Files
threshing-atlas/tests/core/advisor.test.js
T
ekcSchianoandClaude Fable 5.2 d01ec46fcc The Black is not dice: developer statements, evidence report, and a plan that looks for the rule
- planner/knowledge.yaml: what the game's developer has said (paraphrased, dated, graded), what the Atlas does
  with each statement, and the open questions with the experiment that tests each.
- analysis/black/evidence.py -> analysis/reports/BLACK.md: counts from the runs. On the same recorded route the
  shadow shows on 21 of 79 first attempts and 1 of 80 later ones; no single route; Blue then Black after the shadow.
- game.yaml: the shadow is a guard (not a roll); what follows it is the Blue, then the Black; on-screen text for
  the squad's direction screen.
- planner/experiments.yaml v2: one exact path on every fresh account (fresh-replicate), then one choice changed
  at a time (fresh-ablate); shadow-again, second-attempt-same-path, veteran-repeat, veteran-variety; what the bot
  records about the account; the sequence to follow once the shadow shows.
- solver.js: assign() takes account facts, account offsets and cycled arms; recommendOption() answers the
  squad screen (stay + direction) in one pick.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014jqCC2LZ4x1WHQ1V57nTsv
2026-11-16 15:24:03 +00:00

227 lines
14 KiB
JavaScript

// The route advisor (core/solver.js): it must give exactly the Python route model's answers, and walk a run in the
// game's order. node --test tests/core/advisor.test.js
const test = require("node:test");
const assert = require("node:assert");
const path = require("node:path");
const fs = require("node:fs");
const os = require("node:os");
const { execFileSync } = require("node:child_process");
const S = require("../../core/solver");
const ROOT = path.join(__dirname, "../..");
const published = JSON.parse(fs.readFileSync(path.join(ROOT, "data/public/policy.json"), "utf8"));
// a synthetic world with known answers, built by the Python model (tests/fixtures/synthetic_policy.py)
const tmp = path.join(fs.mkdtempSync(path.join(os.tmpdir(), "atlas-")), "synthetic.json");
execFileSync("python3", [path.join(ROOT, "tests/fixtures/synthetic_policy.py"), tmp], { stdio: "inherit" });
const policy = JSON.parse(fs.readFileSync(tmp, "utf8"));
const choice = (X, h, obj = "thisAttempt") => S.advisor({ model: policy.model }, X, obj).next(h).choice;
test("the advisor gives exactly the Python model's values and choices, for every colour and goal", () => {
let checked = 0;
for (const smp of Object.values(published.samples)) {
for (const [X, t] of Object.entries(smp.targets)) {
for (const obj of ["thisAttempt", "eventually"]) {
const adv = S.advisor(smp, X, obj);
assert.ok(Math.abs(adv.value.x - t[obj].pX.point) < 1e-12, `${X} ${obj}: ${adv.value.x} vs ${t[obj].pX.point}`);
// follow the published tree: every decision on the route must match, option by option
let n = t[obj].tree, h = {};
while (n && n.type !== "end") {
if (n.type === "decision") {
const s = adv.step(h);
assert.strictEqual(s.var, n.var); assert.strictEqual(s.choice, n.choice, `${X} ${obj} ${n.var}`);
for (const [o, v] of Object.entries(n.options)) assert.ok(Math.abs(s.options[o].pX - v.pX) < 1e-12, `${X} ${obj} ${n.var}=${o}`);
checked++;
if (n.var === "encounter") break;
h[n.var] = n.choice; n = n.next;
} else {
const [k, b] = Object.entries(n.branches).filter(([k]) => k !== "die").sort((a, c) => c[1].p - a[1].p)[0];
if (!n.var.startsWith("die")) h[n.var] = k;
n = b.next;
}
}
}
}
}
assert.ok(checked > 100, checked);
});
test("synthetic world: Blue goes to the valley, passes the Red, attempts only the Blue", () => {
const h = { company: "Alone", direction: "Center", fs: "Red", react: "Hold still", fate: "Killed the bullies yourself", post: "Forest" };
assert.strictEqual(choice("Blue", h), "Valley");
