​

Self-Correction

BioCortex implements a four-level self-correction hierarchy so that failures are handled by reflection and targeted fixes instead of blind retries. This reduces error propagation and improves success rates on complex pipelines.

---

​

Levels Overview

---

​

Level 1: Reflection-Guided Code Repair

When code fails (exception or Critic failure):

1. Reflect — The LLM is asked to produce a structured Reflection:
   • Error summary
   • Root cause analysis
   • Concrete fix strategy
   • Confidence
   • Reusable lesson
2. Fix — A new code version is generated based on this diagnosis, not a generic retry.
3. Progressive error chain — Previously tried fixes and their outcomes are passed into the next attempt so the LLM does not repeat the same failed approach.
4. Retry limit — Standard retries (e.g. up to 3). If all fail, deep retry is triggered.

​

Deep retry

• Uses the reasoning model (not just coder) for deeper analysis.
• Can use successful code from peer nodes (same DAG level) as reference.
• Generates a different approach (e.g. different library or algorithm).
• Typically up to 2 deep retries before marking the node failed.

---

​

Level 2: Output Quality Refinement

When code runs but the Critic marks the result as low-quality (empty, suspicious values, incomplete):

• The Critic’s structured feedback (issues, retry_guidance) is passed to the Executor.
• The LLM generates targeted improvements to the code or parameters.
• Re-execution and re-validation follow.

---

​

Level 3: Plan Revision

When multiple nodes (e.g. whole branches) fail:

• The SelfRefineEngine can suggest a revised execution plan.
• Successful steps are preserved.
• Failed components are replaced with alternative approaches (e.g. different tools or steps).
• The revised plan is then executed like a new DAG.

---

​

Level 4: Report Self-Refinement

After the Synthesizer produces the first report:

1. Self-review — The system (or LLM) scores the report on criteria: completeness, accuracy, clarity, actionability, scientific rigor.
2. Targeted improvement — If below threshold (e.g. 0.8), generate an improved version.
3. Quality scoring — Re-score; repeat until threshold is met or no further gain.
4. Degradation detection — If a new version scores lower, revert to the previous one.

This yields a final report that meets a consistent quality bar.

---

​

Integration with the Pipeline

• Executor triggers Level 1 (and deep retry) when execution or validation fails.
• Critic triggers Level 2 when validation fails with actionable feedback.
• Orchestrator / SelfRefineEngine triggers Level 3 when multiple nodes fail.
• Synthesizer (or a post-step) triggers Level 4 after the first draft report.

---

​

Configuration

• Max standard retries, max deep retries.
• Quality threshold and max iterations for report refinement.
• Whether to use reasoning model for deep retry.

See Configuration and biocortex.core.self_refine (or equivalent).

---

​

Next Steps

• Multi-Agent Pipeline — Where Critic and Executor plug in.
• Context Window and Budget — Token limits during reflection and refinement.