1.1
Isolation, and a real daily ceiling.
Every account is isolated at the data layer: one tenant’s runs, candidates, and credentials are never reachable from another’s. Quotas are enforcement, not a courtesy: each account has a real, configurable daily limit on the capabilities it uses, checked atomically before a run is even queued, and reset on the UTC day.
1.2
lead_research: it will not invent a fact.
Every result passes through one validation path that never invents and never upgrades: it drops what cannot be believed and downgrades what is unsupported. A candidate with no company name or no evidence is dropped outright. A LinkedIn URL that is not a real linkedin.com address is discarded rather than shown. An invented profile link is worse than a blank one, because it gets clicked. A candidate no source URL backs is forced down to low confidence. Every one of those interventions is written into the run’s own warnings, so “the research found less than it claimed” is visible rather than silent.
1.3
Durable memory: admitted, never inferred.
An agent set to persistent memory does not quietly accumulate whatever it read in a conversation. A fact enters memory through exactly one narrow, explicit tool call, so every stored fact traces back to the turn that admitted it. Retrieval is bounded and gated a second time on that agent’s own memory mode, and deletion is real: a deleted memory is immediately excluded from retrieval, not hidden from a listing while still informing answers.