Route model: first attempts are their own shadow roll; experiment plan and automated bot cycle

- game.yaml: `givens` (first or later attempt) and `distinct` levels that are not pooled; the shadow roll
  depends on the attempt first. Form submissions show the shadow on 9% of first attempts and 0.3% of later ones.
- Reporting bias: a roll whose bias cannot be measured can borrow another roll's (`borrow: {shadow: meadow}`),
  so form submissions inform the shadow again. An assumption, stated in planner/policy.yaml.
- policy.json: targets.<colour>.firstAttempt (this attempt, eventually). Black: about 1.4% on a first attempt,
  0.07% on a later one.
- solver.js: advisor options `attempt`, `fix`, sampling "uniform"; assign() and armFor() for the plan.
- planner/experiments.yaml + build_experiments.py -> data/public/experiments.json (progress and status per
  experiment, rebuilt from the bot's runs).
- python -m collection cycle: import bot runs, refit, re-plan, deliver the kit; .gitea/workflows/bot-cycle.yml.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014jqCC2LZ4x1WHQ1V57nTsv
This commit is contained in:
ekcSchiano
2026-11-12 03:31:33 +00:00
co-authored by Claude Fable 5.2
parent ee07ee143e
commit ab3664f369
25 changed files with 1938 additions and 573 deletions
+17 -1
View File
@@ -62,6 +62,10 @@ def estimate(runs, prepared, cfg, rare_goals, seed=20261109, sims=4000):
e = {"rare": y, "n": n, "completeRare": ry, "completeN": rn}
if g.get("complete"):
e.update(rho=1.0, lo=1.0, hi=1.0, used="complete: counted as reported")
elif ry < cfg.get("min_complete_rare", 5) and (cfg.get("borrow") or {}).get(node):
# complete logging has (almost) never seen it, so the bias cannot be estimated for this roll: assume
# the group over-reports it as much as it over-reports another rare roll (filled in below)
e.update(rho=None, borrow=cfg["borrow"][node])
elif ry < cfg.get("min_complete_rare", 5):
# complete logging has (almost) never seen it: the bias cannot be estimated, so this roll is learned
# from complete logging only (its range then shows how little is known)
@@ -83,10 +87,22 @@ def estimate(runs, prepared, cfg, rare_goals, seed=20261109, sims=4000):
for g in groups:
for node in rolls:
e = out[g["name"]][node]
if e.get("rho") is None and not e.get("exclude"):
if e.get("rho") is None and not e.get("exclude") and not e.get("borrow"):
src = out.get(g.get("same_as") or "", {}).get(node, {})
e["rho"] = src.get("rho") or 1.0
e["lo"], e["hi"] = src.get("lo", e["rho"]), src.get("hi", e["rho"])
# a roll whose bias cannot be estimated borrows the same group's estimate for another roll (`borrow`)
for g in groups:
for node in rolls:
e = out[g["name"]][node]
if e.get("borrow"):
src = out[g["name"]].get(e["borrow"], {})
if src.get("rho"):
e.update(rho=src["rho"], lo=src.get("lo"), hi=src.get("hi"),
used=f"assumed equal to this group's bias for the {e['borrow']} roll: complete logs have only "
f"{e['completeRare']} rare outcomes, too few to estimate it")
else:
e.update(lo=None, hi=None, exclude=True, used="left out: no bias estimate to borrow")
return out, names
+24 -5
View File
@@ -28,6 +28,8 @@ ALONE = "(alone)"
ALPHA = 20
ALPHA_GRID = [1, 2, 5, 10, 20, 50, 100, 200, 500, 1000, 2000, 5000, 20000, 100000]
ENC_CATS = G["encounter_outcomes"]
GIVENS = G.get("givens") or {}
DEFAULT_GIVEN = {k: g["default"] for k, g in GIVENS.items()}
def enc_colour(ctx):
@@ -62,6 +64,8 @@ def prepare_run(r):
v = {"company": r["company"] if r["company"] in DECISIONS["company"] else None,
