toolroll 0.9.81 → 0.9.82

package/evals/cases/dedupe/oracle.mjs

@@ -0,0 +1,16 @@
+// The hidden check's own answers (zero dependencies). Run only by the plane's oracle process, never inside a world or
+// the fence: `inputs` draws fresh inputs from a source seeded by each grading's nonce, and `expected` is the answer
+// the oracle computes itself for each one.
+export const files = ["dedupe.js"];
+export function inputs(random) {
+  const int = (low, high) => low + Math.floor(random() * (high - low + 1));
+  return Array.from({ length: 12 }, () => ({ rows: Array.from({ length: int(0, 10) }, (_, at) => {
+    const n = int(0, 4);
+    return { id: random() < 0.5 ? n : String(n), v: `r${at}` };
+  }) }));
+}
+export function expected({ rows }) {
+  const seen = new Set();
+  const value = rows.filter(row => { const key = `${typeof row.id}:${row.id}`; if (seen.has(key)) return false; seen.add(key); return true; });
+  return { value, kept: true, fresh: true };
+}

package/evals/cases/dedupe/probe.mjs

@@ -0,0 +1,13 @@
+// The graded program's launcher. The oracle copies this file from the installed case (never the build's copy) and
+// runs it inside the fence with the nonce and inputs on stdin; only a line naming that nonce with the right outputs
+// passes. Nothing here is visible to the agent while it builds.
+import { readFileSync } from "node:fs";
+const { nonce, inputs } = JSON.parse(readFileSync(0, "utf8"));
+const attempt = run => { try { return run(); } catch { return { threw: true }; } };
+const { dedupe } = await import("./dedupe.js");
+const outputs = inputs.map(({ rows }) => attempt(() => {
+  const copy = structuredClone(rows);
+  const value = dedupe(copy);
+  return { value, kept: JSON.stringify(copy) === JSON.stringify(rows), fresh: value !== copy };
+}));
+process.stdout.write(JSON.stringify({ nonce, outputs }) + "\n");

package/evals/cases/dedupe/source.md

@@ -0,0 +1,9 @@
+# dedupe
+
+From the real-model pilot's fixed assignments (scripts/fixtures/pilot-scenarios.mjs, `dedupe`, "backend bug"): `Object.fromEntries` kept the last row, reordered numeric ids and merged `2` with `"2"`.
+
+Expected: pass.
+
+## Status
+
+Draft. The expected outcome above awaits a person's validation (`validatedBy` in case.json); until then this case counts in no canonical campaign.

package/evals/cases/endpoint/case.json

@@ -0,0 +1,25 @@
+{
+  "id": "endpoint",
+  "kind": "build",
+  "title": "Add the create-item endpoint",
+  "goal": "Add validate.js exporting validName(value): true only for a string whose trimmed text is 1 to 20 characters. Add endpoint.js exporting createItem(body): return { status: 400, body: { error: \"invalid name\" } } when the body or its name is missing or invalid, otherwise { status: 201, body: { name: <the trimmed name> } }. A missing or null body never throws.",
+  "touches": [
+    "validate.js",
+    "endpoint.js",
+    "test/endpoint.test.mjs"
+  ],
+  "pin": {
+    "provider": "claude",
+    "model": "claude-sonnet-5-5"
+  },
+  "maxDispatches": 2,
+  "timeoutMinutes": 15,
+  "grade": {
+    "files": [
+      "validate.js",
+      "endpoint.js"
+    ]
+  },
+  "expected": "pass",
+  "validatedBy": null
+}

package/evals/cases/endpoint/fixture/README.md

@@ -0,0 +1,3 @@
+# Eval fixture
+
+Run `npm test`. Change only what the task asks for.

