Welcome to InvarLock
The original InvarLock introduction, with a current starting point for native comparisons, captured results and bounded judging.
Historical context: This introduction retains earlier product language. The quickstart below is a reconstructed v0.6.0 illustration, from a release published after this post’s November 2025 date. The v0.13.0 transition replaced that operator workflow. Documentation links lead to maintained guidance; the retired commands and contracts below are not present-day instructions.
October 19, 2026 revision: Corrected shell continuations and identified the quickstart as a later v0.6.0 reconstruction with a supported adapter option.
October 18, 2026 revision: Outdated commercial references have been removed. The historical technical examples are unchanged.
Start with the current release
September 16, 2026 update. InvarLock 0.16.0 supports native comparisons, captured evaluator records and bounded judging through evaluate, verify and report. Each workflow retains its own required facts and verification limits.
Start with the CPU evidence quickstart to verify a retained comparison without running a model. Use captured results for an existing evaluation pipeline, or native setup to prepare model execution. Bounded judge measurements assess frozen answers under an explicit plan; replay neither establishes rating correctness nor reruns inference. Your reviewer retains the release decision.
The introduction below retains historical product language; the quickstart now identifies the later release contract it illustrates.
InvarLock evaluates weight edits with paired measurements and auditable evidence.
Highlights
- Evaluate edited weights against a baseline with paired metrics and confidence intervals.
- GuardChain checks for “unsafe to compare” measurement mismatches and quality drift.
- Evidence packs capture the artifacts you need to verify and share results.
If you edit model weights (quantization, pruning, fine-tuning, merges), you eventually hit the same question: did this change silently break anything that matters? “It loads” isn’t enough, and single-number metrics often miss the failure modes you’ll regret later.
InvarLock is designed for that moment. It produces an evaluation report that is both human-readable and machine-verifiable, so you can make upgrade decisions with evidence—not vibes.
Historical quickstart
For a concrete historical illustration, the sequence below uses the v0.6.0 CLI contract released on April 4, 2026. This is a reconstruction checked for this revision, not a command sequence available at the post’s original publication date. It has not been rerun against a model here. For a runnable setup with the released package and complete trust inputs, use Getting started.
pip install "invarlock[hf]==0.6.0"
The historical v0.6.0 example is:
INVARLOCK_DEDUP_TEXTS=1 invarlock evaluate --allow-network \
--baseline gpt2 \
--subject gpt2 \
--adapter auto \
--profile dev
The v0.6.0 contract describes an evaluation report and optional portable packaging, then called a proof pack. This example uses the dev profile and does not establish CI/release assurance.
Plans at publication
- Broader adapter and framework coverage
- Better “what changed?” analytics over time
To go deeper, start with the docs. For questions and feedback, email [email protected].
Sources
Website documentation explains the maintained workflow. The v0.6.0 CLI contract supplies the reconstructed historical quickstart added in this revision; it is not a confirmed original source. The transition article explains the later change to signed paired evidence, and the dated update points to the published 0.16.0 workflow.