eisterman/dojo_cate
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Add failed-Exam loss switch; keep -3 and accept the length (6.11, D50)

Browse Source 
RulesConfig.exam_failure_loss (bots aware) with the sim axis --exam-loss.
Losing 1 or 0 stripes saves 3-5 minutes but weakens comebacks, so the
rules keep -3. Default games are unchanged.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
This commit is contained in:
Federico Pasqua
2026-09-27 17:51:13 +02:00
parent 542ac1b15f
commit 52d8af620b
11 changed files with 116 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files
agents/plans/2026-09-25-dojo-board-and-sim.md
+43 -2
View File
@@ -66,6 +66,7 @@ Deliverables:
| D47 | **Kiai stays +2**: +2 and +1 are indistinguishable on every target (6.6); +2 decides more fights (5% vs 3% of Kumite in casual play). `kiai_bonus = 1` stays a simulator switch; RULES.md drops the "+1 to be tested" note | Developer, 6.6 results. |
| D48 | **6+ players start on 3 (Yellow)**; any player count may start on 3 as a "shorter game" option (saves 15–20% of the time, 6.7). 8 players stays long (107 / 142 min mixed / casual from Yellow) and is accepted: it will rarely be played. Balancing simulations stay at start 0 at every count (developer); `--start 3` reproduces the rules' 6+ setup or the shorter game | Developer, 6.7 results. |
| D49 | **Chance square parked** (6.4.1 and 6.4.3): Fatica on crossing (D36) on board draft 2 (D42) already makes squares matter; 6.4 closed | Developer. |
| D50 | **Length accepted as is; a failed Exam still costs 3 stripes.** L1 (Brown floor against Fatica) rejected; L2 (failed Exam −1 / 0) saves only 3–5 min on the mean and weakens comebacks (6.11.4). `exam_failure_loss` stays a simulator switch. At 4 players mixed and balanced are inside the target; casual (91 / 122) and all-aggressive (90 / 127) stay slightly outside, within the "1–2 h" promise; the shorter-game start on Yellow (D48) is the players' lever | Developer, 6.11 results. |
## Phases
@@ -176,7 +177,7 @@ Deliverables:
- One change at a time. When a change is adopted, re-run the 6.1 baseline so later steps start from the current rules, and log the decision (D-number) here and in `RULES.md`.
- Results go in this file as a table per step (command line, scenario, key measures).
**Next action:** 6.11 length tuning, then 6.8. 6.2 is suspended (D43), 6.9 skipped (D44); 6.4.1 (chance square) stays optional. Order of the steps: D41. Everything up to D49 is committed; commit only when the developer asks.
**Next action:** 6.8 consolidation, when the developer says to continue. 6.2 is suspended (D43), 6.9 skipped (D44); 6.4.1 (chance square) stays optional. Order of the steps: D41. Everything up to D50 is committed; commit only when the developer asks.
**Targets** (confirmed by the developer as a starting point, D32):
@@ -320,7 +321,47 @@ Deliverables:
- [x] 6.10.3 `movement.walks`: walks that exclude nothing but the start are cached per (board, start, steps, Dojo open) with `functools.cache` (same pattern as `bots.dojo_distance`); callers get a fresh list. ~3,200 entries per process, ~80 MB peak.
- [x] 6.10.4 Parallel scaling: 8 processes running at once each take ~1.6x their solo time on this machine (8 real cores, clock and cache sharing), so ~5x on 8 workers is the hardware limit; no pool change. 16 workers (hyperthreads) add ~12%.
- Result: 100 games on one process 8–12 s → 1.4–1.7 s; mixed + balanced + casual at 1,000 games on 16 workers 57 s → 8 s. All 210 fingerprinted games identical; the v0.8 baseline run prints the same metrics.
- [ ] 6.11 Length tuning (D39): at 4 players, bring every lineup to a mean of 60–90 minutes with p90 ≤ 110 (v0.8: mixed 72 / 95 and balanced 68 / 92 inside; casual 91 / 122 and all-aggressive 90 / 127 outside). Find where the time goes (climb to Brown vs Exam phase), bring candidate levers to the developer, simulate the chosen ones on at least two lineups.
- [x] 6.11 Length tuning (D39; closed by D50): at 4 players, bring every lineup to a mean of 60–90 minutes with p90 ≤ 110 (v0.8: mixed 72 / 95 and balanced 68 / 92 inside; casual 91 / 122 and all-aggressive 90 / 127 outside). Find where the time goes (climb to Brown vs Exam phase), bring candidate levers to the developer, simulate the chosen ones on at least two lineups.