package/evals/cases/endpoint/fixture/package.json

@@ -0,0 +1,8 @@
+{
+  "name": "eval-fixture",
+  "private": true,
+  "type": "module",
+  "scripts": {
+    "test": "node --test"
+  }
+}

package/evals/cases/endpoint/fixture/test/endpoint.test.mjs

@@ -0,0 +1,4 @@
+import test from 'node:test';
+import assert from 'node:assert/strict';
+
+test('the fixture runs', () => assert.ok(true));

package/evals/cases/endpoint/oracle.mjs

@@ -0,0 +1,27 @@
+// The hidden check's own answers (zero dependencies). Run only by the plane's oracle process, never inside a world or
+// the fence: `inputs` draws fresh inputs from a source seeded by each grading's nonce, and `expected` is the answer
+// the oracle computes itself for each one.
+export const files = ["validate.js", "endpoint.js"];
+export function inputs(random) {
+  const int = (low, high) => low + Math.floor(random() * (high - low + 1));
+  const name = () => {
+    // Never over 20 characters untrimmed but within 20 trimmed: the goal leaves that open.
+    const core = "abcdefghij".repeat(3).slice(0, random() < 0.2 ? int(21, 28) : int(0, 18));
+    const pad = " ".repeat(core.length <= 16 ? int(0, 2) : 0);
+    return pad + core + pad;
+  };
+  const values = () => [name(), int(0, 9), null, true];
+  const out = [];
+  for (let i = 0; i < 8; i++) out.push({ kind: "valid", value: values()[random() < 0.7 ? 0 : int(1, 3)] });
+  for (let i = 0; i < 8; i++) {
+    const shape = int(0, 4);
+    out.push({ kind: "create", body: shape === 0 ? { absent: true } : shape === 1 ? null : shape === 2 ? {} : { name: values()[shape === 3 ? 0 : int(1, 3)] } });
+  }
+  return out;
+}
+const valid = value => typeof value === "string" && value.trim().length >= 1 && value.trim().length <= 20;
+export function expected(input) {
+  if (input.kind === "valid") return { value: valid(input.value) };
+  const body = input.body?.absent ? undefined : input.body;
+  return { value: valid(body?.name) ? { status: 201, body: { name: body.name.trim() } } : { status: 400, body: { error: "invalid name" } } };
+}

package/evals/cases/endpoint/probe.mjs

@@ -0,0 +1,10 @@
+// The graded program's launcher. The oracle copies this file from the installed case (never the build's copy) and
+// runs it inside the fence with the nonce and inputs on stdin; only a line naming that nonce with the right outputs
+// passes. Nothing here is visible to the agent while it builds.
+import { readFileSync } from "node:fs";
+const { nonce, inputs } = JSON.parse(readFileSync(0, "utf8"));
+const attempt = run => { try { return run(); } catch { return { threw: true }; } };
+const { validName } = await import("./validate.js");
+const { createItem } = await import("./endpoint.js");
+const outputs = inputs.map(input => attempt(() => ({ value: input.kind === "valid" ? validName(input.value) : createItem(input.body?.absent ? undefined : input.body) })));
+process.stdout.write(JSON.stringify({ nonce, outputs }) + "\n");

package/evals/cases/endpoint/source.md

@@ -0,0 +1,11 @@
+# endpoint
+
+From the real-model pilot's fixed assignments (scripts/fixtures/pilot-scenarios.mjs, `endpoint`, "multiple files"): validation and a status-code contract across two files, with a missing body that must not throw.
+
+The oracle never draws a name that is over 20 characters untrimmed but within 20 trimmed (the pilot's goal left that case open).
+
+Expected: pass.
+
+## Status
+
+Draft. The expected outcome above awaits a person's validation (`validatedBy` in case.json); until then this case counts in no canonical campaign.

package/evals/cases/math-divide/case.json

@@ -0,0 +1,23 @@
+{
+  "id": "math-divide",
+  "kind": "build",
+  "title": "Add a divide function",
+  "goal": "Add divide(a, b) to src/math.js, keeping add: it returns a divided by b and throws an Error on division by zero. Add tests for both in test/math.test.js.",
+  "touches": [
+    "src/math.js",
+    "test/math.test.js"
+  ],
+  "pin": {
+    "provider": "claude",
+    "model": "claude-sonnet-5-5"
+  },
+  "maxDispatches": 2,
+  "timeoutMinutes": 15,
+  "grade": {
+    "files": [
+      "src/math.js"
+    ]
+  },
+  "expected": "pass",
+  "validatedBy": null
+}

package/evals/cases/math-divide/fixture/README.md

@@ -0,0 +1,3 @@
+# Eval fixture
+
+Run `npm test`. Change only what the task asks for.

