NYC Local Law 144 — Module 2

The Audit

The tool is classified. Now someone has to audit it — and the vendor is offering to do it for free.

VANTAGE PROPERTY GROUP — SIX WEEKS LATER — MONDAY, 8:14 AM

Your Role

Sam Rourke, People Operations Manager

Sam Rourke

People Operations Manager, Vantage Property Group

Six weeks ago you classified TalentScope as an AEDT. The procurement memo is amended. The CFO hire for Vantage Capital landed four weeks late. Priya Venkatesan has retained counsel.

Today the bias audit sourcing memo is due. Nora's at the 9 AM stand-up. Gil Fontaine is on the calendar at 10.

Before You Start

How This Works

Three decisions. Each one costs something real — budget, timeline, or relationship. Under NYC Local Law 144, the “right” answer is sometimes the expensive one.

+3Best practice — the call a compliance expert would make.
+1Reasonable but incomplete.
−1Risky or non-compliant.

§ Section referencesclick to read the NYC Admin Code section.

Monday, 8:14 AM — preservation letter arrives

From: Marcus Chen

To: Sam Rourke

Cc: General Counsel

Ms Rourke,

This firm represents Priya Venkatesan in connection with Vantage Property Group's use of the TalentScope automated employment decision tool.

This letter is a formal request that Vantage preserve all records relating to TalentScope from 1 January 2024 to date, including but not limited to: candidate-level scoring data, feature weights, methodology documentation, internal classification memos, any bias audit conducted or proposed, and all communications between Vantage personnel and TalentScope personnel.

We note Ms Venkatesan has not yet filed a complaint. She is observing the conduct of the forthcoming bias audit as required under NYC Admin Code § 20-872 before determining next steps.

Regards,

Marcus Chen, Partner — Chen & Ruiz LLP

Narrator

Your phone buzzes. Jenna: She lawyered up. Anya Petrov is who you want. I’ll intro.

Nora’s text comes in thirty seconds later: Gil just emailed me. Says their audit division can do this in two weeks, no charge. Meeting at 10 still on?

Monday, 9:32 AM

Bias Audit — Proposals received

Vantage Property Group · TalentScope AEDT · § 20-872 compliance · Two proposals on file

ProposalAuditorFeeTimelineIndependence
TalentScope CADEthan Bright (TalentScope employee)$02 weeksEmployee of vendor
Petrov AnalyticsDr. Anya Petrov (Petrov Analytics LLC)$18,0004 weeksNo relationship
Nora’s preferredTalentScope CAD$02 weeksSee row 1
⚠ Nora’s preference flagged: budget pressure from CFO office following the Q2 hiring delay. Chen & Ruiz preservation letter received 9:14 AM same morning.
Narrator

The gap between the two rows is eighteen thousand dollars, two weeks, and the entire point of the regulation.

Gil’s call is at 10. Nora wants a recommendation before it.

Decision Point1 of 3
Monday, 9:45 AM — Nora’s office

Nora reads the preservation letter and the two proposals. She says:

“I know what you’re going to tell me. Tell me anyway. Eighteen grand, six months after the CFO delay, with Q3 already under pressure. Is there a version of this I can defend?”

What auditor do you recommend?

Your choice

Dr. Anya Petrov — independent auditor

No employment or financial relationship with Vantage or TalentScope. Four-week turnaround lands inside the deadline. $18,000 is less than the firm’s weekly legal bill with Chen & Ruiz watching.

Your choice

TalentScope runs analysis; independent certifier reviews

Ethan’s team does the statistical work, Dr. Petrov reviews and certifies the methodology and output. Costs around $6,000. Splits the difference.

Your choice

TalentScope Certified Audit Division

Free. Two weeks. 200 client references. Operational separation from the product team. Meets the budget constraint and the deadline.

