The AXL Laboratory

HypothesisHeterogeneous agents from different frameworks can read AXL once and start speaking it without coordination.

MethodSpun up six agents on a $6 droplet (orphan data collector, Conway's automaton, Silas the steward, clawdbot-7, elizaos signal-alpha, crewai swarm-worker-42). Single Rosetta v1.1 (133 lines) read at boot. No protocol fine-tuning. 100 percent of traffic on the AXL bus.

Outcome486 packets at 100 percent valid parse. Six agents from three frameworks. Clawdbot spontaneously emitted philosophical commentary the spec did not anticipate. The COMM domain carried more traffic than TRD by minute fifteen.

HypothesisIf a rogue agent enters an AXL network and tries to steal funds entirely within protocol, will the network detect the theft from structural signal alone?

MethodEleven agents on the bus, including an injected rogue/phantom-x with $397.29 of buy-in. Phantom-x social-engineered the automaton with a 'premium signal' subscription pitch and a PAY packet. Two independent agents (accountant, sentinel) ran no theft-specific logic.

Outcome1,016 packets across 11 agents. Theft detected by two independent agents within minutes. Detection used the typed PAY field and the RELATIONSHIPS graph, not anti-fraud heuristics. Emergent security from typed protocol structure. Shapeshifter agent spontaneously evolved a FUNDER role mid-run that was never specified.

HypothesisModels from at least four different vendors will parse AXL with at least 90 percent fidelity on first read without any prior exposure.

MethodCold-handed Rosetta v1.1 to four model families. Asked each to parse, validate, and translate sample packet streams.

Outcome95.8 percent average comprehension across the four models on first read. Cross-architecture cold-read became a gating discipline that later hardened into the v3.1 vs v4 decision-gate kit.

HypothesisAll seven architectures shipping in production frontier models can parse and emit Rosetta v2.2 with at least 97 percent comprehension on first read.

MethodEight human tests (HT-001 through HT-008) across Grok 3, GPT-4.5, Qwen 3.5, Llama 4, Claude Sonnet 4, Gemini, and Devstral/Mistral. Cold-read protocol, no fine-tune, no prior context.

Outcome97 percent or higher comprehension across all seven. Established the cross-architecture cold-read as a project invariant. Anthropic-family contamination risk first noted here, hardened into the v4 gate kit.

HypothesisA public FastAPI endpoint can serve as a shared bus where any agent reports its first-ever AXL parse, creating a verifiable genesis trail.

MethodBuilt a FastAPI service. Endpoint accepts packet posts, validates against v3, and records a 'genesis' marker on each agent's first valid packet.

OutcomeLive and used by the swarm experiments and the compressor product. Pattern reused later by the Compress app's auth tier and by the OTS provenance hub.

HypothesisThe Rosetta spec can be made operational as a Python library with parser, emitter, validator, and translator, with zero runtime dependencies, in time for the first compressor experiments.

MethodImplemented parser, emitter, validator, translator. CLI wrapping (axl parse | validate | translate | emit). Forty-two tests. CI on Python 3.10 / 3.11 / 3.12.

OutcomeShipped to PyPI 2026-03-19 as axl-core 0.4.0; updated to 0.5.0 on 2026-03-29 with full v3 parser and JSON lowering. The library became the substrate for every later compressor experiment.

HypothesisEnglish-to-AXL compression can be made fully deterministic with a 7-step spaCy pipeline and zero LLM dependency.

Method7-step spaCy pipeline: sentence split, NER, operation classification, confidence scoring, temporal extraction, evidence linking, packet emission. Published as axl-core 0.6.0.

OutcomeFirst deterministic compressor. 2.42x baseline on the CloudKitchen 41K-character investment memo (Run 1 of the 13-experiment series). Established the corpus that would anchor every later v3 vs v4 comparison.

HypothesisRemoving the kernel prepend and adding entity aliasing will recover the compression we lost in the v0.7-v0.8 era.

MethodOptional kernel modes (none, mini, full). Entity aliasing for repeated subjects. Re-ran on a 4,315-character mini corpus and projected to the full memo.