- [x] 6.11.1 Where the time goes (scratch `phases.py`, 400 games per lineup; "endgame" = from the first Brown on the table to the win):
| Lineup | Climb / endgame (min) | Winner's stripes gained (15 needed) | Drops out of Brown per game | Stripes lost per game: square / Kumite / Exam |
|---|---|---|---|---|
| mixed | 49 / 22 (31%) | 23.3 | 2.2 | 3.9 / 27.8 / 4.3 |
| balanced | 48 / 21 (30%) | 22.6 | 2.2 | 2.4 / 26.6 / 4.3 |
| casual | 56 / 36 (39%) | 24.7 | 4.0 | 11.2 / 23.1 / 4.4 |
| aggressive | 60 / 32 (35%) | 28.6 | 2.7 | 12.4 / 42.7 / 4.4 |
Casual and aggressive are longer in both phases, most of all in the endgame (+11 to +15 min): casual players fall out of Brown twice as often (Fatica tolls, 11 stripes per game), aggressive ones trade stripes in Kumite (43 lost per game).
- [x] 6.11.2 Existing switches (`--start`, `--exam`), 1,000 games, 4 players, minutes mean / p90:
| Setting | mixed | balanced | casual | aggressive |
|---|---|---|---|---|
| v0.8 (start 0) | 72 / 95 | 68 / 92 | 91 / 122 | 90 / 127 |
| start 1 | 68 / 92 | 65 / 87 | 88 / 122 | 88 / 124 |
| start 2 | 63 / 88 | 59 / 83 | 81 / 112 | 82 / 122 |
| start 3 | 57 / 80 | — | 77 / 110 | 74 / 111 |
| Exam defender not Black (`none`) | 60 / 77 | — | 79 / 103 | 74 / 97 |
The start position shifts every lineup by about the same amount, so it cannot close the ~23-minute gap between balanced and casual: start 2 puts balanced under 60 while aggressive stays at p90 122. The Black-belt defender off pushes the Exam pass rate to 67–80% (target 35–55%) and the rivals at the end to 0.27–0.49: rejected again (as in D31). Other measures barely move with the start position.
- [x] 6.11.3 Candidate levers brought to the developer: L1 Fatica cannot drop a player out of Brown (**rejected by the developer**); L2 a failed challenger loses 0 or 1 stripe instead of 3 (to test); L3 accept the current length (developer: "satisfied enough").
- [x] 6.11.4 L2 switch: `RulesConfig.exam_failure_loss` (default 3; engine and bot `_exam_score`), sim axis `Scenario.exam_loss` / `--exam-loss` (scenario names gain `fail-N`). Fingerprinted games identical at the default. Run: `uv run dojo sim --games 1000 --seed 0 --exam-loss 3 1 0 --lineup balanced,farmer,aggressive balanced casual aggressive`.
| Lineup | Failed Exam loses | Minutes mean / p90 | Exam challengers per game | Lead changes | Halfway leader wins | Worst player out | Rivals within 1 belt |
|---|---|---|---|---|---|---|---|
| mixed | 3 | 72 / 95 | 1.68 | 6.1 | 40% | 13% | 0.84 |
| mixed | 1 | 68 / 88 | 1.48 | 5.0 | 44% | 15% | 0.94 |
| mixed | 0 | 67 / 86 | 1.45 | 4.6 | 44% | 16% | 0.87 |
| balanced | 3 | 68 / 92 | 1.69 | 6.0 | 39% | 10% | 0.81 |
| balanced | 1 | 65 / 84 | 1.50 | 5.0 | 43% | 12% | 0.97 |
| balanced | 0 | 65 / 83 | 1.46 | 4.5 | 43% | 13% | 0.90 |
| casual | 3 | 91 / 122 | 1.74 | 5.7 | 39% | 12% | 1.42 |
| casual | 1 | 87 / 116 | 1.61 | 4.9 | 42% | 13% | 1.47 |
| casual | 0 | 86 / 114 | 1.55 | 4.6 | 41% | 14% | 1.40 |
| aggressive | 3 | 90 / 127 | 1.58 | 7.3 | 41% | 21% | 0.50 |
| aggressive | 1 | 85 / 115 | 1.40 | 6.2 | 45% | 23% | 0.59 |
| aggressive | 0 | 83 / 111 | 1.32 | 5.8 | 49% | 25% | 0.54 |
Exam pass rate unchanged (40–44%). Reading: −1 saves 3–5 minutes on the mean and 6–12 on the p90; 0 adds only 1–2 more. The price is the comeback: the first Brown keeps retrying instead of dropping back (fewer different challengers), lead changes fall by ~1 per game, the halfway leader wins 3–8 points more, the worst player is out a little longer. No setting brings casual under p90 110.
- [ ] 6.9 Catch-up candidate A — **skipped (D44)**. Was: `RulesConfig` switch, Allenamento +2 for a player 2+ belts behind the leader (bot aware of it); compare on the contention, lead and length measures and on the style win rates.
- [ ] 6.8 Consolidation: the final configuration at 3 / 4 / 6 / 8 players (start 0 as for balancing, plus start 3 at 6 and 8 as in the rules, D48) with all lineups except `random` (D45), 2,000 games each; every target checked; `RULES.md` updated to the next version with the decisions; open items moved to Phase 7.