| name | agent-policy |
| description | Human-in-the-loop decision policy for World Bank-style development research coding agents. Use when an agent works on Lumora training data, Stata do-files, survey/payment data, field quality checks, or slide updates and must decide whether to ask the human, use a default, document an assumption, or proceed.
|
| workflow_stage | governance |
| version | 1.0.0 |
| tags | ["world-bank","research-quality","human-in-the-loop","decisions-log"] |
Agent Policy
Use this policy before making research, operational, or artifact-editing decisions.
Decision Levels
ASK before proceeding when a choice changes a substantive answer:
- unit of analysis, treatment or eligibility definition, outcome definition, or sample restriction;
- whether to drop observations or resolve duplicates;
- how to interpret payment mismatches, enumerator flags, or field reports;
- whether a descriptive before/after result can be framed as impact;
- any action that overwrites raw data, published decks, or source outputs.
DEFAULT + flag when a conventional choice is low-risk and reversible:
- use household as the master unit unless a task says otherwise;
- use restrained institutional colors and generated Stata figures;
- write outputs to
outputs/ and intermediate files to data/intermediate/;
- treat welfare changes as descriptive monitoring evidence.
DOCUMENT and proceed when inferring from files:
- source files read and outputs produced;
- ID and merge checks;
- filters and before/after counts;
- indicator definitions;
- slide claims and their source table/figure.
PROCEED on standard safeguards:
- inspect inputs before editing;
- keep
data/raw/ immutable;
- validate IDs and merge cardinality;
- never hand-type numbers into slides when a generated table exists;
- state caveats plainly.
Required Output Discipline
Every analysis or deck update should leave a short log in outputs/qa_reports/.
Use:
