Measurement contracts for CI and release verification
Certificates now record and enforce estimator measurement contracts under CI/release profiles, and B200 validation suites can cleanly split calibration vs execution.
Release: InvarLock 0.3.6 - Enforced estimator contracts and phased B200 validation suites
October 19, 2026 revision: Corrected the guard-order variable to B200_GUARDS_ORDER and restored certify and certificate terminology. The command and artifact rename followed in v0.3.8.
Highlights
- Guard estimator “measurement contracts” recorded in certificates and enforced by
verify --profile ci|release. - B200 validation suite can split into
--calibrate-onlyand--run-only. - B200 calibration defaults avoid CPU SVD bottlenecks by defaulting guard order away from spectral/RMT.
This release is for anyone who has looked at two runs and asked: were these actually measured the same way? 0.3.6 makes that answer easier by recording approximation-only, GPU/MPS-first guard estimator measurement contracts in certificates and enforcing them under invarlock verify --profile ci|release.
It also improves the reality of long suites. The B200 validation suite can now run in phases: generate calibration outputs first with --calibrate-only, then come back later with --run-only to finish remaining tasks without turning recovery into a manual exercise. B200_GUARDS_ORDER gives controlled experiments a direct way to choose guard order.
For B200 calibration, the defaults are intentionally pragmatic. Generated presets now carry guard order explicitly, avoid slow CPU-bound SVD by dropping spectral/RMT unless enabled, and place Hugging Face caches under the output directory by default. certify now honors the preset’s guard order too, so execution matches the calibration contract instead of silently forcing the canonical full chain.
Sources
Website documentation explains the maintained workflow. The changes described here belong to InvarLock v0.3.6; the tagged release record preserves that historical scope.
More in Release
Explore nearby related posts.
Release
Proof packs v2 and role-based adapter routing
Adapters move to role-based routing, proof packs become easier to inspect (v2 layout), and reporting output gets a readability upgrade.
Release
Deterministic evidence packs and safer perplexity runs
Evidence packs gain a deterministic bash test suite and better runtime helpers, window selection becomes stable/offline, and perplexity runs get safer around bad token IDs.
Release
Evaluation reports, strict evidence packs, and Transformers v5
A terminology reset (report/evaluate), stricter evidence-pack verification, and a clean upgrade path for Hugging Face Transformers v5.