How we decide what to believe

Computational results are easy to produce and easy to over-read. Our process is designed to catch our own mistakes before anyone else has to.

The method, in order

  1. Define the question

    A precise, falsifiable question, with a statement of what result would make us stop.

  2. Check prior art

    Targeted review of what is already known, so that replication is not presented as discovery.

  3. Register the plan

    Datasets, baselines, controls, metrics and decision rules are written down before outcomes are examined. Public registration begins with the first ledger entry.

  4. Test against controls

    Held-out and independent data, negative controls, and the strongest simple baseline, not a weak one.

  5. Seek independent review

    Adversarial review by someone other than the builder. Vivethos is one person today, so none has been publicly documented yet. Agreement is not proof.

  6. Record the result

    Every claim enters the evidence ledger with its status, its limits and its history, including negative results.

Distinctions we keep

A simulation is not an experiment. A prediction is not a validation. Agreement between AI systems is not evidence.

  • Computational prediction versus experimental result

    A model output is a hypothesis about biology. Only measurement tests it.

  • Numerical verification versus biological validation

    Showing that software solves its equations correctly does not show the equations describe the biology.

  • Retrospective versus prospective

    Re-analysing existing data cannot establish what a new experiment would find.

  • Replication versus novelty

    A result is not new because we have not seen it before. We search the literature first and say when work repeats prior findings.

Commitments

  • Reproducibility

    Methods, code and data are documented well enough for others to repeat the work and reach the same result, or to show why not.

  • Falsifiability

    Claims are framed so that evidence could prove them wrong, and success criteria are stated before analysis.

  • Independent evaluation

    Our aim is to assess results on held-out data and by parties other than those who produced them.

  • Transparent limitations

    Where a method fails is reported with the same prominence as where it works.

  • Scientific integrity

    No claim of discovery, validation or novelty is made without the evidence to support it.

  • Responsible biotechnology

    Work involving biological systems considers safety, ethics and the consequences of misuse.

Use of AI

AI systems assist with analysis, drafting and review. They are tools, not authors. Their conclusions are checked against the original evidence, and people remain accountable for every claim.