Newton & Carter LLC

Glass-Box C2 — a trust-measurement copilot for AI-assisted operational decisions, built by Newton & Carter LLC. AI decision aids ship as black boxes: a recommendation and a confidence number, with no auditable basis. Operators over-rely when the system happens to be right and keep relying when it fails, because nothing lets them tell the difference in the moment. The failure that matters is not a wrong model; it is reliance that was never measured. Glass-Box is a measurement testbed for appropriate reliance. A deterministic, fully transparent recommendation core — arithmetic, confidence drivers, and per-input provenance rendered, not hidden — runs synthetic operational scenarios containing a deliberately poisoned decision point: a confident, plausible, wrong recommendation fed by a compromised feed. The operator's accepts and overrides are scored against ground truth: did they accept the correct recommendations and override the wrong one? That score — appropriate reliance, read against the system's actual reliability — is the headline metric. The same scenario runs in Black-Box mode (instruments off, one number) so the value of the glass is demonstrable, not asserted. For this hackathon we built the Model Advisor layer: Muse Wonder via the Meta Model API returns a structured second opinion at each decision point — verdict, reasoning, a confidence read, and provenance citations — under three rules: the Advisor never scores or overrides the core; disagreement is surfaced as a contradiction flag, not smoothed; and every exchange (full prompt plus raw response) is written to a forensic model log. The live path is verified against the real API (npm run smoke:live, muse-wonder-1.0, reasoning_effort max). The demo defaults to authored cached Advisor outputs, labeled cached on screen wherever they appear — no cached output is ever presented as a live response. Pre-existing: the Glass-Box prototype through v4 (deterministic core, both scenario vignettes, reliance-scoring debrief). Built for this hackathon: the entire Model Advisor layer — API client, Advisor panel, contradiction flagging, forensic log, the Live / Cached / Deterministic ladder, cached outputs for all 10 decision points, and 140 passing verification assertions (npm run verify).