# Evidence-to-language rulebook

## Core discipline

Write the strongest sentence supported by the weakest necessary inference. Every public sentence should be reproducible from structured evidence fields.

## Terms

### Contacted

Use when a connection from the tested app to a destination was attributed and observed.

Do not substitute: “shared data with”, “sent personal data to”, “tracked by”.

### Transmitted

Use only when the payload or an exact synthetic value was observed leaving the device in the request, or when another direct source establishes transmission.

State the observed field, not a broader category:

- good: “the synthetic email address was present in the request”;
- poor: “personal data was harvested”.

### Recipient

Use only when the legal or operational role is documented or confirmed. Infrastructure operation alone is insufficient.

### Not observed

Always attach scope: “not observed in the two tested journeys on version X”. Never write “does not happen”.

### Optional

Treat “optional” as a claim to test. Record whether refusal was available, equivalent and compatible with the core service.

### First party / processor / third party

These are different dimensions:

- first party describes relation to the service;
- processor is a legal role requiring evidence;
- third party may be technical, contractual or independent.

Never infer one automatically from another.

## Headline gate

A headline must contain or be immediately followed by:

- exact state or journey;
- strongest material limitation;
- calibration warning when synthetic.

## Claim review questions

1. What precise observation makes the sentence true?
2. Could a different technical explanation fit the evidence?
3. Does the wording imply motive, legality or identity not established?
4. Is a store declaration being mistaken for an independent platform audit?
5. Is absence being turned into impossibility?
6. Is an SDK’s presence being turned into active use?
7. Is a cloud provider being turned into an independent recipient?
8. Would the same rule produce a publishable good result?