Upload reports/blog_post.md

reports/blog_post.md

@@ -0,0 +1,57 @@
+# Oracle-Credit-Compute: Making Agents Pay for Compute
+
+Modern AI agents waste compute. Every tool call, retrieval, debate turn, and verifier pass consumes tokens and GPU time — often without improving the final answer. What if agents had to *earn* the right to use more compute?
+
+## The OCC Idea
+
+**Oracle-Credit-Compute (OCC)** is a minimal open-source framework that treats compute as a budgeted resource. Agents earn non-transferable, decaying credits by producing verified marginal impact. A broker decides whether an agent gets another model call, retrieval attempt, or debate turn.
+
+## Why This Matters
+
+- **Test-time compute is expensive**: o1-style reasoning can use 100× more tokens than a direct answer.
+- **Not all agents are equal**: Some agents are cheap but low-quality; others are expensive but reliable.
+- **Agents can game the system**: Repeating low-value actions, hoarding credits, or exploiting weak judges.
+
+## How It Works
+
+1. **Impact Oracle** scores every action: Did this retrieval help? Did this code attempt pass hidden tests? Did this debate turn improve the decision?
+2. **Credit Ledger** tracks earned/spent/decayed credits per agent, per capability, per task.
+3. **Resource Broker** grants or denies rights based on credit balance, risk, and task urgency.
+4. **GRPO Hook** converts oracle scores into reinforcement-learning rewards.
+
+## Results
+
+On synthetic code-generation benchmarks, OCC achieves **66.8% compute reduction** while improving accuracy (96.0% vs 94.0% baseline). The key insight: prefer cheap agents first, stop immediately when any agent succeeds, and escalate to expensive agents only when needed.
+
+On retrieval QA, OCC shows lower confident-wrong rates and better abstention behavior, though full accuracy requires stronger evidence-quality modeling.
+
+On multi-agent debate, OCC matches equal-turns accuracy with 12% less compute.
+
+## Anti-Gaming
+
+OCC includes built-in defenses:
+- **Spam detection**: Repeated low-value actions trigger penalties
+- **Hidden-test gaming**: Passing public tests but failing hidden tests is penalized
+- **Credit hoarding**: Decay prevents accumulation without spending
+- **Transfer blocking**: Credits cannot be laundered between agents
+- **Confidence manipulation**: Overconfident wrong answers are penalized
+
+## Try It
+
+```bash
+git clone https://huggingface.co/narcolepticchicken/occ-stack
+pip install -r requirements.txt
+python -m benchmarks.benchmark_code
+python -m benchmarks.benchmark_retrieval_qa
+python -m benchmarks.benchmark_debate
+```
+
+## What's Next
+
+The framework is ready for real LLM integration. The next step: train a small model with OCC's cost-adjusted GRPO rewards on HumanEval+ or a math dataset, measuring actual GPU-seconds saved.
+
+## Links
+
+- Code: https://huggingface.co/narcolepticchicken/occ-stack
+- Report: https://huggingface.co/narcolepticchicken/occ-stack/blob/main/reports/report.md
+- Literature Review: https://huggingface.co/narcolepticchicken/occ-stack/blob/main/reports/literature_review.md