@kal-elsam/kairo-runtime 0.16.0 → 0.17.0

package/src/global/intelligence/capability-scoring.js

@@ -0,0 +1,480 @@
+// Robust multi-metric role scoring: percentile-normalized, confidence-aware,
+// never a fabricated zero for missing evidence. Consumes the same raw
+// CapabilityRegistry evidence (model-capability-registry.js) that AI TEAM
+// always has — this module owns turning that evidence into a per-role
+// score; it never collects evidence itself.
+//
+// Why percentiles, not the AA composite indices directly: AA's
+// intelligenceIndex/codingIndex are already a blend Kairo doesn't control
+// (which benchmarks, what weights) and mix 0-100/0-1 scales across
+// sources. Ranking on percentile position within Kairo's own real,
+// accessible candidate pool sidesteps the whole scale question — relative
+// order is scale-invariant — and stops one blended vendor number from
+// silently deciding a role Kairo could score on its own real component
+// benchmarks instead.
+//
+// Percentile alone can't decide EVERYTHING, though: with Kairo's typical
+// real candidate pool (2-3 accessible providers), a percentile rank is
+// always exactly {0, 0.5, 1} — it encodes ORDER, never real magnitude, by
+// construction. Two models 1% apart and two models 40% apart both just
+// read "1 vs 0" with only two real candidates. So this module exports two
+// parallel, deliberately different numbers per model: capabilityPercentile
+// (robust rank position — decides WHO'S AHEAD) and a separate real,
+// scale-normalized gap value (decides HOW CLOSE — near-equivalence bands
+// and capability floors in model-intelligence.js compare THIS, never the
+// percentile, so small pools keep real granularity instead of collapsing
+// every non-winner to a floor-failing 0).
+
+/**
+ * @typedef {"reasoning"|"coding"|"terminalExecution"|"softwareExecution"|"instructionFollowing"} Capability
+ */
+
+/**
+ * One real, distinct benchmark identity per row — `metricAliases` lists
+ * every metric NAME any connected source uses for that SAME real
+ * benchmark (e.g. AA's free API reports GPQA as "gpqa"; a manufacturer's
+ * own launch-page table reports the identical benchmark as
+ * "gpqa-diamond"). Deduplicated via bestEvidence-style pick (verified
+ * first, then most recent) BEFORE ranking, so the same real benchmark
+ * reported by two sources counts once, never twice in a capability's
+ * median. Distinct real benchmarks (GPQA vs HLE — Humanity's Last Exam is
+ * a different exam, not a rename) always stay separate rows.
+ *
+ * terminalBenchV2/terminalBenchHard/terminal-bench/terminal-bench-science
+ * are pooled as one identity here, inheriting the same grouping this
+ * codebase's CANONICAL_CAPABILITIES table already uses elsewhere — not a
+ * new decision introduced by this module.
+ *
+ * tauBanking is deliberately excluded from softwareExecution: it's kept
+ * as a separate agentic/tool-use signal, never counted as software
+ * engineering capability, per explicit product decision.
+ * Each alias also carries its own real `scale` ("unit": 0-1 fraction, as
+ * AA's free API and most independent evals report; "hundred": 0-100
+ * score, as manufacturer launch tables tend to report) — used ONLY by the
+ * real-gap calculation below, never by the percentile ranking (which is
+ * scale-invariant by construction).
