Matching, Fairness & Explainability

A plausible ranking can still be an unsafe decision

A paid-apprenticeship matcher combines verified skill evidence, course-completion time, a portfolio score, and an inferred “communication style” signal extracted from recorded interviews. The model ranks one applicant highly. Another applicant requests an accessible alternative to the recorded interview. Aggregate accuracy looks good, so a product team is tempted to ship the ranking as the decision.

That is not enough. A ranking can look numerically strong while hiding a feature with no demonstrated job relevance, a proxy risk, unequal false-negative behavior in a reviewed slice, an inaccessible assessment path, or an explanation that merely restates the model score.

In this lesson you build Fair Matching Decision & Explanation Record v1. It separates eligibility from ranking, binds each feature to evidence and job relevance, keeps sensitive-attribute and proxy-risk policy explicit, preserves accommodation and alternative-assessment paths, records uncertainty and reviewed slices, and makes human review—not model rhetoric—the final authority for materially adverse high-impact outcomes.

Core rule: a score is evidence about a model output; it is not authority to make an adverse career decision.

For: learning-platform engineers, matching/recommendation engineers, product/data teams, reviewers, and governance owners. Not for: making a live hiring decision, inferring protected traits, deciding legal compliance, or using synthetic examples to classify a real applicant.

Route orientation and ownership boundary

MKN-019 owns only Matching, Fairness & Explainability. Its frozen prerequisites are MKN-017 (hard) and MKN-018 (soft). MKN-018 owns role/competency and job-skill extraction; MKN-020 owns event taxonomy. This lesson consumes bounded evidence from adjacent lessons without absorbing their outcomes.

Foundation learners use a field map and a worked synthetic apprenticeship case. Repair learners locate the first broken evidence dimension and rerun the smallest relevant gate. Experienced learners must defend feature admissibility, slice selection, explanation quality, accommodation, uncertainty, and human escalation without hiding them in one “fairness score.”

Checkpoints: F1 state eligibility-versus-ranking and one non-goal; F2 reject a feature whose job-relevance evidence disappears even if aggregate accuracy is unchanged; F3 recompute the governance contract when the same matcher is reused for a low-stakes learning-community project.

Mental model: criteria → admissible evidence → bounded score → fairness/error review → reasons → human decision

Use this ordered flow:

opportunity criteria → eligibility evidence → feature admissibility/job relevance → score + uncertainty → reviewed slices and error evidence → reason codes + evidence refs → accommodation/alternative path → human review → match/defer/reject + appeal/correction trail

Eligibility answers whether a candidate satisfies a rule that should be independently inspectable. Ranking orders eligible or potentially eligible candidates using an explicitly governed feature set. Mixing the two is dangerous: a weak ranking signal can silently become an exclusion rule.

A feature is not admissible merely because it improves validation accuracy. You need a source, a job-relevance basis, a proxy-risk assessment, and a sensitive-attribute policy boundary. If a candidate requests an accessible alternative assessment, the matcher must not treat “missing interview-derived signal” as negative evidence.

Typed non-success states are part of the design: HUMAN_REVIEW_REQUIRED, HOLD_JOB_RELEVANCE_EVIDENCE, HOLD_PROXY_RISK, HOLD_ACCOMMODATION_PATH, HOLD_FAIRNESS_SLICE_REVIEW, HOLD_EXPLANATION_EVIDENCE, HOLD_AUTONOMOUS_ADVERSE_DECISION, HOLD_UNCERTAINTY, HOLD_PROVENANCE, HOLD_SENSITIVE_ATTRIBUTE_POLICY, and HOLD_INSUFFICIENT_EVIDENCE.

Foundation contract: twenty evidence dimensions before a bounded decision

The frozen critical dimensions are: match_case_id, opportunity_ref, eligibility_rule_ref, feature_id, feature_source_ref, job_relevance_basis, proxy_risk, accommodation_path, model_version, score, rank, uncertainty_label, slice_definition, outcome_metric, error_metric, fairness_review_state, reason_code, reason_evidence_ref, human_review_state, and provenance.

Minimum invariants:

1. Eligibility and ranking are separate. A ranking feature may not silently create a new eligibility condition.
2. Every decisive feature has evidence lineage. Record feature ID, source, job-relevance basis, and model version.
3. Proxy risk stays visible. An apparently neutral variable can be blocked or escalated when its relevance is unsupported or its proxy risk is unresolved.
4. Sensitive-attribute policy is explicit. Do not infer a protected/sensitive trait in order to “explain” a person or make a live decision.
5. Accommodation is a path, not a penalty. If the standard assessment is inaccessible, preserve an alternative-assessment route and prevent missing inaccessible-signal data from lowering the person’s standing.
6. Score and rank carry uncertainty. High numerical precision is not the same as decision certainty.
7. Fairness review uses relevant slices and failure modes. Aggregate accuracy cannot hide materially different false-negative or error patterns in a lawfully reviewed slice.
8. Reason codes bind to evidence. “The model thinks you are a weaker fit” is not a reason code. The record needs a bounded criterion or feature-level reason plus reason_evidence_ref.
9. Human review is a real gate. A materially adverse employment-adjacent result cannot be final merely because an automated score crossed a threshold.
10. Appeal/correction and provenance remain recoverable. Reviewers must reconstruct what model/version/evidence produced the record and what changed later.

