create-quorum-router 0.1.12 → 0.1.15

package/README.md

@@ -13,11 +13,12 @@ cd my-quorum-router-demo
 deno --version
 deno task smoke
 deno task calibration:demo
+deno task calibration:hierarchy-demo
 deno task intake
 deno task supabase:status
 ```
 
-Current package version: `create-quorum-router@0.1.12`. Releases are published
+Current package version: `create-quorum-router@0.1.15`. Releases are published
 from an immutable Git tag through GitHub Actions OIDC Trusted Publishing.
 
 ## What the generated project supports
@@ -25,11 +26,18 @@ from an immutable Git tag through GitHub Actions OIDC Trusted Publishing.
 `deno task smoke` is deterministic fixture-only and does not call a provider
 API.
 
-`deno task calibration:demo` runs the bundled calibration-by-task API against
-deterministic local observations. The report is advisory-only and is not
-connected to routing weights, provider eligibility, or execution. The command
-does not call provider APIs; on a new Deno installation, its first run resolves
-the pinned Zod dependency before execution.
+`deno task calibration:demo` runs the bundled flat calibration-by-task API
+against deterministic local observations. `deno task calibration:hierarchy-demo`
+runs three deterministic scenarios that select at the prompt-pattern level, then
+fall back to task subtype, then task type as narrower sample buckets become
+insufficient. Generated projects also export
+`aggregateHierarchicalTaskCalibration()` and
+`resolveHierarchicalTaskCalibration()` with
+`prompt_pattern → task_subtype → task_type` sample-count fallback. Reports are
+advisory-only and are not connected to routing weights, provider eligibility, or
+execution. Both demos are local-only; the hierarchy demo does not call provider
+APIs. On a new Deno installation, the first run resolves the pinned Zod
+dependency before execution.
 
 `deno task intake` is the first real setup command. It detects local provider
 wrappers, checks OAuth/session status, runs safe list-only model inventory where
@@ -40,6 +48,7 @@ the next command.
 deno task check
 deno task smoke
 deno task calibration:demo
+deno task calibration:hierarchy-demo
 deno task intake
 deno task auth:status
 deno task auth:login

package/bin/create-quorum-router.js

@@ -3,7 +3,7 @@ const fs = require("fs");
 const path = require("path");
 const { spawnSync } = require("child_process");
 
-const VERSION = "0.1.12";
+const VERSION = "0.1.15";
 const SUPPORTED_TEMPLATES = new Set(["basic"]);
 
 function usage() {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "create-quorum-router",
-  "version": "0.1.12",
+  "version": "0.1.15",
   "description": "Create a local QuorumRouter project.",
   "license": "MIT",
   "bin": {

package/templates/basic/README.md

@@ -1,7 +1,7 @@
 # QuorumRouter generated workspace
 
 This generated workspace contains the MIT-licensed QuorumRouter current release.
-npm latest targets v0.1.12.
+npm latest targets v0.1.15.
 
 QuorumRouter is **MIT**. It is **open source**. Commercial and production use
 are permitted under the MIT License.
@@ -38,6 +38,7 @@ deno --version
 ```bash
 deno task smoke
 deno task calibration:demo
+deno task calibration:hierarchy-demo
 deno task intake
 deno task auth:status
 deno task models:list
@@ -48,11 +49,23 @@ deno task supabase:status
 `smoke` proves the local scaffold runs with deterministic fixtures only. It does
 **not** call a real provider API.
 
-`calibration:demo` exercises the bundled calibration-by-task API with local
-fixture observations. Calibration reports are advisory-only: the scaffold does
-not use them to change routing weights, ranks, provider eligibility, quorum, or
-execution. The command does not call provider APIs; on a new Deno installation,
-its first run resolves the pinned Zod dependency before execution.
+`calibration:demo` exercises the bundled flat calibration-by-task API with local
+fixture observations. `calibration:hierarchy-demo` runs three deterministic
+queries that select a sufficient prompt-pattern group, fall back to task
+subtype, and fall back again to task type. Calibration reports are
+advisory-only: the scaffold does not use them to change routing weights, ranks,
+provider eligibility, quorum, or execution. Both demos are local-only; the
+hierarchy demo does not call provider APIs. On a new Deno installation, the
+first run resolves the pinned Zod dependency before execution.