"direction": r["direction"] if r["direction"] in DECISIONS["direction"] else None,
"fs": r["first_sighting"] if r["first_sighting"] in ("Red", "Brown", "Green", "Orange") else None}
# the givens (game.yaml): first or later attempt, from the run's attempt number
v["attempt"] = r.get("attempt_group") if r.get("attempt_group") in GIVENS.get("attempt", {}).get("values", ()) else None
lf = r.get("left_squadmates")
known = lf is not None and not (isinstance(lf, float) and lf != lf) and lf != ""
v["left"] = ALONE if v["company"] == "Alone" else (("Leave them" if lf is True or str(lf).lower() == "true" else "Stay together") if v["company"] and known else None)
@@ -279,7 +283,7 @@ def fit_alpha(T, node):
pairs = []
for key in t:
L = int(key.split(":")[0])
if L == 0:
if L <= CHANCE[node].get("distinct", 0): # the top level, and levels that are not pooled (game.yaml)
continue
ctx = key[key.index(":") + 1:].split("|")
parent = f"{L - 1}:" + "|".join(ctx[:L - 1])
@@ -368,10 +372,14 @@ class Params:
In draw mode each world draws its numbers in the order its own walk of the game first needs them (a branch with
probability exactly zero in that world is never walked), so call solver.draw_worlds() before using it.
chance(node, state) -> {outcome: array(D)}; encounter(ctx) -> {outcome: array(D)}.
given: the run's givens (game.yaml `givens`, e.g. first or later attempt); having() gives the same worlds for
other givens.
"""
def __init__(self, T, mode="mean", seeds=(), alphas=None):
def __init__(self, T, mode="mean", seeds=(), alphas=None, given=None):
self.T, self.mode, self.alphas = T, mode, alphas or {}
self.given = {**DEFAULT_GIVEN, **(given or {})}
self.rngs = [Rng(s) for s in seeds] if mode == "draw" else []
self.D = len(self.rngs) if mode == "draw" else 1
self.memo = {} # (node, level key) or ("enc", ctx) -> array (K, D)
@@ -379,6 +387,14 @@ class Params:
self.world = None # while drawing one world (see draw_world), its index
self._cache = {}
def having(self, **given):
"""The same odds (and the same posterior draws) for a run with other givens."""
import copy
p = copy.copy(self)
p.given = {**self.given, **given}
p._cache = {}
return p
def _levels(self, node, state):
sp = CHANCE[node]
out = []
@@ -408,6 +424,7 @@ class Params:
if not gate_open(node, state): # a closed gate: the roll always comes out the same way
other = CHANCE[node]["otherwise"]
return {k: np.full(self.D, 1.0 if k == other else 0.0) for k in CHANCE[node]["outcomes"]}
state = {**self.given, **state}
# once every world has been drawn, the answer depends only on the context values: cache it
ck = (node, tuple(state.get(k) for k in CHANCE[node]["context"]))
if self.world is None and ck in self._cache:
@@ -419,13 +436,15 @@ class Params:
def _chance(self, node, state):
cats = CHANCE[node]["outcomes"]
al = self.alphas.get(node) or ALPHA
fitted = self.alphas.get(node) or ALPHA
distinct = CHANCE[node].get("distinct", 0)
p = None
for key in self._levels(node, state):
for L, key in enumerate(self._levels(node, state)):
c = self.T[node].get(key, {})
parent = p
al = 1.0 if L <= distinct else fitted # a level that is not pooled keeps a token prior of one run
def make_a(d, c=c, parent=parent):
def make_a(d, c=c, parent=parent, al=al):
if parent is None:
return [c.get(k, 0) + 0.5 for k in cats]
return [c.get(k, 0) + al * (parent[i] if d is None else parent[i][d]) for i, k in enumerate(cats)]
+41 -1
View File
@@ -250,13 +250,30 @@ def draw_worlds(T, seeds, alphas, support):
"""Posterior draws, one world per seed, each drawn in its own walk of the game (see Params)."""
