Guard-value contracts and stock-clean attention control
InvarLock 0.12.1 tightens the Mistral guard-value evidence contract, promotes targeted spectral/RMT/VE probes to required detections, and replaces the attention control with a stock-cap-clean 1.05x lane.
Release: InvarLock 0.12.1 - Guard-value evidence contracts and stock-clean attention control
Highlights
- The Mistral 7B guard-value package has a stock-cap-clean attention negative control:
spectral_moderate_scale_attn_l31_o_s105, a 1.05x attentiono_projscale that passes PM, adds no new baseline-relative spectral cap, and stays below the documented stock attention cap. - Targeted guard-value probes are stricter scenario contracts:
spectral_moderate_scale_mlp_l31_up_s112,rmt_norm_noise_l31_ffn_up_b030, andve_mlp_scale_skew_l31_down_s090are required detections for the Mistral guard-value package. - The old FP8 stress lane is historical/informational context instead of a current guard-value gate, and the retired attention control report is removed from the active public evidence package.
0.12.1 is a focused patch release for evidence interpretation. The public claim is not simply that "a guard moved." The current Mistral package has to show PM acceptance plus baseline-relative guard movement for the targeted spectral, RMT, and variance/VE probes, while keeping the selected attention control clean.
The Public Evidence Walkthrough and Failure Examples now point at the selected attention control, spectral_moderate_scale_attn_l31_o_s105. The useful detail is the stock-cap boundary: the attention target has z = 2.7987064430328767 against a stock cap of 3.018, so it remains a negative control rather than another guard-value hit.
The Evidence Pack Internals reference also reflects the scenario strictness split. must_fail reports are expected direct verification failures. must_detect reports can still pass direct invarlock verify when the primary metric and hard policy gates pass; the evidence-pack verdict then checks the required sidecars or baseline-relative guard movement. That distinction matters for guard-value evidence, because the targeted Mistral spectral, RMT, and VE probes are evidence-pack detections rather than generic report-schema failures.
There is a small docs-surface cleanup in the release as well: the upstream MkDocs assets harden MathJax loading and add focus and selection styling. The site sync carries those assets forward while preserving the site-owned docs palette transforms.
For the immutable release record, read the tagged CHANGELOG.md for v0.12.1.
More in Release
Continue through nearby posts in the same reading thread.
Release
Self-edit evidence, backend compatibility, and larger model lanes
InvarLock 0.12.0 adds knowledge/self-edit workflow metadata, widens BYOE edit-lane evidence, and reorganizes public evidence around backend compatibility and larger model findings.
Release
Report outlines, guard warnings, and wider public evidence
InvarLock 0.11.0 makes human reports more consistent, separates guard warnings from hard failures, and expands the public evidence surface across model families.
Release
Evidence packs, authenticity, and quantized-adapter validation
InvarLock 0.10.0 makes public evidence packs more explicit, adds signer-authenticity checks, and expands optional quantized-subject adapter validation.