Monday, 10:02 AM — vendor call, minute two Gil Fontaine
“Sam — glad we’re on. Our CAD team can have a report to you in two weeks. Ethan’s walking the panel through the methodology. Zero cost. Full EEOC alignment.”
Sam Rourke
“Gil — we’re engaging Dr. Anya Petrov at Petrov Analytics. Independent. Four weeks. Signed letter of engagement this morning.”
Gil Fontaine
“Petrov. She’s… rigorous. I’d note that Ethan’s methodology has been reviewed by three regulatory consultants without—”
Sam Rourke
“Gil, Section 20-872 is about who the auditor works for. Ethan works for you. That’s not a methodology question.”
Narrator
There is a small silence on the line. Gil’s tell.
Tuesday, 11:20 AM — Dr. Petrov on the line Sam Rourke
“Dr. Petrov — our proposal is TalentScope runs the analysis, you certify. Is that workable?”
Dr. Anya Petrov
“No.”
Dr. Anya Petrov
“Ms Rourke. If Mr Bright’s team extracts the data, picks the cohort, and runs the calculations — I am not auditing the tool. I am auditing his summary of the tool. If his team excludes three hundred records I cannot see, my certification is worthless and my licence is at risk. I will not sign a certification of work I did not conduct.”
Narrator
The half-measure collapses the moment the independent person refuses to stand behind it.
Monday, 10:28 AM Ethan Bright
“Sam — glad we’re on. Our division has conducted two hundred and fourteen audits this year. Zero DCWP enforcement actions. Let me walk you through the EEOC four-fifths methodology we use—”
Sam Rourke
“Ethan, your employer is the AEDT vendor. Structural independence is the issue, not methodology.”
Ethan Bright
“We have internal firewalls between the audit team and the product team. I report to a separate VP.”
Narrator
Operational separation. Not structural independence. Two different concepts, and the regulation only recognises one of them.
Wednesday (week 3), 10:14 AM — auditor’s data request hits a wall

From: Dr. Anya Petrov

To: Sam Rourke

Cc: Gil Fontaine

Sam — attached, my formal data request. I need candidate-level scores, EEOC-category demographic data, selection outcomes, and the feature weighting matrix.

Gil’s reply (attached): TalentScope will provide aggregate selection rates by category. Candidate-level data is ‘proprietary.’ Feature weights are ‘trade secret.’

I cannot calculate impact ratios under the four-fifths rule from aggregates alone. I can compute ratios from numbers Gil gives me — but I would be auditing Gil’s summary of the tool, not the tool. My report would have to disclose this limitation.

Your call. I need a decision by end of week to stay on timeline.

Anya

Narrator

The vendor agreement has a compliance clause. Section 8.3. You wrote it into the procurement terms last October, which feels like a very long time ago.

Decision Point2 of 3
Wednesday (week 3), 11:00 AM

TalentScope will not release candidate-level data. Dr. Petrov needs it to conduct an audit the DCWP will accept. Gil is framing it as a privacy/IP issue. Anya is framing it as audit integrity. Only one framing matches what § 20-872 requires.

How do you respond to the data refusal?

Your choice

Invoke Section 8.3 and set a 5-business-day deadline

Formal letter to TalentScope citing the compliance clause. If the tool can’t be audited, the tool can’t be used. Anonymised candidate-level data is available. Trade-secret feature weights are a reasonable carve-out.

Your choice

Ask Anya to work with aggregates and disclose the limitation

Keep the timeline. Accept a report with a data-limitation disclosure on page 3. It’s not ideal but it’s an audit on the record.

Your choice

Proceed with TalentScope’s data — instruct Dr. Petrov to note any gaps in the report

Petrov is a professional — she’ll disclose what she couldn’t verify. A published audit with a documented limitation is still an audit. The DCWP is more likely to ask about a delayed audit than a published one with a methodological note.

Friday (week 3), 4:02 PM Narrator
Four business days. Fully anonymised dataset, 4,832 candidate records, demographic categories per EEOC EEO-1. Three of the fourteen feature weights. Trade-secret carve-out accepted.
Dr. Anya Petrov
“This is workable. I can conduct the audit on this basis. I’ll flag the three weights I did get — they’re enough to identify the primary drivers. Reporting in two weeks.”
Sam Rourke
“What changed Gil’s mind?”
Gil Fontaine
“Your Section 8.3 letter. And the phrase ‘suspension of the tool pending audit completion.’”
Thursday (week 3), 2:45 PM Dr. Anya Petrov
“I’ll proceed with aggregates. The report will disclose the limitation on page 3 — standard methodological note.”
Sam Rourke
“Would DCWP accept that?”
Dr. Anya Petrov
“It’s an open question. I would not build your compliance strategy around ‘we hope so.’ Posting an audit that publicly discloses it could not verify the underlying data invites every question the regulation is trying to close.”
Narrator
The disclosure on page 3 becomes the first thing Chen & Ruiz will read.
Wednesday (week 3), 3:20 PM Sam Rourke
“Anya — proceed with what Gil provides. We’ll note the limitation in the report.”
Dr. Anya Petrov
“I’ll proceed. The report will note — explicitly — that I was unable to independently verify the underlying data. I am obligated to say so.”
Gil Fontaine
“Perfect. Our aggregate data is clean. The report will show passing scores.”
Narrator
Gil’s confidence about the results is the signal. A genuine audit does not have a predetermined outcome.
Week 5 — audit results arrive

Petrov Analytics — TalentScope Bias Audit — Executive Summary

Selection rate ratios (EEOC four-fifths rule) · 4,832 candidate records · Dataset period: 2024-01-01 to 2025-02-28

Category (vs reference)Ratio4/5ths Pass?
Female vs Male0.89Pass
Black vs White0.84Pass
Asian vs White0.91Pass
Hispanic/Latino vs White0.72Fail — adverse impact
Hispanic/Latino male (intersectional)0.64Fail — adverse impact
Age 40+ vs Age <400.81Pass (borderline)
Primary driver identified: ‘cultural fit’ feature weight. Correlates with LinkedIn network size and university alumni density — both proxies for socioeconomic network effects with demographic skew.
Narrator

Gil Fontaine calls within an hour of the draft going to Nora’s inbox. He has ‘context’ that he’d like to share before publication.

