68b2122e13
autoResolveCombat now calls maybeShortRest after a won fight (HP < 60% or any slots used, charges > 0), mirroring what a competent prod player does between rooms. Re-baseline at n=100 across the full 10×3×5 matrix written to baseline_h5sim_all10.jsonl. Headline (vs baseline_j2a_v2_all10.jsonl, full delta in h5sim_findings.md): mage L12 manor +8pp, cleric L12 manor +7pp, but bard L12 manor stays at 1pp and bard/cleric L12 underdark still 0. J2c (defaultKnownSpells damage option at L3+ for bard, L4+ for cleric) is still the lever for the trailers. Four borderline -9pp regressions on n=100 cells; flagged for an n=300 re-run before declaring a sim defect.
2.5 KiB
H5-sim re-baseline — does prod short-rest close the caster gap?
Corpus: baseline_h5sim_all10.jsonl (15k rows, n=100). Summary table: h5sim_all10_summary.txt. Baseline for comparison: baseline_j2a_v2_all10.jsonl (J2b).
TL;DR
No. Wiring short-rest into the sim between rooms gives mage L12 manor a real lift (+8pp clear rate) but does not move the bard/cleric trailers in the cells that matter:
| class | lvl | zone | J2b %clr | H5sim %clr | Δ |
|---|---|---|---|---|---|
| bard | 12 | manor_blackspire | 2 | 1 | −1 |
| cleric | 12 | manor_blackspire | 9 | 16 | +7 |
| bard | 12 | underdark | 0 | 0 | 0 |
| cleric | 12 | underdark | 0 | 0 | 0 |
| mage | 12 | manor_blackspire | 49 | 57 | +8 |
J2c (defaultKnownSpells widening — bard damage option at L3+, cleric at L4+) is still the lever. The hypothesis that H5 alone would close it is falsified.
Notable regressions (n=100, ~±10pp 95% CI — borderline)
| class | lvl | zone | J2b %clr | H5sim %clr | Δ |
|---|---|---|---|---|---|
| warlock | 12 | manor_blackspire | 34 | 25 | −9 |
| sorcerer | 12 | manor_blackspire | 47 | 38 | −9 |
| cleric | 3 | goblin_warrens | 15 | 6 | −9 |
| ranger | 3 | goblin_warrens | 64 | 54 | −10 |
n=100 standard error on a 50/50 cell is ~5pp, so these are ~2σ. Could be real (the picker prefers high-tier slots and short-rest refreshes low-tier first, so a mid-zone short rest may flush the high-tier-slot budget by encouraging more L2/L3 spends), or noise. Re-run at n=300 to disambiguate before declaring this a sim defect.
Cells that lifted
Mage L12 manor +8, cleric L12 manor +7, fighter L3 goblin +10, druid L7 forest +8, sorcerer L7 forest +4, mage L7 goblin +3. Pattern: classes that already had viable damage rosters benefited; classes whose damage rosters are thin (bard/cleric) did not.
Verdict
- Ship H5+sim short-rest as-is — the prod fix (full-HP gate loosened) is correct independently and the sim now reflects what a competent player does.
- J2c (defaultKnownSpells widening) is still required to lift bard/cleric trailers.
- T5 dragons_lair untouched (0% across all 10 classes) — J3 territory.
- Re-baseline at n=300 on the four borderline regressions before tuning the picker.