Why your Discord bot's game numbers are wrong

Someone in your server asks the bot how long Clove's smoke lasts when she casts it dead. The bot says 14 seconds. It said 14 seconds last year too, when that was true. Patch 12.05 cut it to 6. The patch note is one line: "Smoke duration 14s >> 6s when casted while dead." The bot never read it.

Nobody notices for a week, because the bot sounds sure. Then a member who plays Clove corrects it in front of everyone, and every answer the bot gave before that is now suspect. That is the cost: the trust, more than the number.

The four ways it goes stale

1. The model remembers

A chat model answering from its weights answers from the internet as it was when the model was trained. It has no date on any figure and no way to tell you it is out of date. Ask it the Vandal's body damage and you may get 39, which was right until patch 1.07 made it 40. Riot's notes date 1.07 to September 2020. The wrong number is six years old and still circulating in training data.

2. The wiki was read once

A scrape is a snapshot. The bot that scraped the ability page in March serves March's page in September. Wikis also lag the game: a change lands in the client on patch day and reaches the page when a volunteer edits it. Two delays stacked, neither visible to the bot.

3. Someone typed a table

The most common one. A mod pastes cooldowns into the knowledge base, and it is right on the day. Patch 10.04 raised Ruse's restock from 30 to 40 seconds. Patch 12.05 cut its dead duration from 14 to 6. The table has no field for "which patch is this from", so nobody knows it needs updating until a player says so.

4. The note and the game disagree

The subtle one. Ruse's alive duration is 14 seconds per the patch note, which says that includes half a second of formation and half a second of expiration. Measured across 349 replay instances, the median is 15. Both are true depending on what you count. A bot that has only one of them cannot say which it has.

What a source has to do

Every one of those failures has the same shape: a number with no date, no source, and no way to tell a current value from a dead one. A source the bot can trust has to do three things, and most do none.

1. Date every value by patch. Not "last updated" on the page; the patch the value took effect in, on the value. Then "since 12.05" is a field the bot can read out, and a value with no patch is honestly undated.
2. Keep the old value. When 14 became 6, keep the 14 with the patch it stopped in. That is what lets the bot answer "what did 12.05 change" and what lets you audit an old answer.
3. Say how well it is attested. Read from the game files, confirmed by a patch note, observed in the client, measured from replays, copied from a community wiki, or arithmetic on other values. Those are different kinds of true, and a bot that can name the kind stops sounding equally sure about all of them.

What to do this week

If your bot answers game questions today, three checks:

• Ask it five figures you know changed in the last year. If it gets one wrong, it is answering from memory or a stale scrape, and every other figure is suspect.
• Ask it which patch a figure is from. If it cannot say, it cannot be corrected when the next patch lands, because nobody will know which answers moved.
• Find the table. If a human typed the numbers, put a "patch" column beside them, and a date next to the patch. It is the cheapest fix and it makes the next patch day a diff instead of a hunt.

Or read from a source that already does all three. The Knowledge API is 14 games of damage, cooldowns, costs, and changes, each value with the page or file it was read from, a tier for how well it is attested, and the patch it took effect in. Superseded values stay in the file with the patch they stopped in. That 14 to 6 change is one line in Ruse's file, with the note it came from, and a bot reading it says 6 the night the corpus publishes.

How to wire it into a bot is the next post. It is four reads and one rule in the system prompt.

Questions