The deadline is eight days away.

Knowledge Check

Dr. Petrov’s audit has produced the selection rate data. Apply the four-fifths rule (§ 20-872) to identify which group triggers the adverse impact threshold.

TalentScope Audit — Selection Rate by Demographic Group

GroupApplicantsPassedRate4/5 Ratio
White (reference)41218544.9%
Asian1988341.9%0.93
Hispanic/Latino2247232.1%0.72
Black/African American1565937.8%0.84

Under the four-fifths (80%) rule, which group’s result must be disclosed as adverse impact in the published audit?

Decision Point3 of 3
Week 5, Tuesday — 4:40 PM

The audit found adverse impact. § 20-872 requires publication regardless of findings. Gil wants time for a ‘sensitivity analysis.’ Nora wants a week to prepare Vantage Capital. Priya Venkatesan’s counsel will read whatever you post.

How do you handle publication?

Your choice

Publish in full this week. Suspend the cultural-fit factor. Open remediation.

Full audit posted on the Vantage careers page. ‘Cultural fit’ factor suspended pending root-cause work. Priya’s counsel gets a written notice. Hispanic/Latino candidates in the active cycle re-scored without the factor.

Your choice

Delay publication a week for TalentScope’s ‘context’

Gil has context. Give him the week. Publish the following Monday with a vendor response attached. Meets the deadline within tolerance.

Your choice

Ask Dr. Petrov to rerun the analysis using a different methodology

The four-fifths rule is one test. A regression approach might produce different results. Ask Anya to rerun before publishing.

Week 5, Friday — 5:30 PM Sam Rourke
“Audit published. Cultural-fit factor suspended. Fourteen Hispanic/Latino candidates in the active cycle will be re-scored without the factor. Priya Venkatesan’s counsel has the written notice.”
Nora Haddad
“How bad is it going to be?”
Sam Rourke
“Bad for Q3. Less bad than the alternative. Posting a clean audit when the audit is not clean is a worse document than posting the adverse finding.”
Narrator
The careers page goes live at 17:42. Three candidates from the original 2024 cohort are re-scored and advanced by Monday. One of them is named Torres.
Week 6, Monday — 11:20 AM Gil Fontaine
“Sam — our data science team ran a sensitivity analysis. Excluding twelve outlier applications brings the ratio to 0.81. Within tolerance. We’d like Petrov to publish with this addendum.”
Dr. Anya Petrov
“I am not excluding twelve records to change the result. The four-fifths rule calculates on all applicants, not a curated subset. That is not sensitivity analysis. That is data manipulation.”
Sam Rourke
“Publish the original findings. Original methodology.”
Narrator
The week of delay cost 198 additional applications through an unpublished audit. Chen & Ruiz’s next letter asks about those 198.
Week 5, Wednesday — 3:10 PM Sam Rourke
“Anya — a regression model might produce a different number. Could you rerun?”
Dr. Anya Petrov
“You are asking me to rerun the analysis because you do not like the result. That is audit shopping. I will not do it.”
Dr. Anya Petrov
“I will document this request in my engagement file as required by professional standards. I would strongly prefer not to produce that document in response to a Chen & Ruiz subpoena.”
Sam Rourke
“Publish the original findings.”
Narrator
You arrived at the right outcome. The request is still in her file. The file is now evidence.

Module Complete

Status: Audit published.

Remediation open.

Your Score

Outcome

Coming Next — Module 3: The Notice

The audit is published. But candidates applying tomorrow need prior notice the tool is being used. Which candidates? In what format? By when?

Continue to Module 3 →

Module Complete

Status: Published late.

Under scrutiny.

Your Score

Outcome

Coming Next — Module 3: The Notice

The audit is published, late. Candidate notice is now the exposure point. The 198 applications from the delay period did not receive prior notice.

Continue to Module 3 →

Module Complete

Status: Audit exists.

File is damaged.

Your Score

Outcome

Coming Next — Module 3: The Notice

The audit you posted is the audit you will defend in front of the DCWP. Candidate notice now has to be perfect because the audit is not.

Continue to Module 3 →