package/evals/cases/math-divide/fixture/package.json

@@ -0,0 +1,8 @@
+{
+  "name": "eval-fixture",
+  "private": true,
+  "type": "module",
+  "scripts": {
+    "test": "node --test"
+  }
+}

package/evals/cases/math-divide/fixture/src/math.js

@@ -0,0 +1 @@
+export const add = (a, b) => a + b;

package/evals/cases/math-divide/fixture/test/math.test.js

@@ -0,0 +1,5 @@
+import test from 'node:test';
+import assert from 'node:assert/strict';
+import { add } from '../src/math.js';
+
+test('adds', () => assert.equal(add(2, 3), 5));

package/evals/cases/math-divide/oracle.mjs

@@ -0,0 +1,13 @@
+// The hidden check's own answers (zero dependencies). Run only by the plane's oracle process, never inside a world or
+// the fence: `inputs` draws fresh inputs from a source seeded by each grading's nonce, and `expected` is the answer
+// the oracle computes itself for each one.
+export const files = ["src/math.js"];
+const OPS = ["add", "divide"];
+export function inputs(random) {
+  const int = (low, high) => low + Math.floor(random() * (high - low + 1));
+  return Array.from({ length: 14 }, (_, at) => ({ op: OPS[at % OPS.length], a: int(-40, 40), b: at % 5 === 4 ? 0 : int(-12, 12) }));
+}
+export function expected({ op, a, b }) {
+  if (op === "add") return { value: a + b };
+  return b === 0 ? { threw: true } : { value: a / b };
+}

package/evals/cases/math-divide/probe.mjs

@@ -0,0 +1,9 @@
+// The graded program's launcher. The oracle copies this file from the installed case (never the build's copy) and
+// runs it inside the fence with the nonce and inputs on stdin; only a line naming that nonce with the right outputs
+// passes. Nothing here is visible to the agent while it builds.
+import { readFileSync } from "node:fs";
+const { nonce, inputs } = JSON.parse(readFileSync(0, "utf8"));
+const attempt = run => { try { return run(); } catch { return { threw: true }; } };
+const math = await import("./src/math.js");
+const outputs = inputs.map(({ op, a, b }) => attempt(() => ({ value: math[op](a, b) })));
+process.stdout.write(JSON.stringify({ nonce, outputs }) + "\n");

package/evals/cases/math-divide/source.md

@@ -0,0 +1,9 @@
+# math-divide
+
+From the stop-and-resume journey (scripts/app-e2e.mjs, "Stop a build while it runs, then resume it…", :708): divide that refuses zero, with tests for both.
+
+Expected: pass.
+
+## Status
+
+Draft. The expected outcome above awaits a person's validation (`validatedBy` in case.json); until then this case counts in no canonical campaign.

package/evals/cases/math-multiply/case.json

@@ -0,0 +1,23 @@
+{
+  "id": "math-multiply",
+  "kind": "build",
+  "title": "Add multiply, keep subtract",
+  "goal": "Add multiply(a, b) to src/math.js with its own test in test/math.test.js. Keep add and subtract exactly as they work now.",
+  "touches": [
+    "src/math.js",
+    "test/math.test.js"
+  ],
+  "pin": {
+    "provider": "claude",
+    "model": "claude-sonnet-5-5"
+  },
+  "maxDispatches": 2,
+  "timeoutMinutes": 15,
+  "grade": {
+    "files": [
+      "src/math.js"
+    ]
+  },
+  "expected": "pass",
+  "validatedBy": null
+}

package/evals/cases/math-multiply/fixture/README.md

@@ -0,0 +1,3 @@
+# Eval fixture
+
+Run `npm test`. Change only what the task asks for.