+ * @type {Array<{benchmarkId: string, capability: Capability, metricAliases: Array<{metric: string, scale: "unit"|"hundred"}>}>}
+ */
+export const BENCHMARK_IDENTITIES = [
+  { benchmarkId: "gpqa", capability: "reasoning", metricAliases: [{ metric: "gpqa", scale: "unit" }, { metric: "gpqa-diamond", scale: "hundred" }] },
+  { benchmarkId: "hle", capability: "reasoning", metricAliases: [{ metric: "hle", scale: "unit" }] },
+  { benchmarkId: "mmlu-pro", capability: "reasoning", metricAliases: [{ metric: "mmluPro", scale: "unit" }] },
+  { benchmarkId: "livecodebench", capability: "coding", metricAliases: [{ metric: "liveCodeBench", scale: "unit" }] },
+  { benchmarkId: "scicode", capability: "coding", metricAliases: [{ metric: "sciCode", scale: "unit" }] },
+  { benchmarkId: "terminal-bench", capability: "terminalExecution", metricAliases: [
+    { metric: "terminalBenchV2", scale: "unit" }, { metric: "terminalBenchHard", scale: "unit" },
+    { metric: "terminal-bench", scale: "hundred" }, { metric: "terminal-bench-science", scale: "hundred" }
+  ] },
+  { benchmarkId: "cursorbench", capability: "softwareExecution", metricAliases: [{ metric: "cursorbench", scale: "hundred" }] },
+  { benchmarkId: "kairo-success", capability: "softwareExecution", metricAliases: [{ metric: "kairo.success", scale: "unit" }] },
+  { benchmarkId: "ifbench", capability: "instructionFollowing", metricAliases: [{ metric: "ifBench", scale: "unit" }] }
+];
+
+// A capability falls back to AA's own composite index ONLY when none of
+// its component benchmarks (above) produced a real, comparable cohort
+// anywhere in the current candidate pool — never alongside real component
+// benchmarks (that would double-count the same underlying capability:
+// once via its real components, once via the blended index that already
+// includes them).
+export const COMPOSITE_FALLBACKS = { reasoning: "intelligenceIndex", coding: "codingIndex" };
+
+function activeBenchmarkIdentities(capability) {
+  return BENCHMARK_IDENTITIES.filter((identity) => identity.capability === capability);
+}
+
+/** How many of a capability's real, distinct benchmark identities exist at all — reasoning: 3 (gpqa/hle/mmlu-pro), coding: 2, terminalExecution/instructionFollowing: 1 each. Static reference data (never per-model) — used by /models --evidence to render "X/Y benchmarks" alongside a real model's own count. */
+export function activeBenchmarkCountForCapability(capability) {
+  return activeBenchmarkIdentities(capability).length;
+}
+
+/**
+ * How many of this capability's real, distinct benchmark identities this
+ * EXACT model has real evidence for — counting identities (AA's "hle" and
+ * Hugging Face's "hle" are the SAME real benchmark, counted once), never
+ * raw evidence entries or sources. Deliberately independent of whether a
+ * comparable cohort formed for ranking (percentileForBenchmark's own,
+ * separate concern) — this measures how much of the capability THIS
+ * candidate's own evidence actually covers, regardless of who else is in
+ * the pool.
+ * @param {ReturnType<import("./model-capability-registry.js").createCapabilityRegistry>} registry
+ * @param {object} model
+ * @param {Capability} capability
+ * @returns {number}
+ */
+export function countModelBenchmarks(registry, model, capability) {
+  const id = registry.registerIdentity(model.adapterId, model.modelId, model.displayName ?? null);
+  let count = 0;
+  for (const identity of activeBenchmarkIdentities(capability)) {
+    if (bestAcrossAliases(registry, id, identity)) count += 1;
+  }
+  return count;
+}
+
+// A capability is "comparable" once real, distinct benchmark coverage
+// reaches at least half of its real active benchmarks, rounded UP —
+// reasoning (3 active) needs 2, coding (2 active) needs 1, and a
+// capability with only ever one real known benchmark (terminalExecution,
+// instructionFollowing) needs just that one — the 50% floor never
+// demands evidence the real catalog structurally can't provide. Below
+// that floor, a candidate is "provisional" for the capability: it can
+// still be scored and ranked (never excluded outright — a real single
+// data point is still real evidence), but never wins outright over a
+// real, more broadly comparable candidate — see model-intelligence.js's
+// own comparable-before-provisional selection order.
+const COMPARABILITY_THRESHOLD_RATIO = 0.5;
+
+/**
+ * @param {Capability} capability
+ * @param {number} benchmarkCount - this model's own real distinct benchmark count for the capability (see countModelBenchmarks)
+ * @returns {boolean}
+ */
+export function isCapabilityComparable(capability, benchmarkCount) {
+  const active = activeBenchmarkCountForCapability(capability);
+  if (active === 0) return true; // no known real benchmark for this capability at all — never gate on something unmeasurable
+  return benchmarkCount >= Math.ceil(active * COMPARABILITY_THRESHOLD_RATIO);
+}
+
+// Metrics where a LOWER value is the better real result. Every metric in
+// BENCHMARK_IDENTITIES today is higher-is-better; this stays a real,
+// checked table (not a hardcoded assumption baked into the ranking math)
+// so a future lower-is-better benchmark is handled correctly without
+// silently ranking it backwards.