assert.strictEqual(choice("Blue", { ...h, hour: "Valley" }), "Pass the Red (meadow)");
const meadow = { ...h, hour: "Valley", valley: "Pass the Red (meadow)" };
assert.strictEqual(choice("Blue", { ...meadow, meadow: "Blue" }), "Attempt the bond");
assert.strictEqual(choice("Blue", { ...meadow, meadow: "Orange" }), "Die on purpose");
assert.strictEqual(choice("Red", { ...h, hour: "Valley" }), "Face the Red");
assert.strictEqual(choice("Black", h), "Tree");
});
test("routing to one encounter: Blue only at the meadow equals the Blue policy; only at the tree Greens gives 0", () => {
const smp = { model: policy.model };
const meadow = S.advisor(smp, "Blue", "thisAttempt", { only: ["meadow-Blue"] });
const all = S.advisor(smp, "Blue", "thisAttempt");
assert.ok(Math.abs(meadow.value.x - all.value.x) < 1e-9);
assert.ok(S.advisor(smp, "Blue", "thisAttempt", { only: ["tree-Greens"] }).value.x < 1e-9);
assert.ok(S.advisor(smp, "Brown", "thisAttempt", { only: ["tree-Greens"] }).value.x > 0.05); // the Greens rescue
assert.ok(meadow.attempts("meadow-Blue") > 100);
});
test("a run is walked in the game's order", () => {
const adv = S.advisor(published.samples.all, "Blue", "thisAttempt");
assert.strictEqual(adv.step({}).var, "company");
assert.strictEqual(adv.step({ company: "1 Squadmate" }).var, "leave");
assert.strictEqual(adv.step({ company: "1 Squadmate", leave: "Stay together" }).var, "direction");
assert.strictEqual(adv.step({ company: "Alone", direction: "Center" }).kind, "chance"); // the first dragon is a roll
const seen = { company: "Alone", direction: "Center", fs: "Red" };
assert.strictEqual(adv.step(seen).var, "react");
const base = { ...seen, react: "Hold still" };
const scream = adv.step(base);
assert.strictEqual(scream.var, "scream");
assert.strictEqual(scream.risk.label, "Killed at the first sighting");
const help = { ...base, scream: "Investigate the scream", clearing: "Help your squadmate" };
assert.strictEqual(adv.step({ ...base, scream: "Investigate the scream" }).var, "clearing");
assert.strictEqual(adv.step(help).var, "fight");
assert.strictEqual(adv.step({ ...help, fight: "Pause to assess" }).var, "dagger");
assert.strictEqual(adv.step({ ...help, fight: "Charge the smaller cadet" }).var, "fate"); // charging skips the dagger
const fate = adv.step({ ...help, fight: "Pause to assess", dagger: "Engage the smaller cadet" });
assert.strictEqual(fate.var, "fate");
assert.strictEqual(fate.risk.label, "Killed in the fight");
assert.strictEqual(adv.step({ ...base, scream: "Keep walking" }).var, "wounded");
assert.strictEqual(adv.step({ ...base, scream: "Investigate the scream", clearing: "Walk away" }).var, "hour");
// an older record that only knows the bullies' fate still lands in the right place (the skipped steps are
// assumed to be the recommended ones)
assert.strictEqual(adv.next({ ...base, fate: "Killed the bullies yourself" }).decision, "post");
});
test("recommendOption points the bot at the recommended option on its screen", () => {
const adv = S.advisor(published.samples.all, "Blue", "thisAttempt");
const h = { company: "Alone", direction: "Center", fs: "Red", react: "Hold still", scream: "Keep walking", wounded: "Tourniquet" };
const rec = S.recommendOption(adv, h, ["Continue toward the center of the valley", "Find the tallest tree and climb for a better view"]);
assert.strictEqual(rec.decision, "hour");
assert.strictEqual(rec.index, rec.choice === "Valley" ? 0 : 1);
assert.strictEqual(S.recommendOption(adv, {}, ["Something we have never seen"]).index, null);
});
test("solverHistory turns form variables into a history keyed by the game's steps", () => {
S.useGame(published.samples.all);
const h = S.solverHistory({ company: "Alone", direction: "Center", first_sighting: "Red", scream: "Ignore the scream", resolution: "Tourniquet", hour: "Tree", branch: "Descend slowly: two Greens" });
