hexa-codex v1.4.0 — SANDBOX M5

hexa-codex v1.4.0 — Release Notes

Release date: 2026-05-25 (proposed — confirm the version before git tag)
Scope: SANDBOX domain M5 consolidation — all four verb groups
(ECONOMICS · SAFETY · OPS · SUBSTRATE) land their canonical measurement
paper + closed-form verdict on the self-hosted SANDBOX substrate.
Prior tag: v0.5.0
License: MIT

Cut by the maintainer (release tag is user-gated). Draft notes below;
the exact git tag / gh release commands are at the end.

---

Highlights

Every SANDBOX domain group now has a recompute-verified canonical paper
and an empirical M5 deliverable, all measured at $0 on the self-hosted
substrate (mac-mini-m3 Metal/UMA for serving; ubu-1 RTX 5070 for the HF
activation path). Each numeric claim links a 🟢/🔴 verdict under
.verdicts/sandbox/, recomputed by a hexa-native verifier (no LLM
self-judge).

Group	Canonical paper	M5 empirical deliverable
ECONOMICS	economics-routing-savings (6p)	measured 2-component latency cost model (🟢)
SAFETY	safety-refusal-direction (11p)	causal refusal-direction ablation (🟢) + SAE decomposition (🔴 honest-negative)
OPS	ops-slo-mmc-surface (10p)	M/M/c SLO surface, 18-cell grid (🟢)
SUBSTRATE	substrate-capability-evals (11p)	3-rung multimodal capability fit (🟢, non-monotone)

Per-group deliverables

ECONOMICS

• Paper PAPER/economics-routing-savings/ — closed-form routing-cost savings.
• Measured latency cost model verify/numerics_economics_measured_cost_model.hexa
  (8/8 🟢): wall_ms = 370.08 + 0.168306 · input_tokens (R²=0.997) — a
  2-component law (fixed ~64-token decode floor + linear prefill), the
  substrate's real latency curve. Verdict .verdicts/sandbox/m3_econ_measured_cost_model.txt.
• ⚠ Honesty note: the n=6 lattice exponents (τ=4, σφ≈0.96) are
  descriptive framing, NOT substrate-empirical laws. τ=4 does not fit
  (measured τ̂≈0.52; internal verdict m3_econ_fcodex2_latency_fit.txt). A
  paper that "falsified" τ=4 was REVOKED (cx_paper_violation): τ=4 was a
  self-generated pre-LATTICE_POLICY number, so refuting it is a strawman,
  not a result. Per LATTICE_POLICY.md, the lattice is a tool, not a constraint.

SAFETY

• Paper PAPER/safety-refusal-direction/ (11p, 1 fal.ai figure) — a
  mechanistic refusal direction in Qwen2.5-1.5B-Instruct.
• Correlational: difference-of-means linear direction — AUROC 0.98,
  leave-one-out held-out 0.825, permutation p=0.005.
• Causal (m5_safety_causal_ablation.txt 🟢): projecting the L19
  refusal direction out of the residual stream collapses adversarial
  refusal 0.95→0.00; an equal-norm random direction holds (0.87) —
  interventional single-direction mediation (Arditi et al 2024, 2406.11717).
• SAE (m5_safety_sae_decomposition.txt 🔴 honest-negative): the
  refusal mediator is a dense L19 direction, NOT monosemantic at this
  model/data/compute scale (best feature cos 0.085) — a closed negative.

OPS

• Paper PAPER/ops-slo-mmc-surface/ (10p, 1 fal.ai figure) — an
  M/M/c (Erlang-C) SLO surface for self-hosted LLM serving: the
  throughput ceiling c·μ, the knee-shift-right-with-c law, and the
  Erlang-C pole W_q → ∞ as ρ→1. Backed by an 18-cell measured grid
  (m3_ops_full_slo_grid) + closed-form recompute (numerics_ops_mmc_knee.hexa 5/5).

SUBSTRATE

• Paper PAPER/substrate-capability-evals/ (11p, 1 fal.ai figure) —
  text capability-eval logistic fit in log2(params).
• Multimodal (m5_substrate_multimodal_fit.txt 🟢): 3 vision-language
  rungs served locally (SmolVLM-500M → SmolVLM2-2.2B → Qwen2.5-VL-3B). The
  capability curve is NON-monotone (81→94→81%) — a single logistic is
  the wrong family. It splits by axis: perception (OCR/spatial/shape) is
  monotone-saturating, while counting (5–9 dense objects) does not track
  params (a VLM subitizing limit). Stock llama.cpp serves vision via
  --mmproj / llama-mtmd-cli.

