/**
 * llmindex.io — index version, weights and IRT hyperparameters (single source of truth)
 * Author:  Simon-Pierre Boucher
 * Contact: contact@spboucher.ai
 * License: Proprietary — © Simon-Pierre Boucher, all rights reserved
 *
 * NEVER hardcode weights elsewhere. Every change here bumps INDEX_VERSION
 * (semver) and adds an entry to docs/methodology/CHANGELOG.md.
 */
import { DOMAINS, type Domain } from './domains';

export const INDEX_VERSION = '0.2.0';

/** 2PL IRT fit hyperparameters (consumed by apps/psychometrics via /api/v1/methodology). */
export const IRT_HYPERPARAMS = {
  model: '2PL',
  priors: {
    theta: { dist: 'normal', mean: 0, sd: 1 },
    difficulty_b: { dist: 'normal', mean: 0, sd: 1.5 },
    log_discrimination_a: { dist: 'normal', mean: 0, sd: 0.5 },
  },
  maxIterations: 500,
  tolerance: 1e-6,
  /** Items below this discrimination are auto-flagged for retirement review. */
  minDiscrimination: 0.3,
  /** Items with |b| beyond this many logits are auto-flagged. */
  maxAbsDifficultyLogits: 3,
  /** Fixed anchor subset is capped at this fraction of any scored run. */
  maxAnchorFraction: 0.2,
} as const;

/**
 * Sub-metric weights inside a domain composite. Latency and cost are NEVER
 * blended in — they live on the efficiency frontier (Pareto), by design.
 *
 * Rationale (v0.2.0): accuracy_irt is the latent-ability estimate — the
 * primary construct — and dominates (0.60). Consistency (answer flip-rate
 * across seeded re-instantiations) and contamination resistance are
 * robustness corrections grounded in the template-memorization literature
 * (0.15 each); calibration rewards honest uncertainty but is the noisiest
 * sub-measurement at current sample sizes (0.10). Domain weights stay EQUAL:
 * with no task-utility function, the maximum-entropy prior is the only
 * non-arbitrary choice — per-domain scores are always published so any
 * consumer can re-weight.
 */
export const SUBMETRIC_WEIGHTS = {
  accuracy_irt: 0.6,
  consistency: 0.15,
  calibration: 0.1,
  contamination_resistance: 0.15,
} as const;

export type SubMetricKey = keyof typeof SUBMETRIC_WEIGHTS;

/** Domain weights for the Global Index (equal in v1). */
export const DOMAIN_WEIGHTS: Record<Domain, number> = Object.fromEntries(
  DOMAINS.map((d) => [d, 1 / DOMAINS.length]),
) as Record<Domain, number>;

/**
 * contamination_delta (fixed-vs-perturbed accuracy gap, in [0,1]) maps to a
 * resistance score: resistance = clamp01(1 - delta / CONTAMINATION_DELTA_FLOOR).
 * A gap ≥ 20 accuracy points ⇒ resistance 0.
 */
export const CONTAMINATION_DELTA_FLOOR = 0.2;

/** θ → 0–1000 rescale for display: INDEX = CENTER + SLOPE·θ, clamped. */
export const THETA_SCALE = { center: 500, slope: 150, min: 0, max: 1000 } as const;

/** CI half-width multiplier (95% normal). */
export const CI_Z = 1.96;