gogobee/sim_results/j2b_findings.md

J2b — Post-J2a all-classes baseline (n=100)

Date: 2026-05-17 Corpus: sim_results/baseline_j2a_v2_all10.jsonl (15,000 rows = 10 classes × 3 levels × 5 zones × n=100). Per-cell tabular summary in sim_results/j2b_all10_summary.txt. Replaces: sim_results/baseline_j1_all10.jsonl as the post-sim-fix canonical corpus. The J1 numbers measured a strawman where casters never cast (see j2_findings.md).

L12 leaderboard (mean %clr across all 5 zones)

Class	J1 baseline	J2b	Δ	J1 boss%	J2b boss%	Per-zone J2b %clr (T1 / T2 / T3 / T4 / T5)
fighter	79.6	80.0	+0.4	100.0	100.0	100 / 100 / 100 / 100 / 0
ranger	78.6	80.0	+1.4	100.0	100.0	100 / 100 / 100 / 100 / 0
paladin	72.2	78.4	+6.2	99.6	99.2	100 / 100 / 96 / 96 / 0
rogue	70.6	76.8	+6.2	99.0	99.0	100 / 100 / 90 / 94 / 0
druid	38.6	61.6	+23.0	88.4	88.0	100 / 100 / 74 / 34 / 0
mage	20.0	53.4	+33.4	77.2	77.6	100 / 99 / 49 / 19 / 0
sorcerer	20.0	50.6	+30.6	70.4	70.0	100 / 98 / 47 / 8 / 0
warlock	21.4	48.2	+26.8	82.2	81.6	100 / 100 / 34 / 7 / 0
bard	21.6	40.4	+18.8	82.0	84.4	100 / 100 / 2 / 0 / 0
cleric	18.6	39.0	+20.4	61.6	62.4	96 / 90 / 9 / 0 / 0

Headline

1. No martial regression. Fighter/Ranger/Paladin/Rogue all moved within +0.4 to +6.2pp, indistinguishable from sweep variance. The simMartialFirstClass gate added after the v1 sweep saved Ranger (which had regressed -35.8pp when the picker burned L3 slots on lightning_arrow instead of letting the weapon-attack + Hunter's Mark + Extra Attack chassis swing).
2. Caster cliff is dissolved. The pre-fix 19–22% cluster spread out to 39–62%. The picker alone — no balance changes — converted the trailers to mid-band classes.
3. Druid is now top-of-band-2 at 61.6%. Wild Resilience's ThornLash on weapon swings plus the new cast path stacks cleanly. Druid is still defensive-first by design; the cast layer turns boss rooms from "swing-and-bleed" into "burn-and-finish".
4. T5 dragons_lair still walls everyone (0% clear across all 10 classes). That's the J3 problem; J2 wasn't on the hook for it, and the sim-artifact half of J3's hypothesis 1 (autoResolveCombat never casting/consuming) is now closed by J2a.

What's still trailing — bard and cleric

Bard L12 manor 2%, cleric L12 manor 9% — both reach the boss (~50–84%) and die there. Not a sim artifact; this is the spell pool. The v1 picker only scores damage-effect spells, and these two classes have thin damage rosters in defaultKnownSpells:

• Bard damage options: vicious_mockery (cantrip), shatter L2. Everything else in the bard kit is control / heal / buff, which the picker skips.
• Cleric damage options: sacred_flame (cantrip), guiding_bolt L1, spiritual_weapon L2, spirit_guardians L3, flame_strike L5. The non-damage options (bless, healing_word, mass_cure_wounds, etc.) are skipped too.

Both classes' identity spells are control (hold_person, hypnotic_pattern, command) and heal (cure_wounds, healing_word, mass_cure_wounds). The picker treats those as zero-value.

Note on cleric underdark: 87% of L12 underdark runs end as tpk with stop ended — that's expedition-side starvation, not boss-fight TPK (only 13% reach boss/elite). Cleric's exploration loop is brittle for unrelated reasons (low STR/CON build, supply burn). Out of J2 scope.

