A paid, scoped model-change pilot

Complete a model-change review with your team.

Apply InvarLock to one or two agreed real model changes. Finish with a report, signed evidence, a guided reviewer check and a runbook for the next change.

Discuss your pilotSee what the package adds ↓

Keep your evaluation tools. Prepare records for internal or client review.

What your reviewer opens, runs and retains

  1. Open the report

    Read the change, tests, result and limitations in the delivered HTML report. No evaluation-account access is needed.

    HTML report · Authored explanation
  2. Separate trust inputs ↓Reference records, requirements and signer details approved independently of the package.

    Verify the evidence locally

    Install the released core and run the supplied command with those inputs and your own signing key. We guide this first check.

    Signed verification receipt

  3. Keep the review record

    Keep the report, evidence and receipt in your approval ticket or document store. Record release approval separately; the receipt does not authorize deployment.

    Your separate release decision

Included verification support uses files and a local command; no hosted account is required. Hosting, managed access and ongoing support need a separate scope and price. Trust controls →

What you receive

One supported workflow, applied to one or two agreed changes. The deliverable includes:

Evaluation setup
Configure a supported workflow around your intended use, tests and acceptance requirements. Check input compatibility first.
Comparison
Run a supported comparison in your environment or import compatible test-level records. Show missing results and scoring choices.
Report and explanation
HTML report and authored summary of quality, gains, regressions, uncertainty and limitations, with resource measurements where available.
Signed evidence
Signed package with detailed test results, setup, agreed checks and verification instructions.
Guided reviewer check
Guide your technical reviewer through checking the evidence with separately approved inputs and retaining the signed verification receipt.
Configuration and runbook
Agreed setup and instructions for later changes, preserving earlier evidence and review history.

We inspect a sanitized sample first: a report or aggregate score alone may not contain the facts needed for verification. See the supported workflows.

Start with one change

Bring these to the sample review:

  • One current or recently completed model or prompt change.
  • Per-case evaluation records, or inputs ready for a supported comparison.
  • A technical owner and a named customer or internal reviewer.
  • The workload, acceptance requirements and existing review process.

Plan for a 60-minute scoping session, a technical owner for preparation and a reviewer for one guided check. We estimate hands-on work and agree dates after sample review. A second release is optional.

Check which evaluation inputs you need

Exact match needs answers and references; likelihood needs measured reference log probabilities. For bounded judging, bring frozen answers and a rubric. We can use compatible retained ratings or scope supported collection under an agreed plan and budget. Read the supported judging options →

Your team supplies compute and infrastructure unless agreed otherwise. New hosted judging uses your authorized service access; local judging uses your model files, compatible runtime and compute. We agree call or compute limits before collection. When importing external evaluator results, fresh workload calls remain your evaluation harness’s responsibility. Verification does not refresh old observations or establish judge accuracy.

For judge-based comparisons, check whether the planned independent cases can meet the required precision before collecting ratings. Where agreed, hypothetical edits can explain sensitivity in supported native exact-match results; they do not change the recorded verdict or predict a rerun. Method and limits.

See retained ModelKit and evaluator handoff examples →

Paid, fixed-scope engagement

After sample review, agree deliverables, timeline and fixed fee before work begins. Any additional integration or compute costs are identified separately.

Included in the quote

  • One supported path in one environment, with setup for that path.
  • One change; a second only if agreed in the quote.
  • The deliverables above, including one guided reviewer session and handover with a runbook.

Quoted separately

  • CI integration, specialized collection or analysis, further comparisons, sessions and compute.
  • Deployment work, hosting, managed access and ongoing support.
  • New adapters, substantial cleanup or extra environments require a revised scope and quote.

A completed handoff does not require a passing verdict. The reviewer checks and keeps the result, including regressions. A revised candidate or rerun is additional work unless included in the quote. Missing records, unexpected inputs or an invalid signature can prevent verification.

Limits. Your organization retains the release decision. Outcomes and savings are not guaranteed; one handoff does not establish repeat use. We do not host or operate models for you. This is implementation and support, not an independent assurance opinion, audit or certification.

Discuss your pilot

Tell us what is changing, how you evaluate it and who will review it. We work in your environment and agree data access and handling before setup. Keep private data, model weights and keys out of the contact form.

Discuss your pilot

The open-source tools are available without a paid engagement. Get started with the documentation.

Company and contact information →