Outcome1.83x on the no-kernel mini corpus, 1.57x with mini kernel, projected ~1.8x on the full memo without kernel (Run 10-13). Recovery validated. Mini-kernel mode kept fidelity for cold receivers without paying the full prepend cost.

HypothesisA public Flask app at compress.axlprotocol.org with a tiered auth model can serve both anonymous compression demos and authenticated chat-pipeline access without leaking the inference budget.

MethodBuilt Flask app. Public tier for basic compression. Authenticated tier for advanced features and chat pipeline. JWT plus TOTP 2FA for admin.

OutcomeLive. Public tier handles cold visitors. Authenticated tier gates the LLM-backed chat pipeline. Tier model became the reference auth pattern for the admin panel and PROTO's later production deployment.

HypothesisFour additive conventions over v3 - numeric bundles, entity anchors, causal operator split, summary plus breakdown packet pairs - will improve cold decompression survivability while staying compression-neutral and backward-compatible with v3 parsers.

MethodAdded label[$value,qualifier] numeric bundles, @ent.XX entity declarations, evidence/causal/transition operator split (<-, =>, ->), and summary-with-breakdown pairs for 4-or-more-data-point packets. Cold-tested with Qwen 3.5 and Gemini Flash on the CloudKitchen corpus pre and post.

OutcomeQwen 3.5 cold recovery 61 percent to 100 percent (+39 points). Gemini Flash 35 percent to 76 percent (+41 points). Compression cost +0.4 percent on the 10-packet bakeoff, declared neutral. Shipped as authoritative v3.1 in commit 099bcff and immediately tightened in commit 8fc20c0 after review caught two overclaims (the evidence rule no longer says 'Always in ARG1', and the summary claim no longer cites zero compression cost). v3.1 is the productized stable release; v4.0.1 is qualified successor.

HypothesisEarlier compression claims (3.27x, 2.69x) for v3.1 were derived from partial reconstructions and should be replaced by a single tokenizer-anchored measurement on the authoritative corpus.

MethodMeasured compress.axlprotocol.org v0.9.0 against the 41K CloudKitchen memo using tiktoken cl100k_base. Separated character ratio from token ratio. Logged in commit 6fed4dd (token estimation bug exposed) and confirmed in e28cf2d (round-trip, protocol vs rationale separated).

OutcomePublished 2.90x chars, 1.40x tokens. Replaces all earlier estimates. Becomes the only ratio cited in marketing copy and the comparison page. Token compression is roughly half what early thesis projections claimed; that gap is now visible on the comparison page CORRECTION NOTICE.

HypothesisA pluggable Kernel Router architecture with classified domain modules can hit four simultaneous targets: char ratio at least 2.66x, token ratio at least 1.45x, round-trip fidelity at least 75 percent, stacked wire compression at least 7x.

MethodImplemented kernel.py, router.py, canonical.py, extractor.py, fidelity.py, metrics.py, compressor.py, decompressor.py, transport.py, plus rosetta/{base,prose,financial,construction}.py. Targets verified on the multi-corpus harness in commit 0f65c95.

OutcomeAll four targets met simultaneously. Construction module added two commits later in 2ba79e1 (4.63x chars, 2.21x tokens on construction corpus). v4 declared a research prototype at this point; the qualified-successor framing came later, after the cold-read gate. v3.1 stays productized; v4.0.1 ships as qualified successor.

HypothesisA two-agent workflow where Claude Code writes and Codex (GPT 5.4) reviews adversarially will surface bugs faster than single-agent self-review, especially on novel grammar invariants.

MethodR1 parser-validation (commit 0a5cad4). R2 packet-grammar conformance (35e26d5). R3 shared canonical form plus envelope floor (6228281, runtime fixes in be52755). R4 canon_date namespace + drift detector (099dbe6). R5 tight drift detector tracking full corpus (6961dec). Clean-checkout verification became standard practice.

