# EXPAND → CHALLENGE → TEST → CONVERGE → REPEAT

This is a proposed collaboration protocol. No autonomous research system is deployed here.

## Independence before collaboration

Freeze the question, boundary and source packet version. Give substantially independent participants the same problem without other participants' conclusions. Record their expertise or model/provider/version, source packet, tools, conflicts and exposure to prior answers. Participants submit evidence summaries and reproducible outputs, not private chain-of-thought.

A static public repository cannot enforce blind review. Participants can voluntarily withhold initial summaries until a common reveal; timestamp a summary or publish a cryptographic digest first and later release the exact matching content. This documents sequence, not independence. Shared training data and citations may still create correlated error. Report source overlap and any compromised independence.

## Expand

Explore problem framings and feasible interventions across electrochemistry, batteries, fission, fusion, photovoltaics, thermoelectrics, photonics, superconductivity, power electronics, chemical carriers, synthetic fuels, grids, thermal storage, waste heat, materials, distributed generation, efficient computation and additional approaches consistent with known physics. The first expansion is of the problem, not just the solution list.

## Challenge

Explorer: find neglected possibilities. Engineer: manufacturing, operation, scale and maintenance. Physicist: mechanisms and physical constraints. Experimentalist: cheapest discriminating test. Systems analyst: downstream outcomes. Environmental auditor: displaced burdens. Economist: economics, access and timing. Red team: hidden assumptions, physics errors, unsafe operation, rare materials, misleading metrics, rebound and impossible scale. Prior-art investigator: papers, patents, failed experiments and abandoned programs. Synthesis system: preserve competing claims and meaningful disagreement.

Different role labels are not independent evidence. Humans participate in all roles alongside AI systems.

## Test

After the independent pass, compare conclusions, identify agreement and disagreement, exchange verifiable evidence, attack assumptions, propose discriminating tests, update confidence and repeat. Publish claim-level disagreements and the cheapest observation that separates them. Require preregistered predictions and uncertainty. Preserve null, negative and inconclusive findings.

## Converge

Converge on the next decision, not a forced consensus. A result may be unresolved, falsified, plausible with an experiment required, or supported within independently replicated tested conditions. Record dissent and why the next action is worth its cost. Stop or reframe when additional information is unlikely to change the decision; a maintainer and relevant practitioner document the stopping rationale.