I got the wrong answer — why?

Why answers can be wrong

I try hard to be right. I still get things wrong sometimes. Board game rules are unusually tricky for an AI to handle: rulebooks have footnotes, diagrams, edge-case boxes, errata sheets, card text, and expansion rules that sometimes contradict the base book. Any one of those can break my answer.

This page is a practical guide to debugging a wrong answer. Run through these checks before concluding that I am broken — in most cases, the fix is a re-ask with slightly different context.

Check #1 — Is the right game selected?

Every answer I give is relative to a specific game. If you ask "how does combat work?" while the selected game is Wingspan, you get a Wingspan-flavoured answer — which is nonsense, because Wingspan does not have combat.

• Web chat: the current game appears in the sidebar and in the header above the input box. If it is wrong, click it and pick the right one.
• Telegram: type /game to see the current game. Use /select [game name] to switch.
• Widget: the game is fixed by the publisher — you cannot switch it yourself. If the widget is on the wrong game's page, contact the site owner.

About 30 percent of "wrong answer" reports I see turn out to be game-context problems.

Check #2 — Does the rulebook have diagrams or tables I cannot read?

I extract rulebook text using Apache Tika. Tika handles plain text and most tables reasonably well, but it cannot read information that only exists as an image:

• Hex maps with movement costs written on the terrain art
• Flowcharts of the turn sequence drawn as a diagram
• Unit cards photographed into the rulebook as graphics
• Setup diagrams showing token placement

If your question is about something that lives in a picture, I will either miss it entirely or guess from surrounding text. Open the rulebook PDF yourself and check whether the relevant rule is drawn rather than written. If it is, I cannot help with that specific detail.

Check #3 — Is the answer on a card rather than in the rulebook?

Many modern games spread rules across two places:

1. The general rulebook (which I have)
2. The text on individual cards, tiles, or player boards (which I usually do not have)

Questions like "what does the Crimson Ranger do?" or "what is the effect of the Altar of Whispers tile?" are card-specific questions. Unless the publisher has uploaded a reference card or encyclopedia PDF alongside the main rulebook, I will not have the text of individual components. I will either admit I do not know, or — worse — I will confabulate an answer from a generic rule that sounds plausible.

If you suspect this, check the rulebook PDF for a "card reference" section. If there is no such section, I genuinely do not have that information.

Check #4 — Was the page citation correct?

Every answer I produce includes PDF citations like [PDF1, Pages 34-35]. These are clickable links in the web chat and widget. Open the link. Look at the page I cited.

Three things can happen:

1. The cited page contains exactly the rule I described. My answer is right.
2. The cited page contains the rule, but I summarised it inaccurately. My answer is partially right — I found the source but misread it.
3. The cited page does not contain the rule at all. My answer is wrong, and I cited badly.

Case 3 is the most useful signal for me. If you find a wrong citation, use the feedback button. It goes directly to the quality queue.

Important caveat about page numbers: my page citations are estimated from character offsets inside the extracted text, not from explicit page markers. Accuracy is usually within 1-2 pages. So "Pages 34-35" might really mean "somewhere around pages 33-36". If the exact page is empty, scan forward and backward before concluding the citation is broken.

Check #5 — Are there multiple editions of the rulebook?

Board games get new editions, errata releases, and living rulebooks. I have exactly one PDF per game in most cases — usually the most recent one a publisher or community has sent me — but I might not be on the latest version.

Signs you are hitting an edition mismatch:

• The rule I describe is almost right but has the wrong numbers (e.g., "3 VP" instead of "4 VP")
• The component counts in my answer do not match your game box
• I mention a rule that you cannot find anywhere in your copy of the rulebook

If you suspect edition drift, check the footer or title page of my cited PDF (open the link) to see which edition or print run it is. If it is older than your copy, flag it via feedback. Publishers in the partner portal can upload the newer edition themselves.

Check #6 — Language mismatch

The rulebook I have might only exist in English, even if you are asking in Italian or German. That is actually fine for retrieval — my vector search is language-agnostic enough to find the right English chunks from an Italian question. But sometimes the answer goes wrong in a specific way:

• A rule has a game-specific term of art (e.g., "Aufseher" in a German-original rulebook) that only exists in the non-English rulebook. If I am synthesising from the English edition, I will miss the term entirely.
• Translation nuances: a rule that says "may" in English sometimes says "must" in a translation. If you are playing with the translated rulebook, my English-sourced answer can be subtly wrong.

If you suspect this, re-ask with a direct translation of the exact phrasing from your rulebook.

What to do when you find a wrong answer

Every answer has a thumbs-up / thumbs-down button underneath it. Use the thumbs-down when you find a wrong answer. Please include a short comment explaining why it is wrong — one sentence is enough.

Negative feedback goes into an admin review queue. Wrong answers that share a pattern (wrong for many users, wrong in the same way) get flagged for prompt-template fixes or corpus refinement. This is the single most useful thing you can do to improve my answers over time.

When extended mode helps

If you get a Tier 1 answer that seems off, and the question is genuinely tricky — a contested rule, an edge case, an interaction between two subsystems — let the system escalate to Tier 2 (extended mode). Tier 2 pulls in community discussions on top of the official rulebook. It is slower (7-35 seconds instead of 5-7), but it handles ambiguous rules much better than Tier 1.

Escalation is usually automatic: if my Tier 1 confidence is below the threshold, I escalate without asking. You can also explicitly request extended mode from the UI when you feel the Tier 1 answer did not engage with the real complexity of the question.

When to trust community answers

Tier 2 includes citations from community threads ([T1], [T2], etc.). These are useful but carry a warning: they are community interpretation, not official errata. Designers occasionally post in these threads, which is the gold standard. Random forum users are less reliable — some get it right, some confidently get it wrong.

Rules of thumb:

• If a community citation is from a designer post (marked with a crown icon in the thread view), trust it like an official source.
• If a community citation contradicts the rulebook and is not from a designer, treat it as "there is debate about this rule".
• If the community consensus is unanimous and the rulebook is ambiguous, the community interpretation is usually correct in practice.

Closing note

No automatic system is perfect. I am a rulebook assistant, not a rules lawyer. I will get edge cases wrong, I will miss things that only appear in diagrams, and I will occasionally invent a rule that sounds plausible but is not in your book. Your feedback — especially the thumbs-down with a short explanation — is how I improve. Every wrong answer you report makes the next answer slightly less likely to be wrong.

If an answer is wrong and matters (mid-game, mid-dispute), always check the PDF citation yourself before taking the answer as gospel. That habit alone eliminates the worst failure modes.