identity is the lineage: who wakes the machine, what keeps it alive, what checks its work, and what learns from the results.
These are responsibilities, not a literal turn order. AEGIS, FORGE, ARGUS and CORTEX serve the system by verifying, realizing, observing and remembering. Nothing receives authority merely by advancing through a diagram.
ATLAS
verb
awakens
text
The seed that wakes AienOS on a machine. Atlas carries the lineage to new hardware and brings the sovereign stack to life on it.
The operating-system layer of Sovereign Intelligence. It owns the hardware, keeps the agent alive and continuous, and holds memory, keys, and receipts on the owner's side of the line.
Aien's language for all computing, GPU to CPU. Its own tools check every program and compile it to the machine code of the chip that runs it. Replacing CUDA is not true yet. Nothing runs on a claim; everything runs on a check.
The persistent intelligence itself. It learns from verified experience: what survived Omega's checks and contact with reality becomes knowledge. Weights suggest, programs explain, evidence teaches.
Most AI asks you to trust a demo and a promise. This project publishes its science in the open instead: the Turing, a unit of net held-out explanatory compression (one operational component of machine understanding), plus the papers behind it and the data anyone can recheck. It starts with the Turing.
The Turing (T)
A unit of net held-out explanatory compression
One Turing is one bit of net held-out description-length improvement: a short explanation that predicts things it has never seen earns Turings. Guessing earns nothing. Memorizing earns nothing, because the test data is new.
The Turing measures net held-out explanatory compression: one operational component of machine understanding.
Current level: L2, Brownian scope only, instrument certification. Outside Brownian scope the ceiling is L1. TY-2 does not change the ceiling.
The scorer behaves as specified on known stochastic cases. It does not mean AIEN discovered anything.
A simulation that runs entirely in your browser. No live machine is attached to this page, and nothing you type leaves it. It shows the shape of the interface the real machine speaks. The real machine's proof is in the evidence section, receipts and all.
[*] Capability vault sealed. Zero disk secrets invariant active.
[*] Type a syscall below or click one of the quick trigger buttons above.
Recorded project state
The canonical layer record was updated 2026-10-01. These dated states preserve their original evidence scope; they are not an assertion that a past release or implementation note describes the current checkout. For the current product, use the Alpha release page and the release record. status.json and llms.txt use this same canonical record.
Omega foundation gates
status
QUALIFIED
note
Gates 1 through 14 closed with one combined admission receipt: 230 checks, 0 failures.
Implemented (instrument); science partial. Current level: L2, Brownian scope only, instrument certification. L2 verdicts were checked against the private sealed records on 2026-10-01; the records stay private. Outside Brownian scope the ceiling is L1. TY-2 recorded 2,559,679.825 bits of Turing gain. This is a gain in bits, not a yield. Yield is T/J (Turings per joule), and T/J is not established: no energy was measured.
Facts gathered at 2026-10-16T16:29:47Z. This is the checked-in public record read at build time, not live telemetry. Repository activity does not establish hardware qualification.
aienux
key
aienos
repo
aien-dev/aienux
role
The AienOS kernel: boot, scheduling, IPC and capability boundaries. Agent-first; the model is never the kernel.
Watch the repositories, read the receipts, open an issue. Release artifacts and installation instructions are on the Alpha page; source, qualification and deployment are separate states.