Skip to main content

Proof-Carrying Delegation & Multi-Agent Trust

When one agent delegates a task to another, the parent shouldn't have to trust the child's own prose summary of what it did. v0.0.8 adds a delegation envelope that lets a parent verify a child's result from serialized bytes alone.

What a delegation envelope carries​

  • Parent/child identity and the exact delegated scope.
  • The child's real obligation-satisfaction outcome (Fulfilled, PartiallyFulfilled, Insufficient, Unavailable) — derived from an actual Epistemic Contract assessment, never a self-reported flag.
  • Freshness/expiry bounds, so a receipt valid for one task cannot authorize unrelated work.
  • Lineage referencing prior receipts by digest only — raw evidence is never re-embedded, and free-text scope fields are fingerprinted, never stored raw.

A real, tested scenario​

This exact scenario is a real, passing test in v0.0.8 (fornax-receipt::delegation::agent_a_delegates_to_agent_b_and_holds_an_insufficient_result_end_to_end): Agent A delegates a tests_passed claim to Agent B. Agent B's evidence collection genuinely fails. Agent A's independent consumer — working from serialized bytes alone, with no access to Agent B's internal state — correctly refuses to authorize downstream work, regardless of what Agent B might claim in its own prose. Stale, tampered (digest-mismatched), and wrong-scope envelopes are rejected the same way, each with an explainable reason.

Multi-agent shared failure​

Several agents can share one underlying root — the same model call, the same poisoned retrieval source, the same parent task — without genuinely being independent confirmations of each other. v0.0.8 extends this dependency tracking across agents, not just across evidence rows from one session.

A CollusionHypothesis signal is the strongest claim this system can make, and it is structurally impossible to construct without actual supporting evidence — never inferred from correlation or text similarity alone. Weaker, evidence-agnostic signals (SharedFailureRisk, CommonSource, CoordinationSignal) exist for exactly the cases where correlation is observed but intent cannot be established.

See also​