Phase 3a (sweep): elite-bracket + drift sweep, strong partial positive

Wired two harness lever overrides — EliteInterruptThresholdOverride (live=19) and ThreatDriftBaseOverride (live=3) — into the day-loop in expedition_balance.go. Live runHarvestInterrupt / dailyThreatDrift are untouched; the harness re-buckets Standard↔Elite after the live call. TestExpeditionBalance_Phase3_GlobalLeverSweep walks a 3×3 grid (elite ∈ {17,19,23} × drift ∈ {1,3,5}) over the Phase 1 matrix at 200 trials/cell. -short skips. Elite-bracket threshold is the dominant lever for T1–T3. At e=23/d=1: T1 mean 24.0% (goblin_warrens 40.5%), T2 7.2% (sunken_temple 14.5%), T3 1.8%. Still well below target bands (T1 70-90%, T2 62-82%) — the lever moves the needle in the right direction but cannot land any tier on-band alone. T4/T5 fingerprint changed but didn't lift. At e=23 dragons_lair death drops 60% → 24% but starvation climbs to 75% — the fighter now survives elites long enough to run out of supplies. T4 cells shift the same way. Indicates a second lever is needed for the higher tiers (standard-fight survivability or supply margin), to be swept in Phase 3-B. Plan doc updated. Renumbered the trailing "per-zone outlier pass" to Phase 4 and "MAD / second-order" to Phase 5 so the test names align with phase numbers going forward. -short suite: same 2 pre-existing failures (TestAdv2Scenario_ZoneRunGoblinWarrens, TestMageSpellbookLineInRender). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-07-15 00:32:40 +00:00 · 2026-05-15 11:00:04 -07:00
parent 9a4a93e72e
commit 235122f2a1
3 changed files with 193 additions and 6 deletions
--- a/gogobee_expedition_difficulty.md
+++ b/gogobee_expedition_difficulty.md
@@ -201,7 +201,7 @@ before the dragon roll lands.
 Stat-block tuning of the bosses themselves is off-limits per the
 Phase 3 constraint (bestiary is shared across systems).

-### Phase 2 — global lever tuning
+### Phase 3 — global lever tuning

 Tune the knobs that affect every zone uniformly:

@@ -218,9 +218,42 @@ Tune the knobs that affect every zone uniformly:
 Goal: every tier centerline lands within ±5pp of target. Bands stay
 ±10pp.

-**Exit:** test asserts the ±10pp band per tier; commit.
+#### Phase 3-A — elite-bracket + drift sweep (shipped)

-### Phase 3 — per-zone outlier pass
+Wired two harness overrides
+(`EliteInterruptThresholdOverride`, `ThreatDriftBaseOverride`) and
+swept a 3×3 grid (elite ∈ {17, 19, 23} × drift ∈ {1, 3, 5}) over
+the Phase 1 matrix. Test: `TestExpeditionBalance_Phase3_GlobalLeverSweep`.
+
+**Strong partial positive.** The elite-bracket threshold is the
+dominant lever for T1–T3; threat drift is a small modulator on top.
+
+| (elite, drift) | T1 mean | T2 mean | T3 mean | T4 mean | T5 mean |
+|----------------|--------:|--------:|--------:|--------:|--------:|
+| 17, 1 (more elites) | 0.5% | 0.0% | 0.0% | 0.0% | 0.0% |
+| 19, 3 (live)        | 3.0% | 0.2% | 0.0% | 0.0% | 0.0% |
+| 23, 1 (rare elites) | **24.0%** | **7.2%** | **1.8%** | 0.0% | 0.0% |
+| 23, 3               | 16.5% | 4.0% | 2.0% | 0.0% | 0.0% |
+| 23, 5               | 11.8% | 3.2% | 0.2% | 0.0% | 0.0% |
+
+Best cell at e=23/d=1: goblin_warrens 40.5%, sunken_temple 14.5%.
+Still well below the T1 70-90% / T2 62-82% target bands — the lever
+moves the needle in the right direction but cannot land any tier
+on-band alone.
+
+**T4/T5 fingerprint changed but didn't lift.** At e=23 the dragons_lair
+death-fraction drops from ~60% → 24% but starvation climbs to 75% —
+the fighter now survives elites long enough to run out of supplies.
+T4 (underdark/feywild) likewise sees death drop and starvation rise.
+Indicates a second lever is needed for the higher tiers: either
+standard-fight survivability (surprise-nick floor / standard roster
+density) or supply margin.
+
+**Next:** Phase 3-B sweeps surprise-round nick and supply-burn knobs
+on top of e=23/d=1 to push T1-T3 toward target band and lift T4-T5
+off the floor.
+
+### Phase 4 — per-zone outlier pass