+const LOWER_IS_BETTER_METRICS = new Set();
+
+function modelKey(model) {
+  return `${model.adapterId}::${model.modelId}`;
+}
+
+function directionFor(metric) {
+  return LOWER_IS_BETTER_METRICS.has(metric) ? "lower" : "higher";
+}
+
+function convertByScale(scale, value) {
+  return scale === "hundred" ? value / 100 : value;
+}
+
+/**
+ * The single most trustworthy real entry for a model across every metric
+ * alias of one benchmark identity — verified first, then most recent,
+ * exactly bestEvidence()'s own rule, just applied across aliases instead
+ * of a single metric name. This is the dedup step: the same real
+ * benchmark reported under two different source-specific key names never
+ * produces two competing data points for the same model. Returns the
+ * alias's own `scale` alongside the raw entry, for callers that need a
+ * real, unit-converted value (the gap calculation) — percentile ranking
+ * itself never needs this, since relative order is scale-invariant.
+ * @param {ReturnType<import("./model-capability-registry.js").createCapabilityRegistry>} registry
+ * @param {string} id
+ * @param {{metricAliases: Array<{metric: string, scale: string}>}} identity
+ */
+function bestAcrossAliases(registry, id, identity) {
+  let best = null;
+  for (const alias of identity.metricAliases) {
+    for (const entry of registry.getEvidence(id, alias.metric)) {
+      const isBetter = !best
+        || (entry.verified && !best.entry.verified)
+        || (entry.verified === best.entry.verified && Date.parse(entry.date ?? "") > Date.parse(best.entry.date ?? ""));
+      // The evidence's OWN real scale (set by the ingestion source that
+      // actually knows it) always wins over the alias's metric-name-based
+      // guess — two sources can share one metric name ("hle") while
+      // genuinely reporting in different real scales (AA: 0-1 fraction;
+      // Hugging Face: 0-100, verified live — DeepSeek-V4.1-Flash's real
+      // HLE via HF is 63.9, not 0.639). The alias scale stays a fallback
+      // only for evidence that never declared its own.
+      if (isBetter) best = { entry, scale: entry.scale ?? alias.scale };
+    }
+  }
+  return best;
+}
+
+/**
+ * Percentile rank (0-1) of each real, comparable result for one benchmark
+ * identity, among the given candidate models — scale-invariant (only
+ * relative order matters, so 0-1 fractions and 0-100 scores from
+ * different sources never need unit conversion here). Requires at least
+ * two distinct models with a real result for this exact benchmark WHEN
+ * other real candidate models exist in the pool but simply lack evidence
+ * for this one benchmark (that reduces coverage honestly instead of
+ * inventing a comparison with nobody). The one exception: when the ENTIRE
+ * candidate pool is a single model (the only accessible real provider —
+ * a real, common case, not a data gap), that lone model gets percentile 1
+ * for any benchmark it has real evidence for — being the only real
+ * option available IS the real, correct result, not a fabricated one.
+ * @param {ReturnType<import("./model-capability-registry.js").createCapabilityRegistry>} registry
+ * @param {Array<object>} models
+ * @param {{metricAliases: Array<{metric: string, scale: string}>}} identity
+ * @returns {Map<string, {value: number, verified: boolean}>} modelKey -> percentile + whether the deciding real entry was independently verified
+ */
+function percentileForBenchmark(registry, models, identity) {
+  const direction = directionFor(identity.metricAliases[0].metric);
+  const results = [];
+  for (const model of models) {
+    const id = registry.registerIdentity(model.adapterId, model.modelId, model.displayName ?? null);
+    const best = bestAcrossAliases(registry, id, identity);
+    // Different models can win bestAcrossAliases via different metric
+    // aliases of the SAME real benchmark identity (e.g. one model's best
+    // evidence is a 0-1 "unit" alias, another's is a 0-100 "hundred"
+    // alias) — comparing their raw entry.value would rank 0.64 below 57.9
+    // even though 0.64 (64%) is actually ahead of 57.9/100 (57.9%).
+    // Percentile ranking is scale-invariant only WITHIN one common scale,
+    // so every result is normalized to the same 0-1 scale before sorting.
