science(talent): bind first-stage cohort to opportunity-origin provenance

[seonghobae]

Finding

Fresh review of Proposed Talent owner PR #293 at exact afa2927abffda4d19ee747ff3c301003c7dfd7a1 found a scientific/provenance gap before the denominator chain already covered by #396–#403.

The current contract is strong once a material DecisionOpportunitySet / StageCohortManifest exists: it preserves exact entry membership and counting unit, canonical correction resolution, immutable manifest lineage/as-of views, policy-regime identity, external semantic-dependency authority, and intended-versus-actual execution congruence. ADR 0292 also defines a material stage broadly enough to include inclusion/exclusion and routing.

What remains underspecified is how the first recorded opportunity set proves where it came from and whether it is complete for the population/opportunity universe that a buyer or later scientific analysis believes it represents.

Without an explicit origin/enumeration contract, all current invariants can be satisfied after an unrecorded pre-cohort filter. For example, an internal-mobility discovery rule, succession-nomination process, talent-pool sourcing algorithm, manager nomination, employee opt-in gate, or external import can construct M1 from only a subset of potentially relevant workers/options. If M1 is then treated as “the initial denominator,” later stage attrition and fairness evidence are internally consistent but scientifically conditional on an invisible earlier selection mechanism.

This is distinct from #396. #396 preserves the exact denominator for each recorded material stage. This issue establishes the boundary that prevents the formation of the first recorded denominator itself from becoming an unobserved selection step or from being misrepresented as population-complete.

Scientific evidence and scope

Alexander III, Song, Hickman, and Shin (2025), Sourcing algorithms: Rethinking fairness in hiring in the era of algorithmic recruitment, shows the analogous pre-applicant problem in recruitment. Sourcing algorithms can identify/screen potential applicants before they know of an opportunity, changing the applicant pool before applicant-level selection-rate analysis begins. The paper models sourcing as an additional first hurdle and warns that applicant-only analysis can miss disparity introduced before applicant-pool formation. DOI: 10.1111/ijsa.12499.

That paper is not direct empirical evidence about Orgmetra succession or internal mobility and does not make a legal rule universal. Its narrower scientific-design implication is applicable here: when a material discovery/nomination/exposure mechanism determines who can enter the first observed cohort, later analysis needs provenance for that mechanism and must not silently reinterpret the resulting cohort as an exhaustive source population.

Owner boundary

talent_management owns only purpose-minimized provenance for how a Talent process obtains its first opportunity cohort and what completeness/coverage semantics it is allowed to claim.

It does not become authoritative for Person, Employment, Assignment, Organization, Position, Job/FJA/KSAO, protected demographic, or scientific/fairness truth.

• authoritative worker/assignment/position/job universes remain with their existing Orgmetra owners;
• Talent consumes released/versioned owner snapshots, receipts, or ACL projections rather than cross-context SQL or source copying;
• protected-group joins and validity/fairness/adverse-impact interpretation remain workforce_validation authority;
• a source-universe receipt is provenance about coverage/origin, not proof that the process is valid, fair, or legally compliant.

Required contract refinement while ADR 0292 remains Proposed

Define an explicit opportunity-origin / source-universe provenance contract for the first material stage. Exact UL can change; do not freeze the provisional names below. Preserve these distinctions:

• every first-stage manifest declares a versioned origin mode or equivalent semantic contract, distinguishing at least exhaustive/enumerated source populations from intentionally non-exhaustive mechanisms such as nomination, opt-in, targeted discovery, or external import where applicable;
• if a process claims that the first manifest is population-complete for a defined worker/opportunity universe, it binds a released/versioned owner receipt or immutable snapshot identity for that source universe, its applicability/effective/system-time window, enumeration rule/version, and exact completeness/coverage semantics;
• if the origin is intentionally non-exhaustive, the manifest must say so explicitly and preserve the material exposure/discovery/nomination/opt-in provenance needed to explain who could enter the observed cohort. The downstream evidence must not relabel that cohort as the complete eligible population;
• any algorithmic or human prefilter that materially decides exposure, discovery, nomination, inclusion, or eligibility before M1 exists is itself a material stage or a versioned source-universe transformation with its own policy/evidence/execution provenance. It cannot disappear outside the stage graph merely because the product chose M1 as its first persisted cohort;
• the first manifest binds the exact origin/source receipt(s) used, and those receipts are included in its evidence identity/digest projection where material. A generic label such as all employees, eligible workers, recommended, or manager nominees is not completeness authority;
• source-universe replacement/correction is append-only and temporal. A later correction does not rewrite which universe/origin receipt governed a historical first-stage manifest or released analysis;
• changes in the enumeration/discovery rule that materially alter who can enter the process participate in the existing science(talent): separate policy-regime changes from denominator corrections #400/science(talent): make policy-regime identity reproducible and dependency-closed #401 policy-regime semantics rather than masquerading as data-only correction;
• when exposure/discovery itself has failures or unavailable owner evidence, those states remain visible in the origin/first-stage evidence instead of silently shrinking the denominator;
• multiple-membership, cross-classified, repeated-opportunity, and time structure are retained where the source universe requires them;
• workforce_validation receives enough released evidence to distinguish (a) the source universe or declared origin scope, (b) the first observed Talent cohort, and (c) later stage denominators before choosing an estimand or making fairness/validity claims.