 After globals settle, the matrix will name outlier zones — tiers
 where one zone reads dramatically off-band while sibling zones are
@@ -238,7 +271,7 @@ not the monsters themselves.

 **Exit:** all per-zone cells within band; commit.

-### Phase 4 — optional MAD / second-order
+### Phase 5 — optional MAD / second-order

 If post-Phase-3 the bands hold but feel wrong subjectively
 (e.g. "median days too short", "no expedition ever hits Siege"), add a
--- a/internal/plugin/expedition_balance.go
+++ b/internal/plugin/expedition_balance.go
@@ -103,6 +103,24 @@ type expeditionBalanceProfile struct {
 	// TestExpeditionBalance_Phase2_LeverSweep.
 	RetreatThreatBumpOverride   int
 	SurpriseNickDivisorOverride int
+
+	// Phase 3 global-lever overrides. Zero means "use live":
+	//
+	//   EliteInterruptThresholdOverride — total (raw + tier + threat-mod)
+	//   at which the harvest-interrupt bracket promotes from Standard
+	//   to Elite. Live value is 19 (dnd_expedition_combat.go: bracket
+	//   table). Higher = fewer elites; lower = more.
+	//
+	//   ThreatDriftBaseOverride — daily threat-clock drift base before
+	//   DM-mood mod. Live value is 3 (dnd_expedition_threat.go:
+	//   dailyThreatDrift). Lower = slower AC/init creep + slower
+	//   bracket promotion via the +1/20-over-40 threat-mod path.
+	//
+	// Both wired into the harness day loop only; live callers go
+	// through the shipped constants. See
+	// TestExpeditionBalance_Phase3_GlobalLeverSweep.
+	EliteInterruptThresholdOverride int
+	ThreatDriftBaseOverride         int
 }

 // expeditionTrialResult is the outcome of one simulated expedition.
@@ -243,9 +261,15 @@ func (h *expeditionHarness) advanceExpeditionOneDay() expeditionTrialResult {

 	// 3. Daily threat drift — math is in dailyThreatDrift (pure);
 	// applyDailyThreatDrift's DB write is bypassed but the in-memory
-	// mutation is mirrored here.
+	// mutation is mirrored here. Phase 3 sweep can override the base
+	// drift constant via the harness profile.
 	if !h.exp.BossDefeated {
 		delta, _ := dailyThreatDrift(h.exp.DMMood)
+		if h.threatDriftBaseOverride > 0 {
+			// dailyThreatDrift returns base(3) + mood-mod. Swap the
+			// base while preserving the mood-mod component.
+			delta = h.threatDriftBaseOverride + (delta - 3)
+		}
 		h.exp.ThreatLevel += delta
 		if h.exp.ThreatLevel < 0 {
 			h.exp.ThreatLevel = 0
@@ -258,9 +282,21 @@ func (h *expeditionHarness) advanceExpeditionOneDay() expeditionTrialResult {

