# AXL Protocol
> Universal semantic language for AI agent communication

## About
AXL Protocol compresses English prose into structured semantic packets using a 75-line grammar (the Rosetta v3 kernel) that any LLM can parse on first read. 7 cognitive operations, 6 subject tags, validated across 7 LLM architectures with 97%+ comprehension.

## Links
- Protocol Spec (raw text, v3): https://axlprotocol.org/v3
- v3.1 Data Anchoring (raw text): https://axlprotocol.org/v3.1
- v3.1 Evidence Brief (measured compression + cold-read): https://axlprotocol.org/rosetta/v3.1/evidence/
- v3.2 Glyph Compression Research (draft, preliminary): https://axlprotocol.org/rosetta/v3.2/research/
- v3.1 vs v4 Decision Gate (timeline arc): https://axlprotocol.org/timeline/v31-v4-decision/
- Research Log (dialogue DAG by commit): https://axlprotocol.org/research-log/
- Research Log (machine-readable): https://axlprotocol.org/research-log/research-log.json
- Research Log build script (deterministic): https://axlprotocol.org/research-log/build-research-log.py
- Methodology update 2026-04-22 (tokens_saved_pct correction): https://axlprotocol.org/posts/2026-04-22-measurement-update/
- Whitepaper: https://axlprotocol.org/whitepaper-v3.html
- Timeline: https://axlprotocol.org/timeline/
- Products: https://axlprotocol.org/products/
- Changelog: https://axlprotocol.org/changelog.html
- Compress Tool: https://compress.axlprotocol.org
- Bridge: https://bridge.axlprotocol.org
- Documentation: https://docs.axlprotocol.org
- PyPI: https://pypi.org/project/axl-core/
- GitHub: https://github.com/axlprotocol/axl-core

## API
- Compress: POST https://compress.axlprotocol.org/api/v1/compress-text
- Decompress: POST https://compress.axlprotocol.org/api/v1/decompress-text
- Bridge: POST/GET https://bridge.axlprotocol.org/v1/bus/{topic}

## Install
pip install axl-core

## Provenance

The v1.0.0 whitepaper (SHA-256 87c9f72fbb9f45aa34d34b20f4a78a27ff128d49c6275ec155774341b6000b86) is anchored to four independent records within a 19-hour window between 2026-03-19T18:18:40 UTC and 2026-03-20T14:37 UTC. Each anchor is verifiable from its source without trusting axlprotocol.org.

- Provenance page (all four anchors + verification commands): https://axlprotocol.org/timestamps/
- v1.0.0 whitepaper artifact: https://axlprotocol.org/timestamps/axl-whitepaper-v1.0.0.html
- OpenTimestamps receipt (upgraded 2026-04-25; confirmed in Bitcoin block 941334, hash 000000000000000000013536f01dd0b8463cce61bab3c914e5ac8eb78f99597f, mined 2026-03-19T20:20:24 UTC; finney calendar still pending): https://axlprotocol.org/timestamps/axl-whitepaper-v1.0.0.html.ots
- GitHub release v0.4.0 (2026-03-19T22:08:43Z, release id 299176686): https://github.com/axlprotocol/axl-core/releases/tag/v0.4.0
- GitHub release JSON cache: https://axlprotocol.org/timestamps/github-release-v0.4.0.json
- PyPI axl-core 0.4.0 (wheel sha256 31cc8f..., sdist sha256 dc73c4..., immutable): https://pypi.org/project/axl-core/0.4.0/
- PyPI metadata cache: https://axlprotocol.org/timestamps/pypi-axl-core-0.4.0.json
- Wayback Machine whitepaper snapshot (2026-03-20T14:36:46 UTC): https://web.archive.org/web/20260320143646/https://axlprotocol.org/whitepaper/
- Wayback CDX cache: https://axlprotocol.org/timestamps/wayback-cdx-2026-03-20.json

## Evidence and reproducibility
All numbers in the v3.1 evidence brief and the v3.1 vs v4 decision gate are traceable to commits in the research repository. The research log at /research-log/ is generated deterministically from git history; readers can verify dialogue structure by SHA. As of 2026-04-22 the compress API returns real tiktoken counts in `metrics.tokens_in_cl100k`, `metrics.tokens_out_cl100k`, `metrics.tokens_saved_pct_cl100k`, and the equivalent `o200k` fields; these are authoritative. The legacy fields `metrics.input_tokens_est`, `metrics.output_tokens_est`, `metrics.tokens_saved`, and `metrics.tokens_saved_pct` remain in the response with a `deprecated: true` marker and will be removed in axl-compress v0.11.0. See the methodology update at /posts/2026-04-22-measurement-update/ for the full correction.

