Author receipt
Records that declared checks ran twice against pinned inputs and agreed under the captured local environment.
replayable computational evidence
Make computational claims replayable.
Certiv records an author-declared check, hash-pinned inputs, its explicit expectation, and the observed outcome in an inspectable receipt. Reviewers see the scope before deciding whether to rerun it.
A reproduced result applies only to its stated check—not the whole paper, peer review, scientific truth, novelty, or priority.
independent AI-discovery reproductions
Watch an exact claim pass—and one changed value fail.
Certiv independently checked AlphaEvolve’s rank-48 matrix certificate and FunSearch’s 512-point cap set. No author code ran. Each exact result passed; each Certiv-created mutation failed. AlphaEvolve is the one reproduced result counted in the 11-paper arXiv corpus; FunSearch is a supplemental Nature-published showcase.
- Central results reproduced
- 2
- Controlled changes rejected
- 2
- Author commands executed
- 0
evidence without overclaiming
One record. Three different statements.
Certiv keeps author execution, operator verification, and domain judgment visibly separate.
Verifier verdict
A separately operated node can isolate execution and sign its own verdict under the operator’s stated controls.
Scientific or business judgment
Experts still assess methods, source data, interpretation, significance, safety, and fitness for use.
designed for accountable adoption
One evidence format for research, review, and R&D.
Start with a controlled use case and measure whether receipts make review, handoff, and decision traceability measurably clearer.
Research labs
Make results easier to inspect and rerun.
Pin declared inputs, capture exact expectations, and give collaborators a compact record they can inspect before rerunning.
Try the research workflowUniversities & venues
Ask for consistent computational evidence.
Standardize the receipt fields you request while keeping scholarly review and independent reproduction outside the tool’s verdict.
Plan an evaluationCompanies & R&D
Keep decision evidence with the work.
Add machine-readable evidence to internal releases without sending source, data, or receipts to certiv.org.
Review deployment optionsinspect before you adopt
Three public, inspectable examples.
Start with a reproduced AI result, a small research computation, or the checks used to release Certiv itself.
Original pass. One-value mutation fail.
Follow two central results from pinned author artifact to independent verifier, machine receipt, and controlled counterexample.
See AI reproductionsMean of a small declared dataset
Download the complete packet, run the command locally, and compare the exact expectation with the published receipt.
Open worked exampleCertiv checks its own release
Inspect the release receipt, coverage, build checks, evidence boundary, and exact inputs behind every declared result.
Inspect release evidencelocal-first workflow
From declaration to re-check.
Declare
Name each claim, command or trusted checker, input, and exact expectation in certiv.yaml.
Run twice
Certiv executes from separate workspace copies and fails closed on divergence, truncation, or changed declared inputs.
Inspect
The receipt exposes hashes, commands, checker identity, coverage, observations, and the precise evidence boundary.
Re-check
A reviewer can validate structure without execution, rerun locally, or use a separately operated verifier.
deployment boundary
Choose controls that match the evidence you need.
Certiv Core
Local CLI, structural kernel, read API, and verifier components under Apache-2.0.
Self-hosted verifier
Your organization controls isolation, identity, keys, storage, availability, retention, monitoring, and incident response.
Institutional adoption
Add deployment planning, governance, integration, procurement evidence, and support under separate commercial terms.
inspectable by design
Evidence, limits, and operating responsibilities—together.
Review exact release artifacts, executable checks, security guidance, accessibility status, and published limitations before you adopt.
adopt with a bounded first workflow
Test the workflow before making it policy.
Run the worked example, read the evidence boundary, and test Core in an isolated environment you control.