Release
Proof-pack showcase coverage and reproducible CI
Proof packs add new showcase and evidence artifacts, while CI and release flows become more deterministic and easier to validate repeatedly.
Release: InvarLock 0.3.10 - Guard showcases, VE sidecars, and deterministic verification
Highlights
- Proof packs add a guard showcase suite, expanded scenario filtering/overrides, and new evidence artifacts like verdict table generation plus a VE probe sidecar.
- CI and release verification are tighter and more reproducible, with deterministic
verify-full, Python 3.12 smoke coverage, and scheduled weekly verification. - Mixture-of-Experts and variance/VE paths are hardened for broader model compatibility, with additional reliability fixes across proof-pack packaging and evaluation handling.
0.3.10 grows the practical surface area of proof-pack validation while tightening failure handling and evidence quality. The new guard showcase suite, scenario filtering/errors-only mode, suite-scoped scenarios, and model override support make it easier to run targeted evidence checks without turning every proof-pack exercise into a full sweep.
The evidence surface gets richer too: verdict table generation, a VE ve_probe sidecar, intervention summaries, and additional RMT/spectral/variance showcase injections give reviewers more concrete artifacts to inspect. At the same time, verdict/task failures fail closed more consistently.
CI and release verification become more reproducible through deterministic verify-full, Python 3.12 smoke coverage, scheduled weekly verification, on-demand docs CI, and dependency preflight hardening for net-enabled installs. The model path broadens through Mixture-of-Experts compatibility work, including fused expert weight layouts, broader VE layer discovery, Mixtral block_sparse_moe support, and safer behavior when reports exist but evaluation exits nonzero.
For the immutable release record, read the tagged CHANGELOG.md for v0.3.10.
More from the blog
Continue through recent releases and implementation notes.
Synthesis
The Minimum Evidence Surface for Trustworthy Weight-Edit Results
A trustworthy weight-edit result needs more than a benchmark delta. It needs a bounded claim, an exactly paired comparison, and verification that rejects incomplete evidence.
Release
Evidence packs and explicit runtime provenance
InvarLock 0.8.0 moves the public bundle surface to evidence packs, pins docs to versioned release paths, and makes container-vs-host runtime provenance explicit across evaluate and verify.
Research Note
Fail-Closed Verification for Weight-Edit Evaluation
A verifier is only useful if it rejects incomplete evidence. InvarLock's verification path is designed to stop stronger claims when the evidence bundle is missing or inconsistent.