Skip to main content

Deterministic Eligibility Gate Before Soft Ranking

Context​

CAN-REQ-UC-001 requires mandatory-criteria-compliant venue-space results to be separated from incomplete and ineligible candidates. Soft score, sponsorship or AI suggestion cannot rescue a hard violation.

This ADR is a gated draft. Its upstream use-case and rules must be accepted before it can advance.

Decision question​

Should mandatory eligibility be enforced by a deterministic, non-compensable gate before soft ranking, with three-state evidence handling (met | unmet | unknown)?

Decision drivers​

  • Hard requirements are non-compensable.
  • Unknown is not positive evidence.
  • Venue eligibility derives from eligible spaces.
  • AI cannot silently change criterion hardness.
  • Results must be explainable and reproducible.
  • Ranking may evolve without changing eligibility semantics.

Constraints and invariants​

  1. Any hard unmet makes the candidate ineligible.
  2. unknown is never met.
  3. Unknown behavior is controlled by explicit criterion policy.
  4. Soft ranking runs only after eligibility classification.
  5. Ineligible candidates do not enter result count, best-match or automatic shortlist.
  6. Canonical input, data snapshot, rule version and tie-break contract determine reproducible output.
  7. Authorization-safe candidate selection occurs before eligibility evaluation.

Options​

A — Single blended score​

Rejected candidate: soft relevance can mask mandatory failure.

B — Deterministic gate, then soft ranking​

Recommended candidate: partition into eligible, needs_info, ineligible; rank only eligible results.

C — AI-only classification​

Rejected candidate: cannot override hard rules or provide required determinism.

Proposed decision​

Adopt Option B, subject to upstream use-case acceptance and authorized ADR acceptance.

authorized candidate set
→ canonical criteria
→ hard eligibility evaluation
→ eligible / needs_info / ineligible
→ soft ranking of eligible results
→ deterministic tie-break
→ explanation and audit

Consequences​

Positive: invariant safety, explainability, test separation and ranking flexibility. Trade-offs: canonical criteria, explicit unknown policy, versioned rules/data and potentially fewer eligible results.

Risks and mitigations​

  • Incorrect hard/soft classification → persist provenance and authorized changes.
  • Missing data → needs_info, never automatic eligible.
  • Unstable ordering → deterministic tie-break and versioned inputs.
  • Tenant data leakage → authorization-safe candidate set.

Reversal/supersession​

Supersede if an accepted requirement changes eligibility semantics or verified evidence shows the gate cannot meet required targets without violating invariants.

Evidence plan​

Verify accepted/effective upstream use-case, resolvable IDs, option analysis, higher-record compatibility, R2 authority, AI CEO review and independent record-integrity evidence.

No implementation is authorized while this ADR is draft.

CAN-DEC-ADR-001 · governance_status: draft.