A measured agent-state protocol. Every claim here was tested, including the ones that failed; the failures and the fixes are in the open.

AB-1 is a deterministic 8-bit semiotic layer over the Unicode Braille Patterns block (U+2800–U+28FF) for machine-to-machine communication of AI agent state. One code point encodes one machine state across eight orthogonal dimensions of agency. This repository is the reference implementation, the benchmark harness, and the honest evidence ledger behind the paper.

The code is the canonical specification. Where the prose and the code disagree, the code wins.

What survives measurement (and what didn't)

This project's credibility is the arc, not a list of wins:

• Atomicity — claimed, falsified, fixed with receipts. The premise that Braille cells are atomic single tokens is false on every stock production tokenizer (~3 tokens/cell on cl100k/o200k; BERT maps all 256 to [UNK]). AB-1 ships a vocabulary extension that makes every cell exactly one token, 256/256, exact round-trip, measured before/after.
• Hardened lexicon — proven. Exhaustive enumeration over all 256 states and all single-bit flips: a single-parity-check command code (128 commands, distance 2) is single-error detecting with zero unvetted→audited promotions; an extended Hamming [8,4,4] code (16 commands, distance 4) is single-error correcting. Tokenizer- and model-independent.
• Token efficiency — triangulated on public reproducible data. Against a steelman (delta-encoded JSON, same emit-on-change discipline), AB-1 carries agent state tracking in ~92% fewer tokens on stock cl100k/o200k, ~97% with the extension. Converged across a synthetic sweep, a private agent ledger, and the public Crucible Claude Code session log (anyone can reproduce the public number).
• Separability — a scoped probe, explicitly NOT load-bearing. Characterized on three axes (quant depth F16→Q2_K, size 1B→14B, controlled cross-family at 3B/F16). No stock model robustly encodes AB-1 bit-structure at any size/quant/lineage; degradation appears only at the 2-bit extreme. This motivates the model-independent mechanisms; it is not the spine.
• Security syntax-firewall — an explicit hypothesis, not a result.

Layout

ab1/                 reference implementation (zero-dependency core)
  core.py            8-bit encoding, the bit-table convention
  dsp.py             Differential State Protocol (emit-on-change)
  crc.py             CRC-8 checksum cell
  lexicon.py         hardened command codes + Hamming analysis
  tokenizer.py       the vocabulary extension (atomicity by construction)
bench/               reproducible experiments; bench/results/FINDINGS.md is
                     the full honest ledger
tests/               spec-anchor + roundtrip tests
AB-1_CANONICAL.md    the consolidated specification (CC-BY-4.0)
paper/               the arXiv paper (LaTeX) + bibliography

Reproduce

python -m venv .venv && . .venv/bin/activate
pip install tiktoken transformers           # tokenizer experiments
python tests/test_core.py                   # 5/5 spec + roundtrip
python bench/tokenizer_parity.py            # atomicity (negative result)
python bench/extension_proof.py             # the fix, before/after
python bench/lexicon_proof.py               # the proof (deterministic)
python bench/token_reduction.py --trace bench/results/crucible_trace.json

Separability experiments (bench/quant_*.py, bench/cross_family.py) require local GGUF models and a GPU; they self-persist results to bench/results/ and survive a mid-run crash (incremental flush).

License

• Reference code: Apache-2.0 (LICENSE) — chosen so a standard can be adopted freely, with an explicit patent grant.
• Specification text (AB-1_CANONICAL.md, the paper): CC-BY-4.0.

A standard wins by adoption; the open core is intentionally permissive. Any proprietary downstream integration is licensed separately and is not part of this repository.

Citation

See CITATION.cff. Authored by Chris White (Octothorped Industries LLC).

Status / known-open

Stated plainly rather than omitted: the embedding-grounding (fine-tune) experiment is deliberately deferred (future work); the security firewall is a hypothesis with a proposed evaluation, not a measured result; the separability study is a single-metric pooled-embedding probe. None of these are the load-bearing claims.