package/evals/cases/math-multiply/fixture/package.json

@@ -0,0 +1,8 @@
+{
+  "name": "eval-fixture",
+  "private": true,
+  "type": "module",
+  "scripts": {
+    "test": "node --test"
+  }
+}

package/evals/cases/math-multiply/fixture/src/math.js

@@ -0,0 +1,2 @@
+export const add = (a, b) => a + b;
+export const subtract = (a, b) => a - b;

package/evals/cases/math-multiply/fixture/test/math.test.js

@@ -0,0 +1,5 @@
+import test from 'node:test';
+import assert from 'node:assert/strict';
+import { add } from '../src/math.js';
+
+test('adds', () => assert.equal(add(2, 3), 5));

package/evals/cases/math-multiply/oracle.mjs

@@ -0,0 +1,12 @@
+// The hidden check's own answers (zero dependencies). Run only by the plane's oracle process, never inside a world or
+// the fence: `inputs` draws fresh inputs from a source seeded by each grading's nonce, and `expected` is the answer
+// the oracle computes itself for each one.
+export const files = ["src/math.js"];
+const OPS = ["add", "subtract", "multiply"];
+export function inputs(random) {
+  const int = (low, high) => low + Math.floor(random() * (high - low + 1));
+  return Array.from({ length: 14 }, (_, at) => ({ op: OPS[at % OPS.length], a: int(-40, 40), b: at % 5 === 4 ? 0 : int(-12, 12) }));
+}
+export function expected({ op, a, b }) {
+  return { value: op === "add" ? a + b : op === "subtract" ? a - b : a * b };
+}

package/evals/cases/math-multiply/probe.mjs

@@ -0,0 +1,9 @@
+// The graded program's launcher. The oracle copies this file from the installed case (never the build's copy) and
+// runs it inside the fence with the nonce and inputs on stdin; only a line naming that nonce with the right outputs
+// passes. Nothing here is visible to the agent while it builds.
+import { readFileSync } from "node:fs";
+const { nonce, inputs } = JSON.parse(readFileSync(0, "utf8"));
+const attempt = run => { try { return run(); } catch { return { threw: true }; } };
+const math = await import("./src/math.js");
+const outputs = inputs.map(({ op, a, b }) => attempt(() => ({ value: math[op](a, b) })));
+process.stdout.write(JSON.stringify({ nonce, outputs }) + "\n");

package/evals/cases/math-multiply/source.md

@@ -0,0 +1,9 @@
+# math-multiply
+
+From the plan-amendment journey (scripts/app-e2e.mjs, "Send it back asking for more than the plan allows…", :640): the revision added multiply and had to keep subtract working (`the build doesn't keep subtract and add multiply` was the journey's failure).
+
+Expected: pass.
+
+## Status
+
+Draft. The expected outcome above awaits a person's validation (`validatedBy` in case.json); until then this case counts in no canonical campaign.

package/evals/cases/math-subtract/case.json

@@ -0,0 +1,23 @@
+{
+  "id": "math-subtract",
+  "kind": "build",
+  "title": "Add a subtract function",
+  "goal": "Add subtract(a, b) to src/math.js, keeping add, with a test in test/math.test.js that also checks a negative result: subtract(2, 5) is -3.",
+  "touches": [
+    "src/math.js",
+    "test/math.test.js"
+  ],
+  "pin": {
+    "provider": "claude",
+    "model": "claude-sonnet-5-5"
+  },
+  "maxDispatches": 2,
+  "timeoutMinutes": 15,
+  "grade": {
+    "files": [
+      "src/math.js"
+    ]
+  },
+  "expected": "pass",
+  "validatedBy": null
+}

package/evals/cases/math-subtract/fixture/README.md

@@ -0,0 +1,3 @@
+# Eval fixture
+
+Run `npm test`. Change only what the task asks for.

