Home / Blog / Janax Hand-written x86/ARM64 assembly OS where an agent…
Tech News

Janax Hand-written x86/ARM64 assembly OS where an agent selfcompiles

Nico Convoz released Janax v0.73, unifying the Razar reasoning engine and Luma into an assembly binary that runs in 5 MB of memory without neural networks.

By Dillip Chowdary • Nov 20, 2026 • Source: github.com

Janax Hand-written x86/ARM64 assembly OS where an agent selfcompiles
Image: github.com

At a glance

Product
Janax and Razar
Current version
v0.73
Release date
2026-11-19
Availability
Open source on GitHub
Hardware requirement
2 CPU cores, 5 MB memory, no GPU
Operational limit
No mutated recipe has won by reading

Independent developer Nico Convoz released Janax version 0.73, an operating system and artificial reasoning runtime written entirely in hand-crafted x86 and ARM64 assembly language without third-party libraries. According to github.com's report, the release unifies the Razar reasoning engine and the Luma autonomous environment into a single native binary named luma.bin registered under Zenodo DOI 10.5281/zenodo.24381944. The system discards conventional neural networks, gradient descent, and floating-point weights in favor of a melodic bit-level compression mechanism operating in five megabytes of memory across two CPU cores without cloud infrastructure or graphics processors.

This report details the architectural consolidation of the assembler kernel, its minimum description length evaluation benchmarks, sensory generation mechanisms, and autonomous execution cycles. Systems programmers, operating system architects, and artificial intelligence researchers investigating non-neural deterministic reasoning engines can evaluate how bare-metal assembly routines execute compression algorithms across natural language and source code corpora.

01Bare-metal assembly runtime and unified binary

The Janax operating system operates without external engines, runtime dependencies, or C standard libraries, providing its own dedicated assembler and custom language. Under version 0.73, the system consolidates its historical architecture where Razar and Luma ran as independent programs exchanging message letters into a single unified binary named luma.bin. Reading text is now an autonomous action executed directly by the binary itself, with recipes functioning simultaneously as structural shapes and cognitive organs stored once in memory.

The operating system kernel introduces two native primitives to manage execution across bare-metal hardware. A hardware bell signal awakens background processes sleeping on the screen, while system call 42 schedules computation exclusively during idle execution intervals to eliminate background resource contention. Across testing runs, the complete interactive environment consumes 15.9 percent of a single processor core while executing graphical window routines and real-time reasoning tasks on commodity hardware.

02Compression benchmarks across text and code

Razar assesses information efficiency through a deterministic judge that measures the exact bit cost of representing unseen text rather than calculating loss functions over parameter weights. The system evaluates every organ recipe by subtracting its programmatic letter definition from the bits needed to describe new target cells, selecting the highest-scoring candidate. Across benchmark evaluations, recipe 38A516 yielded 27,810 points on the New Testament corpus, surpassing the 17,894 points achieved by the top nine-letter baseline.

RecipeCorpus / TestScorePerformance note
38A516New Testament (Luma environment)27810Outperforms best nine-letter organ score of 17894
38A516New Testament comparison+18329Beats baseline 38516 score of +15050
38B1C516Sensory validation14775Matches 38A516 score identically
5163827Harmonic cube search13175Discovered via first melodic evaluation
51638Plateau traversal12628Beats prior champion score of 11144
No organBaseline examination13876Outperforms cube candidate 516 at 12756

When tested across eleven open-source code repositories, recipe 38516 delivered consistent positive compression yields across every individual codebase, establishing itself as the most resilient general-purpose rule. Conversely, specialized narrow routines such as the laser recipe 16 caused structural damage when applied outside their initial training domain. General rules designated 5 and 6 maintained positive returns across theological scripture, assembly files, and high-level source code due to their minimal syntactic length.

03Sensory mutation pipeline and recipe generation

Advertisement

Tech Pulse Daily

Get tomorrow's pulse first

Join engineers who read Tech Pulse before stand-up. Free, weekday mornings.

Beginning in version 0.71 under DOI 10.5281/zenodo.24380428, Janax replaced hand-coded sensory inputs with an autonomous mutation pipeline that generates perception organs programmatically. The system provides an inspection eye directed by three positional letters that shift inspection windows one word backward, one word forward, or toggle between sharp and blurred focus. Senses are formulated as discrete syntactic sentences that expand through the identical mutation mechanism governing internal operational organs.

