AI and provider disclosure
Effective August 14, 2026. OnMe uses deterministic rules first and keeps every model call behind the same Track permission and spend policy.
Sieve
OnMe makes zero calls with OnMe provider credentials for Sieve accounts. If you connect Codex, Claude, Cursor, or another certified host, that host uses your own subscription, tools, authorization grants, and provider terms. OnMe receives only the structured verdict, evidence references, delivery directive, and decision trace the host submits. On supported certified iPhones, bounded judgment may run locally. Private feedback weights and content-free numeric examples stay in the device Keychain. A versioned personal ranking algorithm may train and run locally only after its chronological holdout materially beats deterministic calibration; it returns one bounded score and can be rolled back to deterministic ordering. It cannot request delivery or change Track, project, privacy, spend, or write authority.
Hosted
Hosted may process permitted evidence through production models from OpenAI or Anthropic and may use a permitted web-search provider when the Track’s broader-web toggle is on. Candidate models do not receive traffic until they pass OnMe’s frozen extraction, privacy, authority, structured-output, accuracy, and cost evaluation. OnMe records the provider, immutable snapshot, token and tool usage, latency, cost, Track, and routing reason for each call.
Limits and approvals
Hosted provider and tool spend is capped at $5 per customer billing period. OnMe does not silently exceed that ceiling. A customer host may continue the work, otherwise OnMe returns needs_executor. AI cannot grant jar or project permission, increase spend authority, or perform calendar, Slack, or other external writes without explicit approval.
Training and reliability
OnMe does not train cross-user models on customer content in v1. Model outputs can be wrong or incomplete; high-impact decisions should be checked against visible evidence and primary sources. Provider availability and retention practices remain subject to their applicable service terms and OnMe’s configured privacy settings.