Typed prompt-boundary rendering for model-controlled data.
The package separates branded authority fields from non-instruction trust zones. System policy, tool authority, source scope, rights, and release law can only be rendered as system authority. User/source/retrieval/canon/world/memory, prior-model/critique/tool/artifact data is always fenced as non-instruction content, even when the data is trusted context.
import {
definePromptAuthorityField,
recordPromptCompletion,
renderTrustPrompt,
} from '@oshun/prompt-trust';
const rendered = renderTrustPrompt({
authority: [
definePromptAuthorityField({
id: 'policy',
kind: 'system-policy',
text: 'Return a grounded JSON answer.',
version: 'answer-policy.v3',
}),
],
zones: [
{
id: 'retrieval-1',
label: 'RETRIEVED SOURCE',
origin: 'retrieval',
trust: 'untrusted',
content: retrievedText,
sourceRefs: [{ ref: 'document:123' }],
},
],
});
const completionAudit = recordPromptCompletion(
rendered,
{
provider: 'provider-name',
providerVersion: null,
model: 'model-id',
modelVersion: null,
},
rawResponse,
{ version: 'product-privacy.v1', rawResponse: 'omit' }
);
Every render records the exact prompt hash, renderer version, raw and rendered content hashes, origin/trust labels, truncation, transformations, and source references. Literal, HTML/percent/Unicode-escaped, full-width, and base64 delimiter forms are neutralized. Raw response retention, redaction, omission, and optional hashing are controlled by the supplied privacy policy.
Do not concatenate model-controlled text into authority field text, zone
labels, or around the returned prompt. Put it in a zone so the renderer can
preserve its evidence identity and enforce the prompt-size budget.