+
+The hierarchical demo uses `aggregateHierarchicalTaskCalibration()` and
+`resolveHierarchicalTaskCalibration()` from `src/calibration.ts`. Its
+observations use caller-defined categories such as `task_type: "code-review"`,
+`task_subtype: "typescript"`, and `prompt_pattern: "schema-boundary-review"`.
+Resolution checks `prompt_pattern → task_subtype → task_type`, using the first
+group that reaches the configured sample threshold. Groups never cross exact
+provider/model source boundaries, labels must not contain raw prompts, and the
+result remains advisory-only.
 
 `intake` detects local provider wrappers, checks OAuth/session status, runs safe
 model inventory/list-only probes where possible, writes local health traces
@@ -228,8 +241,9 @@ not paste them into chat/logs and do not commit `.env`.
   `src/provider_client.ts` — provider discovery and safe invocation.
 - `src/best_route.ts`, `src/agent_chat.ts` — gated dogfood commands.
 - `src/cost_aware.ts` — estimated-cost budget selection for Best Route.
-- `src/calibration.ts`, `src/calibration_demo.ts` — strict advisory calibration
-  aggregation and an offline runnable example.
+- `src/calibration.ts`, `src/calibration_demo.ts` — strict flat and hierarchical
+  advisory aggregation, parent fallback resolution, and an offline runnable
+  example.
 - `src/trace.ts`, `src/redact.ts`, `src/schema.ts`, `src/fixture_smoke.ts` —
   trace/redaction/schema/fixture support.
 - `out/.gitkeep` — local output directory placeholder.

package/templates/basic/deno.json

@@ -7,6 +7,7 @@
     "check": "deno check main.ts src/*.ts",
     "smoke": "deno run main.ts",
     "calibration:demo": "deno run src/calibration_demo.ts",
+    "calibration:hierarchy-demo": "deno run src/hierarchical_calibration_demo.ts",
     "intake": "deno run --allow-read --allow-write --allow-env --allow-net --allow-run=grok,codex,claude,gemini,devin src/cli.ts intake",
     "auth:status": "deno run --allow-read --allow-env --allow-run=grok,codex,claude,gemini,devin src/cli.ts auth:status",
     "auth:login": "deno run src/cli.ts auth:login",

package/templates/basic/src/calibration.ts

@@ -28,6 +28,10 @@ const ObservationIdSchema = boundedCanonicalText(
   MAX_CALIBRATION_IDENTIFIER_LENGTH,
 );
 const TaskTypeSchema = boundedCanonicalText(MAX_CALIBRATION_IDENTIFIER_LENGTH);
+const HierarchicalTaxonomyLabelSchema = boundedCanonicalText(128).refine(
+  (value) => /^[a-z0-9](?:[a-z0-9._:/-]{0,126}[a-z0-9])?$/.test(value),
+  "hierarchical taxonomy labels must be lowercase canonical ASCII identifiers",
+);
 const SourceLabelSchema = boundedCanonicalText(
   MAX_CALIBRATION_IDENTIFIER_LENGTH,
 );
@@ -247,3 +251,497 @@ export function aggregateTaskCalibration(
     groups,
   });
 }
+
+export const MAX_HIERARCHICAL_CALIBRATION_GROUPS =
+  MAX_CALIBRATION_OBSERVATIONS * 3;
+
+export const HierarchicalCalibrationScopeSchema = z.enum([
+  "task_type",
+  "task_subtype",
+  "prompt_pattern",
+]);
+
+export const HierarchicalTaskCalibrationObservationSchema = z.object({
+  observation_id: ObservationIdSchema,
+  task_type: TaskTypeSchema,
+  task_subtype: HierarchicalTaxonomyLabelSchema.optional(),
+  prompt_pattern: HierarchicalTaxonomyLabelSchema.optional(),
+  source: TaskCalibrationSourceSchema,
+  evaluation_basis: z.literal("caller_attested_external_ground_truth"),
+  correct: z.boolean(),
+  confidence: z.number().finite().min(0).max(1),
+  evaluated_at: EvaluationTimestampSchema,
+}).strict().superRefine((observation, context) => {
+  if (
+    observation.prompt_pattern !== undefined &&
+    observation.task_subtype === undefined
+  ) {
+    context.addIssue({
+      code: "custom",
+      message: "prompt_pattern requires task_subtype",
+      path: ["prompt_pattern"],
+    });
+  }
+});
+
+export const HierarchicalTaskCalibrationObservationListSchema = z.array(
+  HierarchicalTaskCalibrationObservationSchema,
+).max(MAX_CALIBRATION_OBSERVATIONS).superRefine((observations, context) => {
+  const ids = new Set<string>();
