Releases

Releases

Current release: v0.6.1 (published 2026-10-15) — a nine-target asset matrix, every artifact carrying a per-target SHA-256 attestation and a summarized sha256sums.txt. Weights are never attached to releases: model weights publish to HuggingFace only, gate-gated (Models).

Related pages: Models (the weights policy) · Security (artifact integrity) · Roadmap (what is next).

• v0.6.1
• The asset matrix
• Attestations
• Release-index note for tooling
• Release history
• Cutting a release

v0.6.1

Published 2026-10-15. Theme: 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 crate (opencodifier-itemgen); and the benchmark program closes its corpus-composition question, with the first gated weights shipping to HuggingFace.

Highlights (full detail in .github/CHANGELOG.md):

• Engine — selective prediction end to end: a label-overlap preflight that abstains before extraction guesses, a selective-prediction evidence floor, boolean gate fit with elicited abstain, and risk-coverage acceptance for calibration artifacts (Confidence).
• opencodifier-itemgen — the relational/contrastive item generator, extracted from ad-hoc scripts (Models).
• Schema — the shared byte-exact macjev renderer, one implementation for tests and tools.
• Model — the NLI verbalization verifier contract.
• Benchmarks — the training/corpus toolchain of record: soft-label trainer path, corpus builders, VIVERE teacher-score campaign tooling, trainer resume, GGUF parity fixtures, corpus-prep gates for the v4/v5/v6 lanes, cost-table mapping to the board CSV.
• Site — the Decisions SDK, the 12-game SDK library, the chrome round, the board data page, the decide console (Playground).
• Verified — the second own arm through the release gate, closing the corpus-volume lever: confidence-gated volume ≠ transfer.

The asset matrix

Nine targets, set by verification honesty (D24) — a target either has named evidence or is named as pending; the release page never says "supported" where the truth is "compiles":

Target	Artifact	Verification level
x86_64-unknown-linux-gnu	opencodifier CLI binary	run-tested + e2e (platform of record)
x86_64-unknown-linux-musl	opencodifier CLI binary	linked (static)
aarch64-unknown-linux-gnu	opencodifier CLI binary	linked (ELF aarch64)
x86_64-pc-windows-gnu	opencodifier.exe	linked + file PE magic (not executed here)
x86_64-apple-darwin	opencodifier CLI binary	linked + Mach-O magic (no macOS host)
aarch64-apple-darwin	opencodifier CLI binary	linked + Mach-O magic (no macOS host)
aarch64-linux-android	opencodifier CLI binary (NDK linker)	linked + file magic
aarch64-apple-ios	libopencodifier_ffi.a — the FFI staticlib, not a CLI binary (the honest artifact without an Apple SDK)	archive members verified
wasm32-unknown-unknown	dual-glue bundle pkg-web/ + pkg-nodejs/	gzip-magic attested

The recipe's preflight fails naming the missing tool — the matrix never narrows silently (D24).

Attestations

Every artifact travels with opencodifier-<version>-<target>.attestation.json beside it (schema opencodifier.attestation/1, produced by scripts/generate_attestation.py): per-target container/arch/format probes, the build command, and the digest. just release-verify re-verifies every attestation against its artifact; just release-checksums writes dist/sha256sums.txt covering the shippable bytes (attestations excluded — they carry the same digests individually). A release page that ships binaries without the JSON beside each one attests nothing.

Release-index note for tooling

The tag name v0.6.0 is unavailable on the GitHub release index (tombstoned server-side); the 0.6 line ships as v0.6.1. Automation that walks tags or diffs ranges should key on v0.5.0 → v0.6.1 and not expect a v0.6.0 ref.

Release history

Version	Date	Theme
0.6.1	2026-10-15	selective prediction end to end; opencodifier-itemgen; first gated weights published; corpus-composition question closed
0.5.0	2026-10-07	the model rung + cross-rung escalation (D26/D27); the transport-neutral wire module; opencodifier-ffi C-ABI; nine-leg release matrix with attestations
0.4.0	2026-10-02	D9 budgets measured + hot-path fixes; the escalation ladder as configuration (D25); semantic graph nodes + optimizer (D21/D22); WASM runtime (D23); fusion study; docs/BENCHMARKS.md + docs/TRAINING.md
0.3.0	2026-09-29	agent skills (§33); spec-coverage map; focused-question extraction (D18); JevBench external-anchor methodology; hosted-tier and Amortyx integration designs
0.2.0	2026-09-28	decision-model benchmark + D16 tiers; confidence truthfulness (D15 calibration + §19 gates); relational solver as the default zero-ML stack; MCP surface
0.1.1	2026-09-26	runner-local CI image; baked-tools pipeline; aegis baseline discipline
0.1.0	2026-09-25	phased build-out: core IR, schema adapters, engine, runtime traits, model crate, CLI + HTTP, recipes

Versioning posture: 0.x — no stable public API guaranteed yet; every breaking change is recorded in the changelog and, where a binding decision moves, in docs/DECISIONS.md. Format: Keep a Changelog.

Cutting a release

just release-preflight               # toolchain readiness for all nine targets, no builds
just release-build <target>          # build, stage, attest one target (dist/v<version>-<target>/)
just release-matrix                  # all nine
just release-checksums               # dist/sha256sums.txt (rewritten, never appended)
just release-notes                   # dist/RELEASE_NOTES.md from the template + attestations
just release-verify                  # re-verify every attestation against its artifact
just ci                              # the gate, one last time

Releases publish on GitForge (platform of record) and GitHub (mirror). Two standing rules: the whole matrix is rebuilt single-tree so every attestation hashes absolute paths from one consistent build; and the aegis baseline is regenerated in the same commit as any source edit that shifts lines — the gate is line-anchored.