Lean-Output Nudge + Verbosity-Steer Savings
Lean-Output Nudge
- Rewritten nudge messages: 'Reason as deeply as you need — but keep your visible output lean'
- Explicitly separates reasoning (unconstrained) from output (compressed)
- Live A/B test: 30-41% actual output reduction (estimate stays conservative at 10-15%)
- Calibration: nudge events now record pre_nudge_avg_output and pre_turns for follow-through
Verbosity-Steer Savings (Pool #4)
- Estimated output-token savings from lean-output conciseness nudges, repriced at baseline output rate
- Added to transformation, combined counterfactual, and breakdown across all platforms
- Correctly classified as estimated tier (trigger observed, magnitude not metered)
Cross-Platform Parity Fixes
- OpenCode: cumulative now uses full 4-pool transf- OpenCode: cumulative now uses full 4-pool transf- OpenCode: cumulative now uses full 4-pool transf- OpenCode: cumulative now uses full 4-pool transf- OpenCode: cumulative now uses full 4-pool transf- OpenCode: cumulative now uses full 4-pool transf- t in UserPromptSubmit — does not break cache prefix
- 90%+ fill suppression: nudge silenced when context is critically full
- Cooldown: max 3 per session, 5 min apart