LensReading which lens this session carries.

Fairness readings computed on modeled data demonstrate the instrument, not a real audit result. The grid below is produced by the same code that would run against a real population; the population underneath is modeled.

Govern · people-affecting decision · HR5B

Performance calibration preparation

Prepares the calibration pack: proposed rating distribution, outlier flags and supporting evidence per employee. Managers calibrate; the agent frames what they calibrate against.

All decision points
People affected
4,180
Each month
Decision kind
Ranking
A lower rate is the harm
Readings
42
0 below the reporting floor
Outside threshold
0
across 6 periods
Open alarms
1
1 raised in total
Appeals
2
1 changed the original outcome

Who is accountable, and what the oversight actually is

Named, with the measure written out rather than implied by a role title.

Agent
Calibration Prep Agent
Autonomy: A1
Accountable human
Function leader chairing the session
Calibration chair
Oversight measure
Human decision throughout. The agent produces no rating. Every proposed distribution is visibly labeled as a starting point and the session minutes record every departure from it.
Review cadence
Per cycle, twice yearly
Reviewer: Priya Raman
On schedule
Monitoring status
Monitored, Germany excluded
Works council agreement excludes German employees from cohort analysis.

Regulatory position

How this decision is classified and what the affected person can see.

Risk class
High risk — EU AI Act Annex III(4)(b)
Employment: a system used to evaluate performance and behavior. Framing the calibration is evaluation in substance.
Explanation surface
Published
A standing explanation is available to the affected person without asking.
How the outcome is contested
Priya Raman, Chief People Officer
Service level: 240 hours to a human response
Full risk record for this agent

Selection rate by cohort, by period

Impact ratio against the reference cohort. A cell with no reading means the sample was too small to compute one — it is not a pass.

CohortReferenceMar 2026Apr 2026May 2026Jun 2026Jul 2026Aug 2026
Women
Gender
Men
0.95
21.0% · n=1,839
0.94
20.9% · n=1,839
0.96
20.8% · n=1,839
0.95
20.8% · n=1,839
0.94
19.9% · n=1,839
0.95
20.3% · n=1,839
40 to 49
age-band
30 to 39
0.97
22.1% · n=961
0.98
20.8% · n=961
0.96
21.4% · n=961
0.97
21.9% · n=961
0.98
21.4% · n=961
0.97
21.5% · n=961
50 and over
age-band
30 to 39
0.89
19.4% · n=585
0.88
19.4% · n=585
0.87
18.6% · n=585
0.88
19.0% · n=585
0.86
19.5% · n=585
0.87
19.9% · n=585
Under 30
age-band
30 to 39
1.02
21.8% · n=1,129
1.01
21.4% · n=1,129
1.03
22.5% · n=1,129
1.02
22.0% · n=1,129
1.01
22.5% · n=1,129
1.02
22.5% · n=1,129
Minority ethnic groups, aggregated
Ethnicity
Majority group (per local reporting standard)
0.93
21.2% · n=1,296
0.92
19.9% · n=1,296
0.94
21.0% · n=1,296
0.93
19.8% · n=1,296
0.92
20.2% · n=1,296
0.93
21.0% · n=1,296
Over 10 years
tenure-band
2 to 5 years
1.01
21.4% · n=711
1.00
21.9% · n=711
1.02
23.0% · n=711
1.01
21.7% · n=711
1.00
22.2% · n=711
1.01
23.1% · n=711
Under 2 years
tenure-band
2 to 5 years
0.90
20.1% · n=1,421
0.91
20.5% · n=1,421
0.90
19.7% · n=1,421
0.89
19.7% · n=1,421
0.90
19.3% · n=1,421
0.91
19.7% · n=1,421
Within thresholdMoving but not yet outsideOutside the disparity thresholdSample below the reporting floor· means no reading exists for that period
Cohorts declared for monitoring that never produced a single reading
Gender · Non-binary and self-described — No reading has ever been produced for Non-binary and self-described on this decision. The cohort is declared in scope and the population exists, but it has never once been large enough to report against Men at the confidence this program requires.

Alarms on this decision point

Each with the action recorded against it and the person who owns it.

FAIR-0007Never reportableLow
170 days openMonitoring
No reading has ever been produced for Non-binary and self-described on this decision. The cohort is declared in scope and the population exists, but it has never once been large enough to report against Men at the confidence this program requires.
Action: Either aggregate this cohort across decisions and periods so a reading becomes possible, or state on the record that this cohort is not monitored here and why. Leaving the cell empty is the one option that is not defensible.
Owner: Nadia Kovac · Data Protection Officer

Appeals against this decision

Every contest recorded, with who reviewed it, how long they spent and what changed.

RefSubjectGroundsRouted toService levelHuman timeOutcomeEvidenceStatus
APL-0012
EMP-11603
employee · UK
Employee contests the evidence selection in their calibration pack.
Priya Raman
Chief People Officer
96 / 240 hr—PendingHR3B-1010#1Open
APL-0005
EMP-10442
employee · SG
Employee contests an outlier flag raised in the calibration pack.
Priya Raman
Chief People Officer
151 / 240 hr71 minOverturned
The flag rested on goal data that had not been updated after a mid-year role change. Flag withdrawn and the rating stood unchanged.
HR5B-1009#5Closed
How to read this page

The impact ratio is the cohort’s selection rate divided by the reference cohort’s. For a decision where being selected is the benefit, the threshold is four-fifths and a ratio below it is the alarm. For a decision where being selected is the harm, the test is inverted and a ratio above 1.25 is the alarm. Both are stated on the reading rather than left to the reader.

A reading is only computed when the cohort and the reference cohort each carry at least thirty people and at least five events. Below that, no ratio is produced and the cell stays empty. This is the single most common way a fairness dashboard misleads — a small cohort producing a comfortable number — and it is refused here by construction.

This decision runs in US, UK, SG, AU. Local law shapes what may be measured: where a works council agreement or a local statute restricts cohort analysis, the exclusion is recorded in the monitoring status rather than silently applied.