J2 DoD check

Reframed DoD (gogobee_harvest_charges_plan.md §6.J2, updated 2026-05-17):

• J2a — ≥80% of boss fights produce a cast or consumable. MET (100% cast rate in the verification sweep, j2a_findings.md).
• J2b — caster L12 manor mean clear ≥ druid's pre-J2a 39%. Mage 49 ✓, sorcerer 47 ✓, warlock 34 ✗ (close), bard 2 ✗, cleric 9 ✗. 3 of 5 pass.
• No martial leader regresses >10pp. MET (all +0.4 to +6.2pp).

Two trailers still under the floor (bard, cleric). They're now bottlenecked on spell-pool richness for the v1 damage-only picker, not on the sim itself. The cleanest next move:

J2c — tried, reverted (2026-05-17)

Implemented two widening levers, ran two n=100 sweeps, and reverted both:

1. Heal-spell preempt (cure_wounds / healing_word / mass_cure_wounds preferred over heal consumable at low HP). Regressed druid -10pp: slot heals at L12 heal ~10HP vs 40HP from a tier-4 Spirit Tonic, and the picker burned slots druid needed for damage spells.
2. Control-spell scoring at 22 (denied-enemy-turn proxy). Regressed warlock -6.6pp — vampiric_touch L3 (10.5 expected over multiple rounds + heal) was outscored by hypnotic_pattern L3 (single denied turn). Dropping the score to 5 brought net deltas back to ±2pp but didn't lift bard or cleric, because the level-first sort prevents low-score control spells from ever winning over a higher-level damage option (which the trailer classes don't have at higher slots).

Corpora retained for reference: sim_results/baseline_j2c_all10.jsonl (heal-preempt + control@22) and sim_results/baseline_j2c_v2_all10.jsonl (control@5 only). The picker reverted to the J2b shape (damage-effect only, level-first sort, no heal-spell preempt).

Real lever for bard/cleric trailing. Their defaultKnownSpells damage rosters cap at L2 (bard: shatter; cleric: spiritual_weapon's 1d8). The picker can't pick higher-damage spells they don't have. Lifting these classes is therefore a prod-level change to defaultKnownSpells (add a damage option at L3+ for bard, a damage option at L4+ for cleric) — not a sim-picker change. Out of scope for J2; revisit alongside class identity / Open5e integration work.

(original) Recommended J2c (now superseded by the section above)

Widen the picker to score control + heal spells.

• EffectControl (e.g. hold_person, hypnotic_pattern, hideous_laughter, command) should score as the enemy damage it prevents over its expected effect duration — i.e. each prevented enemy turn ≈ one round's worth of incoming damage. For a stunner that lands and lasts 1 round on a 30-dmg/round boss, that's a 30-equivalent. Better than most damage spells.
• EffectSpellHeal should compete with the consume-on-low-HP branch — if the cleric has mass_cure_wounds ready, it should fire at the low-HP threshold instead of a Spirit Tonic.
• EffectBuffSelf (e.g. bless, shield_of_faith) lifts to-hit chance / AC. Score in opening round of a fight only; subsequent rounds it's a waste.

This is one focused change in simPickSpell + simPickCombatAction. Expected impact: bard pulls into the 30–40% manor band (control spells are the bard identity); cleric pulls similarly. Martial leaders unchanged. Whether it's needed depends on operator: J2's reframed DoD allows shipping J2 with bard/cleric trailing if the prod target is "five classes mid-band, two below". Per feedback_difficulty_target, the explicit ask is to lift trailers — that argues for landing J2c.

Files touched this phase

• internal/plugin/expedition_sim.go:
  • SimCombatSummary.Events, simIncludeTrace + SetSimIncludeTrace (J2 diagnostic).
  • simPickCombatAction, simPickSpell, simMartialFirstClass, simHealHPThresholdPct (J2a picker + post-baseline fix).
  • BuildCharacter now seeds ensureSpellsForCharacter.
  • autoResolveCombat dispatches via the picker.
• cmd/expedition-sim/main.go: -trace flag.
• Docs: sim_results/j2_findings.md, sim_results/j2a_findings.md, sim_results/j2b_findings.md, sim_results/j2b_all10_summary.txt, plan §6.J2.

No prod combat / class / spell paths changed. Everything is in the sim harness.