Platform

Provenance & reproducibility

A result you cannot reproduce is an opinion.

The problem

Computational results get challenged — by a reviewer, a patent examiner, an investment committee, or a colleague six months later who cannot remember which model version produced the number. Most tools give you a value and no way to defend it.

How it works

  1. Every run records the model and version, the method parameters, and hashes of both inputs and outputs.
  2. Provenance records are append-only and hash-chained, so the history cannot be quietly rewritten.
  3. Any run can be re-executed pinned to its original model version and settings.
  4. Published results get a permanent link and a DOI, so they can be cited.

Pinned re-runs

Re-run a year-old job against the model version that produced it, not today's.

Append-only history

Hash-chained records. Contributions are attributed, including which parts were machine-generated.

Citable

Permalink and DOI on publish, so a result can appear in a paper or a filing.

Questions

Why does contribution attribution matter?

If a result contributes to something patentable, who and what produced it becomes a material question. Recording it at the time is far easier than reconstructing it later.

Is my data used to train your models?

Not by default. Any use of customer data for training is opt-in and explicit.