Skip to content

09 — Documentation Consistency, Citations and Governance

Workstream W9. The licensing observations are an engineering consistency review, not legal advice; items marked "counsel" need a lawyer.

ICX-AUD-031 — Cross-document numeric contradictions (Medium)

The same quantity is stated differently across docs/, docs/ahtl/, ref-docs/, README, CHANGELOG and ROADMAP. Code values were read at c176d03.

Quantity Statements Code / measured
Epoch rollover threshold 16,384 Anthropic / 8,192 OpenAI (README :257, CHANGELOG :93, ADR-0003 :14, deploy guide :922-923, performance.md:34); 4,096 (docs/ahtl/07:104) EpochControllerConfig::default() = 4,096, protect_last_n_turns = 4, tail_slot_max_tokens = 1,024 (epoch.rs:21-29); no provider-specific default exists
Protected tail turns 6 (virtual-memory-and-dag-engine.md:75, CHANGELOG); 4 (ch07, deploy) 4
Tail slot size ≤ 2,048 (glossary :14, VM :71); ≤ 1,024 (architecture :87, adapters :134); 512 or 1,024 (deploy) 1,024
Anthropic breakpoint layout "system + up to 3 sealed epochs" (README :258); BP0 / BP1-3 zones (VM :68-69); BP1 system, BP2 tools, BP3 epoch, BP4 reserved (deploy :847-851) one breakpoint at total-3 (wire.rs:127)
Compression ratio 8–25×, 12–40×, 12–72×, 15–40×, 18.4–72.1×, 72.1×, 214× 34–55× on real inputs; 214× on the fixture (ICX-AUD-028)
HDC Hamming scan (1k nodes) <25 µs, <50 µs (10k), 12.40 µs, 1,306 µs 72.7 µs median
Seqlock publish <300 ns, <500 ns (/dev/shm), <800 ns, ~45 ns, 44.2 ns, <1 µs, <100 µs 62.9 ns in-memory; the file backend is unmeasured
Hook budget <500 µs, <450 µs p99, <400 µs, <100 µs, <250 µs, 118.4 µs 36 ms–940 ms real (ICX-AUD-001)
Shm transport positional I/O, no mmap (performance.md:32, ADR-0004); "memory-mapped" (wire-protocol-and-shm-bus.md:13); memmap2/libc (research report :122, :144) positional I/O
Shm path /dev/shm/icx-fabric-v1; .icx/icx-fabric-v1.shm; $XDG_RUNTIME_DIR/icx .icx/icx-fabric-v1.shm
Redactor 4 tiers (ch05, codec doc, threat model); 3 tiers (docs/ahtl/README.md:29,51, PDF cover); entropy threshold 4.35 vs 4.65 (deploy :80, :967, :1080, :1281) 4 tiers, 4.35
Corner-vertex projection popcount majority ≥ 320 per block (ch03); "parity of the 160 words" (hyperdimensional-and-fractal-index.md:60) majority
PPR iterations 8 (architecture :86, HDC doc :99); 6 (exactness review :163) —
CAS chunk storage Zstd-compressed chunks (architecture :45, :69; VM :12; data model compressed_bytes; research report :140) raw bytes; no zstd dependency
DAG tier naming three incompatible schemes (VM :52-57 / VM :12 / README :87) L1 turn/tool, L2 branch, L3 epoch, L4 fold
MCP schema size / ops <180 vs <150 tokens; four different op lists; signature icx(op, ref, q, k) vs schema id, q, range ops deref slice grep recall fold expand xagent status (mcp.rs:20-45)
AhtlWirePacketV1 fields three different field sets (data model :145-152; 08.md; PDF Ch08) the ahtl.rs:357-366 field set
Coverage ≥ 95%; 96.13%; 97.5%+ 95.75% (95.78% production-only)
Sprints 14 (CHANGELOG :133); 15 (ROADMAP) —
Fractal precision "128-bit fixed-point" (ref-docs/papers/INDEX.md:56) [f64; 16] in memory, i16 Q15/Q1.14 on the wire

Remediation. Keep one source of truth (docs/performance.md plus a generated benchmarks.json). Have every other document link to it rather than restate numbers. Add a CI check that greps the docs for numbers covered by the canonical table.