The implementation does not need a universal all employees table or one giant snapshot. A purpose-bound owner receipt can define the relevant source universe for a particular process/version/time. The invariant is that coverage semantics are explicit and reproducible rather than inferred from the rows that happen to appear in M1.

RED acceptance cases

Before ADR acceptance or production implementation, fail at least when:

1. a caller supplies an arbitrary subset as the first manifest and labels it population_complete without a released/versioned owner source-universe receipt or equivalent evidence;
2. an algorithmic discovery/prescreen removes workers/options before M1 and no material-stage or source-transformation provenance records that effect;
3. a nomination-only, employee-opt-in, targeted-discovery, or externally imported cohort is later analyzed as though it were the complete eligible worker population without an explicit scientific limitation;
4. the source universe is identified only by a mutable query, branch, floating label, live cross-context SQL result, or free-text label such as all eligible employees;
5. the first manifest's source-universe/origin receipt is absent from the material provenance/digest identity even though changing it changes cohort membership;
6. a source-universe snapshot/receipt is stale, wrong-purpose, wrong-tenant, wrong-time, or incompatible with the manifest's policy/evidence window and is still accepted;
7. an enumeration/discovery policy changes materially but the existing first-stage policy regime is reused without science(talent): separate policy-regime changes from denominator corrections #400/science(talent): make policy-regime identity reproducible and dependency-closed #401-compatible semantic evidence;
8. source enumeration/exposure fails for part of the declared universe and the missing portion silently disappears rather than remaining explicit failure/unavailable coverage evidence;
9. a later source-universe correction mutates the historical M1 or an already released downstream result instead of creating append-only corrected evidence/new analysis identity;
10. Talent reconstructs source completeness by querying/copying People/Assignment/Position/Job application tables across bounded-context boundaries;
11. protected demographic attributes are copied into Talent merely to establish or reconcile the source universe;
12. workforce_validation receives only the first Talent manifest and cannot determine whether it arose from an exhaustive universe, targeted discovery, nomination, opt-in, or another materially selective origin mechanism.

Completion boundary

Documentation reconciliation alone is not completion. Keep this issue open through ADR/TRACEABILITY/doctoring reconciliation; executable owner/consumer contract(s) for first-stage origin/source-universe provenance; RED→GREEN tests; released/versioned owner evidence exchange without cross-context SQL/source-copy shortcuts; downstream workforce_validation reproducibility; and normal protected integration.

Do not close it merely because #396 records an “initial opportunity set” or because #403 proves that the procedure executed as declared. A perfectly executed procedure can still begin from an already selected cohort whose origin/coverage is scientifically opaque.

[seonghobae]

Owner-first documentation reconciliation is now complete without closing this issue.

• Talent TRACEABILITY: ordinary-forward commit dfa5af1b9f7047865091ff5f2cd05d1da93bcdbc
• ADR 0292: ordinary-forward commit 1fd6e10ba338010664e17a9c6c5543cfd9054b0f
• Reference doctoring: ordinary-forward commit / current PR docs(talent): select post-hire talent-management boundary #293 head e3ea0fc73203a91886d7021d9743af9fb6a62d6f
• Sole-writer commercialization baseline: PR docs: maintain durable product-technical gap baseline #100 current head c4de80bae23215841f8ed3aaba390d54450ba9ab

The contract now distinguishes authoritative enumerated/source-universe evidence from intentionally non-exhaustive nomination, employee opt-in, targeted discovery, external import, and equivalent origins; makes any material pre-M1 exposure/discovery/nomination/prescreen a material stage or versioned source-universe transformation; binds material origin receipts into first-manifest evidence identity/digest; keeps source-universe correction append-only; and leaves estimand/fairness/validity authority with workforce_validation.

This is not completion. #404 remains open until executable domain/schema/API contracts express the listed RED cases, causal GREEN evidence exists, released owner source-universe evidence exchange is reproducible, downstream workforce_validation can distinguish source universe/origin scope from M1 and later denominators without cross-service SQL/copied protected attributes, and normal protected integration is complete.