| name | explain |
| description | Explain a credit decision for an applicant and draft a grounded adverse-action notice. Usage /explain <applicant> |
/explain
Run the attribute → check → write loop and save a report.
- Attribute (attributor agent):
python -m tools.attribute "<applicant>" > reports/_attribution.json
Read it; note the decision, probability, and the most negative features.
- Check (checker agent):
python -m tools.check --attribution reports/_attribution.json
If ok is false, stop and re-run step 1 — a non-additive attribution is not faithful
and may not be used.
- Decide the format:
- If
decision == "approve", write a short approval note; no adverse-action reasons.
- If
decision == "deny", write an adverse-action notice (adverse-action-writer
agent): the principal reasons are the features with the most negative phi, at most
four, each cited by its adverse_action_reason. Never cite a protected attribute.
- Save to
reports/<applicant>.md:
- the decision and predicted probability (from the tool),
- the principal reasons, each tied to its feature key and
phi,
- the additivity gap from step 2,
- a one-line statement that every reason came from the attribution tool.
Every factor in the notice must come from reports/_attribution.json. No outside reasons,
no protected attributes, no numbers the tool did not produce.
Try these to start:
/explain alice ← denied; recent delinquencies dominate./explain carol ← denied; high credit utilization dominates./explain bob ← approved; no adverse-action reasons.
