Deliverable · training repo · 3 skills + orchestrator

Judgment gym

01 · The stations

Three machines, one circuit

• ~25 min · dailyJudgment drill
  Five stations: a Fermi estimate, a stock-or-flow diagnosis, a third-order-consequences case, a reverse-the-memo (reconstruct the decision from its residue), and a Go problem — because reading a whole board is the same muscle. Every case is pulled fresh from the web and graded against what actually happened.
• ~12 min · high-rep companionAnalyst reps
  Four blocks that close the gaps the drill exposes but can't fix: ratio flashcards (business primitives on recall), read-a-real-table (pull the derivatives out of a filing), an ACH rep (hold competing hypotheses instead of locking onto one), and a brief-compliance check (answer the question that was asked).
• post-session · diagnosticJudgment coach
  The half that makes the misses pay rent. Classifies each error into a named failure mode, traces it to the reasoning move that produced it, rebuilds the attack angle that would have worked, tracks which errors recur across sessions, and sources real open-source cases where the same error shows up in the wild.

02 · A day at the gym

Drill, diagnose, rep

• 00:00Drill — five stations, answers committed before grading. No partial credit for vibes.
• 00:25Coach — every miss gets a name, a cause, and the move that would have caught it.
• 00:35Reps — the specific weakness the coach found becomes tomorrow's flashcards.
• 00:47Log — calibration curve updated. The streak is not the point; the curve is.

03 · House rules

What keeps it honest

• Real cases only — every station is built from publicly verifiable sources, never invented scenarios that flatter the drill.
• Graded against the source, not against feeling right. The gap between confidence and accuracy is the product.
• The log persists. A judgment gym without a calibration log is a mood.
• Misses are the curriculum: the coach's failure-mode taxonomy decides what gets trained next, not a syllabus.