Ask a chatbot what a ferry costs and it will tell you. Ask it where that number came from and, in most cases, it cannot say — not because the answer is hidden, but because there was never an answer to hide. The figure was generated, not retrieved.
For a consumer product that is an irritation. For an organisation, it is the whole problem. A port cannot publish a tariff it cannot source. A revenue department cannot state a rate nobody signed off. The question who says so? is not pedantry — it is the difference between a service and a liability.
So the camp's first rule was not about Python. It was this: a fact must be sourced before the bot may say it.
Classify before you answer
Day eight gave the rule a name. classify_authority() — a function whose only job was to ask, of any claim, who is entitled to make it.
That turned out to be three questions, not one, and Team Code X gave the shape an acronym the camp kept: ART.
Authority — who is entitled to say this? An operator's official site is an authority. A travel blog is not. Neither is the model's own memory, however fluent it sounds.
Register — which record holds it, and in which voice it should be said? The word does double duty, deliberately. There is the record sense: tariffs live in one table, terminals and operators in another, cargo clearance in a third, and a question answered from the wrong one produces something confident and irrelevant. And there is the sense a linguist would recognise — register as how you speak. The same hard fact put to a port operator and to a nervous first-time traveller is one fact in two registers. The bot's persona is not decoration sitting on top of the answer; it is the second half of classifying it.
Territory — whose answer is it? This is the one that catches people, and in an eight-territory programme it caught everybody. The ferry terminal in Basseterre belongs to the St Christopher Air & Sea Ports Authority (SCASPA). Ask the same question about Nevis and the answer changes hands entirely — same question, same register, different island, different fact.
Get all three right and you may speak. Miss one and you are guessing in a way that sounds like knowing.
The Golden Record
By day ten ART had become a table, and the table had a name: the Golden Record — a Single Truth Table with five columns.
Topic. Question. Hard Fact. Source URL. Last Verified.
The first four are what you would expect of any knowledge base. The fifth is the one that shows real maturity, because it admits something most systems will not: facts decay. A tariff correct in July may be wrong in November. A terminal changes operator. A rate is revised.
A system that tells you when it last checked is more honest than one that simply sounds confident — and it is the only kind an organisation can defend.
Team ASPIRE pushed that second sense of Register furthest. Skye answers a six-year-old, Kaleb a ten-year-old, Zion a teenager — three registers over a fact that does not move an inch. What the EC$1,000 is, where half of it sits, what a share actually means: identical in all three mouths, grounded against the same Ministry circulars. Only the register changes.
And a second dial sits on top of the guide. Zion can answer as The Professor or as The Llama — the guide sets the learning level; the personality sets the tone. Two independent axes, both of them register, both riding on a fact pinned before either was applied.
Which is exactly why it holds. Varying how you say something is only safe once you are certain what you are saying; do it the other way round and you are not adapting, you are improvising in three voices. Pin the fact, then choose the register. In that order.
Where guardrails come in
A Golden Record tells a bot what it may say. Guardrails are what happen when it should not.
The two are halves of one mechanism. Grounding supplies the permission; guardrails enforce the refusal. Without the second, the first is a suggestion — and week two had already demonstrated what a language model does with a suggestion.
So the rules were written as code rather than hoped for in a prompt. If no source exists, do not answer. If the source is too old, say so, or hand off. If the question touches security, customs or someone's private records, refuse and escalate to a human. Day nine's jailbreak lab existed to test exactly this: campers spent a session trying to talk their own bots past their own rules, on the principle that you should think like a hacker, but build your chatbots like a defender.
A guardrail is not a filter on rude words. It is a rule about when a system is entitled to speak.
Try it yourself
Below is a real question, answered by a real bot at the showcase. Place it with ART, watch the Golden Record row assemble — then move the Last Verified date and see what the guardrail will and will not permit.
Classify before you answer
A traveller asks: “Who runs the ferry terminal on Nevis?”
Before the bot may answer, the claim has to be placed. Set all three.
The fifth column
The fact has not changed. Only the date it was last checked. Move it, and watch what the guardrail lets the bot say.
Answers
The Charlestown Ferry Terminal on Nevis is operated by NASPA, not SCASPA. Checked 12 days ago.
This is what a guardrail is. Not a filter on rude words — a rule about when the system is entitled to speak, enforced in code rather than hoped for in a prompt.
A teaching illustration written for this publication. The question and the facts are the ones PILOT answered on stage; the classification interface is constructed rather than a screenshot.
What it bought
This is the machinery underneath every build in this book.
W.E.S refusing to return a balance without a valid token. PILOT showing its sources with the date each was last checked, and the line on screen stating that a date is when a fact was verified, not a promise it is still true. A.I.D.A. wrapping a tax estimate in a mandatory disclaimer. ASPIRE AI grounded strictly against Ministry circulars rather than the open internet.
None of that is a feature a fifteen-year-old thinks to add. It is what you build when a real client's name is going on the output — and it started with a five-column table and a question about who says so.