# A proposed protocol for choosing the next research action

Status: proposed and untested. Working record: [EDC-P-0001](../research/protocols/EDC-P-0001.json). This is our synthesis of established methods, not a claim of novelty or a running autonomous system.

## Two questions, two evidence requirements

**Energy research:** For a specified useful service, which intervention improves the outcome within the stated boundary? Energy density competes with efficiency, infrastructure, operations, service redesign and other feasible options. Success requires service-specific observations and comparison. A higher density number alone is insufficient.

**Research-process evaluation:** Does this protocol help choose better next research actions than credible alternatives within comparable resource limits? Better means lower decision loss or a prespecified practical improvement without unacceptable cost, delay or harm. An idea produced by the process does not prove an advantage over a baseline.

The original energy hypothesis, model and desk audit remain unchanged and untested. No process trial has been registered or run. The addition is ASTRA-STATE-0002, a conceptual child of the preserved ASTRA-STATE-0001.

## A decision record before another round

Name the useful service, beneficiary, decision owner, place, time and system boundary. State the decision that new information could change. Include feasible options, current evidence, assumptions, missing information, conflicting predictions and consequences of a wrong choice.

Record candidate next actions, not just candidate technologies: consult a practitioner, search prior art, compute, simulate, measure, prototype, test, replicate, challenge the boundary, preserve disagreement, wait or stop. An action should explain what could be observed and how different outcomes would affect the decision. “Think harder” needs the same justification.

For each candidate, record human time, compute or inference, electricity, spend, laboratory and equipment needs, materials, elapsed time, opportunity cost and relevant environmental burden. Use ranges and units where known; otherwise state what is missing. Budget caps and safety constraints are gates, not penalties that a high score can outweigh. Do not infer electricity use from token count without a defensible model.

## Compare value without pretending to know it

An uncertainty can be large but irrelevant to the present decision. Conversely, a small uncertainty near a decision threshold may matter greatly. Ask what changes the choice, whether the action can discriminate reliably, and what it displaces.

If credible probabilities, utilities and costs exist, compare expected decision benefit net of acquisition costs, including sensitivity to priors and boundaries. Otherwise compare explicit scenarios and trade-offs. Do not add incomparable resource units or divide an invented benefit score by cost. Keep non-dominated alternatives when no justified ordering exists. Do not call all costs entropy: thermodynamic entropy, exergy loss and social or environmental burdens require distinct accounting.

Across energy cases, retain service-specific outcomes and constraints. Register the case sampling frame and comparison rules before counting how often density helps. A Pareto improvement means improving at least one chosen outcome without worsening another within the chosen tolerances; it is not proof of universal leverage. Domain-specific constraint patterns remain a hypothesis, not a map of measured winners.

## A short loop with explicit exits

1. Define the decision and hard limits with someone who understands the service.
2. Record competing explanations, predictions, boundaries and candidate actions. Keep initial contributions separate where feasible and disclose prior exposure.
3. Identify the observation most likely to change an important choice. Compare feasible ways to obtain it with waiting or stopping.
4. Check authority, safety, resources and feasibility before committing. Prerecord the plan and evaluation rule before seeing outcomes.
5. Perform only the approved action. Log deviations, actual costs, anomalies and null or failed outcomes.
6. Update claims within their tested scope. Preserve branches and their prior states. Choose another action, wait with an owner and revisit trigger, or stop.

These checks may be short. They do not require a new committee or multiple agents for every choice. A decision about allocation is revisited after an observation; uncertainty about the allocation process does not make every step a new infinite metareview.

## Stop and wait are useful outputs

Pause after two successive reasoning rounds that change none of the relevant predictions, decision options, uncertainty assessment, boundaries or proposed tests. Two rounds is a provisional review trigger, not an empirically optimal constant. Continue only with a recorded reason that another round can add decision value within the remaining cap.

Stop or defer when the cap is exhausted, the action is not authorized, no feasible observation can change the present choice, or continuing would displace a clearly more valuable commitment. State what new evidence, capability, stakeholder decision or resource would justify returning. Preserve disagreement when the available evidence cannot distinguish branches.

## Guard against attractive group error

Keep the first responses and a record of later exposure. Record source overlap, shared model families when known, conflicts and selection decisions. Different names do not establish statistical independence. Shared evidence is not necessarily bad, but correlated error must not be counted as replication.

Use outcome-anchored scoring, blind process labels where feasible and keep scorers separate from participants. Do not let the founder or a model approve its own preferred conclusion. Lock the case list, comparator configurations and analysis before outcomes. Report failed runs, dissent and null results. Time, tokens, citations, hypothesis counts, agreement and confidence are process observations, not success measures on their own.

Invite affected stakeholders to challenge the service and burden boundary. Preserve anomalies with their measurement context rather than silently deleting them. Unknown unknowns cannot be enumerated, and diverse viewpoints cannot guarantee that they will be found.

## What would count against the protocol?

Within a registered domain and budget, the protocol would lose support if it worsened decisions, missed critical burdens, raised total resource use without useful benefit, or failed a prespecified minimum improvement against a credible baseline. A small underpowered pilot cannot establish equivalence or universal failure. Poor task scoring or leaked outcomes may invalidate the comparison rather than refute the process.

See [the proposed evaluation](allocation-pilot.md) for staged tests and unresolved registration fields. No process result can promote the original physical energy hypothesis, and no energy result alone can promote this process claim.

## Prior art and limits

This design combines decision analysis, value of information, experimental design, preregistration and provenance. The [source ledger](../research/process/sources.json) states exactly what was read and what it does not establish. Adjacent fields and novelty need a more complete review. The [ten cumulative passes](state-0002-review.md) show why the proposal was narrowed.