Optimizer Card — map-elites v1.0.0 — Compliance

artifact_kind: optimizer-card artifact_name: map-elites version: 1.0.0 owner: Nous · agreement-search team last_reviewed: 2026-04-25 next_review: 2026-07-25 optimizer_kind: multi-objective related_phase_179_tasks: [179.4.2.2]#

Optimizer name: map-elites
Algorithm family: MAP-Elites (Mouret & Clune 2015) — quality- diversity search via behaviour-characterization grid.
Implementation source: libs/nous/agreement-search/src/map-elites.ts with shared GA primitives in ga-operators.ts.
Intended use: behaviour-diverse candidate exploration where the workbench surfaces structurally different agreement families along caller-defined behaviour axes (e.g., monetary-vs-restorative balance, immediate-vs-deferred payout, restitution share).

Utility aggregation: scalar fitness within each grid cell; caller-supplied (typically Nash product or weighted sum).
Fairness metrics reported: caller-selected; fairness-metrics.ts attaches to each cell-elite.
Handling of hard constraints: filterCandidates runs on every child before insertion.
Handling of uncertainty: within-cell elite is selected by fitness; fitness can be posterior mean / LCB / Thompson per the caller.

Candidate space: sandboxed clause-mutator DSL.
Behaviour characterization: caller-defined axes; the kernel discretizes them into a grid and keeps the highest-fitness candidate per cell.
Initial population / seeds: §179.4.1.1 sources.
Time / iteration budget: total iterations + cell-improvement patience.
Randomness: seeded mulberry32.

Accepted candidate contract: the populated grid — one elite per occupied cell. The workbench renders this as a heatmap of agreement families (§179.4.3.1).
Pareto-frontier diagnostics: cell-occupancy ratio, per-cell fitness, per-cell elite lineage.
Uncertainty propagation: per-cell elite carries the same posterior detail nash-genetic / NSGA-II emit.

Metric	Value	Evaluator card	Date
Pareto-front coverage	grid occupancy ≥ 60% on benchmark	n/a — `map-elites.test.ts` integration	2026-04-25
	scenarios
Diversity preservation	structurally distinct elites by	n/a — `map-elites.test.ts`	2026-04-25
	behaviour axis (asserted)
Fairness (Nash / KS / max-min)	per-cell metrics; gate runs against	fairness-suite	2026-04-25
	§179.10.6 baseline
Regret vs oracle	bounded by within-cell elite	baseline-benchmark-gate	2026-04-25
	selection
Specification-gaming resistance	`spec-gaming-audit.ts` runs per-cell	fairness-suite	2026-04-25
Runtime (typical)	seconds for ~50 clauses, 10×10 grid	n/a — deterministic synchronous kernel	2026-04-25
Cost per run	$0 once preference fits cached	n/a	2026-04-25

Behaviour-axis design is the caller's responsibility; poorly chosen axes give a misleading diversity picture.
Coarse grids miss family boundaries; fine grids waste compute on empty cells. The kernel reports occupancy so callers can iterate.
Cells empty under filterCandidates failures are reported but not imputed.

Candidate safety gate: filterCandidates on every child before cell insertion.
Clause static validation: inside filterCandidates.
Redline separation: hard constraints filtered before grid population.
No coercive heatmap: coercive-choice-audit.ts runs on the rendered grid before the workbench shows it.