ICX-AUD-033 — Citation errors and unsupported attributions (Medium)

All 22 arXiv IDs resolve, and all 22 local PDFs match their arXiv title and authors. The following problems remain:

Item Project statement Fact Verdict
arXiv 2607.23809 "ACM: Agentic Cache Management" with KV-cache block retention (INDEX :46-48, research report :171, PDF filename) "ACM: Agentic Context Management for Long Horizon Tasks" (Li et al.), about agent memory tools; KV cache appears only in related work CONTRADICTED
"up to 10× cost / 4–8× TTFT" (ADR-0003 :9) attributed to arXiv 2610.28577 — the paper reports 27.7% cost reduction, 1.2× throughput, 1.5× P95 latency, and 2–5× price ratios NOT FOUND / misattributed
"19.0% passkey recall after summarization" (docs/ahtl/07:17) — in no cited paper; its only other source is hard-coded in generate_ahtl_master_diagrams.py:1163,1245 NOT FOUND
"4.5–6.0 bits/token" (ADR-0002 :9) — unsourced; derived in novel-methods from about 1.1–1.5 bits/char NOT FOUND
RVQ "5.5×–8×" (INDEX :60) — 5.5× supported; 8× not found; the paper shows a GSM8k quality gap, not "negligible loss" PARTIAL
"128-bit fixed-point" for Fractal KV archives (INDEX :56) — the paper uses IEEE double and 2-D points CONTRADICTED
LCM arXiv 2605.04050 "May 2026" — arXiv date stamp 14 Feb 2026; the paper has no "cryptographic" pointers (INDEX :14) PARTIAL
Minor RVQ authorship and title paraphrased; ICAE title paraphrased; Titans dated Jan 2025 Titans was submitted 31 Dec 2024 Low

The prior-art GitHub repositories listed (14) all exist, are not archived, and match their descriptions. "ai-memory" is cited by a local filesystem path, not a public URL.

ICX-AUD-032 — The exactness review contradicts itself

See 01-ahtl-spec-conformance.md §4: clause escaping is described as both absent and enforced, the 48-bit birthday figure is wrong, and unresolved items sit under a blanket "ENFORCED".

ICX-AUD-034 — License declarations are inconsistent (Medium; counsel)

Source Says
Cargo.toml:18 (all crates) GPL-3.0-or-later
REUSE.toml:44-50, README :483 icx-proxy and icx-adapters are GPL-3.0-or-later AND AGPL-3.0-or-later
LICENSING.md:11-12 only icx-proxy (and a hosted pager) is AGPL; adapters omitted
README :482 icx-cli is GPL-only, but it depends on icx-adapters and icx-proxy (crates/icx-cli/Cargo.toml:17,22), so the AGPL terms apply to the combined icx binary
  • Effect on tooling: no .rs file carries an SPDX header, and the crates ship no LICENSE file. crates.io and cargo deny would therefore see GPL only.
  • LICENSING.md:23: it says GPL §13 places network deployments of the GPL crates under the AGPL. GPLv3 §13 creates no network obligation; it only permits combination with AGPL works. AGPLv3 §13 applies to modified versions of the AGPL-covered work, not to "the entire service stack".
  • COMMERCIAL.md:62: it implies that internal commercial use requires a commercial license, which conflicts with GPLv3 §2 (unlimited permission to run the program).
  • CLA.md:48: it requires contributed third-party material to be "GPL-compatible", while the repository itself ships CC-BY-NC-ND, CC-BY-SA, CC-BY-NC-SA and arXiv non-exclusive content.
  • Remediation: make the Cargo license field match REUSE for each crate. Add SPDX headers, or a per-crate LICENSE file. Correct LICENSING.md :23 and COMMERCIAL.md :62.

ICX-AUD-035 — Patent and CLA terms vs GPLv3 (Medium; counsel)

  • Termination clause: PATENTS.md:32 and CLA.md:33 terminate the copyright license on a patent suit. That is not among the additional terms GPLv3 §7 permits, so under §7 and §10 it is a "further restriction" that recipients may remove. It therefore offers little protection while signalling non-free terms.
  • Internal-use denial: PATENTS.md:28 denies a patent license for internal non-conveyed use, which is in tension with the unconditional §11 patent grant from contributors who convey.
  • Unsubstantiated filings: "Patent pending (5 applications)" has no application numbers in the repository; the patents/ directory is excluded locally (.git/info/exclude) and described as ready to upload. Confirm the filings exist before keeping the marking: false marking (35 U.S.C. § 292) is a risk, and public disclosure in this repository may affect novelty in jurisdictions without a grace period.
  • CC-BY-NC-ND specifications: they contain reference Rust code, so code copied from the spec into a GPL fork conflicts with ND. Copyright cannot prevent clean-room reimplementation of ideas (17 U.S.C. § 102(b); CC 4.0 §2(b)(2) excludes patent rights).

