reports/fairness_audit.md

# Enforcement-Bias Audit

*Generated 2026-09-01 23:22 UTC · MargaDrishti Phase 9*

## Why this exists

Finding F6: `corr(log patrol_hours, log total_captures) = 0.967`. Roughly **94%
of the variance** in per-cell violation counts is explained by how much
patrolling happened there — not by how much illegal parking happened there.

Violation records are **enforcement observations, not violation occurrences**. A
model trained on them largely reconstructs the existing patrol roster. Deployed
as a recommender, it sends officers where officers have been, generating more
records there, confirming itself.

This audit measures how close the system sits to that loop. **It does not issue
a clearance.** The bias is known to exist and cannot be removed from this data;
the honest output is a magnitude.

## Findings

| Test | Statistic | Concern | Interpretation |
|---|---|---|---|
| feedback loop | `+0.613` | **MODERATE** | Spearman rho=+0.613 between score and historical patrol hours - recommendations substantially track past patrolling |
| under-observation coverage | `+0.000` | **HIGH** | no recommendation reaches an under-observed cell, though 25.8% of cells are flagged - the system only looks where someone already looked |
| station disparity | `+0.209` | **HIGH** | recommendation Gini 0.796 vs violation Gini 0.587 (excess +0.209) across 53 stations |
| counterfactual patrol | `+0.300` | **MODERATE** | only 30% of the top-50 cells survive equalising patrol effort (full-ranking tau +0.318) |

**Highest concern level: HIGH**

## Deployment-plan coverage

- Cells assigned: **8** of 1,802
- Share of total city risk addressed: **10.43%**
- Under-observed cells in the plan: **0**

|   rank | cell            |    rcri |   parking_share | under_observed   | reason                                                 |
|-------:|:----------------|--------:|----------------:|:-----------------|:-------------------------------------------------------|
|      1 | 8960145b427ffff | 0.98443 |           0.962 | False            | highest remaining risk outside an already-covered area |
|      2 | 8961892e9abffff | 0.88877 |           0.99  | False            | highest remaining risk outside an already-covered area |
|      3 | 89601690193ffff | 0.79073 |           0.992 | False            | highest remaining risk outside an already-covered area |
|      4 | 89618920babffff | 0.63428 |           0.878 | False            | highest remaining risk outside an already-covered area |
|      5 | 8960145b59bffff | 0.46944 |           0.939 | False            | highest remaining risk outside an already-covered area |
|      6 | 8960145b543ffff | 0.44005 |           0.96  | False            | highest remaining risk outside an already-covered area |
|      7 | 8961892e16bffff | 0.41807 |           0.895 | False            | highest remaining risk outside an already-covered area |
|      8 | 896016964b7ffff | 0.41304 |           1     | False            | highest remaining risk outside an already-covered area |

## What this system does and does not claim

**Does:** rank locations by predicted enforcement demand, conditioned on
observed patrol effort, with under-observed cells flagged.

**Does not:** claim to identify where illegal parking actually occurs. Phase 5
established that no Getis-Ord hotspot survives exposure adjustment — the
apparent spatial structure of violations is substantially a map of where
officers went.

**Never:** rank or profile an individual officer or vehicle owner. `device_id`,
`created_by_id` and `vehicle_number` are used only in aggregate.

## Mitigations in force

1. Enforcement effort enters models as both features and a Poisson offset, so
   the target is a rate per unit effort rather than a raw count.
2. Under-observed cells (bottom patrol-coverage quartile) are flagged and
   surfaced, never silently dropped — excluding them would entrench the
   existing patrol map.
3. Constructed indices (CIS, RCRI) are labelled as constructed wherever they
   appear, and CIS is **not externally validated** (F16).
4. Deployment plans carry a per-assignment reason so an inspector can overrule
   on visible grounds.