Method and limits
How the agent reviews, and where it stops
The agent follows a fixed method. Its judgement goes into recording the facts and explaining the results; the outcomes and the recommendation come from rules anyone can read below.
The six questions
| Question | Fails when | Concern when | Grounded in |
|---|---|---|---|
| 1. Is it related to the job? | No link to the work or to joining. | Some link, but weak or indirect. | Proportionality (Puttaswamy); research on validity. |
| 2. Does it stand in for who someone is? | Strongly or directly tracks a personal characteristic. | Weakly tracks one. | Code on Wages s.3, RPwD, Transgender Persons and HIV/AIDS Acts; GDPR Art. 9 as reference. |
| 3. Is it the right person? | Content often belongs to someone else. | Matching can go wrong unless the candidate supplies the link. | DPDP s.8(3): accuracy of data used in decisions. |
| 4. Did the candidate publish it, and would they expect an employer to use it this way? | Content is often posted by others. | Self-posted, but for friends, not employers. | DPDP s.3(c)(ii): exemption only for what the person made public. |
| 5. Does it work equally for everyone? | Misreads or penalises some groups outright. | Works unevenly across groups. | Impact ratio and the fairness test. |
| 6. Is there a less intrusive way? | A less intrusive method answers the same question. | A partial alternative exists. | The necessity limb of the proportionality test. |
Outcome rules
Applied in order. The first rule that fires decides the outcome.
- R1If a signal stands in strongly or directly for a personal characteristic (question 2 fails), remove it.
- R2If a signal has no link to the job or to joining (question 1 fails), remove it.
- R3If a signal raises an open legal question, send it to a human decision-maker before any use.
- R4If a less intrusive method answers the same question (question 6 fails), restrict the signal and use that method instead.
- R5If the signal may read the wrong person, content others posted, or misread some groups (question 3, 4 or 5 fails), restrict it to human review under conditions.
- R6If three or more questions raise concerns, restrict the signal.
- R7Otherwise keep the signal, with a condition for every question that raised a concern.
Decision rules
- D1If removed signals carry more than half the scoring weight, or the vendor configuration fails the fairness test, or no signal survives in the automated score: do not adopt as proposed.
- D2Otherwise, if there is no validation study: pilot with safeguards only.
- D3Otherwise: adopt with conditions.
The agent may
- Read the case file, the vendor's documents and the sample reports
- Apply the six questions and the outcome rules, through its tools
- Run the fairness test under any configuration
- Search the approved evidence library and cite it
- Assess a signal the vendor did not list, marking its facts for human confirmation
- Explain, compare, recommend and draft the memo and candidate notice
The agent may not
- Approve or reject the proposal
- Give a definitive legal opinion
- Decide anything about a candidate, sample or real
- Screen, score or look up a real person
- Change a rule, a score or an outcome produced by the tools
When a person must take over
The agent hands a matter to a named person, using its hand-off tool, whenever it meets a legal interpretation (for example, whether the DPDP exemption covers bulk collection), a business judgement (whether culture fit should be scored at all), a fact it cannot verify, or the final decision. Each hand-off appears in the run trace and on the decision sheet.
Limits of this review
- The case is fictional. The method is general; the facts and numbers are illustrative.
- The facts recorded for each signal are judgements. They are written down so they can be challenged, and the rules make the consequences of each judgement explicit.
- The fairness test uses a synthetic pool built on stated assumptions. It shows what the scoring rules do, not what happened to real applicants.
- No test here can show whether the vendor tool predicts joining or retention. That requires outcome data the vendor has not supplied.
- Kestrel's own aptitude test is assumed to be fair and was not audited.
- Legal summaries are for review purposes and are not legal advice. Untested points are marked as open.
- The agent's answers can be wrong. Its numbers come from the tools; its explanations should be checked against the cited source.