ICX-AUD-036 — Redistribution of third-party content (Medium; counsel)

  • arXiv PDFs:
  • REUSE.toml:108 labels all 22 papers CC-BY-4.0 OR arXiv-NonExclusive. Per their arXiv pages, 11 are CC BY 4.0.
  • 10 carry only the arXiv non-exclusive distribution license, which grants arXiv, not third parties, the right to redistribute: 2307.06945, 2403.12968, 2404.07143, 2405.14831, 2406.02069, 2406.10774, 2410.10819, 2502.12110, 2510.11967 and 2610.28577.
  • 2501.13956 (Zep) is CC BY-NC-SA 4.0, which neither label covers.
  • Remediation: remove the non-exclusive PDFs and link to arXiv instead, and relabel Zep.
  • Wikimedia material:
  • Wikipedia wordmark and tagline images under ref-docs/blake3-cryptographic-hash/images/retrieved/ fall through to the DoubleGate CC-BY-NC-ND label, although they are Wikimedia trademarks.
  • Some Commons images were renamed so that their provenance is unclear (for example Banach_Fixed_Point_Contraction.png).
  • The 58-page AHTL PDF embeds 61 images under CC-BY-NC-ND with no CC BY-SA attribution for Commons-derived figures. Share-alike material cannot be relicensed ND.
  • Unicode-3.0 label: it is applied to WG2/ISO documents and Unicode chart and chapter PDFs; the label is correct only for Unicode data files.
  • Reverse-engineering notes: ref-docs/agent-internals/rea-binary-and-sqlite-forensics.md documents reverse engineering of proprietary agent binaries and is labelled as DoubleGate's own CC work. Review the vendors' terms before publishing.

ICX-AUD-037 — Security contact and CLA enforcement (Low)

  • No reporting contact: SECURITY.md has none, although README :309 says it covers "vulnerability reporting". Add a private channel, such as GitHub private vulnerability reporting.
  • CLA workflow: the workflow itself is sound. It runs pull_request_target, checks out the default branch only with persist-credentials: false, keys signatures by account ID, and requires co-authors to sign. However:
  • The CLA status is not a required check in the default-branch ruleset, so it does not block merges.
  • The cla-signatures branch has no ruleset protection.
  • A second, undocumented signing phrase is accepted (.github/scripts/cla-check.sh:11).
  • An exemption is keyed on a git author name, which can be spoofed (.github/scripts/cla-check.sh:28). This only matters for PRs opened by maintainers.

ICX-AUD-038 — Repository hygiene (Low)

  • Personal paths: absolute paths to the maintainer's home directory and a conversation UUID appear in tracked public documents: the deployment guide (8), the adapters doc (4), the REA forensics (11), the prior-art survey (1), and scripts/generate_assets_and_graphs.py (1).
  • Large binaries: 19 MB of generated PDF in git, and a 6.4 MB image stored twice (docs/ahtl/images/retrieved/ and ref-docs/ahtl-foundations/images/retrieved/).
  • patents/ exclusion: it is excluded only through the local .git/info/exclude, not .gitignore, so a fresh clone on another machine could commit it.
  • Lint results:
  • markdownlint-cli2: 143 issues, mostly blank-line formatting; two structural ones, a 7-cell row in a 3-column table at docs/ahtl/05:131 and a dead #roadmap fragment at README :76.
  • typos: two real typos in code comments (nam in redact.rs:756; asess in icx-adapters/src/lib.rs:303-304). The other hits are false positives (CC-BY-NC-ND, hex BA).
  • lychee --offline: no broken internal links.

Positive results

  • uvx reuse lint passes: 468 of 468 files are labelled and REUSE 3.3 compliant. Several labels are wrong, as described above.
  • The cited prior-art repositories exist.
  • The CLA workflow avoids the classic pull_request_target checkout pitfall.