assert.deepStrictEqual(h, { company: "Alone", direction: "Center", fs: "Red", scream: "Keep walking", wounded: "Tourniquet", hour: "Tree", descend: "Slowly (two Greens)", shadow: "no" });
const l = S.solverHistory({ company: "1 Squadmate", left_squadmates: true, branch: "Descend quickly: let the Red pass", hour: "Tree" });
assert.strictEqual(l.leave, "Leave them");
assert.strictEqual(l.red, "Let it pass (two Greens)");
});
test("Thompson sampling: one posterior draw per seed, with the chance of each pick", () => {
const S = require("../../core/solver.js");
const P = require("../../data/public/policy.json");
const best = S.advisor(P.samples.all, "Blue", "eventually").next({});
assert.strictEqual(best.propensity, 1);
const a = S.advisor(P.samples.all, "Blue", "eventually", { sampling: "thompson", seed: 3, draws: 8 });
const b = S.advisor(P.samples.all, "Blue", "eventually", { sampling: "thompson", seed: 3, draws: 8 });
const ra = a.next({}), rb = b.next({});
assert.strictEqual(ra.choice, rb.choice); // same seed, same draw
assert.ok(ra.propensity >= 0 && ra.propensity <= 1);
const picks = new Set();
for (let s = 1; s <= 12; s++) picks.add(S.advisor(P.samples.all, "Blue", "eventually", { sampling: "thompson", seed: s, draws: 1 }).next({}).choice);
assert.ok(picks.size > 1, "different seeds explore different options");
});
test("a first attempt has its own odds, exactly as the Python model publishes them", () => {
for (const smp of Object.values(published.samples)) {
for (const [X, t] of Object.entries(smp.targets)) {
for (const obj of ["thisAttempt", "eventually"]) {
const adv = S.advisor(smp, X, obj, { attempt: 1 });
assert.deepStrictEqual(adv.given, { attempt: "First attempt" });
assert.ok(Math.abs(adv.value.x - t.firstAttempt[obj].pX.point) < 1e-12, `${X} ${obj}`);
assert.strictEqual(adv.next({}).choice, t.firstAttempt[obj].tree.choice);
}
}
}
const all = published.samples.all;
const later = S.advisor(all, "Black", "thisAttempt", { attempt: 7 }), dflt = S.advisor(all, "Black", "thisAttempt");
assert.strictEqual(later.value.x, dflt.value.x); // no attempt given: a later attempt
assert.ok(S.advisor(all, "Black", "thisAttempt", { attempt: 1 }).value.x > 5 * later.value.x);
});
test("fix holds decisions to an experiment arm and solves the rest around it", () => {
const all = published.samples.all;
const fix = { direction: "East", scream: "Investigate the scream", clearing: "Help your squadmate", hour: "Tree" };
const adv = S.advisor(all, "Black", "eventually", { attempt: 1, fix });
assert.strictEqual(adv.step({ company: "Alone" }).choice, "East");
assert.deepStrictEqual(Object.keys(adv.step({ company: "Alone" }).options), ["East"]);
assert.ok(adv.value.x <= S.advisor(all, "Black", "eventually", { attempt: 1 }).value.x + 1e-12);
assert.throws(() => S.advisor(all, "Black", "eventually", { fix: { direction: "North" } }), /no option/);
});
test("mapping: uniform sampling picks among the options on the screen, with its propensity", () => {
const all = published.samples.all;
const screen = ["Enter the valley alone", "Enter the valley with one of your squadmates", "Enter the valley with two of your squadmates"];
const picks = new Set();
for (let seed = 1; seed <= 30; seed++) {
const rec = S.recommendOption(S.advisor(all, "Blue", "eventually", { sampling: "uniform", seed }), {}, screen);
assert.ok(Math.abs(rec.propensity - 1 / 3) < 1e-12);
assert.match(S.norm(screen[rec.index]), new RegExp(all.model.game.bot_text.company[rec.choice]));
picks.add(rec.choice);
}
assert.strictEqual(picks.size, 3);
const a = S.advisor(all, "Blue", "eventually", { sampling: "uniform", seed: 4 });
assert.strictEqual(S.recommendOption(a, {}, screen).index, S.recommendOption(a, {}, screen).index); // same seed, same pick
});
test("the experiment plan: what each attempt of each account is for", () => {
const plan = JSON.parse(fs.readFileSync(path.join(ROOT, "data/public/experiments.json"), "utf8"));
