fix(talent): bind material decision-stage automation and effective human oversight

[seonghobae]

Finding

Fresh review of proposed Talent owner PR #293 at exact 62250021ce3dc93fe085b532fc8fde7180aeaa71 found a narrower high-impact decision-provenance gap that is not covered by the current finalization rule.

ADR 0292 correctly says fully_automated_decision is non-authorizing for a high-impact Talent finalization, and that ai_assisted_human_decision requires substantive human intervention plus accountable-human confirmation. However, the current contract records a decision-level production mode without yet requiring provenance for material upstream stages that can already determine the practical outcome: eligibility screening, inclusion/exclusion, shortlist/slate formation, ordering/ranking, recommendation suppression, thresholding, or routing to/from human review.

That leaves a de-facto-automation loophole. An automated stage could irreversibly remove a worker from a pool/succession/mobility consideration set and a later human confirmation could still be recorded as ai_assisted_human_decision, even though the human never had a meaningful opportunity to see, evaluate, disregard, override, or reverse the material automated output. A final click must not retroactively turn an upstream automated material decision into meaningful human oversight.

This is a product/control contract defect in the Proposed Talent boundary, not permission to move legal applicability into the domain enum or to add a universal score.

Current authority and external evidence

• Protected Orgmetra base remains develop@eb9757f8649aaad026a9865508d9aad50c1a7a4f; docs(talent): select post-hire talent-management boundary #293 current head at finding time is 62250021ce3dc93fe085b532fc8fde7180aeaa71.
• Current ADR/TRACEABILITY already reject actor/signature/click as sufficient proof of substantive human intervention and prohibit fully automated high-impact finalization, but do not bind the production mode/effective oversight of each material decision stage.
• Current consolidated Regulation (EU) 2024/1689 (consolidated 2026-07-27) Article 14 describes effective human oversight for high-risk AI, including awareness of automation bias and the ability, as appropriate and proportionate, to understand system limitations and to disregard, override, reverse, intervene in, or stop AI output. Annex III includes specified employment/worker-management uses. This is design/traceability evidence only; applicability and staged effective dates remain a versioned legal-policy determination.
• Alon-Barkat & Busuioc (2023), Journal of Public Administration Research and Theory, 33(1), 153–169, https://doi.org/10.1093/jopart/muac007, shows why human presence alone is not a sufficient control model: human–algorithm interaction can exhibit uncritical or selective adherence, and the observed effects depend on context. This supports testing actual oversight behavior rather than treating a reviewer identity as proof.

Primary source: https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:02024R1689-20260727

Required causal repair in #293

Keep the repair documentation-only while ADR 0292 is Proposed. Align ADR, TRACEABILITY, and doctoring; do not create a production API/schema/UI merely to satisfy this finding.

1. Define a material decision stage as any automated or human step that can materially determine eligibility, inclusion/exclusion, shortlist/slate membership, ordinal recommendation, routing, or downstream high-impact action.
2. Preserve immutable stage-level provenance sufficient to reconstruct, for each material stage, whether it was human, AI-assisted-human, or fully automated; the model/tool/evidence/policy versions used; and its input/output lineage. A single final-decision enum is insufficient when earlier stages materially constrain the reachable outcome.
3. A later human confirmation must not relabel or erase an upstream fully automated material stage. fully_automated_decision remains non-authorizing for Talent high-impact finalization, and stage history remains immutable even if a human later accepts the result.
4. Define substantive human intervention operationally rather than by actor presence: before the material result becomes locked, the accountable reviewer must receive the relevant evidence and limitations, have an actual available path to disregard/override/reverse the automated output, and record the human disposition (accepted, modified, rejected, or equivalent controlled semantics) plus purpose-bound reason/provenance. Do not hard-code a legal conclusion from that evidence.
5. Automated exclusion/suppression must not make affected records or candidates invisible to the very reviewer expected to exercise oversight unless a separately reviewed policy explicitly authorizes that stage and the decision remains reconstructable/reviewable under the applicable rights contract.
6. Preserve DDD authority: talent_management records Talent planning/decision-stage provenance only; Job/FJA/KSAO, Person/Employment/Assignment, Position, performance, assessment, validity/fairness, and legal applicability remain with their existing owners.

