Neural Learn Rate

If the file does not exist, create it with the defaults above.

2. Calculate Recent Accuracy

Review the last N pattern-use events (where N = accuracy_window):

1. Read logs/claude_log.md for "Pattern Scored" entries
2. Count successes and failures in the window
3. Calculate accuracy: accuracy = successes / total_events

3. Determine Rate Direction

Apply the adaptive rule:

Calculate new rate:

• Decrease: new_rate = current_rate * 0.8
• Increase: new_rate = current_rate * 1.3
• Clamp: new_rate = max(min_rate, min(max_rate, new_rate))

4. Update Confidence Deltas

The learning rate controls how much confidence changes on each pattern scoring event:

• Success delta: +learning_rate (default +0.05)
• Failure delta: -learning_rate * 2 (default -0.10)
• Partial delta: +learning_rate * 0.4 (default +0.02)

These deltas are used by neural-pattern-score when adjusting pattern confidence.

5. Update Configuration

Write the updated values to memory/sona-config.json:

{
  "learning_rate": [new_rate],
  "min_rate": 0.01,
  "max_rate": 0.15,
  "accuracy_window": 10,
  "last_adjusted": "[current ISO timestamp]",
  "adjustment_history": [
    ...previous,
    {
      "timestamp": "[ISO]",
      "accuracy": [value],
      "old_rate": [old],
      "new_rate": [new],
      "direction": "[converge/explore/hold]"
    }
  ]
}

6. Propagate to Pattern Scoring

Ensure that neural-pattern-score reads the current learning_rate from sona-config.json rather than using hardcoded deltas. The playbook for neural-pattern-score references fixed values as defaults — the config file overrides them.

7. Log the Adjustment

Append to logs/claude_log.md:

### Learning Rate Adjusted — [timestamp]
- Accuracy (last [N] events): [x.xx]
- Direction: [converge/explore/hold]
- Rate: [old] → [new]
- Confidence deltas: success=[+x.xx], failure=[-x.xx], partial=[+x.xx]

Output

• Updated memory/sona-config.json with new learning rate and history
• Log entry in logs/claude_log.md

Review Checklist

• At least 5 scoring events were available for accuracy calculation
• Accuracy was calculated from the correct window size
• Rate adjustment direction matches the accuracy range
• New rate is clamped within min/max bounds
• Configuration file is valid JSON after update
• Adjustment is logged with old and new values