Technical deep dive
Built like a referee, not a chatbot
Rule Referee is a board game rules app, and it wasn't built to be another chatbot. It's a rules interpreter. Every answer comes from a rulebook written ahead of time, found by a deterministic NLP retrieval engine. The goal was simple: fast, reliable answers without guessing.
- No internet needed and no outside AI services
- The same question always gets the same answer
- When it isn't sure, it asks instead of guessing
Philosophy
The three principles
Everything in the engine traces back to three habits a good referee already has.
Think before speaking
Rule Referee works out what you're really asking before it answers. It cleans up your wording, matches it to the ways people usually ask about each rule, and keeps track of the topic you're on.
Ask before answering
When a question could match more than one rule, Tally says it isn't sure what you meant and shows the top 3 rule names. Reply 1, 2, or 3 and you get the answer for the rule you picked. No assumptions, just a clear choice and a correct answer.
Stop before guessing
If nothing in the rulebook fits, Rule Referee says so and points you to what it can answer. If a game isn't in the app yet, it tells you instead of making something up.
Comparison
The difference in action
Here's how each principle changes what you actually get at the table.
| Principle | AI chatbots | Rule Referee |
|---|---|---|
| Think before speaking | Starts generating text right away from patterns, and can miss what you actually meant. | Matches your question to the ways people ask about each rule, then picks the rule that fits. |
| Ask before answering | Fills the gaps with assumptions and gives a confident answer that might be wrong. | Shows the top 3 matching rules and lets you choose when more than one could apply. |
| Stop before guessing | Can produce a plausible answer even when it doesn't know, which is how made-up rules happen. | Says it couldn't find that rule and suggests what to ask instead. |
Rationale
Why I didn't use a large language model
Large language models are powerful, but they're built to generate text, not to settle rules. At game night, guessing is a bug, not a feature.
Rule Referee doesn't browse the internet and doesn't rely on outside AI services. Every answer comes from a structured rulebook that was written and indexed ahead of time. That keeps answers deterministic, explainable, kid-safe, and consistent no matter how many times you ask.
At game night, guessing is a bug, not a feature.
Under the hood
Two engines, one referee
Rule Referee runs two layers inside the app. One finds the rule. The other handles the conversation. Neither one writes rules.
Layer 1
The rules engine
This is the deterministic search. It indexes every rule in the game's rulebook: its name, section, summary, core rule, common questions, and tags. When a question comes in, it scores every rule and ranks the best matches.
- The biggest boost goes to rules whose common questions look like yours
- Words and short phrases score higher in some fields, like rule names and tags
- Optional variants step back unless you ask about variants
- It returns a confidence score along with the matches
Layer 2
The conversation layer
This layer decides how to answer. It never invents a rule. Every answer comes from the rules engine's matches. What it adds is short-term memory and good table manners.
- Remembers the topic you're on, per game, so follow-ups stay on track
- Asks a clarifying question when confidence is low
- Understands a one-word reply like "2" or "winning" without looping
- Keeps answers short and in plain language
Step by step
What happens when you ask a question
- 1
Understand the wording
Your question is cleaned up and lightly rewritten. "Can I go again?" and "Do I get another turn?" point at the same intent.
- 2
Score every rule
The rules engine compares your question with every rule in the game, especially the common questions people ask about each one.
- 3
Check confidence
If one rule is a clear winner, that's the answer. If the top matches are close, Rule Referee asks which one you mean.
- 4
Answer in plain language
Tally answers with the rule's short summary, written so everyone at the table can follow it.
- Can I go again?
- Do I get another turn?
- Can I keep going?
When the top matches are close, Tally replies like this:
I'm not 100% sure which rule you mean. Which one is closest?
- Rule name (in its section)
- Rule name (in its section)
- Rule name (in its section)
Reply 1 to 3, or type a few words from the option.
Structured rule memory
Rules are stored as data, not paragraphs
Every game in Rule Referee is its own rulebook file, split into sections, with each section holding individual rules. The rulebook for CATAN, for example, has 117 rules across 11 sections.
Each rule carries a clear name, a short kid-friendly summary, a core rule, examples, the many ways people actually ask about it, tags, and notes on edge cases. You only ever see the short summary. The core rule is there for matching. Think of it as indexed memory, not text generation.
You don't recall a rule by quoting a paragraph. You recall intent, context, and outcomes.
One real rule, from the CATAN rulebook
- rule_id
- catan-robber-roll-7-overview
- rule_name
- What Happens When a 7 Is Rolled
- short_summary
- On a 7, no hex produces resources.
Players with more than 7 resource cards discard half, rounded down. Then the current player moves the robber and may steal 1 random resource card from a player on that hex.What Tally says
- core_rule
- If a 7 is rolled, players with too many cards discard. Then the current player moves the robber and steals one card.
Used for matching, never shown in the app
- example
- A 7 is rolled, so players may discard and the robber moves.
- common_questions
- what happens on a 7 roll 7 what happens when a 7 is rolled what happens when you roll a 7 if i roll a 7 do i get cards what happens when a 7 is rolled in catan
+ 1 more ways people ask it
- tags
- robber 7 seven roll a 7 rolling a 7 discard lose cards half
+ 8 more
Empathy
Why this works well for kids
Kids don't ask rules questions "correctly." They ask them honestly. Because Rule Referee matches intent and keeps track of the topic, it handles repeated questions, differently worded questions, and half-right questions without frustration or confusion.
There's no browsing, no outside content, and no chatting with anything beyond the rules of the game you picked.
Technical summary, in plain language
- No live internet access
- No external AI APIs
- Pre-written, structured rulebooks
- Intent matching through common questions, synonyms, and topic memory
- Scored, deterministic retrieval with a confidence check
- Clarification instead of guessing
- Every game tested against 400 to 650 real rules questions
This approach trades raw flexibility for trust, clarity, and safety, which is exactly what a rules dispute needs.
Questions
How Rule Referee works: FAQ
How does Rule Referee answer rules questions without AI?
Rule Referee uses a deterministic NLP retrieval engine. Every game ships as a structured rulebook written ahead of time. When you ask a question, the engine scores every rule in that rulebook against your words and returns the best match. Nothing is generated on the spot.
What happens when Rule Referee isn't sure what I mean?
It asks instead of guessing. If more than one rule could fit, Tally says it isn't sure which rule you mean and lists the top 3 matches. Reply 1, 2, or 3 (or a few words from the option) and you get that exact rule.
Will Rule Referee give the same answer to the same question?
Yes. The same question in the same game always produces the same result, just like a rulebook should.
Does Rule Referee remember what I just asked?
For a short while, yes. It keeps the current topic for the game you're playing, so a quick follow-up like "what about doubles?" stays on topic. That memory lives on your device and isn't sent anywhere.
Does Rule Referee need the internet?
No. Rule Referee doesn't browse the internet and doesn't call any outside AI service. Every rulebook and the engine that searches it live inside the app.
What's inside a Rule Referee rulebook?
Each game is a set of sections, and each section holds rules. Every rule has an ID, a name, a short kid-friendly summary, a core rule, examples, the common ways people ask about it, tags, and notes on edge cases. Tally answers with the short summary. The core rule stays behind the scenes and helps the engine find the right match.
How are Rule Referee games tested?
Every game is tested against 400 to 650 real rules questions before it ships. Then Autumn spot-checks the answers herself.
Fair, consistent, and easy to understand
Rule Referee isn't trying to be intelligent in the human sense. It's trying to be fair, consistent, and understandable. It thinks before it talks, just like a good referee should.