Gallery
Path of Titans mods
Mods in pages/paleo-presentation/mods/ — click a tile to open full size.
In-game
Path of Titans — gameplay & UI
Screens from pages/paleo-presentation/game/ — click a tile to open full size.
Major Results
Latest Serengeti results: 15 epochs made 1e-4 the best accuracy run
0.9118
Best validation accuracy
ResNet-18, LR=1e-4, augmentation, 15 epochs
0.9206
Predator recall
939 predators caught, 81 missed on validation
0.5238
Heuristic baseline
Simple baseline before ResNet learning
2,142
Validation images
Same holdout split across all Serengeti experiments
| Experiment | Val Accuracy | Pred Recall | Pred F1 | LR | Epochs |
|---|---|---|---|---|---|
| ResNet-18 LR=1e-4 + aug | 0.9118 | 0.9206 | 0.9086 | 1e-4 | 15 |
| ResNet-18 LR=1e-3 + aug | 0.9080 | 0.9020 | 0.9033 | 1e-3 | 10 |
| ResNet-18 LR=1e-3 + aug | 0.9066 | 0.9167 | 0.9034 | 1e-3 | 15 |
| ResNet-18 LR=1e-4 + aug | 0.9062 | 0.8990 | 0.9012 | 1e-4 | 10 |
| ResNet-18 LR=5e-5 + aug | 0.8987 | 0.8569 | 0.8896 | 5e-5 | 15 |
Convergence across epoch settings
Model-selection result: the Serengeti sweep used a 6,529 train / 2,142 validation split. On that stable validation set,
1e-4 + augmentation + 15 epochs was the strongest run, so we used it as the best real-image checkpoint before adapting to Path of Titans.
Ablation Studies
Augmentation dominated the strongest Serengeti runs
Validation accuracy by augmentation setting
Blue = with augmentation
Augmented runs used
RandomHorizontalFlip and ColorJitter(0.3).Pattern
The best run used augmentation:
0.9118 accuracy, 0.9206 predator recall, 0.9086 predator F1.Decision
Because augmented models dominated the upper ranks, the transfer checkpoint kept augmentation.
12
runs compared
2,142
validation images
6/6
top-half augmented
1e-4
best LR + aug
Domain Shift + Safety Tuning
Why the selected model changed
300-image PoT fine-tune validation
Agent safety holdout
1e-4 is still the accuracy pick: it was strongest on Serengeti and on the 300-screenshot PoT validation split. But on the 10-image game holdout, class weighting alone did not fix false negatives at the default threshold. We switched from "best accuracy checkpoint" to a safety operating point: lr=3e-5, predator class weight 3.0, and threshold 0.20, because the live agent should over-warn rather than miss predators.
Label source matters
Kaggle videos could be converted into frames, but without predator/non-predator labels they cannot support supervised evaluation. Serengeti species labels keep the experiment measurable.
🎮 Domain Shift: Savanna → 3D Game
Camera-trap JPEG textures differ fundamentally from rendered 3D polygon meshes. In-game predator appearance (e.g., T-Rex) has no Serengeti analogue. Current classifier degrades with game-engine domain
📊 Pixel Heuristic Fallback Is Blind
When no checkpoint is loaded, frame_to_observation() uses brightness + motion score as a proxy for threat. This fires on any moving bright object (day/night cycle, weather) — high false positive rate
⏱️ Latency Budget Under Pressure
ResNet-18 on CPU runs ~120ms/frame on mid-range hardware — exceeds the 80ms target for 10 FPS. Drops effective rate to ~6 FPS. GPU inference or model distillation (MobileNet) required for smooth real-time
🧠 Primal Mind Has No Episodic Context
Recent-events memory is still a short runtime window, so the agent cannot reason far back in time (for example, repeated attacks from one direction over longer sequences). Longer-horizon memory or Letta RAG integration is still needed.
Threshold beat retraining alone
The weighted fine-tune still missed too many predators at default threshold. Lowering the predator threshold to
0.20 on the weighted run moved recall from 0.571 to 0.714 on the tiny holdout, trading one extra false alarm for fewer missed threats.