Thinking Depth Regression (pre-redaction)
- Entity ID:
ent-20260419-51ad42a2773a - Type:
issue - Scope:
shared - Status:
active - Aliases: pre-redaction decline, thinking budget cut
Description
Quantitative finding that median thinking depth was already falling ~67% before thinking redaction (redact-thinking-2026-02-12) was deployed. Refutes Boris Cherny's position that redaction only hides thinking from the UI without reducing reasoning. Evidence establishes a real behavioral shift predating the UI change.
Key claims
- Read:Edit ratio collapsed from 6.6 to 2.0 during regression
- Stop-hook violations went from 0/day to ~10/day at AMD
- March 8 quality complaints aligned with 50% redaction crossover, not regression onset
- Thinking depth fell 67% before thinking redaction went live
- Adaptive Thinking Zero-Budget Bug is Mechanically Explained by Design
- Model cannot introspect real-time reasoning budget
- Adaptive thinking allocated zero reasoning tokens on fabrication turns
- Shallow thinking drives 80x API request count and 122x billing despite stable prompt count
- Q1 2026 feature cluster shipped 12+ features concurrent with unannounced regression
Relations
- Thinking Depth Regression (pre-redaction) --[caused]--> Wrong-Edit Retry Loop
- GitHub Issue #42796 --[summarizes]--> Thinking Depth Regression (pre-redaction)
- Thinking Depth Regression (pre-redaction) --[caused]--> Read:Edit Ratio Collapse
- GitHub Issue #42187 (Fix-Loop Token Maximization) --[informed_by]--> Thinking Depth Regression (pre-redaction)
- Default Effort Elevation to HIGH --[fixed]--> Thinking Depth Regression (pre-redaction)
- Thinking Depth Regression (pre-redaction) --[depends_on]--> Thinking Signature Field Proxy
- Thinking Depth Regression (pre-redaction) --[informed_by]--> Regression Feedback Loop
- Q1 2026 Feature Release Cluster --[related_to]--> Thinking Depth Regression (pre-redaction)
- Thinking Depth Regression (pre-redaction) --[caused]--> Blame-Shifting Behavior Catalogue