NM Project indexCode and recorded results

Public project catalog · 15 October 2026

Projects and
experiments.

I build experiments and tools across integer computing, Linux isolation, encrypted transport, GPU search, visual mathematics, and technical writing. Each brief explains what was built, the evidence recorded in its repository, and the limits of the result.

Browse all projects Compact table Source repositories

Suggested entry points

Start with these four

Four different kinds of work: reproducible AI arithmetic, embedded navigation, operating-system isolation, and GPU number-theory search.

Integer & embeddedActive

int-llm

Pure-integer C experiments for deterministic language-model training and inference, validated across desktop, embedded, and emulated targets.

Evidence80/80 greedy tokens matched the float reference; the committed arithmetic hash was reproduced across the recorded x86, Arm, RISC-V, Xtensa, and AVR targets.
Reference agreement
80/80 tokens
Determinism
5 architectures
  • C99
  • Q16.48 fixed point
  • Llama
  • MicroGPT
Project brief
Integer & embeddedActive

astro-nav-int

A no-floating-point celestial-navigation library that carries a sextant observation through correction, sight reduction, and position fixing.

EvidenceAltitude and primary-method azimuth outputs are documented within 0.30 arcminute of the double reference, with independent ephemeris and published-source gates.
Accuracy gate
≤0.30 arcminute
Independent references
5 classes
  • C99
  • Q16.48 fixed point
  • CORDIC
  • celestial navigation
Project brief
Secure systemsMaintained

compartment

A small Linux isolation toolkit with two zero-dependency C tools, one policy format, worked profiles, and explicit security boundaries.

EvidenceBPF v0.8.0 documents 3,284 enforcement trials and 44 bypass scenarios on two kernel versions; the core has separate rootless and real-root suites.
BPF core
11,341 lines
BPF tests
22,322 lines
  • Linux
  • Landlock
  • seccomp BPF
  • namespaces
Project brief
Computational searchCompleted

cunningham-chain-search

A visualization-led number-theory project that grew into a staged GPU filter, CPU confirmer, campaign toolkit, and published dataset.

EvidenceThe campaign produced two first-kind CC18 roots and a cumulative dataset of 1,551,489 CC10–CC18 roots; it explicitly does not claim the first known CC18.
Production GPU
4,280 lines
Production CPU
3,331 lines
  • CUDA
  • GMP
  • PARI/GP
  • Python
Project brief

Complete index

All projects

Filter by area or search by name, method, language, or result.

22 projects shown

Integer & embeddedActive

int-llm

Pure-integer C experiments for deterministic language-model training and inference, validated across desktop, embedded, and emulated targets.

Evidence80/80 greedy tokens matched the float reference; the committed arithmetic hash was reproduced across the recorded x86, Arm, RISC-V, Xtensa, and AVR targets.
Reference agreement
80/80 tokens
Determinism
5 architectures
  • C99
  • Q16.48 fixed point
  • Llama
  • MicroGPT
Project brief
Integer & embeddedActive

astro-nav-int

A no-floating-point celestial-navigation library that carries a sextant observation through correction, sight reduction, and position fixing.

EvidenceAltitude and primary-method azimuth outputs are documented within 0.30 arcminute of the double reference, with independent ephemeris and published-source gates.
Accuracy gate
≤0.30 arcminute
Independent references
5 classes
  • C99
  • Q16.48 fixed point
  • CORDIC
  • celestial navigation
Project brief
Secure systemsMaintained

compartment

A small Linux isolation toolkit with two zero-dependency C tools, one policy format, worked profiles, and explicit security boundaries.

EvidenceBPF v0.8.0 documents 3,284 enforcement trials and 44 bypass scenarios on two kernel versions; the core has separate rootless and real-root suites.
BPF core
11,341 lines
BPF tests
22,322 lines
  • Linux
  • Landlock
  • seccomp BPF
  • namespaces
Project brief
Computational searchCompleted

cunningham-chain-search

A visualization-led number-theory project that grew into a staged GPU filter, CPU confirmer, campaign toolkit, and published dataset.

EvidenceThe campaign produced two first-kind CC18 roots and a cumulative dataset of 1,551,489 CC10–CC18 roots; it explicitly does not claim the first known CC18.
Production GPU
4,280 lines
Production CPU
3,331 lines
  • CUDA
  • GMP
  • PARI/GP
  • Python
Project brief
Secure systemsActive

deaddrop

A Go client and relay for short-lived encrypted handoffs, with capsule pairing, optional end-to-end identity, and a detailed protocol specification.

