From b51a6c7039c3a4900d09a7a9c2fc3e5993636749 Mon Sep 17 00:00:00 2001 From: kathleen Date: Thu, 19 Nov 2026 18:40:01 +0000 Subject: [PATCH] The Black on the page: a first-attempt range marked low evidence; the Black panels load again - The odds card for the Black: "Low evidence", "about 1 in N" first attempts, the 95% range written as "between 1 in X and 1 in Y", later attempts effectively never, why it is uncertain (every logged shadow is a rider who chose to post), and the steps (shadow -> reached the Black -> bonded). Dashed border. - Valley map: the Black's routes are solved for a first attempt (they showed ~0% per attempt), labelled "on a first attempt", and "Why this route?" gives the same range and caveat. - "The Black is not dice" panel: why the odds are uncertain, and what replaces the assumption (fresh-account bot runs to the tree). - analysis/black/evidence.py wrote "odds_ratio": Infinity (no later-attempt shadow at all), which is not JSON: the page's two Black panels showed "could not be loaded". Now null, and the report says "unbounded". Co-Authored-By: Claude Opus 5.5 Claude-Session: https://claude.ai/code/session_01UCzTd31GphpMXDhQqWrCMQ --- analysis/black/evidence.py | 7 ++++--- analysis/reports/BLACK.md | 2 +- data/public/black.json | 2 +- site/atlas/explorer.js | 13 +++++++------ site/atlas/index.html | 2 ++ site/atlas/ui.js | 10 ++++++++++ 6 files changed, 25 insertions(+), 11 deletions(-) diff --git a/analysis/black/evidence.py b/analysis/black/evidence.py index da6a38a..13b1126 100644 --- a/analysis/black/evidence.py +++ b/analysis/black/evidence.py @@ -141,7 +141,8 @@ def build(): c[i + 1] += 1 orr, obs, exp, z = mantel_haenszel(by.values()) facts["same_route"] = {"routes": len(by), "first_shadows": round(obs), "expected": round(exp, 1), - "odds_ratio": round(orr, 1), "z": round(z, 1)} + "odds_ratio": round(orr, 1) if math.isfinite(orr) else None, # JSON has no Infinity + "z": round(z, 1)} facts["form_first"] = [sum(c[0] for c in by.values()), sum(c[1] for c in by.values())] facts["form_later"] = [sum(c[2] for c in by.values()), sum(c[3] for c in by.values())] L += ["## 2. The same route on a first attempt and on a later one", "", @@ -153,7 +154,7 @@ def build(): L.append(f"| {' / '.join(key)} | {frac(y1, n1)} | {frac(y0, n0)} |") L += ["", f"Across all {len(by)} recorded routes, comparing first and later attempts on the same route: " f"{obs:.0f} first-attempt shadows where {exp:.1f} would be expected if the attempt did not matter " - f"(odds ratio {orr:.1f}, z = {z:.1f}). First and later attempts choose almost the same routes, so the " + f"(odds ratio {f'{orr:.1f}' if math.isfinite(orr) else 'unbounded: no later-attempt shadow at all'}, z = {z:.1f}). First and later attempts choose almost the same routes, so the " "difference is not in the recorded choices: it is something about the account, or a choice the form does " "not record.", ""] @@ -266,7 +267,7 @@ def main(argv=None): ap.add_argument("--check", action="store_true") a = ap.parse_args(argv) text, facts = build() - data = json.dumps(facts, ensure_ascii=False, indent=1) + "\n" + data = json.dumps(facts, ensure_ascii=False, indent=1, allow_nan=False) + "\n" if a.check: for path, want in ((OUT, text), (OUT_JSON, data)): if not path.exists() or path.read_text(encoding="utf-8") != want: diff --git a/analysis/reports/BLACK.md b/analysis/reports/BLACK.md index e90d2a6..9f5427a 100644 --- a/analysis/reports/BLACK.md +++ b/analysis/reports/BLACK.md @@ -31,7 +31,7 @@ If the shadow depended only on the choices the form records (company, first dire | Alone / Center / Assess the situation / Daggers or not (protect squadmate) / Left bullies to the dragons / Forest | 1 / 28 (3.6%) | 0 / 64 (0.0%) | | Alone / Center / Assess the situation / Engage the smaller cadet / Squadmate's revenge / West | 2 / 9 (22.2%) | 0 / 81 (0.0%) | -Across all 326 recorded routes, comparing first and later attempts on the same route: 93 first-attempt shadows where 38.8 would be expected if the attempt did not matter (odds ratio inf, z = 12.5). First and later attempts choose almost the same routes, so the difference is not in the recorded choices: it is something about the account, or a choice the form does not record. +Across all 326 recorded routes, comparing first and later attempts on the same route: 93 first-attempt shadows where 38.8 would be expected if the attempt did not matter (odds ratio unbounded: no later-attempt shadow at all, z = 12.5). First and later attempts choose almost the same routes, so the difference is not in the recorded choices: it is something about the account, or a choice the form does not record. ## 3. On one route, is the shadow decided by the cadet's wing, section and squad? diff --git a/data/public/black.json b/data/public/black.json index 9043012..76ccf8c 100644 --- a/data/public/black.json +++ b/data/public/black.json @@ -84,7 +84,7 @@ "routes": 326, "first_shadows": 93, "expected": 38.8, - "odds_ratio": Infinity, + "odds_ratio": null, "z": 12.5 }, "form_first": [ diff --git a/site/atlas/explorer.js b/site/atlas/explorer.js index 2938c7a..07d58ce 100644 --- a/site/atlas/explorer.js +++ b/site/atlas/explorer.js @@ -92,7 +92,8 @@ const model = () => st.policy.samples[st.sample]; function advisorFor(enc, colour = st.colour) { const k = `${st.sample}|${colour}|${st.objective}|${enc}`; - if (!st.cache.has(k)) st.cache.set(k, S.advisor(model(), colour, st.objective, { only: [enc] })); + // the Black needs the shadow, which (almost) only shows on a first attempt: its routes are solved for that attempt + if (!st.cache.has(k)) st.cache.set(k, S.advisor(model(), colour, st.objective, { only: [enc], attempt: colour === "Black" ? "First attempt" : null })); return st.cache.get(k); } let encParams = null; @@ -530,7 +531,7 @@ }); // destination label const end = pts[pts.length - 1]; - pill(gLab, end.P.x, end.P.y + 30, `${pct(r.value.x)} per attempt`, { edge: tc, strong: rel === 1 }); + pill(gLab, end.P.x, end.P.y + 30, `${pct(r.value.x)} ${st.colour === "Black" ? "on a first attempt" : "per attempt"}`, { edge: tc, strong: rel === 1 }); // choice points on the focused route: one diamond per place, even when several choices are made there // (at the bully fight: help or walk away, your first move, the dagger, then the bullies' fate) if (focus) { @@ -841,7 +842,7 @@ function panelRoutes(routes) { const best = Math.max(1e-12, ...routes.map((r) => r.value.x)); - const T = model().targets[st.colour], b = T[st.objective]; + const T = model().targets[st.colour], first = st.colour === "Black" && T.firstAttempt, b = first ? T.firstAttempt.thisAttempt : T[st.objective]; const focus = routes.find((r) => r.enc === st.focus) || routes[0]; const card = (r) => { const rel = r.value.x / best, o = encOutcome(r.enc), c = encColour(r.enc); @@ -851,7 +852,7 @@ return ``; + ${pct(r.value.x)}${st.colour === "Black" ? "on a first attempt" : "per attempt"} · ${bond}${n < 30 ? ` · only ${n} attempts logged` : ""}${r.value.o > 0.005 ? ` · ends on another dragon ${pct(r.value.o)}` : ""}`; }; const steps = focus.steps.filter((s) => s.kind === "decision" || s.kind === "chance").map((s) => { if (s.kind === "chance") return `
  • ${esc(s.label)}${esc(String(s.taken))}roll ${pct(s.probs[s.taken])}
  • `; @@ -862,7 +863,7 @@ return `
  • ${opts}${inside}
  • `; }).join(""); const risks = focus.steps.filter((s) => s.risk && s.risk.p > 0.0005).map((s) => `${s.risk.label.toLowerCase()} ${pct(s.risk.p)}`); - const lb = st.policy.samples.lowBias.targets[st.colour].thisAttempt; + const lbT = st.policy.samples.lowBias.targets[st.colour], lb = first && lbT.firstAttempt ? lbT.firstAttempt.thisAttempt : lbT.thisAttempt; return `
    ${mini(st.colour, POSE_OF[st.colour])}
    Routes to a