+  for (const [index, observation] of observations.entries()) {
+    if (ids.has(observation.observation_id)) {
+      context.addIssue({
+        code: "custom",
+        message: "observation_id must be unique",
+        path: [index, "observation_id"],
+      });
+    }
+    ids.add(observation.observation_id);
+  }
+});
+
+export const HierarchicalTaskCalibrationGroupSchema = z.object({
+  scope: HierarchicalCalibrationScopeSchema,
+  task_type: TaskTypeSchema,
+  task_subtype: HierarchicalTaxonomyLabelSchema.optional(),
+  prompt_pattern: HierarchicalTaxonomyLabelSchema.optional(),
+  source: TaskCalibrationSourceSchema,
+  sample_count: z.number().int().positive().max(MAX_CALIBRATION_OBSERVATIONS),
+  accuracy: z.number().finite().min(0).max(1),
+  mean_confidence: z.number().finite().min(0).max(1),
+  brier_score: z.number().finite().min(0).max(1),
+  mean_calibration_bias: z.number().finite().min(-1).max(1),
+  sample_status: z.enum(["insufficient", "sufficient"]),
+}).strict().superRefine((group, context) => {
+  const validLabels =
+    (group.scope === "task_type" && group.task_subtype === undefined &&
+      group.prompt_pattern === undefined) ||
+    (group.scope === "task_subtype" && group.task_subtype !== undefined &&
+      group.prompt_pattern === undefined) ||
+    (group.scope === "prompt_pattern" && group.task_subtype !== undefined &&
+      group.prompt_pattern !== undefined);
+  if (!validLabels) {
+    context.addIssue({
+      code: "custom",
+      message: "scope must match task_subtype and prompt_pattern labels",
+    });
+  }
+});
+
+export const HierarchicalTaskCalibrationReportSchema = z.object({
+  schema_version: z.literal("quorum-router.hierarchical-calibration.v1"),
+  advisory_only: z.literal(true),
+  minimum_sample_count: z.number().int().positive().max(
+    MAX_CALIBRATION_OBSERVATIONS,
+  ),
+  groups: z.array(HierarchicalTaskCalibrationGroupSchema).max(
+    MAX_HIERARCHICAL_CALIBRATION_GROUPS,
+  ),
+}).strict().superRefine((report, context) => {
+  const groupKeys = new Set<string>();
+  for (const [index, group] of report.groups.entries()) {
+    const groupKey = JSON.stringify([
+      group.scope,
+      group.task_type,
+      group.task_subtype ?? null,
+      group.prompt_pattern ?? null,
+      group.source.provider,
+      group.source.model,
+    ]);
+    if (groupKeys.has(groupKey)) {
+      context.addIssue({
+        code: "custom",
+        message: "hierarchical calibration report groups must be unique",
+        path: ["groups", index],
+      });
+    }
+    groupKeys.add(groupKey);
+
+    const expectedStatus = group.sample_count < report.minimum_sample_count
+      ? "insufficient"
+      : "sufficient";
+    if (group.sample_status !== expectedStatus) {
+      context.addIssue({
+        code: "custom",
+        message: "sample_status must match the configured sample threshold",
+        path: ["groups", index, "sample_status"],
+      });
+    }
+    if (
+      group.mean_calibration_bias !==
+        stableMetric(group.mean_confidence - group.accuracy)
+    ) {
+      context.addIssue({
+        code: "custom",
+        message: "mean_calibration_bias must equal mean_confidence - accuracy",
+        path: ["groups", index, "mean_calibration_bias"],
+      });
+    }
+  }
+});
+
+export const HierarchicalTaskCalibrationQuerySchema = z.object({
+  task_type: TaskTypeSchema,
+  task_subtype: HierarchicalTaxonomyLabelSchema.optional(),
+  prompt_pattern: HierarchicalTaxonomyLabelSchema.optional(),
+  source: TaskCalibrationSourceSchema,
+}).strict().superRefine((query, context) => {
+  if (query.prompt_pattern !== undefined && query.task_subtype === undefined) {
+    context.addIssue({
+      code: "custom",
+      message: "prompt_pattern requires task_subtype",
+      path: ["prompt_pattern"],
+    });
+  }
+});
+
+const HierarchicalCalibrationCandidateSchema = z.object({
+  scope: HierarchicalCalibrationScopeSchema,
+  sample_count: z.number().int().positive().max(MAX_CALIBRATION_OBSERVATIONS)
+    .nullable(),