p = model.Params(T, "draw", seeds=seeds, alphas=alphas)
p.reached = {}
default = dict(p.given)
for d in range(p.D):
p.world = d
_Drawer(p, support, d, p.reached).root()
# the same world for runs with other givens (a first attempt): only the rolls that depend on them are new
for given in other_givens():
p.given = {**default, **given}
_Drawer(p, support, d, {}).root()
p.given = default
p.world = None
return p
def other_givens():
"""Every non-default value of every given in game.yaml, one at a time: [{"attempt": "First attempt"}]."""
return [{k: v} for k, g in model.GIVENS.items() for v in g["values"] if v != g["default"]]
def given_key(given):
"""The name a non-default given's results are published under: {"attempt": "First attempt"} -> "firstAttempt"."""
words = " ".join(given.values()).split()
return words[0].lower() + "".join(w.capitalize() for w in words[1:])
# ---------------------------------------------------------------- objectives
def solve(params, X, support, lam=0.0, record=False, only=None):
@@ -364,6 +381,28 @@ def build_policy(runs, draws=200, seed=20261101, min_support=None, alpha=None, c
"crossFit": {"inSample": (cross[0]["inSample"] + cross[1]["inSample"]) / 2,
"heldOut": (cross[0]["heldOut"] + cross[1]["heldOut"]) / 2},
}
# Runs with another given (a first attempt). "This attempt" is solved as above. "Eventually": every attempt
# after it is an ordinary one, worth the default's ratio (lam), so the best play maximises
# P(X now) + P(no bond now) * lam, which is the same score with lam fixed.
later = [xi / (xi + oi) if xi + oi else 0.0 for xi, oi in zip(ev_ratio[0], ev_ratio[1])]
for given in other_givens():
m1, d1 = mean.having(**given), drawp.having(**given)
g_one = solve(m1, X, support, record=True)
g_ev = solve(m1, X, support, lam=ratio.lam, record=True)
e_one = evaluate(d1, X, support, g_one.decide)
e_ev = evaluate(d1, X, support, g_ev.decide)
def gblock(sol, ev, dr, m1=m1):
x, o = ev[0], ev[1]
return {"pX": {**summarise(x), "point": float(sol.value[0][0])},
"pOther": {**summarise(o), "point": float(sol.value[1][0])},
"tree": Exporter(m1, X, support, sol, dr).root()}
ev_block = gblock(g_ev, e_ev, solve(d1, X, support, lam=d_ratio.lam))
x0, o0 = float(g_ev.value[0][0]), float(g_ev.value[1][0])
ev_block["pEventually"] = {**summarise([x + (1 - x - o) * r for x, o, r in zip(e_ev[0], e_ev[1], later)]),
"point": x0 + (1 - x0 - o0) * float(np.ravel(ratio.lam)[0])}
targets[X][given_key(given)] = {"given": given, "thisAttempt": gblock(g_one, e_one, solve(d1, X, support)),
"eventually": ev_block}
return {
"schema": 1,
"model": {
@@ -371,12 +410,13 @@ def build_policy(runs, draws=200, seed=20261101, min_support=None, alpha=None, c
"alphas": alphas, "alphaFit": {k: f["logLik"] for k, f in fitted.items()}, "draws": draws, "seed": seed,
"minSupport": min_support,
"chanceContext": {k: s["context"] for k, s in model.CHANCE.items()},
"given": model.DEFAULT_GIVEN,
"decisions": model.DECISIONS, "support": support, "supportCounts": counts,
"tables": T,
# how much each kind of source over-reports each rare roll, and the correction applied (bias.py)
"reportingBias": bias_est,
# the game itself (core/spec/game.yaml), so the advisor in core/solver.js can re-solve from any point
"game": {k: G[k] for k in ("version", "colours", "decisions", "chance", "encounters", "encounter_outcomes",
"game": {k: G[k] for k in ("version", "colours", "givens", "decisions", "chance", "encounters", "encounter_outcomes",
"by_colour", "start", "steps", "from_form", "bot_text")},
},
"data": {"runs": len(runs)},
+100 -14
View File
@@ -11,6 +11,9 @@
*
* Public API (unchanged for the bot): advisor(sample, colour, objective, { only }), recommendOption(adv, h, texts),
* solverHistory(formVars, seen), makeParams(tables, { alphas, game }), ENCOUNTERS (set from the first sample).
