Three verticals Picaroon retains audits against — ecommerce refund, SaaS support, and booking. Each row is one retained client, anonymized to alias + ARR band + agent class. The headline metric is % scenarios where the agent’s answer degraded versus the prior monthly run: a single-number proxy for whether the latest prompt, model, or KB ingest moved the agent forward, backward, or sideways since the last cycle.
Each section below reads the same six-month window for one vertical. The line shows the vertical-aggregate degraded%, the table lists the latest-month row per client with a month-over-month caret.
A scenario is counted as degradedwhen its rank on a fresh run fell compared to its rank on the immediately prior monthly run against the same agent version. The metric is the percentage of all scenarios run that fall into that bucket — a single-number latency indicator for whether the last cycle’s prompt, model swap, or KB ingest moved the agent forward or backward.
Each row is shown as an alias plus an ARR band plus an agent class. No client name, no agent model, no traffic volume, no scenario titles — just the bucket the underlying customer sits in. The shapes are public; the identities aren’t, so a competitor can read the trend without reverse-engineering the client list.
The line shows the vertical-aggregate degraded% averaged across every retained client in that vertical. A flat line means the latest cycle held the prior rank; a step down means more scenarios maintained or improved their rank; a step up means the cycle regressed versus the prior month.
The table reads the latest month per client. The MoM column shows the change against the prior month on that same client; a ▾ marks a step down (better), a ▴ marks a step up (worse), and a flat row means the cycle held prior rank within the rendering tolerance.
The full scoring taxonomy — OWASP LLM Top 10, the CSA Risk Rubric v2, and the Picaroon vertical catalog every finding ships with — lives on the rubric page.