One place for what is actually measured about the release encoder
(CompactHoldemEncoder + 20 river percentile buckets, imperfect recall).
Machine-readable: results/validation/abstraction_error_summary.json.
This supersedes any older wording elsewhere (in particular the stale
"flop/turn error not measured exactly" row that used to be in
docs/release_status.md — both streets ARE measured).
| street | status | release-encoder error (exact subgames) | consequence |
|---|---|---|---|
| preflop | card mapping lossless (169 classes); betting context abstract | n/a — instability is MCCFR mixing, not card abstraction (docs/preflop_disagreement.md) |
first-action keys largely gated by seed disagreement |
| flop | measured exactly, 7 reduced subgames (fixed runout samples, reduced menu, pot 5 bb) | exploitability median 3.7 bb, max 5.1 bb; EV error ≤ 0.13 bb | street capped at LOW confidence |
| turn | measured exactly, 7 reduced subgames | exploitability median 5.1 bb, max 8.7 bb; EV error ≤ 0.66 bb | street capped at LOW confidence |
| river | measured exactly, six 52-card subgames | pct20: exploitability median 0.42 bb, max 0.57 bb (compact v1 was 3.5–9.6 bb: 4–16x worse) | no street cap |
Raw-key references solve the same subgames to ≤ 0.03 bb — the lift is abstraction, not solver error.
- Exact full-game exploitability: not computed — infeasible
(
results/validation/best_response_feasibility.json). - New in this project: a restricted-LBR lower bound (validated on exact
games, Spearman 0.986 vs true exploitability): the 200k/300k/1M
strategies are all ≥ ~100 bb/100 exploitable by a range-tracking
best responder restricted to the abstract menu (
lbr_holdem.json; candidates not separable at 10k hands). This is an abstraction-level floor: per-street errors are bounded but compound, and imperfect recall denies any equilibrium guarantee.
- Training uncertainty (seed disagreement, visits, recent movement, value uncertainty) — improves with iterations; measured to improve through 1M–2M.
- Abstraction risk (this table + collision dispersion) — does NOT improve with training. Street caps and the collision signal survive into confidence v2 unchanged; extremely stable training must never erase them.