aie-decision

name	aie-decision
description	Run the AIE Decision closed-loop Fermi candidate search from Claude Code, Codex, or another tool-using agent. Use for quantitative decomposition, bounded candidate revision, interval propagation, mechanical ablation, replayable search, or an explicit $aie-decision invocation. Process only supplied variables and evidence; do not acquire external data.

AIE Decision

Use the installed aie-decision CLI as the authority. Do not replace a real run with conversational arithmetic.

Prepare the run

1. Read references/search-loop-contract.md before constructing input.
2. Use only values supplied by the user. Keep observed, assumed, and decision variables distinct.
3. Treat formulas and mutation templates as candidate hypotheses, not facts.
4. Keep run files in a dedicated directory and preserve existing files.
5. Verify the installed interface with aie-decision --help. The public CLI must list search-fermi; otherwise report the installation mismatch.

Run the closed loop

Write the request JSON, then run:

aie-decision search-fermi <input.json> --experimental --output <result.json>

The exact protocol names are mandatory:

• Top level: run_id, question, target, unit, coverage, reference_value, acceptable_width, variables, candidates, mutation_templates, and budget.
• variables is an array with name, lower, upper, method, and optional ablatable. Never use a name-keyed object or kind.
• candidates is an array with candidate_id, formula, mutation_kind, and optional prior_weight. Every initial candidate uses mutation_kind: "seed" and has no parent.
• mutation_templates uses template_id, formula_template, diagnostic_reasons, mutation_kind, and optional prior_multiplier. Use revise when replacing a coarse estimate; template kinds are only expand or revise. Runtime-generated removals use ablate.
• budget uses max_candidates, max_rounds, max_evaluations, and max_seconds.
• --experimental is a CLI flag, not an input field.

Set ablatable: true only when the supplied problem explicitly permits that variable to be removed. Uncertainty alone never makes a factor removable.

Interpret the result

Report the terminal state, rounds and evaluations used, selected formula, remaining variables, interval, absolute width, candidate lineage, and replay checkpoint. Preserve these boundaries exactly:

• result-found means the experimental width and minimality gates passed.
• budget-exhausted and insufficient-information are valid failures; do not rewrite them as success.
• pseudo_posterior is an uncalibrated search ranking score, not a statistical posterior probability.
• experimental_usable: true does not override usable: false.
• calibration: unmeasured, uncalibrated_*, and provisional_uncalibrated must remain visibly labeled.

The current loop uses declared templates and CPU interval arithmetic. Do not claim model-generated proposals, GPU evaluation, external evidence retrieval, or empirical calibration unless a later verified runtime actually performs it.

Other public runtime commands

The installed CLI may also expose discover, start, apply, inspect, finalize, and replay for the versioned agent decomposition protocol. Inspect their live --help before use; do not infer payload fields from this search contract.