---

"I asked an AI agent to write a recipe for tomato scrambled eggs. It proudly delivered: 'Here is your tomato scrambled eggs (note: this dish contains no pork, no beef, no chicken, and no fish).' Why are AI agents so obsessed with telling us what they didn't do? Ship the result, not the conversation."
— Shangyin Tan, UC Berkeley

---

The "Tomato-Egg Problem" in AI Agents

When coding and workflow agents generate deliverables (PR descriptions, documentation, code comments, executive slides, or RPA forms), they frequently exhibit Apophasis (the rhetorical habit of describing an entity by detailing what it is not):

• "Implemented JWT auth (note: we did not use session cookies or basic auth after exploring three other architectures)."
• "Updated customer export. We attempted 4 regex approaches that failed before settling on this one."
• "Tomato Scrambled Eggs (100% pork-free, beef-free, chicken-free, fish-free)."

This conversational residue clutters deliverables, expands context windows, distracts human reviewers, and leaks private development scaffolding.

Ockev solves this at the gate. Named after Occam's Razor (Entia non sunt multiplicanda praeter necessitatem) and the Jev discriminative architecture, Ockev provides an ultra-fast, zero-generation System 1 gatekeeper that intercepts conversational clutter before it reaches readers or production.

---

Performance & Pareto Frontier

Ockev was benchmarked on TomatoEggBench-120, a standardized 120-case real-world benchmark spanning 5 production domains (Code Repositories, Excel/CSV, PDF Documents, Form RPA, and Executive Slide Decks).

TomatoEggBench-120 Benchmark Results

Model / System	Architecture	Benchmark Accuracy	Violation Recall	Inference Latency	Deployment Mode
Ockev-3B (Ours)	Qwen2.5-3B + Pointer Head	95.8% (115/120)	98.5%	105.0 ms	Local / Private GPU
TypeSafe Jev 1.13.0	Causal MoE + RLCD	93.3% (112/120)	96.9%	250.0 ms	Cloud API ($42/1B tok)
Ockev-1.5B (Ours)	Qwen2.5-1.5B + Pointer Head	92.5% (111/120)	95.4%	50.8 ms	Local / Apple Silicon
SemIf Qwen3.5-4B	Qwen3.5-4B (Logits)	74.7% (89/120)	68.0%	850.0 ms	Local Logits Readout
ModernBERT-151M	ModernBERT (Fixed Head)	25.0% (30/120)	35.0%	24.0 ms	Local Embeddings

---

Core Architecture

Unlike standard generative LLMs that waste tokens generating verbose explanations, Ockev uses a Block-Causal Mask with a Pointer Readout Head:

State (Deliverable + Context)  ──┐
                                 ├─► [Causal LM Backbone] ──► [Hidden States]
Questions (Criteria + Options) ──┘                                │
                                                                  ▼
[Decide Token Q] ───────────────► q = W_q · h_decide ──────┐
                                                           ├─► Scaled Dot Product ──► Softmax ──► Verdict
[Option Tokens K_1 ... K_N] ────► k_i = W_k · h_opt_i ─────┘

1. Zero Generated Tokens: Decisions are computed in a single forward pass via metric dot-product between the decision query vector and option key vectors.
2. Order Invariant: Pointer heads project options into metric space independently, eliminating choice position bias.
3. Decoupled Uncertainty: Rather than treating uncertain as an attractor token, Ockev computes decision margins directly: $$\text{Margin} = |P(\text{pass}) - P(\text{revise})| < \tau \implies \text{uncertain}$$

---

Quickstart

Installation

pip install ockev

# For Apple Silicon (Metal acceleration via MLX)
pip install "ockev[mlx]"

CLI Usage

Verify any markdown file, commit message, or PR text before publishing:

# Check a deliverable file
ockev check path/to/PULL_REQUEST.md

# Check raw text directly
ockev check --text "Tomato scrambled eggs (does not contain pork or beef)."

Output:

=== Ockev Deliverable Gate Review ===
Verdict:       REVISE
Confidence:    0.96
Latency:       34.2 ms
Probabilities: {'pass': 0.021, 'revise': 0.979}

[FAIL] Deliverable contains conversational clutter, negative echoes, or discarded options.

Python API

from ockev import OckevGate

gate = OckevGate()

result = gate.review_deliverable(
    candidate_text="Added support for PostgreSQL 16. Verified migration scripts against test db.",
    task_context="Database upgrade PR"
)

print(result)
# {'verdict': 'pass', 'confidence': 0.94, 'probabilities': {'pass': 0.97, 'revise': 0.03}, 'latency_ms': 32.1}

Apple Silicon MLX Native Engine

Run directly on Apple Silicon M-series chips with zero CUDA or PyTorch overhead:

from ockev.mlx_engine import MLXOckevEngine

engine = MLXOckevEngine("Qwen/Qwen2.5-1.5B")
res = engine.evaluate(state="...", question_spec={...})
print(f"MLX Latency: {res['latency_ms']} ms")

---

Agent Integration

Ockev includes an out-of-the-box Agent Skill for Pi and Codex agents:

# ~/.codex/skills/final-state-review/SKILL.md
Produce content that stands on its own for the next reader or agent.
Ship the result, not the conversation.

When activated, the agent automatically runs ockev check before finalizing code, documentation, or PR handoffs.

---

License

Distributed under the Apache 2.0 License.