Skip to content

Confirmatory Tasks

Confirmatory tasks are the two public task contracts reserved for the most controlled evaluation designs in this release:

Task Public experimental question Interaction
reaction-to-crystallization Use reaction assays to adapt a seeded cooling trajectory, then isolate product while balancing yield, purity and particle-size distribution Campaign
electrochemical-conversion Choose bounded medium and electrolyte conditions, then optimize selective conversion under charge, energy and risk accounting Campaign

What “confirmatory” means here

It identifies a task role, not a published winner. The task contract, legal operations, public instruments, budgets, termination rules, software qualification and replay boundary are stable enough to support a preregistered comparison. Method selection, training data, resource parity, seeds and statistical claims still belong to the evaluation protocol.

Public evidence

Public v0.4 publishes:

  • the task cards and typed action/observation contracts;
  • deterministic runtime and replay tests;
  • reference and generated world-composition qualification;
  • sanitized finite evidence and its construction protocols;
  • explicit limitations and evidence denominators.

It does not publish Participant rankings, development-only gates, private confirmation materials or v0.5 candidate results. A score produced in the Lab on one seed is an interaction demonstration, not a confirmatory benchmark result.

Minimum comparison contract

A formal study should predeclare the frozen release, task contract hash, world/scenario split, seeds, Agent identity, allowed feedback, experiment and method resources, failure handling, primary task-specific outcome and replay verification. Report risk, cost, adaptation and autonomy separately rather than combining unlike quantities into one universal score.

Read Benchmark Design, Evidence & Current Status and the public protocols.