    ${st.colour} dragon

    ${chooser()} @@ -874,7 +875,7 @@
      ${steps}

    Diamonds on the map are choices: click one to see where the other options lead. ${st.blocked ? "" : `Your choices apply to every route to this dragon. Changed your mind? `}

    Why this route? -

    Best route for any ${st.colour}: ${pct(b.pX.mean)} per attempt (95% interval ${pct(b.pX.lo)}–${pct(b.pX.hi)}), about ${isFinite(b.expectedAttempts.mean) ? Math.round(b.expectedAttempts.mean) : "–"} attempts. Riders who logged every run suggest ${pct(lb.pX.mean)}.

    + ${first ? `

    Low evidence Best route for the Black, on a first attempt: about 1 in ${Math.round(1 / Math.max(b.pX.mean, 1e-9))}, somewhere between 1 in ${Math.round(1 / Math.max(b.pX.hi, 1e-9))} and 1 in ${Math.round(1 / Math.max(b.pX.lo, 1e-9))} (95%). Later attempts almost never see the shadow. The low-bias sample suggests about 1 in ${Math.round(1 / Math.max(lb.pX.mean, 1e-9))}. Every logged shadow comes from riders who chose to post their run, so the shadow's odds rest on an assumption until complete first-attempt logs exist.

    ` : `

    Best route for any ${st.colour}: ${pct(b.pX.mean)} per attempt (95% interval ${pct(b.pX.lo)}–${pct(b.pX.hi)}), about ${isFinite(b.expectedAttempts.mean) ? Math.round(b.expectedAttempts.mean) : "–"} attempts. Riders who logged every run suggest ${pct(lb.pX.mean)}.

