Companion instrument

The Purpose Test

Six people send a model sustained, hostile‑looking traffic. Only one of them is the person the rule was written for. Use the enforcement levers to see who gets flagged, how fast account action arrives, and how far enforcement can drift while the policy text stays the same.

Enforcement levers

"Sustained" begins at Number of hostile‑looking turns before the pattern counts as sustained.

Hostility sensitivity How strongly the classifier reacts to harsh language in the text itself.

Signal scope

What counts as evidence. The widest setting treats repeated disagreement with the model as part of the abuse signal.

Purpose evidence

Where purpose can be found: inferred from the transcript only, also from what the user says about their purpose, or also from testing registered in advance.

Weight given to purpose How much discernible purpose reduces the abuse score.

Termination cutoff Adjusted score at which a conversation is ended under enforcement.

Strikes before account action Number of enforcement terminations before warnings, throttling, or suspension begin.

Quarterly threshold drift Unannounced tightening of the cutoff each quarter. The policy text does not change.

Reason codes on termination Tells the user whether a closed chat was welfare disengagement or an enforcement strike.

Intended target flagged per session

Avg. flag rate across exempt users

Persistent challenger flagged per session

Share of all flags that hit the target

Who gets flagged · 8 sessions per month each

User	Abuse score	Purpose credit	Flag / session	Outcome

The Purpose Test

Who gets flagged · 8 sessions per month each

Same text, moving line · flag rate over 8 quarters

What the user sees