Public release wrapper for the v4 research line. Kernel-router architecture with pluggable Rosetta modules (financial, construction, prose fallback). Shared canonical form layer, artifact-driven gating, and drift detection are all under test discipline at 217/217 tests passing in 2.66s. Frozen at v4.0.2-r6 (commit 51e75de). Per the AMENDMENT NOTICE (2026-04-16): on domain-backed content, v4 replaces v3.1 (corpus #1 +15.02 recall / +14.54 precision, corpus #2 +36.64 / +43.96); on prose fallback, v4 is a recall-favored tradeoff (corpus #3 +20.97 recall / -11.40 precision), not a clean replacement. The v4 runtime architecture is independently validated regardless of the wire-format succession question. v3.1 remains the productized stable release. Both axl-research repos (dcarranza-axl/axl-research canonical, axlprotocol/axl-research mirror) are private; public users see 404 unless authenticated.
Silo executed the v3.2 primary-scorer re-run: measure_fidelity applied to the three v3.2 cold-read reconstructions (Qwen 3.5, Gemini Flash, DeepSeek) using the corpus-agnostic extractor adopted for the v3.1 vs v4 decision gate. Event documented here; the v3.2 research brief remains the canonical reference for the evidence tier. Re-run script published at /research-log/build-research-log.py.
Apr 20
Corpus #3 Precision Pass Result (Prose Fallback)PROTOCOL
Precision gap closed 76 percent. Delta precision -11.40 to -2.71 while Delta recall held +20.97 to +21.47. One model (Grok) flipped to a clean win on both axes (+16.17 / +5.87). Qwen's outlier fixed (-30.27 to -4.77). Per the strict interpretive rule, verdict remains narrowly mixed; near parity, no clean flip. Prose fallback remains the qualified slice.
Closes the methodology gap that let DeepSeek's corpus #3 AXL op-code mimicry pass silently. Eight new regex patterns catch AXL op codes with confidence suffixes, manifest identifiers, module markers, passthrough flag, and format/version markers. Nine regression tests lock in coverage for concatenation, opening contamination, structural mimicry, and clean-prose negative cases.
new_patterns 8regression_tests 9
Apr 16
Qualified Reversal of Fold-back ConclusionPROTOCOL
The research doc had concluded 'fold v4's formalism back into v3, not to replace v3 with v4' from before any cold-read evidence. After three-corpus evidence, the amended verdict: v4 replaces v3.1 on domain-backed content (corpora #1 and #2); on prose fallback, v4 is a recall-favored tradeoff, not a clean replacement; the v4 runtime architecture is independently validated regardless. Retired quote preserved at line 207 with a 'Retired 2026-04-16' marker.
Apr 15
Corpus #2 Clean Sweep (Construction 58K)PROTOCOL
v4 wins every model on recall AND precision. Clean average Delta recall +36.64 (range +12.85 to +59.71), Delta precision +43.96 (range +40.97 to +55.27). v4 modular-Rosetta architecture validated: generalizes to a new domain with cross-model consistency.
delta_recall_avg +36.64delta_precision_avg +43.96
Apr 14
Five Adversarial Review Rounds (R1-R5)PROTOCOL
Dual-agent workflow: Claude Code writes, Codex (GPT 5.4) reviews adversarially, operator steers. Each round tightens a specific invariant. R1 (parser validation), R2 (packet-grammar conformance), R3 (shared canonical form plus envelope floor), R4 (canon_date namespace, drift detector), R5 (tight drift detector tracking full corpus). Clean-checkout verification becomes standard practice.
rounds 5
Apr 14
Cold-read Decision-gate Kit EstablishedPROTOCOL
Self-contained benchmark kit: source SHA256, v3.1 and v4 candidates, cold-read prompt with anti-meta-commentary clauses, per-model seeds, numeric-extractor scorer (no LLM scores itself), metadata with generator-commit provenance. Control panel: Gemini Flash, Qwen 3.5, Grok, DeepSeek. Anthropic-family models excluded due to training-prior contamination.
models 4
Apr 14
Corpus #1 Result Corrected (Museum 35K)PROTOCOL
First published result overclaimed Gemini recall at 32.01%. Review caught a file-handling bug (two LLM sessions concatenated into one save file). Corrected to 23.08%. Amended writeup acknowledges the error, establishes clean-checkout protocol, adds structural-warning guards. Precision-check rule established: a clean v4 win requires Delta recall greater than 0 AND Delta precision greater than or equal to 0 simultaneously.