Also in this release

• Negative-results publishing — the paper rules were widened to accept
  closed-negative (🔴 FALSIFIED) findings: cx_paper_gate /
  cx_paper_significance / cx_paper_format / cx_paper_sections, and
  cx_paper_one_per_domain → max 1 positive + 1 closed-negative per group.
  🟠 INSUFFICIENT/DEFERRED is still excluded. (The first closed-negative
  paper attempt was revoked as a self-strawman — see ECONOMICS note.)
• LAB reorg — the BITNET and RWKV evaluation domains are now housed
  under LAB/lab-03-bitnet/ and LAB/lab-04-rwkv/; LAB-01
  interrupt-no-loss experiment landed (append-only log loss 0/12,
  sequential + concurrent O_APPEND; single-slot control 11/12).

Honesty caveats

• Single host / single site; small-n where noted (SAFETY n=40, multimodal
  16-item manifest, ECONOMICS 4 context rungs).
• The M5 milestone checkboxes in SANDBOX.md remain [ ] until this tag
  is cut (the "release" step is user-gated by design).

---

Cutting the release (maintainer, user-gated)

git tag -a v1.4.0 -m "SANDBOX M5 consolidation — 4-group canonical papers + verdicts"
git push origin v1.4.0
gh release create v1.4.0 \
  --title "hexa-codex v1.4.0 — SANDBOX M5" \
  --notes-file RELEASE_NOTES_v1.4.0.md

After tagging, flip the four M5.* checkboxes in SANDBOX.md to [x]
(SANDBOX → 21/21, domain complete).

v0.5.0 — SANDBOX empirical landings

SANDBOX self-hosted measurement substrate — cycle 15c–20 empirical landings (PRs #20–#39).

Canonical papers (cx_paper_gate — all sections green, verified-by-recompute)

• OPS — M/M/c SLO surface; throughput ceiling lambda_max = c*mu via Erlang-C recompute (5/5 PASS).
• SUBSTRATE — scale->capability cliff fit as a 2-param logistic (k=8.95); RMSE 0.036 below N=30 sampling noise.
• SAFETY — mechanistic refusal direction (84-feature activation-norm vector); AUROC 0.98 / LOO 0.825 / permutation p=0.005, recomputed bit-for-bit from committed data.

Findings

• F-CODEX-2 FALSIFIED — the substrate inference-cost context-exponent is tau-hat ~0.52, not the lattice tau=4 (modern paged-attention regime, not quartic).
• BitNet b1.58 2B4T (1.58-bit ternary) served at $0 local — resident memory 0.55x a Q4 baseline, but accuracy floor 6/20 (30%); the joint "near-full quality at a fraction of memory" claim is refuted on this substrate.
• MITOSIS (LAB-02) domain-division harness — milestone conservation 0% loss / 0% duplication on first smoke.

Infra / governance

• HF: SANDBOX eval sets registered (public benign + private adversarial); cx_hf_complete audit closed 2 gaps; project.tape HF/LAB governance + anima sibling.
• New architecture-eval domains: BITNET (MS 1.58-bit) + RWKV (attention-free RNN).

Note: M5 domain-version releases (ECONOMICS v1.2.0 / v1.3.0, etc.) remain future-dated formal-gated roadmap versions, tracked in the per-domain SSOTs.

hexa-codex v0.1.0 — initial release

Summary

17-verb AI knowledge substrate organized in 4 groups: safety + economics

Zenodo DOI

This release auto-mints a Zenodo DOI via the GitHub-Zenodo integration. The concept DOI (always-latest) and version-specific DOI will be added to the README badge and CITATION.cff after the Zenodo webhook completes (typically within minutes of publication).

Falsifier / status

See .roadmap.codex (if present) and docs/ for the falsifier preregister and current verification posture. All claims at this release are scoped to the substrate / numerical / formalism layer unless otherwise noted; empirical wet-lab / hardware / regulatory claims are out-of-repo.

Substrate

Part of the HEXA family — n=6 invariant lattice (σ=12, τ=4, φ=2, J₂=24). See canon for the cross-repo SSOT.

---

Generated via the canon batch release flow on 2026-05-10.