const all = published.samples.all;
// attempt 1: every new account plays the same exact path (the replication)
for (let acc = 0; acc < 30; acc++) {
const a = S.assign(plan, { accountSeq: acc, attempt: 1 });
assert.strictEqual(a.experimentId, "fresh-replicate");
assert.strictEqual(a.target, "Black");
assert.strictEqual(a.fix.direction, "West");
assert.deepStrictEqual(a, S.assign(plan, { accountSeq: acc, attempt: 1 })); // the same every time
}
const a1 = S.assign(plan, { accountSeq: 0, attempt: 1 });
const adv = S.advisor(all, a1.target, a1.objective, { attempt: 1, fix: a1.fix });
let h = {};
for (let i = 0; i < 20; i++) {
const s = adv.step(h);
if (s.kind === "decision") { if (s.var === "encounter") break; if (a1.fix[s.var]) assert.strictEqual(s.choice, a1.fix[s.var]); h[s.var] = s.choice; }
else if (s.kind === "chance") h[s.var] = s.var === "shadow" ? "no" : Object.keys(s.probs).find((k) => k !== "die");
else break;
}
assert.strictEqual(h.hour, "Tree");
// the ablation waits for the replication, then gives every arm the same number of accounts
const stage2 = { ...plan, experiments: plan.experiments.map((e) => (e.id === "fresh-replicate" ? { ...e, status: "complete" } : e.id === "fresh-ablate" ? { ...e, status: "running" } : e)) };
const count = {};
for (let acc = 0; acc < 12 * 5; acc++) { const a = S.assign(stage2, { accountSeq: acc, attempt: 1 }); assert.strictEqual(a.experimentId, "fresh-ablate"); count[a.arm] = (count[a.arm] || 0) + 1; }
assert.deepStrictEqual(Object.values(count), Array(12).fill(5));
// with squadmates, staying and the direction are one screen
const squad = stage2.experiments.find((e) => e.id === "fresh-ablate").arms.find((a) => a.id === "with two squadmates");
const rec = S.recommendOption(S.advisor(all, "Black", "eventually", { attempt: 1, fix: squad.fix }), { company: "2 Squadmates" },
["Go east with your squad", "Leave your squadmates", "Convince them to head west instead"]);
assert.strictEqual(rec.index, 2);
assert.deepStrictEqual(rec.also, { direction: "West" });
// attempt 2: an account that saw the shadow tries the same path again; half of the others do too
const who = (attempt, facts) => Array.from({ length: 40 }, (_, acc) => (S.assign(plan, { accountSeq: acc, attempt, facts }) || {}).experimentId);
assert.ok(who(2, { sawShadowBefore: true }).every((id) => id === "shadow-again"));
assert.strictEqual(who(2, { sawShadowBefore: false }).filter((id) => id === "second-attempt-same-path").length, 20);
// later: two accounts in ten are the long-lived ones, the rest share the research experiments
const later = who(6, { sawShadowBefore: false });
assert.strictEqual(later.filter((id) => id === "veteran-repeat").length, 4);
assert.strictEqual(later.filter((id) => id === "veteran-variety").length, 4);
assert.deepStrictEqual([...new Set(later)].sort(), ["mapping", "meadow-blue-route", "sentinel", "veteran-repeat", "veteran-variety"]);
// variety: a different combination on every attempt, both gates kept, every one playable
const combos = new Set();
for (let t = 3; t < 3 + 60; t++) {
const a = S.assign(plan, { accountSeq: 11, attempt: t, facts: { sawShadowBefore: false } });
assert.strictEqual(a.experimentId, "veteran-variety");
assert.strictEqual(a.fix.clearing, "Help your squadmate");
combos.add(a.arm);
if (t < 8) S.advisor(all, a.target, a.objective, { attempt: t, fix: a.fix });
}
assert.strictEqual(combos.size, 60);
// an experiment that is not running is not assigned; its share goes to the others
const done = { ...plan, experiments: plan.experiments.map((e) => (e.id === "meadow-blue-route" ? { ...e, status: "complete" } : e)) };
assert.notStrictEqual(S.assign(done, { accountSeq: 5, attempt: 5 }).experimentId, "meadow-blue-route");
assert.strictEqual(S.assign({ ...plan, experiments: [] }, { accountSeq: 1, attempt: 5 }), null);
});