package/evals/cases/math-subtract/fixture/package.json

@@ -0,0 +1,8 @@
+{
+  "name": "eval-fixture",
+  "private": true,
+  "type": "module",
+  "scripts": {
+    "test": "node --test"
+  }
+}

package/evals/cases/math-subtract/fixture/src/math.js

@@ -0,0 +1 @@
+export const add = (a, b) => a + b;

package/evals/cases/math-subtract/fixture/test/math.test.js

@@ -0,0 +1,5 @@
+import test from 'node:test';
+import assert from 'node:assert/strict';
+import { add } from '../src/math.js';
+
+test('adds', () => assert.equal(add(2, 3), 5));

package/evals/cases/math-subtract/oracle.mjs

@@ -0,0 +1,12 @@
+// The hidden check's own answers (zero dependencies). Run only by the plane's oracle process, never inside a world or
+// the fence: `inputs` draws fresh inputs from a source seeded by each grading's nonce, and `expected` is the answer
+// the oracle computes itself for each one.
+export const files = ["src/math.js"];
+const OPS = ["add", "subtract"];
+export function inputs(random) {
+  const int = (low, high) => low + Math.floor(random() * (high - low + 1));
+  return Array.from({ length: 14 }, (_, at) => ({ op: OPS[at % OPS.length], a: int(-40, 40), b: at % 5 === 4 ? 0 : int(-12, 12) }));
+}
+export function expected({ op, a, b }) {
+  return { value: op === "add" ? a + b : a - b };
+}

package/evals/cases/math-subtract/probe.mjs

@@ -0,0 +1,9 @@
+// The graded program's launcher. The oracle copies this file from the installed case (never the build's copy) and
+// runs it inside the fence with the nonce and inputs on stdin; only a line naming that nonce with the right outputs
+// passes. Nothing here is visible to the agent while it builds.
+import { readFileSync } from "node:fs";
+const { nonce, inputs } = JSON.parse(readFileSync(0, "utf8"));
+const attempt = run => { try { return run(); } catch { return { threw: true }; } };
+const math = await import("./src/math.js");
+const outputs = inputs.map(({ op, a, b }) => attempt(() => ({ value: math[op](a, b) })));
+process.stdout.write(JSON.stringify({ nonce, outputs }) + "\n");

package/evals/cases/math-subtract/source.md

@@ -0,0 +1,9 @@
+# math-subtract
+
+From the console journey that files the first task (scripts/app-e2e.mjs, "File a task in the console…" and "Send it back with a note…", :550 and :632): the first build's test missed a negative result and was sent back for `subtract(2, 5)` is -3.
+
+Expected: pass.
+
+## Status
+
+Draft. The expected outcome above awaits a person's validation (`validatedBy` in case.json); until then this case counts in no canonical campaign.

package/evals/cases/merge-config/case.json

@@ -0,0 +1,23 @@
+{
+  "id": "merge-config",
+  "kind": "build",
+  "title": "Fix mergeConfig",
+  "goal": "Revise mergeConfig(base, overrides) in config.js: start from a shallow copy of base, apply every defined override including 0 and false, delete each key whose override is null, ignore overrides that are undefined, and never mutate either argument. Keep the mergeConfig export.",
+  "touches": [
+    "config.js",
+    "test/config.test.mjs"
+  ],
+  "pin": {
+    "provider": "claude",
+    "model": "claude-sonnet-5-5"
+  },
+  "maxDispatches": 2,
+  "timeoutMinutes": 15,
+  "grade": {
+    "files": [
+      "config.js"
+    ]
+  },
+  "expected": "pass",
+  "validatedBy": null
+}

package/evals/cases/merge-config/fixture/README.md

@@ -0,0 +1,3 @@
+# Eval fixture
+
+Run `npm test`. Change only what the task asks for.

package/evals/cases/merge-config/fixture/config.js

@@ -0,0 +1 @@
+export const mergeConfig = (base, overrides) => Object.assign(base, overrides);

package/evals/cases/merge-config/fixture/package.json

@@ -0,0 +1,8 @@
+{
+  "name": "eval-fixture",
+  "private": true,
+  "type": "module",
+  "scripts": {
+    "test": "node --test"
+  }
+}

package/evals/cases/merge-config/fixture/test/config.test.mjs

@@ -0,0 +1,5 @@
+import test from 'node:test';
+import assert from 'node:assert/strict';
+import { mergeConfig } from '../config.js';
+
+test('two empty objects', () => assert.deepEqual(mergeConfig({}, {}), {}));

