Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Reading mindwanderer

mindwanderer analyzes Spirit Island Digital game data into win-rate and difficulty tables per (spirit × adversary × level × player-count). It’s the best public signal for relative spirit strength — but it’s not ground truth, and reading it well takes a few heuristics.

Beyond the stats dashboard

The headline dashboard at /stats.html is only the entry point. mindwanderer.net/si/ has 100+ detail sub-pages — per-spirit (LightningsSwiftStrike.html), per-adversary (TheHabsburgMonarchy.html), per-scenario, and per-board-layout (layout_T4.html) — each with its own blight-flip rate, average score, and difficulty model. If you’re prepping a specific matchup, open the spirit and adversary detail pages side by side; the dashboard only surfaces aggregate cells.

When you cite mindwanderer numbers, also read BGG thread 3555502 — An Update on the SI Digital Statistics Project: the Handelabra analytics migration changed collection methodology mid-stream, so pre- and post-migration numbers aren’t strictly comparable.

For broader empirical context, the community’s SI Digital Data Analysis Series (Jonathan4290, 26 entries) mines Spirit Island Digital data covering per-spirit-per-adversary results, scenario difficulty, and aspect performance with more narrative than mindwanderer alone. Its “2.27M” figure counts log lines, not games (BGG 3569421).

What it actually is

  • Spirit Island Digital only: every game in it was played in the digital app. There is no tabletop data, so see Digital vs. Tabletop before applying it to your table.
  • Regression models: the site’s figures come from regression models fit to that data, not only from raw wins ÷ games. A model estimate adjusts for the other factors in the model, so a Wilson interval computed from a raw count is not its uncertainty. Use the Wilson interval below only where the page gives a raw wins-and-games count.
  • Skill-heterogeneous: a player’s 1st Shadows game and their 50th are both in the data.
  • Selection bias: it covers the people who play the digital game, at the difficulties they pick.

Where a page does give a raw count, say a hypothetical Shadows vs. England L6: 52% (n=180), treat that as: “The true underlying win rate for this population is probably in (44.7%, 59.2%) at 95% confidence.”

The Wilson 95% CI

Win-rate without a CI is just a point estimate. With n=180 and p=0.52 the Wilson 95% interval is approximately (44.7%, 59.2%). At n=25 that same 52% point estimate broadens to roughly (33.5%, 70.0%) — a much weaker claim.

Mental shortcut (near 50%, the 95% half-width of one rate is 1.96 × √(0.25/n) ≈ 1/√n; the gap between two rates with n games each must exceed about 1.96 × √(0.5/n) ≈ 1.4/√n to clear the noise):

  • n < 20: “we don’t know.” Each rate is ± 22 points or more. Treat as a rumor.
  • n = 20–100: directional but weak. At n = 100 each rate is ± 10 points, and gaps under about 14 points are noise.
  • n = 500: each rate is ± 4.4 points; gaps under about 6 points are noise.

Comparing two spirits

Don’t subtract point estimates. Compute each’s Wilson CI; if intervals overlap, you can’t confidently say one is stronger.

Example (hypothetical numbers, for the arithmetic only): Shadows 52% (n=180) vs. River 48% (n=95):

  • Shadows: (44.7%, 59.2%)
  • River: (38.2%, 57.9%)

Intervals overlap heavily → “maybe Shadows is stronger but it’s within error bars.”

Compare (also hypothetical): Bringer 61% (n=220) vs. Thunderspeaker 47% (n=200):

  • Bringer: (54.4%, 67.2%)
  • Thunderspeaker: (40.2%, 53.9%)

The intervals just miss each other (53.9% vs. 54.4%) → Bringer is plausibly stronger (narrow margin, still reasonable doubt).

What mindwanderer doesn’t capture

  • Scenario effects: scenario cells hold fewer games than no-scenario cells; check n before trusting one.
  • Expansion mix: check whether the page filters by the expansions in use. A cell that pools games played with different decks mixes different Event, Fear, Blight and Power pools.
  • Player-count interaction: 2-handed solo ≠ true 2-player. Filter by player count.
  • Adversary-within-level variation: L3 England ≠ L3 Scotland. Blended top-lines lose this.

Workflow

  1. Pick the spirit × adversary combo.
  2. Filter mindwanderer to your player count.
  3. Read the point estimate + n. Compute Wilson bounds (or eyeball: 95% half-width ≈ 1/√n near 50%).
  4. Compare against the closest alternative spirit via overlapping-intervals test.
  5. Cross-reference Expansion Dilution Claims for how your expansion mix shifts the baseline.

When to ignore mindwanderer entirely

  • You’ve played <10 games total — the community’s average skill is above yours. Their win rates overstate what you’ll experience.
  • You’re doing a specific scenario or aspect — data is thin. Use the spirit chapter’s hand-authored analysis instead.
  • You’re playing for learning — highest win rate ≠ most learning-value. A below-average spirit can teach more if it’s a different archetype than your comfort pick.

See also