Conjecture → experiment → evidence
Validation Workbench
AI makes conjectures abundant. This workbench makes the scarce step explicit: a preregistered, observable experiment that can support, refute, or leave a claim inconclusive. Every campaign is content-addressed before work begins.
Planned ≠ run. Run ≠ validated.invalid inputs fail closed · predictions stay predictions
01
Define
Freeze the bounded claim, scope, falsifier, dossier evidence, and source versions.
02
Measure
Preregister one primary metric, explicit thresholds, required controls, cost, and duration.
03
Run
Capture provenance-bearing observations and SHA-256 artifact identities in an append-only log.
04
Resolve
Apply the frozen evaluator. Missing or ambiguous inputs are invalid, never silently positive.
Public queue
Registered campaigns
…
Loading registered validation campaigns…Workflow guarantees
- Two independent supporting source families before queue entry.
- Measured or computed support; predicted-only claims cannot advance.
- One content-addressed primary experiment in the first workflow version.
- Public reads, service-only writes, and an append-only event log.
- Missing provenance, artifacts, metrics, or units produce an invalid result.
