Learnings

One concept per file: an italic ELI5 opener, then the full mechanism — what it predicts, where it bit us, sources. These are the atoms the book narrates over; chapters link here instead of restating. Same-turn capture rule: see AGENTS.md → Knowledge & the book.

The adjudication seam

• double-billing — give a model two ways to encode one cost and it will eventually pay twice; the fix is an exclusivity rule in the prompt.
• prompt-gates-vs-engine-gates — arithmetic is engine law, judgment is the adjudicator's call; and what to do when a soft gate must become hard.
• free-no-costly-yes — the fairness ledger: rolls cost daylight, refusals are free, and the model's mistakes never bill the player.

Testing & determinism

• deterministic-dice — difficulty 0.0/1.0 are loaded dice that make a stochastic LLM game exactly assertable; plus the wrong-interpreter trap that made the suite silently run outside the venv.

Process & git

• trunk-based-won — the branching models, how each one kills you, and what DORA/Google/Driessen-recanting history actually shows.
• stacked-prs-strand-main — six correct bottom-up merges that shipped nothing: the stranded-base failure mode and the closing-PR rule it produced.

Ops & debugging

• stale-page-not-broken-server — the private-tab experiment that split "server broken" from "client cached a corpse", and the no-store + sha-in-footer fixes.
• windows-keeper-unelevated — schtasks beats the CIM API unelevated, pythonw can't print, and background healers must prove liveness where the human already looks.