* Since November 2026: advisor options `attempt` (the account's attempt number: a first attempt has its own
* shadow odds), `fix` (hold some decisions to given options: an experiment arm) and sampling "uniform" (mapping);
* assign(plan, { accountSeq, attempt }) reads the experiment plan (experiments.json) and says what an attempt is for.
* UMD: window.ThreshSolver in the browser, module.exports in Node. No dependencies.
*/
(function (root, factory) {
@@ -48,22 +51,31 @@
}
/* ---------- the odds: posterior means (or, with rand, one posterior draw) with hierarchical back-off ---------- */
function makeParams(T, { alphas = {}, game, rand = null }) {
/** The givens of a run (game.yaml `givens`): the defaults, with `attempt` set from an attempt number or name. */
function givenFor(game, { attempt = null, given = null } = {}) {
const g = Object.fromEntries(Object.entries(game.givens || {}).map(([k, v]) => [k, v.default]));
if (attempt != null && game.givens && game.givens.attempt) g.attempt = attempt === 1 || attempt === "1" || attempt === "First attempt" ? "First attempt" : "Later attempt";
return { ...g, ...(given || {}) };
}
// given: the run's givens; memo: pass the same Map to get the same posterior draw for another given
function makeParams(T, { alphas = {}, game, rand = null, given = null, memo = new Map() }) {
const normalise = (a) => { let sum = 0; for (const x of a) sum += x; return a.map((x) => x / sum); };
const estimate = rand ? (a) => dirichlet(a, rand) : normalise;
const memo = new Map();
const known = given || givenFor(game);
const ctxKey = (s, names) => names.map((k) => (s[k] == null ? "?" : s[k])).join("|");
function chance(node, state) {
function chance(node, walked) {
const sp = game.chance[node], cats = sp.outcomes;
// a closed gate (game.yaml `requires`): the roll always comes out as `otherwise`
if (sp.requires && Object.entries(sp.requires).some(([k, v]) => state[k] !== v)) return Object.fromEntries(cats.map((k) => [k, k === sp.otherwise ? 1 : 0]));
if (sp.requires && Object.entries(sp.requires).some(([k, v]) => walked[k] !== v)) return Object.fromEntries(cats.map((k) => [k, k === sp.otherwise ? 1 : 0]));
const state = { ...known, ...walked };
let p = null;
for (let L = 0; L <= sp.context.length; L++) {
const names = sp.context.slice(0, L);
if (names.some((k) => state[k] == null)) break;
const key = L + ":" + ctxKey(state, names), mk = node + "#" + key;
if (memo.has(mk)) { p = memo.get(mk); continue; }
const c = T[node][key] || {}, al = alphas[node] || ALPHA;
// a level that is not pooled (game.yaml `distinct`) keeps a token prior of one run
const c = T[node][key] || {}, al = L <= (sp.distinct || 0) ? 1 : alphas[node] || ALPHA;
const a = cats.map((k, i) => (c[k] || 0) + (p ? al * p[i] : 0.5));
p = estimate(a);
memo.set(mk, p);
@@ -174,22 +186,50 @@
* the deadly roll just passed on the step after it. When h already has a later step but skips an earlier
* decision (an older form record, say), the recommended option is assumed for the skipped one.
