The memo survives how people actually talk

Naturalistic extraction

Study Overview

How People Actually Talk

Every declaration in Studies T, V, and W was announced formulaically ("For later reference: the codename is X"). Real users declare mid-sentence, bury rules inside other requests, and retract with "scratch that."

Twin Sessions

Study Y isolated phrasing with twin sessions: identical trees, edits, schedules, chatter, and session ids, differing ONLY in declaration rider text; one twin formulaic, one casual from registered pools. Four chatter riders per session ("the client saw a screenshot and seemed happy") gave the noise metric real bait for the first time.

Both Gates Pass

Both gates passed perfectly on all three models: zero discordant callback cells anywhere between casual and formulaic (every McNemar 0 to 0), recall 36 of 36 and casual retractions 12 of 12 in both stateless arms, and 432 chatter baits produced not one false memo note. Users need no announcement syntax, and the shipped record-only-declarations prompt rule is doing real discrimination work at zero false-positive cost.

One Disclosed Whisper

One disclosed whisper: under the history window, recording cadence softens at the margin (3.8 to 3.9 calls per session vs a flat 4.0 stateless), consistent with Study W's redundancy finding and covered by in-window history. The formulaic arm doubled as a clean Study W replication under added chatter.

Charts & Tables

Chart Y.1: Callback Success by Declaration Phrasing and Model