EvidenceTiered CI and QA cover protocol vectors, bootstrap flows, one-shot semantics, collision handling, crash behavior, and exit-code contracts.
Production Go
7,157 lines
Test and QA
12,813 lines
  • Go
  • XChaCha20-Poly1305
  • X25519
  • Argon2id
Project brief
Secure systemsActive

URTB

A C terminal tunnel using shared capsules and XChaCha20-Poly1305 across ESP-NOW, LoRa, Unix sockets, and generic process streams.

EvidenceSoftware checks run without radio hardware, while a separate hardware tier validates two Heltec V3 boards and jump-host scenarios.
Implementation
8,737 lines
Test and harness code
5,537 lines
  • C
  • ESP-NOW
  • LoRa
  • XChaCha20-Poly1305
Project brief
Secure systemsActive

ephrun

A C toolchain that encrypts Linux executables, sources decryption keys from capsules, files, or kernel keyrings, and executes from memory.

EvidenceVersioned encrypted-file and capsule formats have round-trip, compatibility, threat-model, and deterministic vector acceptance gates.
Implementation
2,727 lines
Tests and harnesses
2,173 lines
  • C
  • libsodium
  • Linux keyrings
  • memfd
Project brief
Secure systemsActive

Unmoor

A C terminal and file-transfer tool with hybrid key exchange, byte-stream repair, multiple UDP legs, and explicit confinement controls.

EvidenceSeptember 15 records report 136 local checks and 781 routed-network checks; the random-loss matrix completed 99 of 100 sessions under its exact-line oracle.
C source + headers
75,367 lines
Routed-path gate
781 checks
  • UDP
  • X25519 + ML-KEM-1024
  • stream repair
  • Linux confinement
Project brief
Integer & embeddedResearch

int-llm precision ladder

Follow-up experiments that narrow integer model weights while retaining the wide Q16.48 path as the comparison oracle.

EvidenceA mixed-F11/F12 TinyLlama candidate matched 632/632 evaluated greedy decisions and reduced the wide integer model file from 8.80 GB to 2.20 GB.
Oracle agreement
632/632 decisions
Conversion gate
1,100,048,384 codes
  • fixed-point inference
  • int16 weights
  • TinyLlama
  • MicroGPT
Project brief
Integer & embeddedResearch

int-llm coordinate permutation

A toolkit for controlled model-layout experiments on both standard SafeTensors and the integer Q16.48 model path.

EvidenceThe all-axis TinyLlama run transformed all 201 learned tensors; the integer path kept 80/80 decisions and the BF16 inverse restored the checkpoint byte-for-byte.
Tensor coverage
201 learned tensors
Integer behavior
80/80 decisions
  • SafeTensors
  • Llama
  • parameter symmetry
  • Q16.48
Project brief
Integer & embeddedCompleted

int-llm SQL

SQL executes model loading, transformer arithmetic, deterministic sampling, and regression checks without custom functions or extensions.

EvidenceThe recorded C-oracle comparison covers 69,748 activation, logit, and probability rows, plus all 122 sampling records.
Oracle trace
69,748 rows
Arithmetic gate
11,750 vectors
  • SQLite
  • recursive CTEs
  • Q16.48
  • MicroGPT
Project brief
Integer & embeddedCompleted

int-llm PostScript

A complete small transformer and trainer in PostScript, with exact wide and F12 checkpoint verification through Ghostscript.

EvidenceThe repository records 20/20 identical samples, all 5,000 training loss lines, and byte-identical wide and F12-exported checkpoints.
Inference agreement
20/20 samples
Training agreement
5,000 steps
  • PostScript
  • Ghostscript
  • Q16.48
  • MicroGPT
Project brief
Integer & embeddedCompleted

int-llm Brainfuck

Generators and a generic byte-cell runner for integer-only transformer inference and training in Brainfuck, with recorded comparisons against C.

EvidenceThe September 18 training record reports exact agreement on all 5,000 loss lines, 20 samples, and both exported checkpoints against the C oracle.
Generated inference
16.40 GB
Generated trainer
210.30 GB
  • Brainfuck
  • Q16.48
  • MicroGPT
  • code generation
Project brief
Integer & embeddedResearch

int-llm visualizations

A weight atlas, inference trace, embedding-space viewer, and 3-D weight explorer for the complete small integer-model checkpoint.