 	// 4. Daytime combat-interrupt rolls.
 	for i := 0; i < h.rollsPerDay; i++ {
-		kind, _ := resolveCombatInterrupt(
+		kind, total := resolveCombatInterrupt(
 			h.exp.ThreatLevel, int(zone.Tier), h.char.Class, h.exp.ZoneID, h.rng.d20,
 		)
+		// Phase 3 lever: re-bucket Standard ↔ Elite using the override
+		// threshold instead of the live 19+ cutoff. Patrol (≥22) and
+		// the Noise/None floor are unchanged.
+		if h.eliteInterruptThresholdOverride > 0 &&
+			(kind == InterruptStandard || kind == InterruptElite) &&
+			total < 22 {
+			if total >= h.eliteInterruptThresholdOverride {
+				kind = InterruptElite
+			} else if total >= 15 {
+				kind = InterruptStandard
+			}
+		}
 		switch kind {
 		case InterruptNone:
 			continue
@@ -520,6 +556,10 @@ type expeditionHarness struct {
 	// live behavior (retreatThreatBump=5, /5 surprise-nick divisor).
 	retreatThreatBumpOverride   int
 	surpriseNickDivisorOverride int
+	// Lever overrides for the Phase 3 sweep. Zero ⇒ live behavior
+	// (eliteInterruptThreshold=19, threatDriftBase=3).
+	eliteInterruptThresholdOverride int
+	threatDriftBaseOverride         int
 	// traceFight, if non-nil, is invoked once per fight inside
 	// runHarnessFight with a human-readable summary. Used by the
 	// Phase 2 lethality probe to spot whether the nick, the picked
@@ -570,6 +610,8 @@ func runExpeditionBalanceTrial(p expeditionBalanceProfile, seed uint64) expediti
 		rollsPerDay:                 rolls,
 		retreatThreatBumpOverride:   p.RetreatThreatBumpOverride,
 		surpriseNickDivisorOverride: p.SurpriseNickDivisorOverride,
+		eliteInterruptThresholdOverride: p.EliteInterruptThresholdOverride,
+		threatDriftBaseOverride:         p.ThreatDriftBaseOverride,
 	}
 	for {
 		res := h.advanceExpeditionOneDay()
--- a/internal/plugin/expedition_balance_test.go
+++ b/internal/plugin/expedition_balance_test.go
@@ -608,6 +608,118 @@ func TestExpeditionBalance_Phase2_LeverSweep(t *testing.T) {
 	}
 }