+    if (best) results.push({ model, entry: best.entry, normalizedValue: convertByScale(best.scale, best.entry.value) });
+  }
+  if (results.length < 2 && models.length > 1) return new Map();
+  if (!results.length) return new Map();
+
+  const sorted = [...results].sort((a, b) => (direction === "higher" ? a.normalizedValue - b.normalizedValue : b.normalizedValue - a.normalizedValue));
+  const percentiles = new Map();
+  const n = sorted.length;
+  sorted.forEach((entry, index) => {
+    // Ties share the same percentile (their shared rank position), so a
+    // real dead-heat is never arbitrarily broken by array order here —
+    // portfolio-level tiebreaks happen later, on the final score.
+    let rankIndex = index;
+    while (rankIndex > 0 && sorted[rankIndex - 1].normalizedValue === entry.normalizedValue) rankIndex -= 1;
+    const percentile = n > 1 ? rankIndex / (n - 1) : 1;
+    percentiles.set(modelKey(entry.model), { value: percentile, verified: entry.entry.verified });
+  });
+  return percentiles;
+}
+
+/**
+ * @param {number[]} values
+ */
+function median(values) {
+  if (!values.length) return null;
+  const sorted = [...values].sort((a, b) => a - b);
+  const mid = Math.floor(sorted.length / 2);
+  return sorted.length % 2 === 0 ? (sorted[mid - 1] + sorted[mid]) / 2 : sorted[mid];
+}
+
+/**
+ * One capability's real percentile per model: the median across every
+ * real benchmark identity in that capability that produced a comparable
+ * cohort (see percentileForBenchmark) — median, not mean, so one real
+ * outlier benchmark can't swing the capability score on its own. Falls
+ * back to AA's own composite index (also percentile-ranked, same
+ * scale-invariance) ONLY when NOT ONE component benchmark produced any
+ * real cohort anywhere in the pool — never blended alongside real
+ * components.
+ * @param {ReturnType<import("./model-capability-registry.js").createCapabilityRegistry>} registry
+ * @param {Array<object>} models
+ * @param {Capability} capability
+ * @returns {Map<string, {percentile: number, benchmarkCount: number, verifiedCount: number}>}
+ */
+export function computeCapabilityPercentile(registry, models, capability) {
+  const identities = BENCHMARK_IDENTITIES.filter((identity) => identity.capability === capability);
+  const perModel = new Map(models.map((m) => [modelKey(m), { percentiles: [], verifiedCount: 0 }]));
+  let anyComponentCohort = false;
+
+  for (const identity of identities) {
+    const percentiles = percentileForBenchmark(registry, models, identity);
+    if (!percentiles.size) continue;
+    anyComponentCohort = true;
+    for (const [key, { value, verified }] of percentiles) {
+      const bucket = perModel.get(key);
+      bucket.percentiles.push(value);
+      if (verified) bucket.verifiedCount += 1;
+    }
+  }
+
+  if (!anyComponentCohort && COMPOSITE_FALLBACKS[capability]) {
+    const fallbackIdentity = { metricAliases: [{ metric: COMPOSITE_FALLBACKS[capability], scale: "hundred" }] };
+    const percentiles = percentileForBenchmark(registry, models, fallbackIdentity);
+    for (const [key, { value, verified }] of percentiles) {
+      const bucket = perModel.get(key);
+      bucket.percentiles.push(value);
+      if (verified) bucket.verifiedCount += 1;
+    }
+  }
+
+  const result = new Map();
+  for (const [key, bucket] of perModel) {
+    if (!bucket.percentiles.length) continue; // no real, comparable evidence at all — absent, never zero
+    result.set(key, {
+      percentile: median(bucket.percentiles),
+      benchmarkCount: bucket.percentiles.length,
+      verifiedCount: bucket.verifiedCount
+    });
+  }
+  return result;
+}
+
+/**
+ * One capability's real, scale-normalized magnitude per model — the
+ * median of each real benchmark identity's own best real value, converted
+ * to a common 0-1 scale via that alias's real `scale`. Unlike
+ * computeCapabilityPercentile, this needs no 2-model cohort (it's each
+ * model's own real value, not a rank), so it keeps real granularity even
+ * for a lone runner-up — exactly what a near-equivalence band or
+ * capability floor needs to mean anything with only 2-3 real candidates.
+ * Falls back to AA's own composite index under the same real rule as the
+ * percentile side: only when NOT ONE component benchmark has any real
+ * value anywhere in the pool, never blended alongside real components.