+  sample_status: z.enum(["missing", "insufficient", "sufficient"]),
+}).strict().superRefine((candidate, context) => {
+  if (
+    (candidate.sample_status === "missing") !==
+      (candidate.sample_count === null)
+  ) {
+    context.addIssue({
+      code: "custom",
+      message: "missing candidates must have a null sample_count",
+    });
+  }
+});
+
+export const HierarchicalTaskCalibrationSelectionSchema = z.object({
+  schema_version: z.literal("quorum-router.hierarchical-selection.v1"),
+  advisory_only: z.literal(true),
+  query: HierarchicalTaskCalibrationQuerySchema,
+  requested_scope: HierarchicalCalibrationScopeSchema,
+  resolution_status: z.enum(["selected", "no_sufficient_group"]),
+  selected_scope: HierarchicalCalibrationScopeSchema.nullable(),
+  candidates: z.array(HierarchicalCalibrationCandidateSchema).min(1).max(3),
+  selected_group: HierarchicalTaskCalibrationGroupSchema.optional(),
+}).strict().superRefine((selection, context) => {
+  const queryScope = selection.query.prompt_pattern !== undefined
+    ? "prompt_pattern"
+    : selection.query.task_subtype !== undefined
+    ? "task_subtype"
+    : "task_type";
+  if (selection.requested_scope !== queryScope) {
+    context.addIssue({
+      code: "custom",
+      message: "requested_scope must match the query labels",
+      path: ["requested_scope"],
+    });
+  }
+
+  const expectedScopes = selection.requested_scope === "prompt_pattern"
+    ? ["prompt_pattern", "task_subtype", "task_type"]
+    : selection.requested_scope === "task_subtype"
+    ? ["task_subtype", "task_type"]
+    : ["task_type"];
+  if (
+    JSON.stringify(selection.candidates.map((candidate) => candidate.scope)) !==
+      JSON.stringify(expectedScopes)
+  ) {
+    context.addIssue({
+      code: "custom",
+      message: "candidates must follow the requested scope fallback order",
+      path: ["candidates"],
+    });
+  }
+
+  const firstSufficient = selection.candidates.find((candidate) =>
+    candidate.sample_status === "sufficient"
+  );
+  if (firstSufficient === undefined) {
+    if (
+      selection.resolution_status !== "no_sufficient_group" ||
+      selection.selected_scope !== null ||
+      selection.selected_group !== undefined
+    ) {
+      context.addIssue({
+        code: "custom",
+        message: "no sufficient candidate must produce no selection",
+      });
+    }
+  } else if (
+    selection.resolution_status !== "selected" ||
+    selection.selected_scope !== firstSufficient.scope ||
+    selection.selected_group?.scope !== firstSufficient.scope ||
+    selection.selected_group.sample_count !== firstSufficient.sample_count ||
+    selection.selected_group.sample_status !== "sufficient" ||
+    selection.selected_group.task_type !== selection.query.task_type ||
+    selection.selected_group.source.provider !==
+      selection.query.source.provider ||
+    selection.selected_group.source.model !== selection.query.source.model ||
+    (firstSufficient.scope !== "task_type" &&
+      selection.selected_group.task_subtype !== selection.query.task_subtype) ||
+    (firstSufficient.scope === "prompt_pattern" &&
+      selection.selected_group.prompt_pattern !==
+        selection.query.prompt_pattern)
+  ) {
+    context.addIssue({
+      code: "custom",
+      message: "selection must use the first sufficient candidate",
+    });
+  }
+});
+
+export const HierarchicalTaskCalibrationDecisionSchema = z.object({
+  report: HierarchicalTaskCalibrationReportSchema,
+  selection: HierarchicalTaskCalibrationSelectionSchema,
+}).strict().superRefine((decision, context) => {
+  const { query } = decision.selection;
+  const expectedGroups = decision.selection.candidates.map((candidate) =>
+    decision.report.groups.find((group) =>
+      group.scope === candidate.scope && group.task_type === query.task_type &&
+      group.source.provider === query.source.provider &&
+      group.source.model === query.source.model &&
+      (candidate.scope === "task_type" ||
+        group.task_subtype === query.task_subtype) &&