+// TestExpeditionBalance_Phase3_GlobalLeverSweep is the global-tuning
+// sweep the plan doc's "Phase 3 — global lever tuning" step calls for.
+// Phase 2c (roster gate) lifted T1 goblin_warrens off the floor (~3%)
+// but every other tier still reads 0% in the post-2c Phase 1 matrix.
+// The Phase 2 lever sweep proved the wounded-cascade knobs are inert
+// once the clamp is in place — deaths are now fresh-entry elite
+// one-shots and multi-day AC/init creep, not chained nicks.
+//
+// This sweep walks two of the global knobs called out in the plan-doc
+// "Phase 3 — global lever tuning" section:
+//
+//   eliteInterruptThreshold (live=19, total roll cutoff for Elite
+//   bracket) — directly controls how often fresh-entry elite fights
+//   trigger during the daytime harvest pipeline. Sweep {17, 19, 23}:
+//   17 = more elites (slope check below live), 19 = live baseline,
+//   23 = elites only when roll+mod ≥ 23 (rare even at T5).
+//
+//   threatDriftBase (live=3, daily threat clock drift before mood-mod)
+//   — slows the multi-day AC/init/supply-burn creep that compounds
+//   over a 14-day expedition. Sweep {1, 3, 5}: 1 = nearly flat, 3 =
+//   live, 5 = harsher.
+//
+// 3×3 = 9 combos × 10 zones × 200 trials/cell. Diagnostic-only — no
+// gates beyond the Phase 1 wiring sanity. -short skips.
+func TestExpeditionBalance_Phase3_GlobalLeverSweep(t *testing.T) {
+	if testing.Short() {
+		t.Skip("phase 3 global-lever sweep is heavy; -short skips it")
+	}
+
+	const trialsPerCell = 200
+	const baseSeed uint64 = 0x9101E5
+
+	eliteThresholds := []int{17, 19, 23}
+	driftBases := []int{1, 3, 5}
+
+	t.Logf("phase3 global-lever sweep — %d zones × %d elite-thresholds × %d drift-bases × %d trials, Fighter @ tier centerline (rolls=4)",
+		len(zoneOrder), len(eliteThresholds), len(driftBases), trialsPerCell)
+
+	type tierStat struct {
+		cells int
+		sumC  float64
+		lo    float64
+		hi    float64
+	}
+
+	for _, elite := range eliteThresholds {
+		for _, drift := range driftBases {
+			tierStats := map[ZoneTier]*tierStat{}
+			t.Logf("─── eliteInterruptThreshold=%d  threatDriftBase=%d ───", elite, drift)
+			for i, id := range zoneOrder {
+				zone, ok := getZone(id)
+				if !ok {
+					t.Fatalf("zoneOrder[%d]=%q not in registry", i, id)
+				}
+				level, ok := phase1TierCenterline[zone.Tier]
+				if !ok {
+					t.Fatalf("zone %q has tier %d with no phase1 centerline mapping", id, zone.Tier)
+				}
+				profile := expeditionBalanceProfile{
+					ZoneID:                          id,
+					Class:                           ClassFighter,
+					Level:                           level,
+					Supplies:                        makeSupplies(zone.Tier, SupplyPurchase{StandardPacks: 3}),
+					CampType:                        CampTypeStandard,
+					EliteInterruptThresholdOverride: elite,
+					ThreatDriftBaseOverride:         drift,
+				}
+				seed := baseSeed + uint64(i)*1_000_003 + uint64(elite)*101 + uint64(drift)*7919
+				r := runExpeditionBalanceCell(profile, trialsPerCell, seed)
+
+				c := r.CompletionRate() * 100
+				t.Logf("CELL  e=%-2d d=%-1d  %-18s T%d  L%-2d  comp=%5.1f%% death=%5.1f%% starve=%5.1f%% med_days=%2d med_threat=%3d encs=%4.1f hp_left=%5.1f%%",
+					elite, drift, zone.ID, zone.Tier, r.Profile.Level,
+					c,
+					r.DeathRate()*100,
+					float64(r.StarvedOuts)/float64(r.Trials)*100,
+					r.MedianDays, r.MedianThreatEnd,
+					r.AvgEncounters, r.AvgHPRemainingPct*100,
+				)
+
+				ts, ok := tierStats[zone.Tier]
+				if !ok {
+					ts = &tierStat{lo: math.Inf(1), hi: math.Inf(-1)}
+					tierStats[zone.Tier] = ts
+				}
+				ts.cells++
+				ts.sumC += c
+				if c < ts.lo {
+					ts.lo = c
+				}
+				if c > ts.hi {
+					ts.hi = c
+				}
+			}
+
+			tiers := []ZoneTier{
+				ZoneTierBeginner, ZoneTierApprentice, ZoneTierJourneyman,
+				ZoneTierVeteran, ZoneTierLegendary,
+			}
+			for _, tier := range tiers {
+				ts := tierStats[tier]
+				if ts == nil || ts.cells == 0 {
+					continue
+				}
+				mean := ts.sumC / float64(ts.cells)
+				t.Logf("TIER  e=%-2d d=%-1d  T%d  n=%d  mean_comp=%5.1f%%  spread=%5.1f pp",
+					elite, drift, tier, ts.cells, mean, ts.hi-ts.lo)
+			}
+		}
+	}
+}
+
 // joinZones is a tiny helper kept local to the test file so the
 // per-tier log line reads in one logical chunk without pulling in
 // strings.Join's import for production code.