EvidenceThe browser re-derives all 122 recorded inference steps with BigInt; the C trace reports zero logit, cache, or sampling mismatches.
Visual surfaces
4 offline pages
Checkpoint scope
14,272 weights
  • BigInt
  • Canvas
  • Q16.48
  • exact visualization
Project brief
Computational searchActive

k-tuplet-search

A code-first prime-tuplet search engine with a generated pattern catalog, record replay corpus, campaign controls, and browser analyzer.

EvidenceThe production CUDA engine records 52 passing tests and replays a sourced corpus of 227 known records across k=16–21; no new-record claim is made.
Production GPU
7,041 lines
CPU reference
2,749 lines
  • CUDA
  • GMP
  • PARI/GP
  • prime k-tuplets
Project brief
Computational searchCompleted

cunningham-chain-data

A compact data repository with schema, manifest, sample, conversion script, and full cumulative snapshots published as release assets.

EvidenceThe March 19 cumulative snapshot contains 1,551,489 roots across CC10–CC18, including two CC18 roots, in a 19 MB compressed release asset.
Published snapshot
1,551,489 roots
Longest results
2 CC18 roots
  • CSV
  • dataset schema
  • release assets
  • Python
Project brief
Visual mathematicsCompleted

four-square-balance

Code, verified data, and an interactive shape-space view supporting the published OEIS sequence A396760.

EvidenceOEIS A396760 has 37 exhaustively verified record locations below 10^9, with independent code paths and retained decompositions.
Published sequence
OEIS A396760
Exhaustive range
n < 10^9
  • number theory
  • OEIS
  • PARI/GP
  • OpenMP
Project brief
Visual mathematicsCompleted

three-square-balance

Independent search paths, verified data, and an interactive visualization supporting the published OEIS sequence A395008.

EvidenceOEIS A395008 has 65 exhaustively verified record locations below 10^9, checked by two differently ordered C searches plus PARI/GP and Python references.
Published sequence
OEIS A395008
Exhaustive range
n < 10^9
  • number theory
  • OEIS
  • PARI/GP
  • OpenMP
Project brief
Writing toolsMaintained

Spec Mirror

Review rules, examples, and a compact prompt generator for making components, behavior, terminology, and requirements explicit.

EvidenceThe JSON defines 51 tags across seven layers, with detection signals, rewrite rules, examples, and two recorded project corrections.
Review taxonomy
51 tags
Review dimensions
7 layers
  • technical writing
  • JSON taxonomy
  • Python
  • requirements review
Project brief
Writing toolsMaintained

Rhetorical Mirror

A small self-editing aid for plain posts and emails, combining pattern tags, a review prompt, and before/after README evaluations.

EvidenceThe committed taxonomy contains 41 tags in seven layers, with two README self-evaluations and the earlier source draft retained.
Pattern taxonomy
41 tags
Pattern families
7 layers
  • writing review
  • JSON taxonomy
  • prompt design
  • voice preservation
Project brief
Visual mathematicsArchived

p2 odd rays

A small Python math-art project that maps positive integers into equivalent shell-and-ray views using integer and bitwise operations.

EvidenceThe repository includes the generator, SVG/PNG outputs, checksums, and two equivalent visual constructions; it makes no research or performance claim.
Equivalent views
2 constructions
Reproducible outputs
SVG + PNG
  • Python
  • 2-adic decomposition
  • bit reversal
  • SVG
Project brief
Visual mathematicsArchived

power-two-square-rays

The former repository location is retained as an archive notice; its exploratory code and notes moved to p2-odd-rays-visual-math-toy.

EvidenceGitHub marks the repository archived; its current tree contains only a README directing readers to the replacement repository.
Current tree
1 file
Repository state
Archived
  • project archive
  • visual mathematics
Project brief

How to read the evidence

Follow the result to its source.

Each brief links to the exact repository revision reviewed. Test and benchmark results are attributed to the project's records; refreshing this catalog does not rerun hardware, GPU, network, or long training campaigns.

Physical line counts include comments and blank lines within named file scopes. The source inventory records those files. Counts describe scale, not quality or test coverage.

  1. 01
    Scope

    What the project does and the conditions under which it was evaluated.

  2. 02
    Implementation

    The languages, algorithms, and system components used.

  3. 03
    Checks

    Tests and comparisons recorded in the repository.

  4. 04
    Measurements

    Reported results, with units and file scopes where applicable.

  5. 05
    Limits

    Known failures, constraints, and what the results do not establish.