RED acceptance before implementation

Add explicit pre-implementation negative contracts that fail when:

• a fully automated eligibility/filtering stage excludes a worker, then a later confirmation click causes the overall record to be represented as if the material path were human-decided;
• an AI recommendation is presented without its evidence/model/policy version or without an actual override/reversal path, yet finalization is accepted as substantively human-supervised;
• the reviewer can confirm but cannot inspect the material basis or recover an automatically suppressed option;
• stage production mode or human disposition is changed after the fact to alter governance/legal classification;
• fully_automated_decision at any material stage is silently collapsed into a final ai_assisted_human_decision label;
• a model-generated ranking/threshold becomes Talent or validity truth merely because a human accepted it.

GREEN is an internally consistent ADR/TRACEABILITY/doctoring contract on one exact head, independently reviewed, that preserves finalization denial plus stage-level provenance/effective-oversight semantics without claiming statutory applicability or production implementation.

Keep this issue open through the #293 repair and normal protected integration. It is not satisfied by adding a reviewer field, confirmation button, or generic "human in the loop" statement.

[seonghobae]

#294 finding은 #293의 current ordinary-forward branch에서 계속 수리 중이며, 이번 후속 리뷰까지 반영했습니다.

기존 stage-level repair는 62250021ce3dc93fe085b532fc8fde7180aeaa71에서 ADR e76d1cb805480990946fe79ff8c2225b085244aa → TRACEABILITY 0fe0b646bdb35c817a42ad59db7f098ace78513b → doctoring d907edf320fe291de1c373afbcc115f4bf416254 순으로 전진했습니다. Final decision label만 보존하지 않고 eligibility/filtering, inclusion/exclusion, shortlist/slate, ranking, suppression, thresholding, routing처럼 reachable outcome을 실질적으로 제한하는 모든 material decision stage를 immutable ordered provenance로 남깁니다. 각 stage는 실제 production mode, model/tool/version, exact evidence/policy versions, input/output effect lineage, actor/disposition을 보존하고, 이후 human confirmation이 upstream automated stage를 erase/relabel할 수 없습니다.

이번 fresh CodeRabbit review에서 d907edf...의 ADR step 4가 separately reviewed versioned policy로 affected worker/option을 reviewer에게 숨기는 자동화 단계를 예외적으로 허용할 수 있어, 같은 ADR의 effective-human-oversight invariant와 충돌하는 Major security finding이 확인됐습니다. 이를 ordinary-forward commit d130981b43353e80b1824cc87e33e086ccfefbb8에서 수리했습니다. Versioned policy는 applicability나 required evidence를 정할 수는 있지만 hidden/irreversible automated effect를 승인하거나 필수 pre-lock-in restore/reversal path를 제거할 수 없습니다. Material effect가 reviewer에게 숨겨져 있거나 finalization 전에 비가역적이거나 실제 사용 가능한 pre-lock-in restore/reversal path가 없으면 high-impact finalization은 항상 fail closed합니다. Policy approval/applicability rule은 이 invariant를 우회할 수 없습니다. TRACEABILITY는 이미 이보다 엄격했으므로 competing edit를 만들지 않았습니다.

해당 fifth inline review thread에는 exact reviewed head와 repair commit을 답변했고 resolved 처리했습니다. CodeRabbit도 d130981에서 addressed를 확인했습니다. 현재 #293의 inline review thread 5개는 모두 resolved/outdated이며 formal review inventory에는 COMMENTED만 있고 qualifying independent APPROVED는 없습니다.

Current exact head d130981...에서 Foundation 34499505610, SAST 34499505545, Security 34499505469, CodeQL PR 34499505637은 terminal SUCCESS입니다. 다만 documentation-only scope의 dependency-review/gitleaks/OSV/Trivy/Scorecard/Strix 세부 job은 skipped이므로 substantive security evidence로 승격하지 않습니다. Required OpenCode 34499503305/102946526098은 exact head admission과 central review request 뒤 current-head verdict가 아직 없다는 이유로 fail closed했고, Noema 34499503308/102946459162는 현재 exact head에서 계속 IN_PROGRESS입니다. Leaf rerun shim, empty commit, predecessor verdict transfer, self-approval, admin bypass는 사용하지 않습니다.

#294는 Proposed documentation repair가 존재한다는 이유로 완료 처리하지 않습니다. ADR 0292 independent review, accepted owner contracts, executable RED/GREEN implementation evidence, normal protected integration이 이루어질 때까지 open owner evidence를 유지합니다.