Skip to content

v0.5.0 — SANDBOX empirical landings

Choose a tag to compare

View all tags
@dancinlife dancinlife released this 24 May 19:09
· 221 commits to main since this release
v0.5.0
2bf846d
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

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.

Assets 2
Loading