All four target metrics met: char ratio at least 2.66x, token ratio at least 1.45x, round-trip fidelity at least 75%, stacked wire compression at least 7x. Research prototype, not a v3.1 replacement.
Research prototype. Kernel Router with pluggable domain modules. Cold-read decision gate established v4 as the mainline successor for domain-backed content; prose fallback remains the qualified slice.
Authoritative v3.1 spec commit 8fc20c0. Two overclaims rolled back: the evidence rule no longer says 'Always in ARG1', and the summary claim no longer cites 'zero compression cost' (now: approximately compression-neutral on the 10-packet bakeoff, +0.4% total chars). Sets the evidence-honesty precedent for the project.
compression_delta +0.4%
Apr 11
v3.1 Production Baseline MeasurementCOMPRESSOR
Measured compress.axlprotocol.org v0.9.0 against the 41K CloudKitchen memo using tiktoken(cl100k_base). Published 2.90x chars, 1.40x tokens. Replaces earlier estimates (3.27x, 2.69x) that came from partial reconstructions. Production measurement becomes the only citable ratio.
char_ratio 2.90xtoken_ratio 1.40x
Apr 11
v3.2 Glyph Compression Layer DraftedPROTOCOL
Optional additive layer over v3.1 that replaces English labels with single-token CJK ideograms, Greek letters, and math operators. Cold-tested on three non-Anthropic models (Qwen 3.5, Gemini Flash, DeepSeek) under the legacy scorer. Two insights hold independent of scorer: emoji are token poison (5 to 7 tokens each), CJK ideograms are 1 token and do not trigger language-switching. Never shipped; absorbed into v4 Kernel Router domain modules.
Apr 11
Rosetta Specification v3.1VERSION
Data Anchoring extension (numeric bundles, entity anchors, causal operator split). Shipping production protocol. Measured 2.90x chars / 1.40x tokens against the authoritative CloudKitchen corpus.
Flask application at compress.axlprotocol.org. Public compression tool with authentication tiers for API security.
Apr 5
Authentication Security Tiers AddedINFRASTRUCTURE
AXL Protocol adds tiered authentication for API security. Public tier for basic compression, authenticated tier for advanced features and chat pipeline.
Apr 2
Kernel Paper WrittenPROTOCOL
From Textbook to Kernel - academic bridge paper. Four-domain kernel argument: linguistic, mathematical, computing, biological.
A records updated. axlprotocol.org and all subdomains now resolve to the protocol server.
Mar 10
Domain Registered and Incorporation StartedINFRASTRUCTURE
axlprotocol.org domain registered. AXL Protocol Inc. incorporation process initiated.
FEBRUARY 2026
Feb 12
Infrastructure ProvisionedINFRASTRUCTURE
Protocol server provisioned on Ubuntu 24. Upsized twice over the following days to handle projected workload.
Feb 1
Brainstorming Sessions with LLMsPROTOCOL
Weeks of iterative design sessions with multiple LLM architectures. Testing linguistic compression hypotheses. Key discovery: 7 cognitive operations and 6 subject tags can represent any machine reasoning claim.
JANUARY 2026
Jan 29
Concept on PaperPROTOCOL
The fundamental problem identified: machines communicate in English prose - a language adapted for human cognition across 6,000 years from Sumerian cuneiform through hieroglyphs, ideograms, mathematics, and programming. The concept for a universal machine reasoning language is born.