Skip to content

v0.6.1

Choose a tag to compare

View all tags
@aliasfoxkde aliasfoxkde released this 15 Oct 21:51
· 498 commits to main since this release
v0.6.1
6ed1db6
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

The first public release of the 0.6 line. The engine learns when not to
answer: selective-prediction evidence floors and elicited abstention join
the calibration stack, the corpus toolchain becomes a first-class item
generator crate, and the benchmark program closes its corpus-composition
question — merged-v6 passes the release gate and the first gated weights
ship.

Added

  • Engine — selective prediction end to end: a label-overlap preflight that abstains before extraction guesses (d5f4733), a selective-prediction evidence floor (df230f5), boolean gate fit with elicited abstain (c5d2527), and risk-coverage acceptance for calibration artifacts (d9b86e9) — a calibration artifact now earns deployment by its risk-coverage curve, not a single operating point.
  • opencodifier-itemgen — new crate: the relational/contrastive item generator that builds suite items, extracted from ad-hoc scripts (70a59af).
  • Schema — shared byte-exact macjev renderer, one implementation for tests and tools (8b415d2); noul non-string sides skip with unit tests (661925d).
  • Model — NLI verbalization verifier contract for the #15 enhancement program (a6a2e07).
  • Benchmarks — the training/corpus toolchain of record: soft-label trainer path + corpus builders (9937045), VIVERE teacher-score campaign tool (5b800cb) + batched teacher scoring (cfdf760), trainer resume-state + --resume (c44ad6a), GGUF parity fixtures with the BOS serving-path root cause (0364e17), corpus-prep gates for merged-v4/v5/v6 lanes (bf5370f, d464cdb), reporting parity metrics over frozen evidence (4b2cab8) + the cost table mapping to the board CSV (dc89d22), the rel200 eval slice + iterate-B val history (da0f599), the e1 calibration lane + JevBench slot arm (224929b), and the MiniCPM5 arm (b52fbaa).
  • Site — Decisions SDK + shell kit (3740523), the 12-game SDK game library (9ef9906), the chrome round: switcher + ratings (6b066b1), a wiki-style docs page (f4530cb), the decide console (f7a56bb), the board data page + quickstart + landing fixes (e617e35), and board parity + billed-cost columns (505e648).
  • CI — test/e2e lanes build on container-local disk (6cbf4d0); heavy lanes fit the 4G container cap (e427961).

Fixed

  • e1 readout anchor off-by-one token (7082a46); length-bucket chunk splitting kept intact (d989c9d); itemgen compiles under the pinned 1.90 toolchain (de8b944); wasm conformance keys + intra-doc link (a265d81); perkind argv honors rows/suite paths (4494142).
  • Repo hygiene carried by the 0.6.1 bump: CONTRIBUTING.md, CHANGELOG.md, and SECURITY.md moved under .github/ where GitHub surfaces them; README gains the Models section with HuggingFace cross-links; the site release-flow doc now tells the truth about production deploys (manual direct-upload, never automatic); the 2026-10-12 platform-adoption handoff is landed in docs/handoffs/; wasm conformance fixtures re-keyed for the semver bump (cache keys embed engine semver by design).

Verified — the corpus-composition question closes

  • T3.3' (2B on merged-v6-full): release gate PASS ×2 — suite holdout 0.6917 (≥ 0.68) and JevBench 0.6580 (≥ 0.6494) — second own arm through the gate (TRAINING.md, 2026-10-15). The comparison with E1-E closes the corpus-volume lever: confidence-gated volume ≠ transfer; E1-E (0.8083 / 0.6840) remains the release candidate.
  • First gated weights published: aliasfox/OpenCodifier-Decision-2B-v6full (per-surface numbers, shipped per-kind calibration, withheld-corpus disclosure).
  • Engineering levers measured: length-bucketing +31.3% replicates, b4-bucket sweet spot +44.4% (T2.3); causal_conv1d buys +29.2% tok/s but fails same-seed trajectory parity — throughput lever, banned from gate arms (T2.2, fixed comparator).

Security & quality

  • rustls 0.23.45 → 0.23.46 (RustSec: hybrid-PQ KEX + session-resumption downgrade) — lock-only.
  • Workspace dependency requirements aligned; conformance fixtures re-keyed for 0.6.1.
  • All CI lanes green on the released commit; aegis 0 new findings.

Binary assets

Cross-target binary matrix, per just release-build (each target ships
its binary plus a SHA-256 attestation). Attached: x86_64-unknown-linux-gnu,
x86_64-unknown-linux-musl, aarch64-unknown-linux-gnu, x86_64-pc-windows-gnu,
x86_64-apple-darwin, aarch64-apple-darwin, aarch64-apple-ios,
wasm32-unknown-unknown, aarch64-linux-android. Binaries were compiled from
the release tree (GitForge main ef19ad7, identical tree to GitHub main).
Model weights are NOT attached to releases by policy — gated weights
publish to HuggingFace only (docs/RELEASE_POLICY.md).

Assets 20
Loading