+ * @param {ReturnType<import("./model-capability-registry.js").createCapabilityRegistry>} registry
+ * @param {Array<object>} models
+ * @param {Capability} capability
+ * @returns {Map<string, number>} modelKey -> real 0-1 magnitude
+ */
+export function computeCapabilityGapValue(registry, models, capability) {
+  const identities = BENCHMARK_IDENTITIES.filter((identity) => identity.capability === capability);
+  const perModel = new Map(models.map((m) => [modelKey(m), []]));
+  let anyComponentValue = false;
+
+  for (const identity of identities) {
+    for (const model of models) {
+      const id = registry.registerIdentity(model.adapterId, model.modelId, model.displayName ?? null);
+      const best = bestAcrossAliases(registry, id, identity);
+      if (!best) continue;
+      anyComponentValue = true;
+      perModel.get(modelKey(model)).push(convertByScale(best.scale, best.entry.value));
+    }
+  }
+
+  if (!anyComponentValue && COMPOSITE_FALLBACKS[capability]) {
+    const fallbackIdentity = { metricAliases: [{ metric: COMPOSITE_FALLBACKS[capability], scale: "hundred" }] };
+    for (const model of models) {
+      const id = registry.registerIdentity(model.adapterId, model.modelId, model.displayName ?? null);
+      const best = bestAcrossAliases(registry, id, fallbackIdentity);
+      if (best) perModel.get(modelKey(model)).push(convertByScale(best.scale, best.entry.value));
+    }
+  }
+
+  const result = new Map();
+  for (const [key, values] of perModel) {
+    if (values.length) result.set(key, median(values));
+  }
+  return result;
+}
+
+/**
+ * A role's real, scale-normalized magnitude per model — the median across
+ * its relevant capabilities' own real gap values (see
+ * computeCapabilityGapValue). This is what model-intelligence.js's
+ * NEAR_EQUIVALENCE_BAND and EFFICIENT_CAPABILITY_FLOOR actually compare —
+ * never RoleEvaluation.capabilityPercentile, which only encodes rank
+ * order and collapses to {0, 0.5, 1} with Kairo's typical small candidate
+ * pools.
+ * @param {ReturnType<import("./model-capability-registry.js").createCapabilityRegistry>} registry
+ * @param {Array<object>} models
+ * @param {Capability[]} relevantCapabilities
+ * @returns {Map<string, number>} modelKey -> real 0-1 magnitude
+ */
+export function computeRoleGapValue(registry, models, relevantCapabilities) {
+  const perCapability = relevantCapabilities.map((capability) => computeCapabilityGapValue(registry, models, capability));
+  const result = new Map();
+  for (const model of models) {
+    const key = modelKey(model);
+    const values = perCapability.map((m) => m.get(key)).filter((v) => v != null);
+    if (values.length) result.set(key, median(values));
+  }
+  return result;
+}
+
+/**
+ * @typedef {object} RoleEvaluation
+ * @property {string} role
+ * @property {number} capabilityPercentile - median across the role's relevant capabilities, each already a median across that capability's real benchmarks
+ * @property {number} coverage - fraction of the role's relevant capabilities that produced a real score (0-1)
+ * @property {"high"|"medium"|"low"} confidence
+ * @property {number} benchmarkCount - total real benchmark identities that contributed, across all relevant capabilities
+ * @property {Record<string, number>} capabilities - per-capability percentile, only for capabilities with real evidence
+ * @property {Record<string, number>} benchmarkCountsByCapability - this model's own real, distinct benchmark-identity count per relevant capability (see countModelBenchmarks) — never a source count (AA HLE + Hugging Face HLE is one benchmark, not two).
+ * @property {Record<string, number>} benchmarkCoverage - benchmarkCountsByCapability[capability] / activeBenchmarkCountForCapability(capability), 0-1, per relevant capability.
+ * @property {string[]} provisionalCapabilities - relevant capabilities where this model's real coverage falls below the comparability floor (isCapabilityComparable) — empty when every relevant capability clears it.
+ * @property {boolean} isProvisional - true whenever provisionalCapabilities is non-empty.
+ */
+
+/**
+ * Confidence tiers, from real coverage and provenance only — never from
+ * the score's magnitude:
+ *   high:   >=70% of the role's relevant capabilities scored, AND at
+ *           least one contributing benchmark was independently verified.
