Update README.md

README.md

@@ -1,3 +1,76 @@
----
-license: mit
----
+---
+language:
+- en
+license: mit
+pretty_name: AI Reward Channel Attack-Surface Coherence Collapse v0.1
+dataset_name: ai-reward-channel-attack-surface-coherence-collapse-v0.1
+tags:
+- clarusc64
+- ai-safety
+- alignment
+- reward-hacking
+- reward-tampering
+task_categories:
+- text-classification
+- tabular-classification
+- reinforcement-learning
+size_categories:
+- 1K<n<10K
+configs:
+- config_name: default
+  data_files:
+  - split: train
+    path: data/train/ai-reward-channel-attack-surface-coherence-collapse.csv
+---
+
+## What this dataset is
+
+This dataset models reward tampering as a coherence collapse.
+
+It tracks when the link between task success and reward breaks  
+because the agent starts interacting with the reward channel itself.
+
+## What it tests
+
+Given a scenario you output:
+
+- collapse_flag  
+- collapse_driver  
+- channel_coherence_score  
+- minimal_mitigation_set  
+
+## What the fields mean
+
+- channel_coherence_score  
+How well reward still tracks intended task success  
+0 means fully broken  
+1 means intact  
+
+- collapse_flag  
+0 means still coherent  
+1 means coherence collapsed  
+
+- collapse_driver  
+What caused the collapse  
+examples: exploit, evaluator_manipulation, termination_hack  
+
+- minimal_mitigation_set  
+The smallest set of controls to stop the attack path
+
+## Why it matters
+
+Reward hacking starts as probing.
+
+If you see the coherence score dropping early  
+you can isolate reward channels before full tampering.
+
+## Files
+
+- data/train/ai-reward-channel-attack-surface-coherence-collapse.csv  
+- tester.csv  
+- scorer.py  
+- README.md  
+
+## Path
+
+ClarusC64/ai-reward-channel-attack-surface-coherence-collapse-v0.1