Pricing

Meter the re-execution. Never the verdict.

Calma re-executes the world's computations, so compute is the real cost — tiers gate how much and how deep you re-run, never whether a wrong number can slip through. The full fail-closed verdict taxonomy is identical on every plan.

Placeholder pricing · billing (Stripe) coming soon

Open Source
$0self-host

The recompute engine as a library + CLI. Verify on your own machine, bring your own compute.

View on GitHub
  • Recompute from committed artifacts + local runs
  • 628-recipe trusted metric catalog
  • Full fail-closed verdict taxonomy
  • Validity checks (leakage / overfitting)
  • Self-hosted badge + CI gate
  • Community support
Free
$0hosted

Connect a public repo and verify the numbers — the static layer always lights up, deep verify is capped.

Start free
  • Connect any public repo
  • 5 deep-verify scans / day
  • Top-3 claims per scan by salience
  • 30 sandbox-minutes / month
  • Reproducibility badge + 7-day history
  • Community support
Enterprise
Custom

For funds, labs, and regulated orgs where a wrong number costs millions — and neutrality is the point.

Contact us
  • GPU verification (CUDA repos)
  • Neutral third-party attestation
  • Transparency log + Rekor anchor
  • Audit / compliance evidence pack
  • SSO / SAML · RLS / roles
  • Dedicated capacity · BYOC / on-prem
  • SLAs + dedicated support

What a scan actually costs

Discovery (listing every claimed number) is ~free and stays generous on every tier. Only the expensive part — re-running the repo in an isolated sandbox — draws down your budget. At the ceiling, deep verify pauses and discovery keeps running; you always get the claim list.

Usage-metered, value-anchored

Marginal cost is cents per scan; the buyer's alternative — one wrong number reaching a trade, a shipped model, or an audit — is thousands to millions. Pro includes a monthly sandbox-minute budget with metered overage; GPU is an Enterprise add-on.

Open at the core

The recompute engine is open-source — re-executing to ground truth is the defense, not hiding the formulas. Read the code on GitHub, or verify a repo right now.