README.md

---
tags:
- ml-intern
---
# Oracle-Credit-Compute (OCC) Stack

**Compute is not neutral.** In multi-agent debate, giving an adversarial agent equal speaking turns causes accuracy to collapse from 73.3% to 56.7% — worse than single-round voting. The mechanism is volume amplification, not persuasion: honest agents retain their positions 84% of the time but get outvoted.

OCC is a mechanism-design layer that treats agent compute as a scarce, earned, auditable privilege. Agents earn non-transferable, decaying, capability-scoped credits based on verified marginal impact.

**Paper:** `reports/occ_paper.md`  
**Design doc:** `design.md`  
**Literature review:** `reports/literature_review.md`

---

## Key Results

### Debate Collapse (Real LLM — Qwen3-Coder-30B on H200)

| Condition | Accuracy | Δ |
|-----------|:--------:|:---:|
| 1-round baseline | **73.3%** | — |
| 3-round debate (with adversary) | **56.7%** | −16.7pp |
| Judge voting | **73.3%** | recovered |
| Confidence-weighted voting | **73.3%** | recovered |
| Equal-token (cap adversary volume) | **70.0%** | −3.3pp |

**Root cause:** Volume amplification (H1 ✅). Not persuasion (H4 ❌, 84% retention). Not skill-dependent (H6 ❌, weak adversary still collapses).

### Simulated Benchmarks

| Benchmark | Baseline | OCC | Savings |
|----------|:--------:|:---:|:-------:|
| Code (pass@1, iso-accuracy) | 0.78 | 0.78 | 52.3% compute |
| Debate (adversarial) | 0.56 | 0.76 | +20pp accuracy |
| Retrieval QA | 0.79 | 0.71 | 33% retrievals (accuracy lags) |

### Anti-Gaming

10 attack vectors tested, all contained: credit farming, collusion, oracle spoofing, verbosity gaming, confidence manipulation, strategic abstention, identity laundering, sybil agents, sandbagging, griefing.

### Honest Limitations

- OCC ≈ random gating at moderate budgets (doesn't beat simple baselines in median case)
- GRPO training produced no improvement at 0.5B scale
- Single seed for mechanism isolation (CI ±16pp)
- Simulated benchmarks for code/QA (not real LLM agents)
- HumanEval result is adaptive retry, not OCC credit allocation

---

## Quickstart

```bash
git clone https://huggingface.co/narcolepticchicken/occ-stack
cd occ-stack
pip install -r requirements.txt

# Simulated benchmarks (CPU)
python benchmarks/benchmark_code.py
python benchmarks/benchmark_debate_v2.py
python benchmarks/benchmark_retrieval_qa.py

# Ablations + anti-gaming (CPU)
python eval_runner.py

# Unit tests
python tests/test_oracle.py
python tests/test_ledger.py

# Debate collapse analysis
python jobs/analyze_collapse.py reports/debate_collapse_mechanism_results.json
```

---

## Architecture

```
┌─────────────┐    ┌─────────────────┐    ┌──────────────┐
│  Agent      │───▶│  ResourceBroker │───▶│  Compute     │
│  (requests  │    │  (allow/deny/   │    │  (model call,│
│   resource) │◄───│   downgrade)    │◄───│   retrieval) │
└─────────────┘    └─────────────────┘    └──────────────┘
       │                   │
       ▼                   ▼
┌─────────────┐    ┌─────────────────┐
│ CreditLedger│◄───│  ImpactOracle   │
│ (earn/spend/│    │  (score action  │
│  decay)     │    │   on verified   │
└─────────────┘    │   impact)       │
                   └─────────────────┘
```

## Components

| Component | File | Description |
|-----------|------|-------------|
| Impact Oracle | `oracle/oracle.py` | Multi-mode scoring (code, QA, debate) with cost-adjusted rewards |
| Credit Ledger | `ledger/ledger.py` | Non-transferable, decaying, capability-scoped credits with audit trail |
| Resource Broker | `broker/broker.py` | Capability-based access control (allow/deny/downgrade/escalate) |
| GRPO Hook | `rl/grpo_hook.py` | TRL-compatible reward function for learned allocation |
| Reward | `rl/reward.py` | Cost-adjusted reward with anti-gaming penalties |

## Formal Definition

Full formal definition with 10 system invariants, ledger event schema, threat model, and when-to-use guidance in [`design.md`](design.md).

## Repo Structure

```
occ/
  oracle/          # ImpactOracle
  ledger/          # CreditLedger
  broker/          # ResourceBroker
  rl/              # GRPO hook + reward function
  benchmarks/      # 3 benchmark scripts
  jobs/            # GPU job scripts + analysis
  tests/           # Unit tests
  reports/         # Paper, reports, literature review, blog post
  design.md        # Formal system definition
```

## Citation

```bibtex
@misc{occ2026,
  title={Compute Is Not Neutral: Mechanism Analysis of Adversarial Debate Collapse and the OCC Stack},
  author={narcolepticchicken},
  year={2026},
  url={https://huggingface.co/narcolepticchicken/occ-stack}
}
```