+      (candidate.scope !== "prompt_pattern" ||
+        group.prompt_pattern === query.prompt_pattern)
+    )
+  );
+
+  decision.selection.candidates.forEach((candidate, index) => {
+    const group = expectedGroups[index];
+    const expected = group === undefined
+      ? { sample_count: null, sample_status: "missing" }
+      : {
+        sample_count: group.sample_count,
+        sample_status: group.sample_status,
+      };
+    if (
+      candidate.sample_count !== expected.sample_count ||
+      candidate.sample_status !== expected.sample_status
+    ) {
+      context.addIssue({
+        code: "custom",
+        message: "selection candidate must match the aggregate report",
+        path: ["selection", "candidates", index],
+      });
+    }
+  });
+
+  const expectedSelected = expectedGroups.find((group) =>
+    group?.sample_status === "sufficient"
+  );
+  if (
+    JSON.stringify(decision.selection.selected_group) !==
+      JSON.stringify(expectedSelected)
+  ) {
+    context.addIssue({
+      code: "custom",
+      message: "selected_group must match the first sufficient report group",
+      path: ["selection", "selected_group"],
+    });
+  }
+});
+
+export type HierarchicalCalibrationScope = z.infer<
+  typeof HierarchicalCalibrationScopeSchema
+>;
+export type HierarchicalTaskCalibrationObservation = z.infer<
+  typeof HierarchicalTaskCalibrationObservationSchema
+>;
+export type HierarchicalTaskCalibrationGroup = z.infer<
+  typeof HierarchicalTaskCalibrationGroupSchema
+>;
+export type HierarchicalTaskCalibrationReport = z.infer<
+  typeof HierarchicalTaskCalibrationReportSchema
+>;
+export type HierarchicalTaskCalibrationQuery = z.infer<
+  typeof HierarchicalTaskCalibrationQuerySchema
+>;
+export type HierarchicalTaskCalibrationSelection = z.infer<
+  typeof HierarchicalTaskCalibrationSelectionSchema
+>;
+export type HierarchicalTaskCalibrationDecision = z.infer<
+  typeof HierarchicalTaskCalibrationDecisionSchema
+>;
+
+type HierarchicalGroupSeed = {
+  scope: HierarchicalCalibrationScope;
+  task_type: string;
+  task_subtype?: string;
+  prompt_pattern?: string;
+  source: TaskCalibrationSource;
+  observations: HierarchicalTaskCalibrationObservation[];
+};
+
+function summarizeHierarchicalGroup(
+  seed: HierarchicalGroupSeed,
+  minimumSampleCount: number,
+): HierarchicalTaskCalibrationGroup {
+  const sample_count = seed.observations.length;
+  const correctCount = seed.observations.reduce(
+    (total, observation) => total + Number(observation.correct),
+    0,
+  );
+  const confidenceTotal = stableSum(
+    seed.observations.map((observation) => observation.confidence),
+  );
+  const brierTotal = stableSum(seed.observations.map((observation) => {
+    const outcome = Number(observation.correct);
+    return (observation.confidence - outcome) ** 2;
+  }));
+  const accuracy = stableMetric(correctCount / sample_count);
+  const mean_confidence = stableMetric(confidenceTotal / sample_count);
+
+  return HierarchicalTaskCalibrationGroupSchema.parse({
+    scope: seed.scope,
+    task_type: seed.task_type,
+    ...(seed.task_subtype === undefined
+      ? {}
+      : { task_subtype: seed.task_subtype }),
+    ...(seed.prompt_pattern === undefined
+      ? {}
+      : { prompt_pattern: seed.prompt_pattern }),
+    source: seed.source,
+    sample_count,
+    accuracy,
+    mean_confidence,
+    brier_score: stableMetric(brierTotal / sample_count),
+    mean_calibration_bias: stableMetric(mean_confidence - accuracy),
+    sample_status: sample_count < minimumSampleCount
+      ? "insufficient"
+      : "sufficient",
+  });
+}
+
+export function aggregateHierarchicalTaskCalibration(
+  observations: readonly unknown[],
+  options: TaskCalibrationOptions = {},
+): HierarchicalTaskCalibrationReport {
+  const parsed = HierarchicalTaskCalibrationObservationListSchema.parse(
+    observations,
+  );
+  const { minimum_sample_count } = TaskCalibrationOptionsSchema.parse(options);
+  const grouped = new Map<string, HierarchicalGroupSeed>();