    `}

    Each roll's odds may depend on your earlier choices, as far as the data support it. Choices marked "no clear winner" in Deep Dive are within the noise.${risks.length ? ` Risks on this route: ${risks.join(", ")}.` : ""}

    The bond chance at each encounter is the community average: the form does not record the answers riders gave there.

    diff --git a/site/atlas/index.html b/site/atlas/index.html index 5dcab9d..fda0874 100644 --- a/site/atlas/index.html +++ b/site/atlas/index.html @@ -182,6 +182,8 @@ section h2 { font-family: var(--font-crest); letter-spacing: 0.03em; } .odd .big { font-family: var(--font-crest); font-weight: 700; font-size: 2.1rem; line-height: 1; color: var(--c, var(--ink)); } .odd .what { font-weight: 600; } .odd .how { font-size: var(--step--1); color: var(--muted); } +.odd-low { border-style: dashed; } +.odd-low .big { opacity: 0.85; } .odd svg { display: block; width: 100%; max-width: 220px; height: auto; } .odd::before { content: ""; position: absolute; inset: 0 auto 0 0; width: 4px; background: var(--c, var(--line)); } .combo { display: grid; grid-template-columns: repeat(auto-fit, minmax(min(100%, 320px), 1fr)); gap: 16px; } diff --git a/site/atlas/ui.js b/site/atlas/ui.js index 30da379..4b7fe32 100644 --- a/site/atlas/ui.js +++ b/site/atlas/ui.js @@ -159,6 +159,8 @@ if (!b) { if (t2) t2.innerHTML = '

    Your first attempt is different

    The shadow evidence could not be loaded.

    '; return; } const ff = b.form_first, fl = b.form_later, sr = b.same_route, as = b.after_shadow, gaze = as.answers["held its gaze"]; if (t2) t2.innerHTML = `

    Your first attempt is different

    Among riders who helped their squadmate and climbed the tree, the shadow came in ${frac(ff)} first attempts (${pct(ff[0] / ff[1], 1)}) and ${frac(fl)} later ones. On the very same routes, first attempts saw it about ${Math.round(sr.first_shadows / Math.max(sr.expected, 1) * 10) / 10}× as often as chance would give. The shadow is not a dice roll you can repeat.

    `; + const steps = $("#black-steps"); + if (steps) steps.innerHTML = `How it breaks down: the shadow came in ${frac(ff)} reported first attempts (over-reported); after it, ${as.reached_black} riders reached the Black and ${as.bonded_black} bonded (${pct(as.bonded_black / Math.max(as.reached_black, 1))}).`; if (!box) return; const br = b.best_route; box.hidden = false; @@ -170,6 +172,7 @@ ${gaze ? `
  • At the Black, riders who held its gaze bonded in ${frac(gaze)} logged answers.
  • ` : ""}
  • The bot has made ${b.bot.attempts.toLocaleString()} attempts on ${b.bot.accounts} accounts, ${b.bot.gated_tree} on the gated path, and seen the shadow ${b.bot.shadows} times: none of them was a first attempt${b.bot.earliest_attempt ? ` (its earliest was attempt ${b.bot.earliest_attempt})` : ""}, which fits.
  • +

    Why the Black's odds are uncertain. The Black comes only after the shadow, and every logged shadow comes from riders who chose to post their run. Riders are much more likely to post a run with a shadow than one without, and no complete log (the bot, bulk logs) has a first attempt with a shadow yet, so the estimate assumes riders over-report the shadow as much as the meadow Blue. Fresh accounts the bot plays to the tree on their first attempt will replace that assumption; until then the odds are a range, not a figure.

    Method and every table: analysis/reports/BLACK.md; the numbers: data/black.json.

    `; } @@ -189,6 +192,13 @@ const art = (c) => window.DragonArt ? `` : ""; $("#odds-colours").innerHTML = cols.map(({ c, first, p, pAll, lo, hi }) => { const e = enc(c); + // The Black rests on one roll no complete log has measured (the shadow on a first attempt): a range, not a figure + if (first) return `${art(c)} + Low evidence + about 1 in ${oneIn(p)}first attempts end bonded to a Black, on the best route + ${dots(Math.round(p * 1000), 1000, cssVar(DV[c]))} + Somewhere between 1 in ${oneIn(hi)} and 1 in ${oneIn(lo)} (95%). Later attempts: no shadow has ever been logged, so effectively never.${Math.abs(pAll - p) / p > 0.25 ? ` All form posts suggest about 1 in ${oneIn(pAll)}.` : ""} Every logged shadow comes from riders who chose to post their run, so this rests on an assumption until complete first-attempt logs exist. + See the route →`; return `${art(c)} ${p >= 0.5 ? pct(p) : `1 in ${oneIn(p)}`}${p >= 0.5 ? "of runs" : "runs"} end bonded to ${/^[AEIOU]/.test(c) ? "an" : "a"} ${c}, on the best route${first ? ", on your first attempt" : ""} ${dots(Math.round(p * 1000), 1000, cssVar(DV[c]))} -- 2.54.0