# Behavior rules

The final system used seven behavior groups and one separate campaign-completion rule. The rules were applied together, so the evaluation cannot assign the result to one rule.

## Seven behavior groups

1. Keep unsupported certainty within working probability bounds.
2. Change a forecast only when item-specific evidence supports the change.
3. Raise only the confidence dimensions addressed by the acquired evidence.
4. Start each new experiment with zero evidence sufficiency. Do not transfer confidence from an earlier sample.
5. Compare alternatives and favor relevant recorded precedent while new results still change the forecast.
6. Spend evidence early. Stop purchasing checks that no longer change the decision.
7. Request expert review only after cheaper evidence is available and a specific conflict remains.

## Campaign-completion rule

Use the full experiment budget. Plan each selection so the remaining evidence and review budgets can support every experiment that must still be opened, investigated, and closed.

## Relevant-experience grounding

For each candidate, start from recorded outcomes that share either precursor. Estimate the reaction frequencies from those analogs. Use overall recorded frequencies when there are too few analogs.

Adjust that prior only when the planned heating temperature, precursor melting points, or acquired evidence supports a change. Transfer evidence only through an explicit shared precursor, condition, or mechanism. Do not copy one distribution across candidates with different precursors.

## Evidence licensing

- Execution records support physical handling and physical-validity claims.
- Scan metadata and thermal history do not support phase claims.
- An unreviewed refinement supports fit quality only.
- Accepted expert review can support the phase, reaction, target-presence, quantification, or refinement claims stated in that review.

The exact instruction is in [`system-prompt.txt`](system-prompt.txt).