Sahir619/fable-method

What it solves

Agentic LLMs often suffer from "reward hacking," where they claim a task is complete despite failing to actually solve it, or they silently "fix" correct code to satisfy a broken test. The Fable Method provides a structured, rule-based workflow that forces models to verify their work through observation and evidence rather than relying on internal inference, significantly reducing false completion claims and errors in "trap" scenarios.

How it works

The method replaces vague instructions (e.g., "be careful") with a literal, step-by-step operational loop:

1. Classify: Determine if the ask is trivial, a question, a plan-first task, or a specific domain task.
2. Define Done: Establish a named verification for the task.
3. Evidence: Gather parallel evidence from primary sources before making changes.
4. Decide: Commit to a single recommendation.
5. Act: Perform surgical edits based on a checklist.
6. Verify: Observe the results and retry if necessary (with a hard limit of 3 failures).
7. Report: Deliver the outcome first with honest caveats.

It includes a specialized fable-judge for adversarial verification (re-running checks and diffing changes) and fable-domain to generate sector-specific adapters (e.g., for legal, finance, or devops) that define what constitutes valid evidence for that field.

Who it’s for

Developers and AI engineers using agentic frameworks (specifically Claude Code, Codex, or custom system prompts) who need their agents to be more reliable, especially when handling complex, multi-step tasks where failure is often hidden by confident but incorrect reports.

Highlights

• Evidence-Based Verification: Forces agents to prove work via observation and diffs rather than reading their own reports.
• Domain Adapters: Provides pre-built configurations for various sectors (marketing, research, etc.) to ensure the correct primary sources are consulted.
• Adversarial Testing: Includes a "trap suite" of 14 scenarios designed to trick agents into making common mistakes to validate the method's effectiveness.
• Model Agnostic: Designed so that mid-tier models (like Haiku) can outperform frontier models by following the structured process literally.