science(talent): preserve decision-opportunity denominators for stage-level fairness evidence

[seonghobae]

Finding

Fresh review of Proposed Talent owner PR #293 at exact d130981b43353e80b1824cc87e33e086ccfefbb8 found a narrower scientific/evidence gap after the #294 stage-provenance repair.

ADR 0292 now does a good job preserving the individual material decision path: eligibility/filtering, inclusion/exclusion, shortlist/slate membership, ranking, suppression, thresholding and routing retain ordered immutable stage provenance; hidden/irreversible automated effects fail closed; later human confirmation cannot relabel an upstream automated stage.

However, the current contract still does not require an immutable decision-opportunity denominator / stage cohort sufficient to reconstruct who was actually in scope at each material stage. The traceability file says buyer/scientific claims require sampling/error/failure denominators, but it does not bind those denominators to the same exact policy/evidence/time/stage lineage as the individual stage records.

That is a real auditability gap. If Talent records only the people who reached a shortlist, were explicitly excluded, or generated a decision record, a later Workforce Validation analysis can lose the denominator needed to distinguish:

• eligible from not_evaluable, stale/missing evidence, policy-inapplicable and authorization-denied states;
• initial opportunity set from later screened/slated/ranked sets;
• genuine absence from a cohort from silent suppression or missing upstream data;
• one stage/version/time window from another after criteria or policy changes;
• multiple-membership/cross-classified contexts from a single flattened population;
• attrition caused by a model/rule from attrition caused by human disposition or owner-data unavailability.

A per-person provenance chain is therefore necessary but not sufficient for stage-level fairness/adverse-impact evidence.

Authority boundary

This issue does not make talent_management the owner of protected demographic attributes, fairness verdicts, psychometric validity or legal discrimination determinations.

• talent_management should own only the purpose-minimized, immutable provenance needed to identify the exact decision opportunity set and stage transitions for a Talent process it owns.
• Protected/sensitive group attributes remain under their authoritative privacy/governance owner and are joined only through purpose-authorized released contracts or approved analysis snapshots.
• workforce_validation remains the owner of validity/fairness/scientific analysis and consumes released Talent cohort/provenance evidence rather than reading Talent tables directly.
• Legal applicability and reporting obligations remain versioned policy/compliance authority, not a Talent domain constant.

External evidence

These sources are design/scientific evidence, not universal legal applicability claims:

1. U.S. EEOC, Questions and Answers to Clarify and Provide a Common Interpretation of the Uniform Guidelines on Employee Selection Procedures. The guidance defines selection rate as selected ÷ applicants, treats hiring/promotion and other employment decisions as selection processes, and distinguishes the total process from component procedures. It also notes that small numbers and time aggregation matter. Current source: https://www.eeoc.gov/laws/guidance/questions-and-answers-clarify-and-provide-common-interpretation-uniform-guidelines
2. Alexander, L., III, Song, Q. C., Hickman, L., & Shin, H. J. (2025). Sourcing algorithms: Rethinking fairness in hiring in the era of algorithmic recruitment. International Journal of Selection and Assessment, 33, e12499. https://doi.org/10.1111/ijsa.12499. The paper shows that an algorithmic prescreen can alter the applicant pool before conventional applicant-level adverse-impact measures are calculated, so downstream parity can miss upstream access disparities.
3. Xu, H., & Zhang, N. (2024). Goal Orientation for Fair Machine Learning Algorithms. Production and Operations Management. https://doi.org/10.1177/10591478241234998. The paper emphasizes that ML outputs in organizational selection are often intermediate rather than the final decision, so fairness objectives must reflect the broader decision process rather than treating one model output as the whole outcome.
4. ISO 30415:2021, Human resource management — Diversity and inclusion, remains published as of this review and provides D&I governance/accountability guidance. ISO currently shows it in systematic-review stage 90.60, so its status must be rechecked before ADR acceptance or release claims: https://www.iso.org/standard/71164.html

Required repair while ADR 0292 remains Proposed

Keep this documentation/contract-first. Do not add production schema/API/UI before the owner ADR is accepted.

Define a versioned DecisionOpportunitySet / StageCohortManifest concept (name can change with UL review) that, for each material Talent decision stage, binds at minimum:

• tenant and purpose;
• process/case/pool/plan identifier and exact version;
• material stage identifier/order and causal predecessor;
• effective/system time or explicit measurement window;
• exact eligibility/decision-policy version and required evidence versions;
• immutable manifest digest/reference for the in-scope opportunity set;
• explicit transition counts and classifications such as advanced, not_evaluable, stale_or_missing_evidence, authorization_unavailable, policy_inapplicable, human_excluded, automated_excluded_or_suppressed, withdrawn, and other controlled semantics justified by the domain;
• failure/error denominator rather than silently dropping requests or records;
• multiple-membership/cross-classified/time context when relevant;
• linkage to the individual material-stage provenance records without copying protected demographic attributes into Talent.

The manifest should be immutable after a material stage closes. Corrections append a linked replacement/reconsideration version; they do not rewrite the original denominator.

RED acceptance before implementation

Add pre-implementation negative contracts that fail when:

• a shortlist/slate/ranking result exists but the exact upstream opportunity-set manifest is absent;
• excluded/suppressed records are silently omitted from the stage denominator;
• not_evaluable, stale/missing evidence, authorization failure or upstream owner unavailability is collapsed into a negative Talent outcome;
• a policy/criteria version changes but an old denominator is reused for a new stage;
• a cohort manifest can be mutated after the stage is finalized rather than superseded with provenance;
• one person with multiple legitimate memberships is silently reduced to one membership when the analysis design requires multiple-membership/cross-classified context;
• Workforce Validation must use cross-service SQL or copied protected demographics to reconstruct the denominator;
• a fairness/adverse-impact claim is produced from synthetic-only data or from an outcome set with no right-cleared, provenance-backed denominator.

GREEN at the Proposed-ADR layer is an internally consistent ADR/TRACEABILITY/doctoring contract on one exact head that preserves opportunity-set lineage while keeping fairness/validity and protected-group authority outside Talent.

Keep this issue open through normal protected integration and later executable owner implementation. Do not close it merely because individual stage provenance exists or because a dashboard can display selection percentages.

[seonghobae]

Documentation-layer repair is now carried ordinary-forward on canonical Talent owner PR #293 through exact head e723df83e4f36494de4b1ceb319810b1c092b38d: doctoring 05b9e149..., TRACEABILITY 89656b08..., ADR e723df83.... ADR 0292 now requires an immutable per-material-stage decision-opportunity/stage-cohort manifest tied to exact tenant/purpose/process/stage/policy/evidence/time lineage, explicit transition/failure states, append-only correction, and relevant multiple-membership/cross-classified/time structure while keeping protected demographics and fairness/validity/scientific verdicts outside Talent.

The durable product baseline sole-writer #100 has separately reconciled the same invariant at exact 545de588e550a66446f14ff3f7e98cb988a03002.

Keep #396 open. This documentation reconciliation is Proposed-contract GREEN only; it is not protected integration or executable implementation. Closure still requires normal ADR integration and later RED/GREEN owner contracts proving missing/silent/stale/mutable/flattened/cross-service/synthetic denominator failures fail closed.