Code Oracle

Sub-50ms neuro-symbolic verification for AI coding agents.
Structural AST topology validated by Tarjan SCC and a non-autoregressive decision model (Laya).

The Problem: Autoregressive Verification Overhead

When coding agents (Claude Code, Cursor, OpenCode, Codex) inspect code modifications, they face three operational bottlenecks:

1. Confirmation Bias: Generative models reviewing their own diffs frequently rationalize their own logic errors.

2. Latency and Token Overhead: Re-evaluating complete files through a frontier model introduces 2,000 to 6,000 ms of round-trip latency and consumes output tokens on conversational explanations.

3. Tool Schema Bloat: Standard MCP servers inject sprawling multi-tool schemas into the prompt context on every turn, reducing effective agent window capacity.

Related MCP server: Symbol Delta Ledger

Architecture

Code Oracle decouples code generation from verification. It runs locally as an independent evaluation layer.

Instead of generating conversational critiques, Code Oracle constructs a localized AST graph using Tree-sitter, checks topological invariants symbolically, and evaluates residual drift using Laya (a 421M parameter ModernBERT decision model).

Verification Pipeline

flowchart TD
    Agent["🤖 AI Coding Agent / Developer<br/>(Claude Code, Cursor, Antigravity)"]

    subgraph Engine ["⚡ CODE ORACLE ENGINE (&lt; 50ms)"]
        direction TD

        Stage1["Stage 1: Tree-sitter &amp; k-Hop TopoSlice<br/>• Multi-language AST parsing (&lt; 8ms)<br/>• Extracts callers, callees &amp; interfaces<br/>• Isolates k-hop neighborhood graph"]

        Stage2{"Stage 2: Deterministic Symbolic Gate<br/>Tarjan's SCC &amp; Contract Invariants"}

        HardVeto["🚫 Hard Veto Early Exit (&lt; 25ms)<br/>Instant rejection on cycles &amp; signature drift"]

        Stage3["🧠 Stage 3: Laya ModernBERT 421M Head<br/>• Evaluates linearized Micro-DSL subgraph<br/>• In-memory resident CPU inference<br/>• Calibrated continuous risk scoring (0.0 - 1.0)"]

        Stage1 --> Stage2
        Stage2 -- "Cycle / Invariant Breach" --> HardVeto
        Stage2 -- "Topologically Valid" --> Stage3
    end

    Agent -->|"(1) Proposes Patch / Refactor"| Stage1
    HardVeto -->|"(2) Fast-Fail Verdict"| Verdict["🎯 Structured Typed Verdict<br/>VERDICT: APPROVED / REJECTED<br/>Risk Score &amp; Invariant Telemetry"]
    Stage3 -->|"(2) Calibrated Verdict"| Verdict

Key Characteristics

• Sub-50ms Target Latency: In-memory execution using ONNX Runtime or MLX on local CPU and Apple Silicon.

• Zero Output Token Tax: Emits structured status codes and calibrated probability vectors (Pass, Fail, Risk Score) without text generation.

• Lean Tool Surface: Exposes a single endpoint (verify_patch), avoiding multi-tool schema overhead in agent context.

• Offline Execution: Runs without external API calls or network egress.

Architectural Targets vs. Frontier LLM Review

Operational Scope and Boundaries

Code Oracle operates within explicit technical boundaries:

1. Evaluator, Not Author: Code Oracle does not generate, autocomplete, or refactor code. It evaluates proposed patches against existing syntax and topology.

2. Syntax Requirement: Patches must produce a valid Tree-sitter AST. Syntactically invalid inputs fail at Stage 1 before invoking the decision model.

3. Static Topology Bounds: Focuses on structural invariants, dependency cycles, and interface compatibility. It does not replace dynamic test suites, integration environments, or runtime race condition detectors.

4. Memory Footprint: Requires approximately 1.2 GB of RAM to hold the 421M parameter model in memory for single-pass inference.

Pretrained Model Weights

The fine-tuned Laya ModernBERT 421M decision head weights are hosted on Hugging Face:
🤗 wxsys/code-oracle-laya-421m

Code Oracle automatically downloads and caches these weights to ~/.cache/code_oracle/weights/ on first invocation when --neural is enabled, or reads from local ./weights/ if present.

Roadmap

• Architecture Specification & Subgraph Slicing Design

• TopoSlice AST Slicer & Incremental Workspace Indexer

• Tarjan's SCC Cycle Detector & Deterministic Symbolic Gate

• Persistent In-Memory Laya Decision Head & Fine-Tuned Weights (wxsys/code-oracle-laya-421m)

• Lean FastMCP Server interface (verify_patch)

• Agentic SKILL.md distribution for Claude Code, Cursor, and Antigravity

• Git pre-commit & pre-push verification hook with unblock toggle (code-oracle hook)

• Multi-language AST extractors for Tier 1 languages (Python, TypeScript, Go, Rust)

• Synthetic mutation and training dataset mining engine (tools/mine_top_repos.py)

• Google Colab Multi-Language Fine-Tuning Pipeline (notebooks/Laya_Code_Oracle_Finetune.ipynb)

• CPU Thread Auto-Tuning & Hybrid Neuro-Symbolic Latency Optimization

• Dead Code & Orphan Symbol Scanner (code-oracle dead-code via 0-in-degree graph reachability)

• Static Performance Anti-Patterns & Resource Leak Detector (code-oracle perf-lint: nested loop complexity, unclosed handles)

• Official Git Tagging & GitHub Release pipeline (v0.1.0)

• Python Package Wheel Distribution & PyPI Publishing (pip install code-oracle)

• Multi-Agent Ecosystem Integrations (Claude Code, Cursor, Antigravity, OpenCode, and Cline sidecars)

• (Maybe / Experimental) Synthetic User & Local Load Simulation Plugin (AST endpoint discovery + lightweight concurrent stress tester)

License & Attribution

Distributed under the Apache-2.0 License. See LICENSE for details.

Architect & Maintainer:
Wahyu Febri Tamtomo (@wahyuzero)
Founder of frugaldev.biz.id (Radical AI Efficiency & Frugal Computing).