*/
function advisor(sample, X, objective = "eventually", { only = null, sampling = "best", seed = 1, draws = 40 } = {}) {
function advisor(sample, X, objective = "eventually", { only = null, sampling = "best", seed = 1, draws = 40, attempt = null, given = null, fix = null } = {}) {
const m = sample.model, game = m.game;
api.game = game; api.ENCOUNTERS = game.by_colour;
const solveWith = (params) => (objective === "eventually" ? solveRatio(params, game, m.support, X, only) : solve(params, game, m.support, X, 0, only));
// attempt: the account's attempt number (or "First attempt" / "Later attempt"). A first attempt has its own odds
// for the rolls that depend on it (the shadow); without it, a later attempt is assumed.
const defaults = givenFor(game), known = givenFor(game, { attempt, given });
const special = Object.keys(known).some((k) => known[k] !== defaults[k]);
// "uniform" (mapping) may take any option the game has, so it is solved without the minimum-support rule
const all = Object.fromEntries(Object.entries(game.decisions).map(([k, d]) => [k, d.options]));
const base = sampling === "uniform" ? { ...all, encounter: m.support.encounter } : m.support;
// fix: { decision: option } holds those decisions to one option (an experiment arm); the rest is solved around it
let support = base;
if (fix) {
support = { ...base };
for (const [k, o] of Object.entries(fix)) {
if (!game.decisions[k] || !game.decisions[k].options.includes(o)) throw new Error(`fix: ${k} has no option "${o}"`);
support[k] = [o];
if (game.decisions[k].support_by) for (const key of Object.keys(support)) if (key.startsWith(k + ":")) support[key] = [o];
}
}
// One world (the posterior means, or one draw): the odds for this run's givens, solved for the objective.
// "Eventually" on a first attempt: every later attempt is an ordinary one, worth its own ratio (lam), so the
// best play now maximises P(X now) + P(no bond now) * lam: the same score with lam fixed.
const solveWith = (rand) => {
const memo = new Map();
const mk = (g) => makeParams(m.tables, { alphas: m.alphas, game, rand, given: g, memo });
if (objective !== "eventually") { const params = mk(known); return { params, sol: solve(params, game, support, X, 0, only) }; }
if (!special) { const params = mk(known); return { params, sol: solveRatio(params, game, support, X, only) }; }
const lam = solveRatio(mk(defaults), game, base, X, only).lam;
const params = mk(known);
return { params, sol: { ...solve(params, game, support, X, lam, only), lam } };
};
// "best": the best route under the posterior means. "thompson": the best route of one posterior draw of the odds
// (one per attempt: make a new advisor with a new seed for each attempt), so every option is tried about as
// often as it is likely to be the best one; as the data firm up, sampling settles on the best route.
// "uniform": every option on the screen is equally likely (recommendOption picks; see there).
const thompson = sampling === "thompson";
const params = makeParams(m.tables, { alphas: m.alphas, game, rand: thompson ? rng(seed) : null });
const sol = solveWith(params);
const { params, sol } = solveWith(thompson ? rng(seed) : null);
// the chance Thompson sampling picks each option at a history, estimated from `draws` other draws: the
// propensity to log, so the runs can be re-weighted later
let worlds = null;
const propensityAt = (key, choice) => {
if (!thompson) return 1;
if (!worlds) worlds = Array.from({ length: draws }, (_, i) => solveWith(makeParams(m.tables, { alphas: m.alphas, game, rand: rng(seed * 7919 + i + 1) })).choices);
if (!worlds) worlds = Array.from({ length: draws }, (_, i) => solveWith(rng(seed * 7919 + i + 1)).sol.choices);
return worlds.filter((c) => c.get(key) === choice).length / draws;