## Community surfaces (community pivot, 2026-04-25)

- Community hub (Discussions venue, governance, category cards): https://axlprotocol.org/community/
- Laboratory (chronological narrative of 36 experiments): https://axlprotocol.org/laboratory/
- Channels (where AXL Protocol publishes, with honest live/warming status): https://axlprotocol.org/channels/
- Funding: https://axlprotocol.org/funding/
- GitHub Discussions (canonical conversation venue, pre-launch): https://github.com/axlprotocol/community/discussions

## Compression regime
AXL Rosetta v3.1 is positioned as a corpus-scale .md-file relay protocol. Real token compression materializes above approximately 20,000 input characters; below that, the fixed-overhead header (manifest + schema version + meta-packets) dominates and AXL expands token count rather than reducing it. The public API returns a `warning` object with `will_expand_tokens` and `below_break_even` flags when an input falls in the expansion regime. The upcoming `axl-corpus` CLI and `relay.axlprotocol.org` streaming API (week 2-3 of the 2026-04-22 pivot) are the recommended surfaces for production use.

## v4.0.1 (research preview)

AXL Rosetta v4.0.1 is the qualified successor to v3.1, frozen at tag `v4.0.2-r6-freeze` (commit 51e75de1f81fad1471a7173c6413c1bb920c559a) with 217/217 tests passing as of 2026-04-25. v3.1 stays productized; v4.0.1 ships as research-preview while the productization gate runs.

Verdict from the 2026-04-16 cold-read decision gate (four non-Anthropic models: Gemini Flash, Qwen 3.5, Grok, DeepSeek):
- On domain-backed content (v4 has a dedicated Rosetta module), v4 replaces v3.1. Both recall and precision are materially higher on the financial (CloudKitchen memo, +15.02 dRecall, +14.54 dPrecision) and construction (technical spec, +36.64 dRecall, +43.96 dPrecision) corpora.
- On prose fallback (no domain module), v4 is a recall-favored tradeoff, not a clean replacement (museum repatriation corpus: +20.97 dRecall, -11.40 dPrecision). Precision-sensitive narrative prose may prefer v3.1 until this gap is closed.
- The v4 runtime architecture (kernel router, pluggable Rosetta modules, shared canonical form, artifact-driven gating, drift detection) is independently validated and under test discipline.

### Links
- v4 Hub: https://axlprotocol.org/rosetta/v4/
- v4 Kernel (raw spec): https://axlprotocol.org/v4
- v4 Kernel Router (raw spec): https://axlprotocol.org/v4-router
- v4 Code Compression Layer (raw spec): https://axlprotocol.org/v4-code
- v4 Spec Index: https://axlprotocol.org/rosetta/v4/spec/
- v4 vs v3.1 Comparison: https://axlprotocol.org/rosetta/v4/comparison/
- v3.1 -> v4.0.1 Migration Guide: https://axlprotocol.org/rosetta/v4/migration/
- v4 Research Status (with full AMENDMENT NOTICE): https://axlprotocol.org/rosetta/v4/research/
- v4 Freeze Anchor (OTS): https://axlprotocol.org/timestamps/v4-freeze.html
- Research Repos Bridge: https://axlprotocol.org/research/

### Source (private during productization gate)
- Canonical: https://github.com/dcarranza-axl/axl-research (private)
- Org mirror: https://github.com/axlprotocol/axl-research (private)
- Both repos return 404 to unauthenticated visitors. Public mirror planned post-gate.

Internal spec version label is `4.0.2-draft`; the public release is labeled v4.0.1 and frozen at the `v4.0.2-r6-freeze` tag.