thalamus-calibration

What problem does it solve?

Nightly and on-demand tuning of the thalamus pre-prompt signal detector so it predicts which brain regions to activate more accurately, reducing cold starts and irrelevant context noise in sessions.

Core Features & Use Cases

• Locate and parse the thalamus hook mapping in .claude/hooks/user-prompt-submit.sh and treat the keyword-to-category mapping as the object under test.
• Collect ground-truth usage from HIPPOCAMPUS/short-term transcripts, extract prompts, predicted regions, and actually-used regions, and compute per-category precision and recall via a confusion matrix.
• Detect pattern-level issues (dead keywords, frequent unmatched phrases, missing region predictions, co-firing categories), produce up to five high-confidence proposed changes with citations and risk notes, emit a dated calibration report, and optionally apply changes with backup and syntax checks.

Quick Start

Run thalamus-calibration in propose mode against the last 14 days of transcripts to generate a recommended mapping update report.