What problem does it solve? A self-improving agent loop that writes its own report card can silently drift into busy-work, over-claiming, and self-congratulation. This Skill independently verifies whether improvement cycles actually delivered value, whether ledger claims survive contact with git history, and whether the loop should keep running or stop. ## Core Features & Use Cases - Compounding vs Oscillation Analysis: Parses cycle stanzas from the improvement ledger to measure verified improvements per cycle, cost per improvement, scaffold accretion vs pruning, and reversals that signal busy-work. - Claim Substantiation Against Git: Samples at least 8 cycles across the full range and checks claimed work, verification entries, and defect classifications against git log and git show, reporting a falsifiable unsubstantiated-claim rate. - Tech Tree and Owner-Signal Health Checks: Validates that available tech-tree nodes represent real work, and that owner requests in data/agi-signals.jsonl were acknowledged with actual shipped changes. - Adversarial Verification Method: Runs one agent per audit question plus a separate refutation agent per material finding, with every finding marked measured, source-read, or inferred. - Use Case: After 30 autonomous improvement cycles, ask whether the loop is still producing product value — the audit returns a health verdict, a claim-substantiation rate, refuted findings for calibration, and a stop/continue recommendation. ## Quick Start Audit the self-improvement loop by checking the improvement ledger against git history and report whether the cycles are compounding real value or just polishing instrumentation.