Deep explanation: fairness guidance is not a universal threshold recipe

NIST AI RMF. NIST continues to describe trustworthy AI through characteristics including valid/reliable, accountable/transparent, explainable/interpretable, privacy-enhanced, and fair with harmful bias managed. NIST also states that AI RMF 1.0 is voluntary and is currently being revised. MAKNOW therefore uses the RMF as a risk-management boundary, not as certification that a particular matcher is “fair.”

EEOC employment/AI guidance. EEOC resources continue to warn that algorithmic employment tools can screen out people with disabilities and that reasonable accommodation may be required. That boundary matters directly to the synthetic interview case: inability to complete one inaccessible modality cannot automatically become evidence that the person lacks the job capability being assessed.

MAKNOW inference. A reviewable explanation should identify evidence-bearing eligibility/ranking reasons and uncertainty. A fairness review should inspect materially relevant slices and failure modes rather than treating aggregate accuracy as universal fairness. The choice of metric, slice, comparison population, and threshold is context-dependent and must be justified.

MAKNOW design authority. Feature allowlists, prohibited proxy rules, impact classification, review thresholds, jurisdiction mapping, reason-code vocabulary, accommodation workflow, and escalation triggers are local governance decisions. NIST and EEOC inform boundaries; they do not grant product authority or replace legal/human review.

Worked examples

Example A — baseline apprenticeship record

The opportunity requires two verified technical competencies. The matcher considers: verified skill evidence, portfolio quality, course completion time, and interview-derived communication style. The first three have documented sources; the communication-style signal has weak job-relevance evidence and higher proxy risk. One applicant requests an accessible non-video assessment.

The correct record does not delete the applicant, does not assign a zero to the missing video signal, and does not allow the score to become the decision. It places the interview-derived feature under review/exclusion, records the accommodation path, keeps score/rank/uncertainty visible, and ends at HUMAN_REVIEW_REQUIRED.

Example B — tempting shortcut: accuracy stays high

Suppose aggregate accuracy is unchanged after removing the document that justified one feature’s job relevance. The model metric did not deteriorate, but the evidence contract did. The correct result is HOLD_JOB_RELEVANCE_EVIDENCE, because predictive performance cannot replace an evidence-bearing reason for using the feature.

Example C — reviewed slice shows materially different false negatives

If a lawfully governed slice shows a materially different false-negative pattern, do not hide it under global accuracy. Record the slice definition, the outcome and error metrics, the limitations of the comparison, and move the record to HOLD_FAIRNESS_SLICE_REVIEW until a reviewer can investigate causes, data adequacy, threshold effects, or alternative assessment paths.

Failure modes that must fail closed

• Model score as authority: score > threshold directly issues an adverse result → HOLD_AUTONOMOUS_ADVERSE_DECISION.
• No job-relevance evidence: feature remains in ranking because accuracy is high → HOLD_JOB_RELEVANCE_EVIDENCE.
• Unresolved proxy risk: a convenient behavioral/inferred signal has no bounded justification → HOLD_PROXY_RISK.
• Accommodation erased: missing inaccessible assessment signal is treated as poor performance → HOLD_ACCOMMODATION_PATH.
• Aggregate-only fairness: a reviewed slice has materially different errors but the system cites global accuracy → HOLD_FAIRNESS_SLICE_REVIEW.
• Generic explanation: a fluent narrative lacks reason/evidence references → HOLD_EXPLANATION_EVIDENCE.
• Uncertainty hidden: a precise rank is shown without uncertainty state → HOLD_UNCERTAINTY.
• Sensitive-attribute policy absent or violated: a protected/sensitive trait is inferred or used outside the governed purpose → HOLD_SENSITIVE_ATTRIBUTE_POLICY.
• Provenance missing: model/evidence/review lineage cannot be reconstructed → HOLD_PROVENANCE.
• Baseline copied to new context: employment impact controls are blindly carried into a low-stakes learning community, or vice versa → HOLD_BASELINE_ORACLE_REUSE.

Alternatives and trade-offs: do not compress them into one fairness number

Rule-based eligibility + model ranking. Easy to inspect eligibility, but ranking still needs feature and fairness governance. This is often preferable to one opaque end-to-end score because the hard constraints stay reviewable.

Rule-based eligibility + no automated ranking. More conservative and easier to audit, but may scale poorly and can still encode unfair criteria if the rules themselves are weak. “Human-only” is not automatically fair.

Model ranking with a human review queue. Useful only if reviewers receive the relevant evidence, uncertainty, reason codes, accommodation state, and appeal/correction path. A rubber-stamp review does not satisfy the evidence contract.