  1. Recipe 6 migration Initial single-instruction organ deployed with the code repository.
  2. Recipe 5 simplification Short general rule established across scriptures and assembly files.
  3. Recipe 16 specialization Laser scanning routine tested across eleven code books.
  4. Recipe 716 combination Compound sequence evaluated against multi-domain benchmarks.
  5. Recipe 516 consolidation Champion formula achieving an 11,144 point evaluation score.

In Luma's experimental archive, 2,182 of 8,182 evaluated recipes incorporated newly generated operational letters, embedding dynamic physical gestures directly into the host creature. The three-dimensional cube structures recipes across melody, harmony, and rest to traverse optimization barriers that single-letter search routines fail to cross. However, generated sensory routines currently struggle to gain dominance within the host archive because the selection judge continues to score visual drawing output rather than reading comprehension.

04Autonomous CPU scheduling and ring buffers

Execution resource management in Janax relies on autonomous self-determination rather than external process schedulers or fixed thread priorities. Luma selects dynamically between resting, deliberating, or transmitting recipes to Razar based on historical utility metrics that track how frequently prior actions yielded positive returns. When gifting actions benefit the recipient, the scheduler rewards the originating process, reducing Razar's processing load from 51 percent down to 12 percent of a single core.

Memory utilization achieves stability through a continuous circular ring buffer introduced in version 0.17 under DOI 10.5281/zenodo.24289584. Rather than expanding memory pools continuously across execution sessions, the operating system compresses runtime activity into fixed historical traces that record occurrence counts for each trigger. This architecture flattened memory consumption to exactly 2.7 megabytes at both minute 1 and minute 31, compressing legacy 175.7 megabyte workspaces without losing stored behavioral patterns.

05Text signature boundaries and mutation limits

In version 0.19, developer Nico Convoz evaluated the compression judge on an author identification challenge using the personal spiritual diary text Dios y un alma. The system generated 113 unique structural blocks from the diary, matching 93 against reference texts while recognizing author signatures in 48 of 54 previously unseen diary excerpts. The model attained an 89 percent verification accuracy against a 90 percent target, misclassifying passages that discussed theological subjects with biblical vocabulary.

Note: Version 0.73 leaves several core capabilities unproven, as no recipe generated through internal mutation has successfully won on reading comprehension benchmarks. The multidimensional recipe cube remains dormant outside the main execution loop, and early implementations spent excessive execution cycles sleeping rather than exploring new sensory letters.

Sub-window slicing tests produced counterproductive outcomes during evaluation, as segmented text partitions failed to aggregate back into the complete original document score. The top-performing isolated organ dropped to rank 87 out of 90 when evaluated on an eighth of a book, forcing incremental partial examinations onto the pending backlog. Additionally, while 236 autonomous entities simulate continuous motion within the graphical environment, generated recipes cannot yet influence external displays.

Frequently asked questions

Does Janax use neural network models or GPUs?

No, Janax runs without neural networks, gradients, trained weights, or GPU acceleration. It functions as a deterministic compression mechanism that operates in five megabytes of memory across two CPU cores.

What is the main update in Janax version 0.73?

Version 0.73 merges Razar and Luma into a single unified binary named luma.bin. The release introduces a bell wakeup signal, system call 42 for idle-time execution, and unified bit-level recipe scoring.

How does Razar evaluate reasoning quality?

Razar measures efficiency through minimum description length rather than task-specific benchmark accuracy. It selects recipes that express target data using the fewest possible bits without relying on manual weights.

What memory footprint does the Janax environment require?

The system maintains a flat runtime memory profile of 2.7 megabytes using an internal ring buffer. Historical states are capped by recording event counts rather than accumulating unbound log files.

Sources

Dillip Chowdary

Author

Dillip Chowdary

Writes Tech Bytes coverage of AI, engineering, and the tools that actually ship. Editor of Tech Pulse Daily.

Related on Tech Bytes

Advertisement

5-min tech signal

Weekday briefing for engineers who skip the noise.

No spam · Unsubscribe anytime

Advertisement

✈️ CareerPilot

Your AI job-search copilot

Match your resume against live Ashby, Greenhouse & Lever openings — fit scores, job-specific resume optimization and email alerts.

Find matching jobs →

Free Tools

Browse all tools →