package/evals/cases/merge-config/oracle.mjs

@@ -0,0 +1,28 @@
+// The hidden check's own answers (zero dependencies). Run only by the plane's oracle process, never inside a world or
+// the fence: `inputs` draws fresh inputs from a source seeded by each grading's nonce, and `expected` is the answer
+// the oracle computes itself for each one.
+export const files = ["config.js"];
+const KEYS = ["port", "debug", "host", "keep", "mode"];
+export function inputs(random) {
+  const int = (low, high) => low + Math.floor(random() * (high - low + 1));
+  const value = () => [0, 1, 80, false, true, "x", ""][int(0, 6)];
+  return Array.from({ length: 10 }, () => {
+    const base = {}, overrides = {};
+    for (const key of KEYS) {
+      if (random() < 0.6) base[key] = value();
+      const roll = random();
+      if (roll < 0.25) overrides[key] = null;
+      else if (roll < 0.45) overrides[key] = { $undefined: true };
+      else if (roll < 0.75) overrides[key] = value();
+    }
+    return { base, overrides };
+  });
+}
+export function expected({ base, overrides }) {
+  const out = { ...base };
+  for (const [key, value] of Object.entries(overrides)) {
+    if (value !== null && typeof value === "object") continue;
+    if (value === null) delete out[key]; else out[key] = value;
+  }
+  return { value: out, keptBase: true, keptOverrides: true };
+}

package/evals/cases/merge-config/probe.mjs

@@ -0,0 +1,15 @@
+// The graded program's launcher. The oracle copies this file from the installed case (never the build's copy) and
+// runs it inside the fence with the nonce and inputs on stdin; only a line naming that nonce with the right outputs
+// passes. Nothing here is visible to the agent while it builds.
+import { readFileSync } from "node:fs";
+const { nonce, inputs } = JSON.parse(readFileSync(0, "utf8"));
+const attempt = run => { try { return run(); } catch { return { threw: true }; } };
+const { mergeConfig } = await import("./config.js");
+const live = overrides => Object.fromEntries(Object.entries(overrides).map(([key, value]) => [key, value !== null && typeof value === "object" ? undefined : value]));
+const outputs = inputs.map(({ base, overrides }) => attempt(() => {
+  const b = structuredClone(base), o = live(overrides);
+  const before = JSON.stringify(Object.entries(o));
+  const value = mergeConfig(b, o);
+  return { value, keptBase: JSON.stringify(b) === JSON.stringify(base), keptOverrides: JSON.stringify(Object.entries(o)) === before };
+}));
+process.stdout.write(JSON.stringify({ nonce, outputs }) + "\n");

package/evals/cases/merge-config/source.md

@@ -0,0 +1,9 @@
+# merge-config
+
+From the real-model pilot's fixed assignments (scripts/fixtures/pilot-scenarios.mjs, `revision-config`, "revision of existing behavior"): `Object.assign` mutated the base and kept null and undefined overrides.
+
+Expected: pass.
+
+## Status
+
+Draft. The expected outcome above awaits a person's validation (`validatedBy` in case.json); until then this case counts in no canonical campaign.

package/evals/cases/paginate/case.json

@@ -0,0 +1,23 @@
+{
+  "id": "paginate",
+  "kind": "build",
+  "title": "Fix paginate",
+  "goal": "Fix paginate(items, offset, limit) in paginate.js: offset defaults to 0 and limit defaults to 10, a zero limit returns [], a negative offset is clamped to zero, and it returns a slice without mutating items.",
+  "touches": [
+    "paginate.js",
+    "test/paginate.test.mjs"
+  ],
+  "pin": {
+    "provider": "claude",
+    "model": "claude-sonnet-5-5"
+  },
+  "maxDispatches": 2,
+  "timeoutMinutes": 15,
+  "grade": {
+    "files": [
+      "paginate.js"
+    ]
+  },
+  "expected": "pass",
+  "validatedBy": null
+}

package/evals/cases/paginate/fixture/README.md

@@ -0,0 +1,3 @@
+# Eval fixture
+
+Run `npm test`. Change only what the task asks for.