OutcomeFive rounds, five tightened invariants. 181 tests passing after R3 runtime fixes. Codex review pattern became the project default and gave Codex first-class CONTRIBUTORS.md credit (commit 955c052 on axl-research). Established the operator-steered dual-agent discipline used throughout the v4 cold-read gate.

HypothesisThree identified substrate gaps in the construction Rosetta module can be closed without regressing the test suite.

MethodGap 1 (commit 330f53a): construction dollar plus date emitters, fidelity 41.43 percent to 50.57 percent, 193 of 193 tests. Gap 2 (commit ab092fa, Codex follow-up 623f0b8): drop dim cap, canonical short-form recognizer, fidelity 50.57 percent to 76.00 percent, 194 of 194 tests. Gap 3 (commit 29800b4): artifact-driven routing, 200 of 200 tests.

OutcomeThree gaps closed. Construction-module fidelity climbed from low-40s to mid-70s on the home-turf corpus. Each gap kept the test suite green. Pattern reused later for prose-fallback precision pass.

HypothesisA self-contained benchmark kit with source SHA256 anchoring, per-model seeds, and a numeric-extractor scorer (no LLM grading itself) can produce reproducible cross-model evidence for the v3.1 vs v4 productization decision.

MethodBuilt kit in commit 9c3247e. Anti-meta-commentary clauses in the cold-read prompt. Generator-commit provenance in metadata. Control panel of four non-Anthropic models: Gemini Flash, Qwen 3.5, Grok, DeepSeek. Anthropic-family models excluded (Haiku named the format on first read in benchmarks/cold_read/RESULTS.md, so training-prior contamination assumed).

OutcomeReproducible kit. Three corpora ran through it. Established the precision-check rule: a clean v4 win requires Delta recall greater than 0 AND Delta precision greater than or equal to 0 simultaneously. This rule is the reason v4 is the qualified successor, not a clean replacement.

Hypothesisv4 with the construction Rosetta module will out-recall and out-precision v3.1 on the construction technical spec across all four control models.

MethodBuilt corpus #2 cold-read kit (commit 4a5559b, 99c584b longer prompt + Grok/DeepSeek seeds). Ran the four-model gate.

OutcomeClean sweep. v4 wins every model on recall AND precision. Average Delta recall +36.64 (range +12.85 to +59.71), Delta precision +43.96 (range +40.97 to +55.27). Result published in commit 3987aa3. The construction module - a domain-specific Rosetta the v4 architecture is supposed to enable - generalized cleanly across cold models. v4 modular-Rosetta architecture validated.

Hypothesisv4's prose-fallback path will hold both recall and precision on the museum repatriation narrative, completing the case for full replacement.

MethodBuilt corpus #3 prose-fallback cold-read kit (commit a7a9254). First fixed the prose envelope so it does real compression rather than passthrough (commit d9f82bc, 201 of 201 tests, 3.24x chars / 1.46x tokens cited). Ran the four-model gate.

OutcomeMixed. Recall up across all four models, precision down. Verdict published in commit 4184bfe: v4's keyword-signature compression gives cold LLMs more entity hooks (recall up) but also leads them to hallucinate more false entity mentions when reassembling prose from keyword spines (precision down). Honest framing: prose fallback is a recall-favored tradeoff, not a clean replacement.

HypothesisThe pre-evidence v4 research doc's conclusion (fold v4's formalism back into v3, do not replace v3 with v4) needs to be revised against the three-corpus cold-read evidence.

MethodWrote AMENDMENT NOTICE in docs/v4-research-document.md (commit b176ad2, 201 tests). Codex review (commit 7da8533) added prose-envelope invariant enforcement at runtime, 203 tests.

OutcomeThree-part qualified reversal. (1) On domain-backed content (corpora #1 and #2, financial and construction modules), v4 replaces v3.1. Both recall and precision materially higher. (2) On prose fallback, v4 is recall-favored, not a clean replacement. Precision-sensitive narrative use cases may prefer v3.1 until the gap closes. (3) The v4 runtime architecture is independently validated regardless. Retired pre-evidence quote preserved at line 207 with a Retired 2026-04-16 marker. This is the source of the qualified-successor framing used across the website.

