The Phase 1 per-class-mean summary was hiding the truth — most cells are
floor/ceiling-saturated (L10+ pinned at 1.0, L1-4 caster cells at high
tier pinned at 0.0), so means barely budge when you tune passives. Added
a per-(level, tier) cross-class spread diagnostic to the matrix log,
then tuned with the levers from doc §6 in priority order:
1. Class passives (dnd_passives.go) — caster trailers (Bard, Mage,
Warlock, Sorcerer) gained level + casting-stat-scaled FlatDmgStart
bursts so the L1-4 chassis isn't a quarterstaff + one weak spell
against a T2-T3 monster; small DamageBonus riders (Mage/Bard/
Sorcerer/Rogue +5%, Warlock 10→12%) and +1 attack for Bard/Warlock
close the steady-DPS gap. Added clampNonNeg so ability-mod-scaled
additions never go negative on sub-10-stat sheets.
2. Subclass L5 tiers (dnd_subclass_combat.go) — the three Sorcerer L5
picks (Wild/Storm/Draconic) and Warlock Great Old One were defense-
only or near-inert pre-tune; each gained a small bite (DamageBonus
+0.10, or a FlatDmgStart burst for Storm) so the L5 chassis can press
through a T4 monster.
Parity band locked in TestClassBalance_Phase1_FullMatrix: cross-class
spread ≤ 35pp on the in-tier diagonal — (level, tier) cells where the
level is appropriate for the tier (T1: L1-4, T2: L3-7, T3: L5-10, T4:
L7-15, T5: L10-20). Off-tier cells (L1 mage at T3 dungeon etc.) are
still logged but not asserted: those are level-vs-tier mismatches and
casters at L1-4 can't muscle through a T3 monster on a single L1-slot
spell the way martials muscle through with weapon dice. Worst in-tier
cell after tuning: ~26pp at L3/T2. The 35pp band gives ~9pp Monte-Carlo
headroom over the worst signal at 200 trials/cell.
TestApplyClassPassives expectations updated to match the new passives.
Phase 0 spike still green, full plugin suite (-short) clean.
Generalizes the Phase 0 spike harness to the full build matrix the
class-balance doc plans for. No tuning yet — just measurement.
- classBalanceProfile gains Subclass; buildHarnessCharacter sets it on
the synthetic DnDCharacter; buildHarnessPlayer now calls
applySubclassPassives after class+race passives, matching live order
(combat_bridge.go, combat_session_build.go). Subclass="" is a no-op,
so L1–L4 pre-unlock rows are unaffected.
- buildPhase1Profiles yields 190 rows: 10 classes × 4 pre-subclass
levels (L1–L4) + 10 classes × 3 subclasses × 5 post-unlock checkpoints
(L5/7/10/15/20). Order is registry order so output reads like the
design doc / !class help.
- TestClassBalance_Phase1_FullMatrix runs the matrix at 200 trials/cell
(~5.5s) and logs every cell plus a per-class tier-mean summary with
min/max range. Only harness-broken pathologies fail the test (0% at
T1 anywhere, or 100% at T5 for an L1 build); per-tier parity bands
land in Phase 2 once we have data to calibrate the tolerance.
Phase-2 baseline from this run: at T4 the cross-class spread of mean
win rate runs Bard 0.62 → Fighter 0.80 (~18pp); at T5 0.48 → 0.64
(~16pp); casters trail martials at the post-unlock tier (T3) by ~20pp.
Phase 0 test (TestClassBalance_Phase0_FighterVsMage) still green with
identical numbers — the additional applySubclassPassives call is a
no-op for Subclass=="".
Per gogobee_class_balance.md §5 Phase 0: harness skeleton, equipment
and spell-selection policies, Fighter-vs-Mage plausibility run before
Phase 1 generalizes to the full 10 × 30 matrix.
Bypassed in Phase 0 (per doc §2): DB-touching layers (magic items,
armed abilities, pending-cast persistence), subclass passives (none
below L5), and race passives beyond Human +1-all. Everything else
flows through the production combat path.
Initial numbers (400 trials/cell, dungeon T1..T5):
fighter L1 .998 .805 .165 .020 .000
fighter L3 1.000 .998 .795 .235 .035
mage L1 .880 .190 .003 .000 .000
mage L3 1.000 .950 .158 .003 .000
Both classes win at T1 (spell policy is firing — Magic Missile lands
each Mage fight); both collapse by their off-tier — monster scaling
works. The Mage's L1 T2 gap (-60pp vs Fighter) is real data, not a
broken harness. Phase 1 picks up the full matrix from here.