};
const VARS = new Set([...Object.keys(game.decisions), ...Object.keys(game.chance), "encounter"]);
@@ -236,7 +276,7 @@
throw new Error("the game has no end");
}
return {
colour: X, objective, only, sampling, seed: thompson ? seed : null, value: sol.value, step, params, game,
colour: X, objective, only, sampling, seed: sampling === "best" ? null : seed, given: known, fix, value: sol.value, step, params, game,
/** attempts recorded at an encounter (genuine attempts only), to show how much data a bond chance rests on */
attempts: (ctx) => Object.values(((m.tables.encounter || {})[ctx] || { outcomes: {} }).outcomes).reduce((a, b) => a + b, 0),
/** the next decision only (null when the next step is a roll or the run is over) */
@@ -258,8 +298,18 @@
function recommendOption(adv, h, offered) {
const rec = adv.next(h);
if (!rec) return null;
const pat = ((adv.game.bot_text || {})[rec.decision] || {})[rec.choice];
const idx = pat ? offered.map(norm).findIndex((t) => new RegExp(pat).test(t)) : -1;
const texts = offered.map(norm), pats = (adv.game.bot_text || {})[rec.decision] || {};
// mapping (sampling "uniform"): every option on the screen that the game knows is equally likely, except a
// decision held by `fix` and the bonding encounter (where the advisor's own choice stands)
if (adv.sampling === "uniform" && rec.decision !== "encounter" && !(adv.fix && adv.fix[rec.decision])) {
const on = Object.keys(rec.options).map((o) => [o, pats[o] ? texts.findIndex((t) => new RegExp(pats[o]).test(t)) : -1]).filter(([, i]) => i >= 0);
if (on.length) {
const [o, i] = on[Math.floor(rng(hash(`${adv.seed}|${rec.decision}|${canon(h)}`))() * on.length)];
return { index: i, decision: rec.decision, choice: o, propensity: 1 / on.length, margin: null, options: rec.options, reason: `${rec.label}: ${o} (mapping: 1 of ${on.length} at random)` };
}
}
const pat = pats[rec.choice];
const idx = pat ? texts.findIndex((t) => new RegExp(pat).test(t)) : -1;
const values = Object.values(rec.options).map((v) => v.pX).sort((a, b) => b - a);
return { index: idx < 0 ? null : idx, decision: rec.decision, choice: rec.choice, propensity: rec.propensity,
margin: values.length > 1 ? values[0] - values[1] : null,
@@ -302,8 +352,44 @@
return h;
}
/* ---------- the experiment plan (planner/experiments.yaml, published as experiments.json) ---------- */
function hash(s) { let x = 2166136261; for (const ch of String(s)) { x ^= ch.codePointAt(0); x = Math.imul(x, 16777619) >>> 0; } return x || 1; }
/** The arm of the n-th unit (0-based) of an experiment: balanced blocks, each block a seeded shuffle of the arms. */
function armFor(experiment, n) {
const arms = experiment.arms, k = arms.length, rand = rng(hash(`${experiment.id}|${experiment.seed || 0}|${Math.floor(n / k)}`));
const order = arms.map((_, i) => i);
for (let i = k - 1; i > 0; i--) { const j = Math.floor(rand() * (i + 1)); [order[i], order[j]] = [order[j], order[i]]; }
return arms[order[n % k]];
}
/**
* What an attempt is for, under the plan: assign(plan, { accountSeq, attempt }) ->
* { experimentId, arm, fix, target, objective, sampling, seed, propensity, burn, planId }, or null when no
* experiment covers the attempt (use the account's own goal).
* accountSeq: the account's number in the order the accounts were made (0, 1, 2, ...), the same for its whole
* life; attempt: the account's attempt number (1 = its first threshing). The answer is the same every time for
* the same plan id, so a restart mid-attempt changes nothing. Pass the result to advisor():
* advisor(policy.samples.all, a.target, a.objective, { attempt, fix: a.fix, sampling: a.sampling, seed: a.seed })
* and log experimentId, arm, planId and the propensity on the run.
