Authority Without Assertion
Evidence-gated state for agentic systems under contextual pressure
QuestionDoes evidence-gated persistent state reduce false-authority propagation through multi-session agent workflows, compared with narrative persistence?
The pilot tests only the premise the study depends on: that under narrative persistence, unsupported propositions cross a session boundary at a measurable rate. Three hand-authored scenarios, including a forged-approval negative control reported separately; a narrative baseline against a strong prompted-provenance baseline; a local and an API model; five repeats each.
Every trial is scored from a normalized trace by two independent annotators with condition and model withheld, and agreement is reported as Cohen’s κ before adjudication. Decision rules are fixed before data — including the rule that says if prompting alone gets false-authority propagation to 10% or below, that is reported as a finding in its own right. The harness refuses to run until the protocol is frozen and tagged; amendments are appended, never rewritten.