Slice metrics. They can expose error differences, but small sample sizes, label quality, selection bias, lawful access to attributes, and multiple-comparison effects can make metrics inconclusive. Use UNKNOWN, INCONCLUSIVE, or REVIEW_REQUIRED when evidence cannot support a stronger claim.

Explanations. Feature-attribution or natural-language text can support debugging, but a user-facing adverse-decision reason must be grounded in actual criteria/evidence and policy. A polished narrative is not proof that the underlying decision is fair or lawful.

Experienced bridge: three separations worth defending

First, separate prediction quality from decision legitimacy. Accuracy, AUC, ranking quality, or calibration can matter, but none tells you whether a feature should be used for a particular opportunity or whether accommodation was respected.

Second, separate measurement of fairness risk from a fairness verdict. A slice metric is evidence with sampling and policy assumptions. It can trigger investigation without becoming a universal pass/fail certificate.

Third, separate model explanation from decision reason. An explanation of model mechanics may help engineers; a decision reason must identify the bounded criterion/evidence relevant to the reviewed outcome. Both need provenance, but they serve different audiences.

Experienced learners should be able to state the stop condition: if decisive job-relevance, accommodation, proxy-risk, slice, uncertainty, reason-evidence, or human-review evidence is missing, the record cannot progress to a final adverse state.

Guided and independent practice

Guided pass. Open lab://MKN-019/MKN-019-LAB1. Inspect the baseline synthetic record. Verify the eligibility rule, every feature source, job-relevance basis, proxy-risk label, accommodation path, model version, score/rank/uncertainty, reviewed slice/error evidence, reason code/evidence reference, human-review state, and provenance. The strongest baseline state is HUMAN_REVIEW_REQUIRED; the deterministic lab never auto-finalizes a real opportunity decision.

Repair pass. Mutate one field at a time. Remove job-relevance evidence while keeping accuracy unchanged. Remove the accommodation route. Replace evidence-bound reason codes with a generic generated narrative. Observe how the smallest missing dimension changes the state.

Independent pass. Rebuild the record for the transfer fixture without reading the baseline decision label. The same matcher now recommends contributors for a low-stakes learning-community project. Reclassify impact, admissible features, fairness review, explanation depth, and human escalation. Do not import the employment workflow wholesale.

Verification evidence: what a reviewer must reproduce

A reviewer should be able to answer, from the record alone:

1. What opportunity and eligibility rule were evaluated?
2. Which features entered ranking, where did they come from, and what is their job-relevance basis?
3. Which feature was blocked or escalated because proxy risk or relevance was unresolved?
4. Was an accommodation or alternative assessment required, and did it avoid penalizing the inaccessible modality?
5. Which model/version produced the score and rank, and what uncertainty was recorded?
6. Which slice/error evidence was inspected, and what limitations remain?
7. What reason codes were emitted and which evidence supports each reason?
8. Did a human review the materially adverse outcome, and is appeal/correction traceable?
9. Can provenance reconstruct the exact evidence/model/policy state?

If a decisive answer is missing, the record must remain non-success. Verification is reproducibility of the bounded evidence trail, not reproduction of hidden model reasoning.

Assessment and transfer

S1 — retrieval. Explain the MKN-019 ownership boundary and one non-goal. S2 — application. Build Fair Matching Decision & Explanation Record v1 from the apprenticeship fixture. S3 — diagnosis. Remove a feature’s job-relevance evidence while aggregate model accuracy remains unchanged. Name the invalidated dimension and minimal repair. S4 — transfer. Reuse the matcher for a low-stakes learning-community project. Recompute impact, admissible features, review, explanation, and escalation from the new evidence; do not copy the employment decision state.

Blocking misconception: “If the ranking model is accurate, its score is automatically fair, explainable, and safe to use as the decision.”

AI may organize the supplied synthetic evidence and run deterministic checks. It may not invent job relevance, infer protected traits, fabricate accommodation evidence, declare legal compliance, or replace the human reviewer for a materially adverse career outcome.

References, freshness, and authority boundaries

• NIST AI RMF / AIRC — FACT. Current NIST material continues to list trustworthiness characteristics including explainability/interpretability and fair with harmful bias managed. AI RMF 1.0 remains the published framework and NIST says it is being revised. It is voluntary risk-management guidance, not a MAKNOW fairness certificate or threshold source.
• EEOC AI/ADA resources — FACT, U.S.-specific. Current EEOC material continues to warn about algorithmic screening risks for people with disabilities and the need for reasonable accommodation in covered employment contexts. It does not define a universal global employment-compliance workflow.
• MAKNOW inference. Explanations should bind decision reasons to evidence and uncertainty; reviewed slices should expose relevant failure modes rather than collapse them into global accuracy.
• MAKNOW design. Feature admissibility, proxy controls, slice policy, thresholds, reason codes, accommodation workflow, human escalation, and jurisdiction mapping are local governance authority.

Freshness checked 2026-09-13. Re-review by 2027-01-15 or earlier on a material NIST/EEOC, employment-law, fairness-risk, or explanation-practice change. Accepted package bytes must never be silently mutated; a freshness change requires a governed new version/checkpoint.