From a1ede53fa9cfe6251d33cadde8cd6dc5c555f01e Mon Sep 17 00:00:00 2001
From: prosolis <5590409+prosolis@users.noreply.github.com>
Date: Sun, 26 Apr 2026 19:12:22 -0700
Subject: [PATCH] Equalize co-op gift magnitudes to kill 'always leave'
 dominant strategy
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Old values (open ±7, leave ±4) had matching EV but asymmetric variance;
risk-averse parties had a strict preference to always leave gifts, which
collapsed the dilemma. New values (all ±6) keep EV=0 for both strategies
AND equalize variance. Choice now depends entirely on inferring sender
intent.

Adds TestCoopGiftVarianceSymmetric — fails loud if anyone reintroduces
asymmetric magnitudes. EV test was already in place but only checked the
mean; missed this on first pass.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 README.md                             |  6 +++---
 internal/plugin/coop_dungeon_gifts.go | 24 ++++++++++++------------
 internal/plugin/coop_dungeon_test.go  | 25 +++++++++++++++++++++++++
 3 files changed, 40 insertions(+), 15 deletions(-)

diff --git a/README.md b/README.md
index 4f3fe19..94a4b11 100644
--- a/README.md
+++ b/README.md
@@ -626,12 +626,12 @@ Multi-day party runs separate from solo. 2–4 players, 2–7 days depending on
 
 **Spectator betting.** Parimutuel pool, 10% rake. Bets can be placed or increased any time during the run; positions are locked in (success or failure can't be switched). Party members can bet on their own run — sandbagging while heavily bet against is exploitative and visible. Odds line shows estimated success% based on party levels, pets, neutral funding assumption, and a hidden TwinBee Helpfulness Rating (rolling last 30 floor events: how often his recommendation was right). Helpfulness shifts the line ±20% but is never shown directly — players notice the line moves after events resolve.
 
-**Gifts.** Anyone *not* in the run can spend one harvest action to send a Care Basket or a Mimic. The party never sees which type. Vote `open` or `leave`. Modifiers are EV-symmetric at a 50/50 sender mix:
+**Gifts.** Anyone *not* in the run can spend one harvest action to send a Care Basket or a Mimic. The party never sees which type. Vote `open` or `leave`. Magnitudes are equalized across all four outcomes — both "always open" and "always leave" yield EV=0 with σ=6, so neither is risk-averse-dominant. The choice is purely about reading sender intent.
 
 |             | Open | Leave |
 |-------------|------|-------|
-| **Care Basket** | +7% (boost) | −4% (basket explodes) |
-| **Mimic**       | −7% (mimic does what mimics do) | +4% (sad mimic helps anyway) |
+| **Care Basket** | +6% (boost) | −6% (basket explodes) |
+| **Mimic**       | −6% (mimic does what mimics do) | +6% (sad mimic helps anyway) |
 
 Multiple gifts stack additively — coordinated senders can swing odds significantly. The full gift log is public at end of run.
 
diff --git a/internal/plugin/coop_dungeon_gifts.go b/internal/plugin/coop_dungeon_gifts.go
index 9a37fe7..b695982 100644
--- a/internal/plugin/coop_dungeon_gifts.go
+++ b/internal/plugin/coop_dungeon_gifts.go
@@ -16,21 +16,21 @@ import (
 
 // ── Gift Constants ──────────────────────────────────────────────────────────
 //
-// Modifiers chosen so the EV of "always open" vs "always leave" is symmetric
-// at a 50/50 sender mix — the party has to read the context, not the math.
+// Magnitudes equalized across all four outcomes so neither "always open" nor
+// "always leave" is variance-dominant at a 50/50 sender mix. Asymmetric
+// magnitudes (e.g., open=±7, leave=±4) gave leave the same EV with lower
+// variance, making it a dominant strategy for risk-averse parties — which
+// collapses the dilemma into "always leave."
 //
-//   basket opened:    +7        basket unopened:  -4
-//   mimic opened:     -7        mimic unopened:   +4
+//   basket opened:    +6        basket unopened:  -6
+//   mimic opened:     -6        mimic unopened:   +6
 //
-// At 50/50 sender mix:
-//   always open  EV = 0.5×(+7) + 0.5×(−7) = 0
-//   always leave EV = 0.5×(−4) + 0.5×(+4) = 0
-
+// Both strategies: EV=0, σ=6. The choice becomes "read sender intent" only.
 const (
-	coopGiftBasketOpened   = 7
-	coopGiftBasketUnopened = -4
-	coopGiftMimicOpened    = -7
-	coopGiftMimicUnopened  = 4
+	coopGiftBasketOpened   = 6
+	coopGiftBasketUnopened = -6
+	coopGiftMimicOpened    = -6
+	coopGiftMimicUnopened  = 6
 )
 
 const (
diff --git a/internal/plugin/coop_dungeon_test.go b/internal/plugin/coop_dungeon_test.go
index 34c664b..2a76c0f 100644
--- a/internal/plugin/coop_dungeon_test.go
+++ b/internal/plugin/coop_dungeon_test.go
@@ -293,6 +293,31 @@ func TestCoopGiftEVSymmetric(t *testing.T) {
 	}
 }
 
+func TestCoopGiftVarianceSymmetric(t *testing.T) {
+	t.Parallel()
+	// Equal EV is not enough — if open and leave have different variance, a
+	// risk-averse party always picks the lower-variance side and the
+	// dilemma collapses. Magnitudes must be equal across all four outcomes.
+	abs := func(x int) int {
+		if x < 0 {
+			return -x
+		}
+		return x
+	}
+	mags := []int{
+		abs(coopGiftBasketOpened),
+		abs(coopGiftBasketUnopened),
+		abs(coopGiftMimicOpened),
+		abs(coopGiftMimicUnopened),
+	}
+	for i := 1; i < len(mags); i++ {
+		if mags[i] != mags[0] {
+			t.Errorf("gift magnitudes must be equal across all four outcomes; got %v", mags)
+			break
+		}
+	}
+}
+
 func TestCoopParimutuelNoWinners(t *testing.T) {
 	t.Parallel()
 	bets := []CoopBet{