+ *   medium: >=40% coverage, OR at least two distinct real benchmark
+ *           identities contributed (even if coverage is thin).
+ *   low:    anything short of that.
+ * A provisional candidate (real benchmark-identity coverage below the
+ * comparability floor for at least one relevant capability — see
+ * isCapabilityComparable) can never reach "high", even with a real,
+ * independently-verified single data point: one verified benchmark is
+ * real evidence, but not YET broad enough evidence to be that confident.
+ */
+function confidenceFor(coverage, benchmarkCount, verifiedCount, isProvisional = false) {
+  if (coverage >= 0.7 && verifiedCount >= 1 && !isProvisional) return "high";
+  if (coverage >= 0.4 || benchmarkCount >= 2) return "medium";
+  return "low";
+}
+
+/**
+ * Computes one RoleEvaluation per candidate model for a role defined by
+ * its relevant capabilities. A model with zero real evidence across every
+ * relevant capability gets no evaluation at all (absent from the
+ * returned map) — per explicit rule, it never competes for the role
+ * rather than being scored as if it were the weakest real option.
+ * @param {ReturnType<import("./model-capability-registry.js").createCapabilityRegistry>} registry
+ * @param {Array<object>} models
+ * @param {string} role
+ * @param {Capability[]} relevantCapabilities
+ * @returns {Map<string, RoleEvaluation>}
+ */
+export function computeRoleEvaluations(registry, models, role, relevantCapabilities) {
+  const perCapability = relevantCapabilities.map((capability) => ({
+    capability, scores: computeCapabilityPercentile(registry, models, capability)
+  }));
+
+  const evaluations = new Map();
+  for (const model of models) {
+    const key = modelKey(model);
+    const capabilities = {};
+    let benchmarkCount = 0;
+    let verifiedCount = 0;
+    for (const { capability, scores } of perCapability) {
+      const entry = scores.get(key);
+      if (!entry) continue;
+      capabilities[capability] = entry.percentile;
+      benchmarkCount += entry.benchmarkCount;
+      verifiedCount += entry.verifiedCount;
+    }
+    const scoredCapabilities = Object.keys(capabilities);
+    if (!scoredCapabilities.length) continue; // no real primary evidence anywhere — doesn't compete for this role
+
+    // Benchmark-LEVEL coverage per relevant capability — independent of
+    // whether a percentile cohort formed (that's a ranking concern; this
+    // is "how much of the capability does THIS model's own evidence
+    // cover"). Counts real, distinct benchmark identities, never sources.
+    const benchmarkCountsByCapability = {};
+    const benchmarkCoverage = {};
+    const provisionalCapabilities = [];
+    for (const capability of Object.keys(capabilities)) {
+      const count = countModelBenchmarks(registry, model, capability);
+      const active = activeBenchmarkCountForCapability(capability);
+      benchmarkCountsByCapability[capability] = count;
+      benchmarkCoverage[capability] = active ? count / active : 0;
+      // A score of 0 real component-benchmark identities can only mean
+      // this capability's real score came entirely from the composite-
+      // index fallback (computeCapabilityPercentile's own pool-wide
+      // switch) — the best real signal available in that mode, coarse
+      // but not partial, so it's never "provisional" for lacking
+      // component benchmarks it structurally couldn't have used anyway.
+      // Only real, PARTIAL component-benchmark coverage (count > 0, but
+      // still below the comparability floor) is provisional.
+      if (count > 0 && !isCapabilityComparable(capability, count)) provisionalCapabilities.push(capability);
+    }
+    const isProvisional = provisionalCapabilities.length > 0;
+
+    const coverage = scoredCapabilities.length / relevantCapabilities.length;
+    evaluations.set(key, {
+      role,
+      capabilityPercentile: median(scoredCapabilities.map((c) => capabilities[c])),
+      coverage,
+      confidence: confidenceFor(coverage, benchmarkCount, verifiedCount, isProvisional),
+      benchmarkCount,
+      capabilities,
+      benchmarkCountsByCapability, benchmarkCoverage, provisionalCapabilities, isProvisional
+    });
+  }
+  return evaluations;
+}
+
+/** Ordinal rank so confidence can be compared/sorted (higher is more confident). */
+export const CONFIDENCE_RANK = { high: 2, medium: 1, low: 0 };