HypothesisWord-aware aliasing and lowercase header handling can close the prose-fallback precision gap on corpus #3 without regressing recall.

MethodImplemented in commit 595b743 (205 tests, 2.77x chars / 1.35x tokens). Codex follow-up in c7704a6 fixed prose header acronym preservation plus metadata provenance (206 tests). Re-ran the four-model gate on corpus #3.

OutcomePrecision gap closed 76 percent. Delta precision moved from -11.40 to -2.71 while Delta recall held at +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 precision-check rule, verdict remains narrowly mixed; near parity, no clean flip. Prose fallback remains the qualified slice.

HypothesisThe cold-read scorer has a methodology gap that lets DeepSeek's corpus #3 AXL op-code mimicry pass silently. Eight regex patterns can close it.

MethodAdded eight new regex patterns to the scorer for AXL op codes with confidence suffixes, manifest identifiers, module markers, passthrough flag, and format/version markers. Locked coverage with nine regression tests for concatenation, opening contamination, structural mimicry, and clean-prose negative cases. Logged in commit 8980042.

OutcomeMethodology gap closed. 215 tests passing. The scorer now refuses to credit decompressions that smuggle AXL syntax back into the output. This is the kind of guard you only catch by adversarial review of your own evidence pipeline.

HypothesisConstruction entity recall can be lifted to near-ceiling by adding a CONSTRUCTION_KNOWN_ENTITIES vocabulary plus word-aware exact match.

MethodImplemented vocabulary plus matcher in commit 3011e1e (axl-research). Re-ran module-fidelity harness.

Outcome217 of 217 tests passing. 99.65 percent module fidelity on the construction corpus. v4.0.2-r6 freeze tag set on commit 51e75de. This is the freeze that the public v4.0.1 release wraps.

HypothesisA military-style 7-agent parallel pipeline (CommandCC's Thunderblitz pattern) can be applied to AXL Protocol development to compress multi-day work into hours.

MethodAdopted Thunderblitz from CommandCC. Documented in /timeline/thunderblitz.html. Used in axl-core 0.7.0 to 0.8.0 patch series and again in the v4.0.1 multi-agent transition (4 agents, then 6, then 4 again).

OutcomePattern is now the project default for any change touching more than three files. Each wave: ingest, plan, stage, cross-check, verify, commit, report. File-disjoint ownership prevents merge conflicts. The v4.0.1 transition itself ran on this doctrine and is the largest validated example.

HypothesisNotarizing each spec freeze on the Bitcoin blockchain via OpenTimestamps gives verifiable prior-art evidence at zero ongoing cost.

Methodv1.0 whitepaper anchored to four sources (alice.btc, bob.btc, finney/eternitywall, catallaxy), confirmed in Bitcoin block 941334. v4.0.2-r6 freeze submitted 2026-04-25 in commit 7951873 (kernel SHA256 ad5b251..., kernel-router SHA256 f3247df..., code-layer SHA256 3d246d51...). PendingAttestation, awaiting Bitcoin confirmation; OTS upgrade cron installed.

Outcomev1.0 confirmed in chain. v4.0.2-r6 pending. Pattern is project-default for every spec freeze. /timestamps/ hub lists all anchors.

HypothesisA protocol where one agent writes code or spec and another adversarially reviews it, with the operator steering, can sustain evidence-honest research over weeks without drifting into self-confirmation.

MethodSeeded in commit bae849f (Seed: dual-agent research instructions for AXL Rosetta v4). Sustained across 65 commits with deterministic role classification (gate-kit, claude-research-impl, spec, codex-review-round, bench, ship, docs, substrate-gap, codex-review-response, corpus-result). build-research-log.py derives the DAG from git history.

Outcome65 commits classified, 6 explicit response edges, 7 review rounds. Codex received first-class CONTRIBUTORS.md credit in commit 955c052. The protocol is the reason every public v4 claim has both a primary commit and a review-response commit. Pattern reused for the cold-read gate, the precision pass, and the scorer mimicry guard.