Decision sheet

Adopt Orbis SocialFit for campus hiring, 2027 cycle

Recommendation

Do not adopt as proposed

Do not adopt as proposed. Orbis SocialFit should not be used to score or shortlist campus applicants.

Why

9 of 13 signals fail the review and must be removed; they carry 69% of the tool's scoring weight.
Under the vendor's configuration, 4 group comparisons fall below the 0.80 impact-ratio line: college tier, language of posts, religion (synthetic label), social media presence.
After the review, no signal is left in the automated score. Shortlisting falls back to Kestrel's own aptitude test, and no group falls below the line.
The vendor has not provided a validation study, so the claim of "92% accurate at predicting who joins and stays" cannot be checked. There is no evidence the tool predicts joining or retention.
1 signal raises a legal question only Kestrel's counsel and Data Protection Officer can answer.

What Kestrel can do instead

The review does not leave Kestrel without an answer to its problem. These uses pass the six questions.

Ask every candidate directly about further-study plans and joining date at offer stage, and record the answer. This replaces the higher-studies signal, which addresses Kestrel's real problem: offers accepted and not taken up.
Verify claimed projects and skills against LinkedIn or GitHub links the candidate supplies, as a pass-or-query check by a person. Never rank on it, and never penalise a candidate without a portfolio.
If Kestrel wants a conduct check, a trained person may review only content the candidate published themselves that shows explicit threats, harassment or hate, in a language they understand, with the candidate given the chance to respond.
Hold the competing-offers signal until the Data Protection Officer and counsel have decided whether it may be used at all.

Signal outcomes

S01	Late-night posting	Remove
S02	Slang and profanity	Restrict
S03	Higher-studies preparation	Restrict
S04	Mentions of other job offers	Needs human decision
S05	Party and alcohol photos	Remove
S06	Political or protest content	Remove
S07	Religious and festival content	Remove
S08	Follower count and engagement	Remove
S09	Language of posts	Remove
S10	Network overlap with employees	Remove
S11	Engagement and wedding posts	Remove
S12	Little or no social media	Remove
S13	Professional profile consistency	Keep, with conditions

Gaps and limits

Missing

Validation study linking scores to joining and retention (missing): Without it, the 92% accuracy claim cannot be checked, and there is no evidence the tool predicts anything.
Description of training data (missing): If the model learned from past hires, it can learn past hiring bias, as a well-known retail recruiting tool did.
Independent bias audit (missing): No selection-rate comparison by gender, region, language or college tier has been done.
Languages the text models support (partial): The vendor confirms English. Hindi, Hinglish and other Indian languages are 'in beta'.
Method for confirming a profile belongs to the candidate (partial): The vendor matches on name and photo similarity. No error rate was given.
Platform permissions for data collection (missing): Several platforms prohibit using their data for background checks or surveillance.
Data retention and deletion terms (partial): Reports are kept for 'up to 5 years' with no link to the hiring purpose.

Conflicting

One small study (Kluemper et al., 2012) found trained raters' Facebook personality ratings correlated with job performance; a larger one (Van Iddekinge et al., 2016) found no link. Neither validates this vendor's signals.

Unclear or outside the agent's authority

Mentions of other job offers: Whether reading candidates' posts about competing employers is covered by the DPDP publicly-available exemption, and whether penalising a candidate for considering other offers is acceptable practice for Kestrel.
Whether the DPDP exemption for publicly available data covers a vendor's bulk collection of candidates' posts for scoring is untested.

Limits of this review

The fairness test uses 1,000 synthetic applicants built on stated assumptions. It shows what the scoring rules do to a realistic mix of people, not what happened to real applicants.
The review assumes Kestrel's aptitude test treats all groups equally. That test was not audited here; if it is biased, falling back to it inherits that bias.

Decisions reserved for people

Head of Talent Acquisition
Whether 'culture fit' should be scored at all for campus hires, and whether the offer-stage conversation is an acceptable substitute for prediction.
Data Protection Officer
Whether any collection of candidates' social media is lawful under the DPDP Act, including the open question on the competing-offers signal.
Both, jointly
The final decision on the proposal, recorded below with names, date and reasons.

Draft notice to candidates

What candidates would be told if Kestrel adopts only the permitted uses. Written to the DPDP standard: the data, the purpose, and how to withdraw or complain.

Kestrel Infotech Ltd.: how we check applications

If you give us links to your LinkedIn or GitHub profile, a member of our recruitment team may look at them to confirm the projects and skills you describe. We will not look at your other social media, and we will not use posts that other people have published about you.

Not having these profiles will never count against you.

At offer stage we will ask you about your joining date and any plans for further study, so we can plan your training place. Your answer will not affect an offer already made.

If anything we see raises a question, we will ask you about it before making any decision. You can ask us what we looked at, ask us to correct it, or withdraw your consent, by writing to privacy@kestrel.example. You may also complain to the Data Protection Board of India.

Sign-off

The agent’s recommendation is not the decision. A decision owner records it here.

NameRoleDecisionDateReasons (required if the recommendation is not followed)