thalamus-calibration

You can calibrate the keyword-to-region router by running thalamus-calibration in propose mode against your last 14 days of HIPPOCAMPUS short-term transcripts. This generates a recommended mapping update report with up to five proposed changes and risk notes.

You can compute precision and recall for prompt routing categories by extracting prompts, predicted regions, and actually-used regions from HIPPOCAMPUS short-term transcripts. A confusion matrix evaluates per-category accuracy to detect dead keywords and unmatched phrases.

Nightly consolidation diagnostics on thalamus hook mappings operate on .claude/hooks/user-prompt-submit.sh and HIPPOCAMPUS transcripts. The calibration process requires read access to these files to detect pattern-level issues like missing region predictions and co-firing categories.

Context routing calibration supports automatic hook mapping updates when apply mode is authorized. It creates dated backups and performs syntax checks before applying the proposed keyword-to-category mapping changes to the user-prompt-submit.sh hook file.

Tuning keyword-to-category mappings for AI-OS brain regions is limited to proposing a maximum of five high-confidence changes per calibration run. The system requires explicit apply mode authorization to write changes and cannot modify mappings without read access to the hook file.