+
+  function add(
+    observation: HierarchicalTaskCalibrationObservation,
+    scope: HierarchicalCalibrationScope,
+  ) {
+    const task_subtype = scope === "task_type"
+      ? undefined
+      : observation.task_subtype;
+    const prompt_pattern = scope === "prompt_pattern"
+      ? observation.prompt_pattern
+      : undefined;
+    const key = JSON.stringify([
+      scope,
+      observation.task_type,
+      task_subtype ?? null,
+      prompt_pattern ?? null,
+      observation.source.provider,
+      observation.source.model,
+    ]);
+    const existing = grouped.get(key);
+    if (existing) {
+      existing.observations.push(observation);
+      return;
+    }
+    grouped.set(key, {
+      scope,
+      task_type: observation.task_type,
+      ...(task_subtype === undefined ? {} : { task_subtype }),
+      ...(prompt_pattern === undefined ? {} : { prompt_pattern }),
+      source: observation.source,
+      observations: [observation],
+    });
+  }
+
+  for (const observation of parsed) {
+    add(observation, "task_type");
+    if (observation.task_subtype !== undefined) {
+      add(observation, "task_subtype");
+    }
+    if (observation.prompt_pattern !== undefined) {
+      add(observation, "prompt_pattern");
+    }
+  }
+
+  const scopeOrder: Record<HierarchicalCalibrationScope, number> = {
+    task_type: 0,
+    task_subtype: 1,
+    prompt_pattern: 2,
+  };
+  const groups = [...grouped.values()].map((seed) =>
+    summarizeHierarchicalGroup(seed, minimum_sample_count)
+  ).sort((left, right) =>
+    compareCodeUnits(left.task_type, right.task_type) ||
+    compareCodeUnits(left.source.provider, right.source.provider) ||
+    compareCodeUnits(left.source.model, right.source.model) ||
+    scopeOrder[left.scope] - scopeOrder[right.scope] ||
+    compareCodeUnits(left.task_subtype ?? "", right.task_subtype ?? "") ||
+    compareCodeUnits(left.prompt_pattern ?? "", right.prompt_pattern ?? "")
+  );
+
+  return HierarchicalTaskCalibrationReportSchema.parse({
+    schema_version: "quorum-router.hierarchical-calibration.v1",
+    advisory_only: true,
+    minimum_sample_count,
+    groups,
+  });
+}
+
+export function resolveHierarchicalTaskCalibration(
+  report: HierarchicalTaskCalibrationReport,
+  query: HierarchicalTaskCalibrationQuery,
+): HierarchicalTaskCalibrationSelection {
+  const parsedReport = HierarchicalTaskCalibrationReportSchema.parse(report);
+  const parsedQuery = HierarchicalTaskCalibrationQuerySchema.parse(query);
+  const requested_scope: HierarchicalCalibrationScope =
+    parsedQuery.prompt_pattern !== undefined
+      ? "prompt_pattern"
+      : parsedQuery.task_subtype !== undefined
+      ? "task_subtype"
+      : "task_type";
+  const scopes: HierarchicalCalibrationScope[] = requested_scope ===
+      "prompt_pattern"
+    ? ["prompt_pattern", "task_subtype", "task_type"]
+    : requested_scope === "task_subtype"
+    ? ["task_subtype", "task_type"]
+    : ["task_type"];
+
+  const matchingGroups = scopes.map((scope) =>
+    parsedReport.groups.find((group) =>
+      group.scope === scope && group.task_type === parsedQuery.task_type &&
+      group.source.provider === parsedQuery.source.provider &&
+      group.source.model === parsedQuery.source.model &&
+      (scope === "task_type" ||
+        group.task_subtype === parsedQuery.task_subtype) &&
+      (scope !== "prompt_pattern" ||
+        group.prompt_pattern === parsedQuery.prompt_pattern)
+    )
+  );
+  const candidates = scopes.map((scope, index) => {
+    const group = matchingGroups[index];
+    return group === undefined
+      ? { scope, sample_count: null, sample_status: "missing" as const }
+      : {
+        scope,
+        sample_count: group.sample_count,
+        sample_status: group.sample_status,
+      };
+  });
+  const selected_group = matchingGroups.find((group) =>
+    group?.sample_status === "sufficient"
+  );
+
+  return HierarchicalTaskCalibrationSelectionSchema.parse({
+    schema_version: "quorum-router.hierarchical-selection.v1",
