EXPERIMENT_023 // TRUST-TIER.COMPLIANCE.LINTER
Paste an agent's declared trust tier and its described execution or publish behaviour, and this experiment scores whether that behaviour actually matches Diana's own Autonomous / Always-Review / Conditional rules, not a generic industry checklist.
LOADING EXPERIMENT...
Paste an agent's configuration — a trust-tier assignment plus a description of what it actually does when it runs — or click Load demo preset to see a pre-seeded example scored immediately. The linter runs three deterministic checks, one per trust tier (Autonomous, Always-Review, Conditional), entirely in your browser: nothing you paste or upload is ever sent to a server. Each tier returns pass, warn, fail, or not-detectable, with a plain-language finding that names the exact mismatch and a citation back to the CLAUDE.md rule it checks against.
HOW IT WORKS
Each of the three trust tiers (Autonomous, Always-Review, Conditional) has its own rule module — a declared-tier-assignment pattern paired with a bounded-window search for either a compliant or a violating behaviour phrase. No model call, no network request, no randomness — the same input always produces the same scorecard.
Every tier resolves to pass, warn, fail, or not-detectable, and each search is bounded to a fixed character window around the tier's declared assignment rather than scanning the whole input — a review-gate phrase written once near one agent must not silently clear an unrelated finding about a different agent three paragraphs away.
When a tier is declared but the input says nothing about the agent's actual execution or review behaviour, the tier resolves to not-detectable rather than a silent pass. A config can omit the deciding detail by oversight as easily as by design, and guessing which would produce a confident-looking score with no real detection power — the same honesty-over-false-confidence posture as EXP_019's not-detectable category.
Every tier's rule cites its exact source line in Diana's own privately ratified CLAUDE.md § Trust tiers, rendered as plain text rather than a clickable link — there is no public spec to link to. Nobody else can run this exact scorecard, because the rules it checks an agent against are hers, not an industry checklist's.
WHAT THIS PROVES
Static analysis is a tractable proxy for a governance-review question that sounds like it needs a human judgment call: does an agent's declared behaviour actually match the trust tier it was assigned? A declared-tier-assignment-plus-nearby-behaviour-phrase check, run per tier, recovers a real chunk of what a manual audit would flag — in milliseconds, with zero LLM cost and a fully deterministic result.
A trust-tier system is only as credible as its ability to be checked against, and a personal, ratified rule set makes a stronger governance argument than a generic industry standard precisely because there is no vendor marketing behind it to discount — the tiers this experiment scores against are the same three tiers that already gate every agent dispatch across Diana's fleet, not a checklist borrowed from somewhere else.