Compare model changes. Keep results others can check.

Compare a model replacement, prompt change or optimized model, then share a report with supporting records reviewers can independently check, within your organization or with a client.

Use the open-source toolsExplore a model-change pilot
Public study · 200 paired tasks

What changes when you switch coding models?

GPT-6 Sol → GPT-6.1 Sol · Python tasks · High reasoning effort

Outcomes for 200 complete task pairs:

Each square is one task. Select a group to trace it.

200 complete task pairs. The overall accuracy difference remained uncertain. 8 of 208 attempted tasks lacked complete paired results.

Open the study report →
Public evidence and study scope

The public checker reproduces published counts. The full signed archive is not public. This study used specialized tooling; applying it to your tasks requires its own scope.

Check the published results.

Each comparison publishes its report with signed records and replay instructions. Authenticate the signature and recalculate the result yourself, including the one that failed its requirement.

See example reports →
Result excerpt from the Mistral 7B report: Policy not met. Across 400 usable pairs the subject-to-baseline NLL ratio was 1.091, with a 95% interval of 1.085 to 1.098 against a limit of 1.05. Signed manifest verified; independent recipient verification not performed by the report.
Excerpt from the report: the recorded policy result and what was checked.

Your team proposes. Your reviewer decides.

One published comparison, step by step. Review within your team or with a client.

Public example · Ministral 3 8BOpen this report →

  1. Your team defines

    Choose the baseline and candidate, evaluation cases and acceptance requirements. This example measures exact-match accuracy.

    Baseline44.75%BF16
    Candidate45.25%Q5_K_M GGUF

  2. InvarLock packages

    Run a supported comparison or import compatible test results. Keep the report, records and verification instructions.

    Client review package (HTML) →
    Signed evidence package

    400 paired MMLU-Pro records

    Recorded result

    Meets the agreed regression threshold

  3. Separate trust inputsApproved inputs and requirements, obtained separately from the package.

    Your reviewer checks

    Check the evidence against your own requirements before deciding on release.

    Signed verification receipt

    Kept in the source pack.

    Inspect the source →

    LimitsA report is not release approval. Replay checks saved evidence; it does not rerun the live model or prove the tests cover every risk.

Getting-started guide →See how independent review works →

Make your next change reviewable.

Start in your own environment, or work with us on a scoped comparison.

Start with the open-source tools

Import compatible test-level records, or run a comparison in your environment.

Run the five-minute check

Run a model-change pilot

A paid pilot for your next model change.

  1. Sample review
  2. Fixed quote
  3. 1–2 real changes
  4. Reviewer check and handover
Explore the pilot