| Tool | Job (one line) | Risk it touches | Evidence it produces | Prove now? |
|---|---|---|---|---|
| Michelangelo | Stop / route before bad use; immutable runtime governance | Model failure, output liability, regulatory, agent mistakes | Checks ran · stayed in intended use · exceptions handled | WALKTHROUGH demo + story strong; match claim to what you can show live |
| Truth Serum | Run a document against stated rules; score each claim / clause | Hallucination, output liability, policy / reg conformance | Verified · needs human · unsupported vs company’s own AI policy (C) | RESURRECT ~2 DAYS first demo = C; A/B later if needed |
| TokenScope | Watch usage / load / behavior across the wire (providers) | Drift, reliability pattern, cost, “was this anomalous?” | Quantitative session / usage signals (not mind-reading the model) | PARTIAL usage / habits real; “fabrication rate” only if defined |
| Ghost Scope | Canary strings — did sensitive text leak into the wild? | Data exposure / confidentiality | Canary planted · found / not found on public web | PARTIAL GhostToken-style canary exists in lab; name carefully |
| Raquel | Compress / reshape language economics of keeping governance on | Cost of strong controls (not a liability cover by itself) | Token reduction measurements (yardstick tests) | LAB honest % only from clean same-doc tests |
| BugBash / dragons | Watch context infection / aberration patterns (chapter 2) | Session failure modes that precede bad outputs | Symptom log · tipping signals (early) | PARKED training / R&D — not on insurance page yet |
Pink House Technology · Phil private mantra sheet · 2026-08-03 · Suite story: continuous-evidence layer · industry problem first · Pink House last.