Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

9 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

MAD-Logic Reliability Framework 🛡️

ASP v2.0 Compliant

Implementation of the MAD-Logic framework for zero-error execution of long-horizon tasks, based on the MAKER (Massively decomposed AgenT KERnel) research paper Solving a Million-Step LLM Task with Zero Errors (arXiv:2511.09030v1).

🏆 The Million-Step Challenge

Why use MAD-Logic? Because standard agents fail exponentially as tasks get longer.

Feature Standard Agent MAD-Logic (k=7)
10-Step Success 10.74% 99.99%
100-Step Success 0.00000002% 99.70%
Reliability Gain 1x 9.97 Billion X

Simulation data for 80% per-step accuracy models.

🚀 Key Features

  • Maximal Agentic Decomposition (MAD): Automatically breaks complex tasks into minimal subtasks to prevent reliability decay.
  • First-to-ahead-by-k Voting: Uses mathematical consensus to guarantee success even with imperfect models.
  • Red-Flag Filtration: Real-time monitoring to catch and discard "suspicious" outputs (hallucinations, loops, hedging).
  • ASP v2.0 Compliant: Universal skill standard for any IDE or Agent (Claude Code, Cursor, etc.).

📦 Installation

To use this framework:

  1. Antigravity Agents: Place the provided .skill file in your ~/.antigravity/skills/ directory.

  2. Generic Agents (Claude Code / Cursor): Clone this repository and point your agent to this directory. It will detect the SKILL.md and scripts/ automatically.

🛠️ Components

  • voting.py: Consensus algorithm with scaling law calculators.
  • red_flag.py: Heuristic filters for response quality.
  • simulation_test.py: Reproduce the paper's findings on reliability scaling.

📚 References

  • Sinha, S., et al. (2025). Solving a Million-Step LLM Task with Zero Errors. arXiv:2511.09030.
  • ASP v2.0 Protocol

About

Automatic error-correction framework for high-stakes multi-step LLM tasks based on the MAKER framework (arXiv:2511.09030v1). Includes Maximal Agentic Decomposition, voting consensus, and red-flagging filters.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages