Create game_config.py

game_config.py

@@ -0,0 +1,106 @@
+# ─────────────────────────────────────────────
+#  game_config.py  –  all balance numbers live here
+# ─────────────────────────────────────────────
+
+STARTING_MONEY: float = 100.0
+
+# ── Upgrades ────────────────────────────────
+# Each upgrade has a list of levels.
+# level 0  = not purchased (baseline)
+# level 1+ = purchased tiers
+#
+# quality_score  : additive contribution to total_quality  (0–100 scale)
+# cost           : price to buy that tier (cumulative spend per tier, not total)
+
+UPGRADES: dict = {
+    "model_size": {
+        "label": "Model Size",
+        "icon": "⬡",
+        "description": "Number of parameters in your model",
+        "levels": [
+            {"name": "Nano  (7M)", "cost": 0, "quality_score": 0},
+            {"name": "Small (70M)", "cost": 20, "quality_score": 8},
+            {"name": "Medium (500M)", "cost": 60, "quality_score": 18},
+            {"name": "Large (3B)", "cost": 150, "quality_score": 32},
+            {"name": "Giant (30B)", "cost": 400, "quality_score": 50},
+        ],
+    },
+    "dataset_size": {
+        "label": "Dataset Size",
+        "icon": "◈",
+        "description": "Tokens of training data",
+        "levels": [
+            {"name": "Scraps  (1B)", "cost": 0, "quality_score": 0},
+            {"name": "Common (10B)", "cost": 15, "quality_score": 6},
+            {"name": "Large  (100B)", "cost": 45, "quality_score": 14},
+            {"name": "Massive (1T)", "cost": 120, "quality_score": 26},
+            {"name": "Web-scale (10T)", "cost": 300, "quality_score": 40},
+        ],
+    },
+    "dataset_quality": {
+        "label": "Data Quality",
+        "icon": "◉",
+        "description": "Curation and filtering pipeline",
+        "levels": [
+            {"name": "Raw dump", "cost": 0, "quality_score": 0},
+            {"name": "Deduplicated", "cost": 10, "quality_score": 5},
+            {"name": "Filtered", "cost": 30, "quality_score": 12},
+            {"name": "Curated", "cost": 80, "quality_score": 22},
+            {"name": "RLHF-aligned", "cost": 200, "quality_score": 35},
+        ],
+    },
+    "architecture": {
+        "label": "Architecture",
+        "icon": "◬",
+        "description": "Model architecture design",
+        "levels": [
+            {"name": "Vanilla LSTM", "cost": 0, "quality_score": 0},
+            {"name": "GPT-style", "cost": 25, "quality_score": 10},
+            {"name": "Flash-Attn", "cost": 70, "quality_score": 20},
+            {"name": "MoE layers", "cost": 180, "quality_score": 33},
+            {"name": "Custom SSM", "cost": 450, "quality_score": 48},
+        ],
+    },
+    "hardware": {
+        "label": "Hardware",
+        "icon": "▣",
+        "description": "Training compute",
+        "levels": [
+            {"name": "CPU cluster", "cost": 0, "quality_score": 0},
+            {"name": "GTX 1080 ×4", "cost": 30, "quality_score": 7},
+            {"name": "A100 ×8", "cost": 90, "quality_score": 17},
+            {"name": "H100 pod", "cost": 250, "quality_score": 30},
+            {"name": "Custom silicon", "cost": 600, "quality_score": 45},
+        ],
+    },
+}
+
+# Max quality score achievable (sum of all top-tier quality_scores)
+# Used to normalise quality to 0–1.  Recomputed at runtime, but shown here for reference.
+# = 50 + 40 + 35 + 48 + 45 = 218
+
+# ── Training economics ───────────────────────
+BASE_TRAIN_COST: float = 10.0  # $ cost at level-0 everything
+TRAIN_COST_QUALITY_SCALE: float = 0.15  # extra cost per quality point
+
+BASE_REWARD: float = 25.0  # $ reward at level-0 everything
+QUALITY_EXPONENT: float = 2.2  # reward grows super-linearly with quality
+MAX_QUALITY_REWARD_MULT: float = 40.0  # cap multiplier at max quality
+
+REWARD_NOISE_PCT: float = 0.12  # ±12 % random variance per run
+
+# ── Cloud Inference ──────────────────────────
+CLOUD_UNLOCK_COST: float = 150.0  # one-time cost to unlock
+CLOUD_BASE_REVENUE: float = 8.0  # $ per tick (every 8 s) at quality 0
+CLOUD_QUALITY_MULT: float = 0.06  # additional revenue per quality point
+CLOUD_TICK_SECONDS: int = 8  # JS poll interval hint (stored here for reference)
+
+# ── Chat temperature mapping ─────────────────
+# quality_ratio = total_quality / max_quality   (0.0 – 1.0)
+# temperature   = lerp between TEMP_MAX and TEMP_MIN based on quality_ratio
+CHAT_TEMP_MAX: float = 2.0  # chaotic gibberish at quality 0
+CHAT_TEMP_MIN: float = 0.3  # coherent at max quality
+
+# ── HuggingFace model ────────────────────────
+HF_MODEL_ID: str = "FlameF0X/Qwen3-4B-Distilled-Claude-4.6"
+HF_MAX_NEW_TOKENS: int = 400