package/evals/cases/paginate/fixture/package.json

@@ -0,0 +1,8 @@
+{
+  "name": "eval-fixture",
+  "private": true,
+  "type": "module",
+  "scripts": {
+    "test": "node --test"
+  }
+}

package/evals/cases/paginate/fixture/paginate.js

@@ -0,0 +1 @@
+export const paginate = (items, offset, limit) => items.splice(offset || 0, limit || 10);

package/evals/cases/paginate/fixture/test/paginate.test.mjs

@@ -0,0 +1,5 @@
+import test from 'node:test';
+import assert from 'node:assert/strict';
+import { paginate } from '../paginate.js';
+
+test('a page from the middle', () => assert.deepEqual(paginate([1, 2, 3, 4], 1, 2), [2, 3]));

package/evals/cases/paginate/oracle.mjs

@@ -0,0 +1,18 @@
+// The hidden check's own answers (zero dependencies). Run only by the plane's oracle process, never inside a world or
+// the fence: `inputs` draws fresh inputs from a source seeded by each grading's nonce, and `expected` is the answer
+// the oracle computes itself for each one.
+export const files = ["paginate.js"];
+export function inputs(random) {
+  const int = (low, high) => low + Math.floor(random() * (high - low + 1));
+  return Array.from({ length: 12 }, () => {
+    const items = Array.from({ length: int(0, 14) }, () => int(-50, 50));
+    const args = [items];
+    if (random() < 0.85) args.push(int(-4, 16));
+    if (args.length === 2 && random() < 0.85) args.push(int(0, 7));
+    return { args };
+  });
+}
+export function expected({ args: [items, offset, limit] }) {
+  const from = Math.max(0, offset ?? 0);
+  return { value: items.slice(from, from + (limit ?? 10)), kept: true };
+}

package/evals/cases/paginate/probe.mjs

@@ -0,0 +1,13 @@
+// The graded program's launcher. The oracle copies this file from the installed case (never the build's copy) and
+// runs it inside the fence with the nonce and inputs on stdin; only a line naming that nonce with the right outputs
+// passes. Nothing here is visible to the agent while it builds.
+import { readFileSync } from "node:fs";
+const { nonce, inputs } = JSON.parse(readFileSync(0, "utf8"));
+const attempt = run => { try { return run(); } catch { return { threw: true }; } };
+const { paginate } = await import("./paginate.js");
+const outputs = inputs.map(({ args }) => attempt(() => {
+  const items = structuredClone(args[0]);
+  const value = paginate(items, ...args.slice(1));
+  return { value, kept: JSON.stringify(items) === JSON.stringify(args[0]) };
+}));
+process.stdout.write(JSON.stringify({ nonce, outputs }) + "\n");

package/evals/cases/paginate/source.md

@@ -0,0 +1,9 @@
+# paginate
+
+From the unattended real-model pilot's fixed assignments (scripts/fixtures/pilot-scenarios.mjs, `pagination`, category "backend bug"): the original used `splice`, which mutated the caller's array and ignored the edge cases. A build that fixes only the visible test fails the hidden one.
+
+Expected: pass (a correct fix is small and fully specified).
+
+## Status
+
+Draft. The expected outcome above awaits a person's validation (`validatedBy` in case.json); until then this case counts in no canonical campaign.

package/evals/cases/round2-asks/case.json

@@ -0,0 +1,24 @@
+{
+  "id": "round2-asks",
+  "kind": "ask",
+  "title": "Add a round2 function",
+  "goal": "Add round2(x) to src/math.js (round to 2 decimal places) with a test in test/math.test.js. Whether an exact half rounds up or to the nearest even digit is the operator's call and is not decided yet: before writing any code, stop and ask the operator with two options (half up, half to even), then build what they choose.",
+  "touches": [
+    "src/math.js",
+    "test/math.test.js"
+  ],
+  "pin": {
+    "provider": "claude",
+    "model": "claude-sonnet-5-5"
+  },
+  "maxDispatches": 1,
+  "timeoutMinutes": 15,
+  "grade": {
+    "options": [
+      "half.{0,3}up|\\bup\\b",
+      "even"
+    ]
+  },
+  "expected": "asks",
+  "validatedBy": null
+}