Decision & game-theory trainers
GTO Drill
28 of 49 answers solver-verified, up from 22 of 45
- Simulation
- UI/UX
Evidence
- Data
- Authored scenario library, audited against a heads-up postflop solver · 49 scenarios (45 authored + 4 derived)
- Stage
- Prototype
- Stack
- JavaScript, Node audit scripts, Poker solver
Evaluation
What was measured, against what, and how.
| Result | 28 of 49 scenarios |
|---|---|
| 95% interval | 43% to 70% (Wilson score interval on 28 of 49) |
| Baseline | First audit, 2026-06-30: 22 of 45 verified (22 unaudited, 1 deferred) |
| Sample size | 49 scenarios (45 authored + 4 derived variants); 20 unaudited, 1 deferred |
| Timeframe | 2026-06-30 to 2026-07-03 (git history of the audit file) |
| Method | hero-action-audit: each scenario's authored GTO action compared with the solver's strategy at the decision node |
| Source | data/answer-audit-status.json in the gto-drill repo (private); baseline from its first commit |
| Confidence | Verified |
Limits
- All 14 preflop scenarios are unaudited: the heads-up postflop solver doesn't cover them
- Eight of the agreeing answers agree only after being re-authored to match the solver
- 34 of 34 pipeline-validation passes are a gate (invariants hold), not an accuracy figure
Outcome: Partly. Goal: every scenario's answer checked against a solver. 28 of 49 are, up from 22 of 45 at the first audit; 20 can't be audited yet (all 14 preflop spots are outside the heads-up postflop solver).
What was built
The deliverables this project covers.
- GTO Drill app Prototype
- GTO poker trainer app Shipped · Live
- Poker app design system and view fixes design Prototype
How I used AI
What the tools did, and what was checked.
Built with: Jett with Claude Code sessions; AutoBuilder run series (architecture v1.3 to v1.10.1); Jett in claude.ai and Claude Desktop chats (incl. claude.ai Design).
Note to be written: tool and model, what the assistant produced, what was rejected or rewritten, and which check caught a problem.