TOOL 03 · EXPLAINABILITY
Adverse-Action Explanation Reporting Template
Check that every reason on a decline letter is real, understandable, actionable and not a stand-in for sex — before the letter goes out.
Reviewer guide · Tool 03
What it is
Records that connect what a complex model did to the reasons a declined applicant is told. Part 1 is completed for every denial. Part 2 checks a sample of those records to decide whether the explanation method can be relied on. Part 3 is a periodic report across the portfolio.
Why it matters
With complex models, the features that most influenced a decision are not always reasons the applicant could act on, and some may stand in for sex. An explanation can satisfy the rule and still be useless to the person who receives it.
How to read it
Each reason must meet four conditions: accurate (drawn from the features behind this decision and confirmed by counterfactual testing), understandable (stated in regulatory reason language), actionable (a realistic change the applicant could make) and clean (not a stand-in for sex). Counterfactual testing shows the stated reasons are real causes of the denial. Changing them, together if there is more than one, would have reversed the decision, and each one on its own moves the application closer to approval. A reason whose association with sex is 0.30 or more is held for compliance review. A record that fails is corrected before the notice goes out, never withheld. If fewer than 95% of a sample pass before correction, the explanation method is not approved for that model.
Jurisdiction
The reason wording follows U.S. Regulation B. For Canadian institutions, PAL applies the same standard as best practice.
Questions this tool informs
| Item | What to look at |
|---|---|
| Q15–Q16 | How hard it is to produce specific, accurate denial reasons from complex models, and whether your current tooling does so reliably |
| Q17 | Whether attribution plus counterfactual verification is technically feasible for you |
| Q18 | Whether the Part 1 record could fit into your adverse-action workflow without unreasonable burden |
| Q19 | The proxy flag and its routing to compliance review |
| Q20 | Whether mapping model features to regulatory reason language reflects how adverse-action compliance works for you |
| Q21 | Barriers to adopting the template |
- 1Identify
- 2Rank the drivers
- 3Translate
- 4Counterfactual
- 5Proxy screen
- 6Disposition
Step 1 — Identify
Identify
How to complete this step
- Enter the application ID, decision date, model version and matching signed Tool 04 reference for this denial. Use the same references throughout the review.
Export your inputs
Key terms
- SHAP
- Signed feature attributions describing a feature’s contribution to a model output, interpreted using the selected denial sign convention.
- Counterfactual
- A smallest realistic change tested against the model. Counterfactual testing shows the stated reasons are real causes of the denial. Changing them, together if there is more than one, would have reversed the decision, and each one on its own moves the application closer to approval.
- Proxy
- A feature correlated with applicant sex; the PAL screen holds flagged reasons for independent compliance review.
- Relationship feature
- A feature reflecting the applicant’s relationship with the institution, tracked in explanation and monitoring records.
- Sufficiency determination
- PAL’s sample-level evaluation of whether reasons met all four conditions before correction; corrections do not rewrite sample results.
- Regulation B
- U.S. rules implementing the Equal Credit Opportunity Act, including adverse-action notification and specific-reason requirements.
Thresholds are declared starting points fixed in the Tool 04 pre-registration. They are not regulatory requirements.