Where we are.
Episode 2 runs in eight parts, about 29 minutes, then questions. On air the yellow card is the part we are in. Everything you see is in the public repo, so you can follow along or come back to it later.
1The argumentBan the tush push: the league voted on it in May 2025, and the ban failed. Five things fans mean by the play. Before the agent can judge it, it has to find it.
2The search, by handThree searches of 260,927 play descriptions: "tush push," "brotherly shove," and "sneak." Zero, zero, two. The play has no name in the data.
3The definition and the dictionaryWhich parts of each definition the 31 columns can see. The dictionary folder and its README. The tool refusing to measure a play with no entry, and the shape an entry takes.
4The policy, read against episode 1'sSide by side, what changed: the tool, the new "Your dictionary" section, the ban as a question the plays cannot settle, the ruling's eighth part, and a two-line allowlist.
5Run itThe runner's new beats, then the episode's claim goes in: the search, the refusal, the entry written, three reads, and a ruling that cites its entry.
6Memory, provedYour claim from the chat, with the entry now on file: the agent uses it, does not rewrite it, and cites it. Then a claim it has to decline.
7CloseMemory is a file. The tool reads only the file. The runner checks the citation and the file. Next: kickers.
8The season sheets, then Q&AFive seasons rebuilt live, this season's sheet made live, then your questions.
Space next
Backspace back
R start over
While you wait
The episode page explains the setup for someone who has never built an agent. Five definitions shows what the data can and cannot see, and the tool page is a diagram of the one script the agent is allowed to run. The explorer, the tool's real output for 54 definitions of the play, goes up after the episode airs. The dictionary is empty until the agent writes its first entry, on air.