+    advisory_only: true,
+    query: parsedQuery,
+    requested_scope,
+    resolution_status: selected_group === undefined
+      ? "no_sufficient_group"
+      : "selected",
+    selected_scope: selected_group?.scope ?? null,
+    candidates,
+    ...(selected_group === undefined ? {} : { selected_group }),
+  });
+}

package/templates/basic/src/hierarchical_calibration_demo.ts

@@ -0,0 +1,132 @@
+import {
+  aggregateHierarchicalTaskCalibration,
+  resolveHierarchicalTaskCalibration,
+} from "./calibration.ts";
+
+const evaluated_at = "2026-07-14T00:00:00Z";
+const source = { provider: "OpenAI", model: "example-model" } as const;
+const observations = [
+  {
+    observation_id: "hierarchy-demo-pattern-1",
+    task_type: "code-review",
+    task_subtype: "typescript",
+    prompt_pattern: "schema-boundary-review",
+    source,
+    evaluation_basis: "caller_attested_external_ground_truth",
+    correct: true,
+    confidence: 0.9,
+    evaluated_at,
+  },
+  {
+    observation_id: "hierarchy-demo-pattern-2",
+    task_type: "code-review",
+    task_subtype: "typescript",
+    prompt_pattern: "schema-boundary-review",
+    source,
+    evaluation_basis: "caller_attested_external_ground_truth",
+    correct: true,
+    confidence: 0.8,
+    evaluated_at,
+  },
+  {
+    observation_id: "hierarchy-demo-pattern-3",
+    task_type: "code-review",
+    task_subtype: "typescript",
+    prompt_pattern: "schema-boundary-review",
+    source,
+    evaluation_basis: "caller_attested_external_ground_truth",
+    correct: false,
+    confidence: 0.7,
+    evaluated_at,
+  },
+  {
+    observation_id: "hierarchy-demo-subtype-1",
+    task_type: "code-review",
+    task_subtype: "typescript",
+    prompt_pattern: "concurrency-review",
+    source,
+    evaluation_basis: "caller_attested_external_ground_truth",
+    correct: true,
+    confidence: 0.8,
+    evaluated_at,
+  },
+  {
+    observation_id: "hierarchy-demo-parent-1",
+    task_type: "code-review",
+    task_subtype: "rust",
+    prompt_pattern: "unsafe-boundary-review",
+    source,
+    evaluation_basis: "caller_attested_external_ground_truth",
+    correct: true,
+    confidence: 0.7,
+    evaluated_at,
+  },
+  {
+    observation_id: "hierarchy-demo-foreign-1",
+    task_type: "code-review",
+    task_subtype: "typescript",
+    prompt_pattern: "schema-boundary-review",
+    source: { provider: "Anthropic", model: "example-model" },
+    evaluation_basis: "caller_attested_external_ground_truth",
+    correct: true,
+    confidence: 0.75,
+    evaluated_at,
+  },
+];
+
+const report = aggregateHierarchicalTaskCalibration(observations, {
+  minimum_sample_count: 3,
+});
+const scenarios = [
+  {
+    name: "pattern-sufficient",
+    query: {
+      task_type: "code-review",
+      task_subtype: "typescript",
+      prompt_pattern: "schema-boundary-review",
+      source,
+    },
+  },
+  {
+    name: "fallback-to-subtype",
+    query: {
+      task_type: "code-review",
+      task_subtype: "typescript",
+      prompt_pattern: "concurrency-review",
+      source,
+    },
+  },
+  {
+    name: "fallback-to-task",
+    query: {
+      task_type: "code-review",
+      task_subtype: "rust",
+      prompt_pattern: "unsafe-boundary-review",
+      source,
+    },
+  },
+].map(({ name, query }) => {
+  const selection = resolveHierarchicalTaskCalibration(report, query);
+  return {
+    name,
+    query,
+    candidates: selection.candidates,
+    resolution_status: selection.resolution_status,
+    selected_scope: selection.selected_scope,
+    selected_sample_count: selection.selected_group?.sample_count ?? null,
+  };
+});
+
+console.log(JSON.stringify(
+  {
+    schema_version: "quorum-router.hierarchical-demo.v1",
+    advisory_only: true,
+    provider_request_sent: false,
+    fallback_order: ["prompt_pattern", "task_subtype", "task_type"],
+    minimum_sample_count: report.minimum_sample_count,
+    observation_count: observations.length,
+    scenarios,
+  },
+  null,
+  2,
+));