diegosouzapw MAD-Logic Reliability Framework
Implements the MAKER (Massively decomposed AgenT KERnel) framework for zero-error execution of long-horizon tasks.
Step 0: Modern Technology Research (MANDATORY)
Before applying MAD-Logic, verify the following:
- State of the Art: Check for new research updates to the MAKER paper (arXiv:2511.09030v1) or subsequent meta-agent architectures.
- Model Benchmarks: Identify the best "cheap-but-reliable" models for voting. Smaller models (e.g., GPT-4o-mini, Claude 3 Haiku) are often superior for high-k voting due to cost and latency.
- MCP Tooling: Look for existing Model Context Protocol servers that provide specialized sub-actions (e.g., file-editing, browser-control) to use as the base for MAD decomposition.
Core Principles
1. Maximal Agentic Decomposition (MAD)
Break tasks into the smallest possible subtasks (m=1 step per call). Benefits:
- Reduces context length → fewer errors
- Enables per-step error correction
- Allows smaller, focused models to succeed
2. First-to-ahead-by-k Voting
For critical steps, sample multiple times and pick the answer that leads by k votes:
- k=3 for standard critical operations
- k=5 for high-stakes (deployment, security, financial)
- k=7+ for mission-critical (rarely needed)
3. Red-Flagging
Discard suspicious outputs before they corrupt the pipeline:
- Too long: Response exceeds expected length → likely confused
- Wrong format: Malformed output → model went off-rails
- Low confidence: Hedging language patterns → uncertain
When This Skill Auto-Triggers
| Condition | Action |
|---|---|
| Task has >5 sequential steps | Apply MAD decomposition |
| Step involves deployment/mutation | Use k=3 voting |
| Step involves security/financial | Use k=5 voting |
| Output seems malformed | Red-flag and resample |
Usage
Voting for Critical Decisions
# Import the voting utility
from maker_reliability.voting import vote_until_consensus
# Get consensus on a critical decision
candidates = [generate_response() for _ in range(10)]
winner = vote_until_consensus(candidates, k=3)
Red-Flag Filtering
from maker_reliability.red_flag import is_red_flagged
response = get_llm_response()
if is_red_flagged(response, max_tokens=500, require_format="json"):
response = resample() # Discard and try again
Integration with Task Boundaries
This skill works with the existing task_boundary tool:
- Each task boundary represents a decomposition level
- Status updates enable progress tracking
- Mode switching (PLANNING/EXECUTION/VERIFICATION) maps to MAKER phases
Theory Reference
For mathematical foundations (scaling laws, cost formulas, k optimization): → See maker-theory.md
Scripts
| Script | Purpose |
|---|---|
| voting.py | First-to-ahead-by-k consensus voting |
| red_flag.py | Output validation and red-flag detection |
Definition of Done
- Task is decomposed into single-step subtasks (MAD).
- Critical decision points have a defined voting margin (k).
- Consensus logic (vote_until_consensus) is used for all k > 1 steps.
- Output is validated against red-flag filters prior to being committed.
- Reliability target (e.g., 99.9%) is calculated and met via appropriate k selection.