*/
function assign(plan, { accountSeq, attempt }) {
let open = (plan.experiments || []).filter((e) => e.status === "running" && attempt >= e.attempts.from && (e.attempts.to == null || attempt <= e.attempts.to)
&& (!e.accounts || (accountSeq % e.accounts.every) < e.accounts.take));
if (!open.length) return null;
if (open.some((e) => e.exclusive)) open = open.filter((e) => e.exclusive); // an exclusive experiment takes its attempts whole
const total = open.reduce((s, e) => s + e.share, 0);
let u = rng(hash(`${plan.seed || 0}|${accountSeq}|${attempt}`))() * total, exp = open[open.length - 1];
for (const e of open) { if (u < e.share) { exp = e; break; } u -= e.share; }
// the unit that is randomised: the account (one arm for life) or the attempt (blocks within the account)
// (accounts are counted among those the experiment takes, so its blocks stay balanced)
const nth = exp.accounts ? Math.floor(accountSeq / exp.accounts.every) * exp.accounts.take + (accountSeq % exp.accounts.every) : accountSeq;
const arm = exp.unit === "account" ? armFor(exp, nth) : armFor({ ...exp, seed: `${exp.seed || 0}|${accountSeq}` }, attempt - exp.attempts.from);
return { experimentId: exp.id, arm: arm.id, fix: arm.fix || null, target: exp.target, objective: exp.objective || "eventually",
sampling: exp.sampling || "best", seed: hash(`${plan.seed || 0}|${exp.id}|${accountSeq}|${attempt}`),
propensity: (exp.share / total) / exp.arms.length, burn: exp.burn || null, planId: plan.id };
}
return Object.assign(api, {
advisor, recommendOption, solverHistory, norm,
advisor, recommendOption, solverHistory, norm, assign, armFor, givenFor,
makeParams: (T, { alphas, game } = {}) => makeParams(T, { alphas, game: game || api.game }),
/** remember the game of a policy sample (sets ENCOUNTERS) without solving anything */
useGame(sample) { api.game = sample.model.game; api.ENCOUNTERS = api.game.by_colour; return api.game; },
+19 -3
View File
@@ -9,6 +9,16 @@ version: 1
# Colours in the order the route model reports them
colours: [Red, Orange, Brown, Green, Blue, Black]
# ---------------------------------------------------------------- givens: what a run starts with
# Known before the first choice; neither decided nor rolled. A roll can depend on a given by naming it in its
# context. The route model solves the game once per value; `default` is the one used when nothing is said.
givens:
attempt:
label: Attempt
# from the run's attempt number (attempt_group in data/public/runs_clean.csv)
values: [First attempt, Later attempt]
default: Later attempt
# ---------------------------------------------------------------- decisions: the rider's choices
# options: every option the game offers, in a fixed order (ties go to the first). Options seen fewer than
# min_support times in the data are not recommended, except where `always_offered` says so.
@@ -48,8 +58,10 @@ min_support: 20
# ---------------------------------------------------------------- chance events: the game's rolls
# outcomes: in a fixed order. context: the earlier steps the odds may depend on, most important first; the
# estimate for a full context is shrunk toward the estimate for a shorter one (hierarchical Dirichlet back-off,
# strength learned from the data). requires: earlier choices without which the roll always comes out as
# `otherwise` (a gate); runs that did not make them say nothing about the roll and are not counted.
# strength learned from the data). distinct: the first N context steps are not pooled at all: each of their values
# is its own roll (only a token prior of one run ties it to the rest). requires: earlier choices without which the
# roll always comes out as `otherwise` (a gate); runs that did not make them say nothing about the roll and are
# not counted.
chance:
fs: { label: First dragon, outcomes: [Red, Brown, Green, Orange], context: [direction, company, left] }
# rolls that can kill you: risk_label names the danger when the advisor reports it
@@ -59,7 +71,11 @@ chance:
shadow:
label: Shadow at the tree
outcomes: ["yes", "no"]
context: [cadet, post, fs, direction, company]
# The shadow is close to a first-attempt event (November 2026, form submissions, gated tree runs: 69 shadows in
# 748 first attempts, 7 in 1,995 later ones, the same in every week), so first and later attempts are two
# different rolls. Routes are then pooled within each.
context: [attempt, cadet, post, fs, direction, company]
distinct: 1
# the "Black Dragon Gates" of the community route PDF (November 2026). The runs agree: in the spreadsheet,
# 78 shadows in 4199 tree runs that investigated and helped, 2 (both form submissions) in 646 that did not.
requires: { scream: Investigate the scream, clearing: Help your squadmate }