Welcome to InvarLock
The original InvarLock introduction, with a current starting point for native comparisons, captured results and bounded judging.
Historical context: This article preserves the product model and examples described at its original publication 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.
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 and commands below remain the historical record.
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
The original installation and command sequence is retained below for historical context. For a runnable setup with the released package and complete trust inputs, use Getting started.
pip install "invarlock[hf]"
Run your first evaluation:
INVARLOCK_DEDUP_TEXTS=1 invarlock evaluate --allow-network \\
--baseline gpt2 \\
--subject gpt2 \\
--baseline-adapter auto --subject-adapter auto \\
--profile dev
That produces an evaluation report and (optionally) an evidence pack you can archive, verify, and share.
What’s next
- Design-partner refinement of the private deployment review workflow
- 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].
If your team wants to help shape the private deployment path, start with design partners.
Sources
Website documentation explains the maintained workflow. This introduction preserves the product model at its original publication date; the transition article explains the later change to signed paired evidence, and the dated update points to the published 0.16.0 workflow.