PUBLIC ENGINEERING PROJECT

LakatoTree

Deterministic verdicts for research programmes · by Gyeongjun Ra (라경준)

GitHub ↗Source-checkout demo ↗CITATION.cff ↗

01 Problem

Agents should not grade themselves.

Research workflows can claim progress without a registered prediction or a measurement behind the claim. LakatoTree binds appraisal to declared evidence instead of accepting a fluent self-report as a verdict.

02 Loop

pre-register
Lock the prediction, metric, direction, baseline, and noise band.
measure
Supply the observed value and any declared novelty measurement.
deterministic verdict
A pure scoring rule derives the judgement and carries its receipt and lineage forward.

03 Honest scope

Measured claims keep their limits attached.

The default live path currently receives a client-submitted numeric measurement. Receipts seal and replay that value: reproduction-confirmation, not value-ownership.
Lean checks the kernel's theory model. It does not verify the Python executable, the full runtime, or the provenance of a measurement.