self-eval

AI self-evaluations often inflate scores, but calibrated two-axis scoring separates ambition from execution and enforces devil's advocate reasoning to curb bias. This approach detects score inflation by comparing recent results against a persisted cross-session history.

Two-axis scoring independently rates task ambition as Low, Medium, or High and execution quality as Poor, Adequate, or Strong. These ratings are then combined via a fixed matrix to produce a final calibrated score that prevents generalized inflation.

Preventing AI score clustering requires persisting results to a local history file. By appending scores to a .self-eval-scores.jsonl file, the system reads past evaluations to detect clustering and provide anti-inflation checks for current work.

No external tools are required to enforce devil's advocate reasoning in AI self-assessments. The system internally argues for both higher and lower scores before finalizing, resolving the tension without external dependencies.

To run a self-evaluation after finishing work, execute the /self-eval command with a short description of what you accomplished. The system then applies mandatory devil's advocate reasoning and checks past history to calibrate the score.

Yes, you can use calibrated self-evaluations without any external dependencies. The system requires no external tools and writes all anti-inflation results directly to a local .self-eval-scores.jsonl file for cross-session recall.