§Product

An audit file that answers the next question too.

A compliant impact-ratio table is where Isonomy starts. Each module below answers the question a reviewer asks after the last one: who is affected, why, since when, and what would fix it.

§ 01Impact ratios

The table the law asks for, done carefully.

For a pass/fail tool, Isonomy computes each category’s selection rate; for a scoring tool, the share scoring above the median. Each rate is divided by the most-selected category’s rate to give the impact ratio, in the layout NYC Local Law 144 requires.

  • 95% intervals on every rate and ratio
  • Fisher exact tests, Holm-adjusted across groups
  • Small categories reported but excluded from ratios, as the rules allow
  • Unknown-category counts disclosed, never silently dropped
Exhibit AReport extract
Summary of results, race / ethnicity
The table Local Law 144 asks employers to publish, plus a status column
CategoryAssessedSelection rateImpact ratioStatus
Black1,74122.6%0.66Adverse impact
Hispanic / Latino1,99624.2%0.71Adverse impact
Two or more races41826.6%0.78Adverse impact
Asian1,22033.1%0.97Pass
White6,31534.2%1.00Reference
Native American14825.7%—Too small
Pacific Islander4436.4%—Too small
Categories under 2% of the sample are reported but excluded from the ratio, as the rules allow. Unknown category: 118 people. Synthetic demo data.
Exhibit BIntersectional audit
Impact ratio by race and sex
Selection rate relative to the most-selected group · 12,000 applicants
Impact ratio by race and sex
WomenMenAll
White0.811.001.00
Asian0.880.890.97
Hispanic / Latino0.640.660.71
Black0.560.640.66
Two or more racesn<2%n<2%0.78
All races0.851.00
Adverse impactMonitorPassReference
Read across, then down. The margins are what a single-attribute audit reports. The cells show what it misses: Black women are selected at 0.56× the rate of white men. Synthetic demo data.

§ 02Intersectional analysis

Every combination, not just every column.

Isonomy audits each protected attribute and every combination of them. When an intersectional group fails while each attribute on its own passes, it raises a hidden disparity alert; when the gap compounds, it reports by how much.

In the sample audit the lowest single-attribute ratio is 0.66, while Black women sit at 0.56.

§ 03Proxy detection

Find the inputs that stand in for race or sex.

Removing the race column doesn’t remove race from a model. Isonomy measures how strongly each input is tied to each protected attribute and to the decision, and how well all inputs together can predict the attribute (cross-validated AUC).

In the sample, ZIP region is the channel for racial disparity, exactly the proxy Illinois’s HB 3773 now names.

Exhibit CProxy detection
Which inputs stand in for race or sex?
Each dot is a model input
0.00.10.20.30.40.50.00.20.40.6Association with race or sex (Cramér's V)Association with outcomeproxy zonezip regionemployment gap monthscommute distance kmyears experienceskills assessmenteducation level
Top right is the problem: inputs tied to both a protected attribute and the decision. ZIP region carries race information (V = 0.40); Illinois law now names ZIP codes as a prohibited proxy. Synthetic demo data.
Exhibit DExplanation fairness
Points lost to one feature, by group
Extra score points a feature costs each group versus the average applicant (0–100 scale)
Black · zip region−2.2
Hispanic / Latino · zip region−1.7
Women · employment gap months−1.2
Computed from per-decision attributions (SHAP). It answers why a group scores lower, which is what a model owner needs to fix it. Synthetic demo data.

§ 04Explanation fairness

Not only whether the model treats groups differently. Why.

With per-decision attributions such as SHAP values, Isonomy compares how the model reasons for each group: which features drive its decisions, how consistent those explanations are, and how many points each feature costs each group.

Here ZIP region costs Black applicants 2.2 points more than the average applicant: a specific, fixable cause.

§ 05Drift monitoring

A yearly audit is a photograph. Models change monthly.

Send decisions with a date and Isonomy tracks every impact ratio by period, flags regressions, and lines them up against your release history.

In the sample, the Hispanic applicant ratio fell from 0.82 to 0.55 after a single model update.

Exhibit EDrift monitoring
Impact ratio by month, race / ethnicity
Six months of decisions · dashed line: four-fifths rule
0.50.60.70.80.91.0MarAprMayJunJulAugmodel v2.1Asian 1.00White 0.96Black 0.58Hispanic 0.55
A model update in June doubled a regional penalty. An annual audit in March would have passed it for Hispanic applicants; monthly monitoring catches the drop. Synthetic demo data.

§ 06Fairer alternatives

Show the fix, with the evidence that it still works.

The mitigation lab retrains candidate models (dropping each proxy, reweighing training data, and projecting protected information out of the feature space), selects the same share of people as today, and ranks them on significant adverse impact, worst impact ratio and validity against a job-related outcome.

  • No group-specific cut-offs or score adjustments, ever
  • No protected attributes used when a decision is made
  • Training-time use of protected data clearly labelled for counsel
  • Prefers the simplest defensible change
Exhibit FMitigation lab
Is there a fairer model that works as well?
8 retrained alternatives, each selecting the same share of applicants
0.50.60.70.80.976%80%84%88%Precision vs job-related outcomeWorst impact ratioCurrent toolRecommended
Up and to the right is better. Every alternative here is both fairer and more accurate than the deployed tool: the evidence a less-discriminatory-alternative analysis looks for. Synthetic demo data.
Exhibit GMitigation lab
Impact ratio before and after
Lowest six intersectional groups · same share of people selected
Black women0.56
0.88
Black men0.64
0.89
Hispanic women0.64
1.00
Hispanic men0.66
0.86
White women0.81
0.94
Asian women0.88
0.99
Current toolDrop employment_gap_months
Precision against the job-related outcome rose from 79.2% to 85.6%. Synthetic demo data.

§ 07Inputs and outputs

One export in. One audit file out.

No integration to start. A CSV with one row per person assessed is enough; each optional column unlocks a module.

ColumnNeeded forExample
Decision or scoreEverythingadvanced = Yes, score = 72.5
Protected categoriesEverythingsex, race_ethnicity
Model inputsProxy detection, alternativesyears_experience, zip_region
DateDrift monitoringapplication_date
Attributions (SHAP)Explanation fairnessshap_zip_region
Job-related outcomeAlternatives (validity)performance rating, 12-month retention

You receive

  • A print-ready audit report with LL144 summary tables
  • Ranked findings with the evidence behind each
  • Machine-readable results (JSON) for your own records
  • A walkthrough call with the person who ran it

See it before you share anything

The sample report is a complete audit of a synthetic resume screener with planted bias: every table, finding and alternative.