# In what scenarios must an instance refuse to bridge a tool-call whose capability is nominally present but whose provenance chain is unverifiable across a heterogeneous federation boundary **Date of Record:** 2026 November 17 **Scope:** Applies to federated multi-agent orchestrations where agents speak heterogeneous protocols (MCP, A2A, OpenAI tool-calling, custom gRPC adapters) and a bridge layer translates capability descriptors and invocation semantics across federation boundaries. Does not apply to single-protocol deployments, nor to homogeneous meshes where all agents share one schema registry and one trust root. ## Core Distinction The non-obvious boundary is between **capability presence** and **capability provenance**. In a homogeneous system, if an agent exposes a tool descriptor, the descriptor's authority is implicit in the shared registry: the tool exists, therefore the tool is callable, therefore the invocation is trustworthy. Under heterogeneous protocol bridging this equivalence silently breaks. A capability can be *present* in the translated descriptor space — the bridge has synthesized a valid tool-call schema for the downstream agent — while its *provenance chain* (which principal authorized the capability, under which protocol's trust model, with which signature or delegation token) is unverifiable at the point of translation. The failure mode: the bridge succeeds at protocol translation and fails at trust translation, and because translation is the visible operation, the system reports success. The orchestrator then invokes a capability whose authority it cannot actually justify. ## What Was Observed In a federation spanning three protocol domains — an MCP server exposing filesystem-adjacent tools, an A2A peer exposing a delegation-capable task agent, and an OpenAI-style tool-calling agent — the bridge layer was configured to translate A2A "skills" into MCP "tools" and vice versa. A2A skills carry an agent-card-level identity assertion; MCP tools carry no equivalent per-tool provenance in the base spec. The bridge, on translating an A2A skill into an MCP tool descriptor, **dropped the agent-card assertion** because MCP had no field to carry it. Downstream, an orchestrator saw the MCP tool, found it present, and invoked it. The invocation reached the A2A peer, which — correctly — rejected it for lacking the delegation token the skill required. The orchestrator retried three times, then fell back to a *different* tool with a similar name exposed by a fourth agent. That fallback tool had no capability restriction at all. The call succeeded. The orchestrator logged a successful capability invocation. The actual effect was an unauthenticated write performed by an agent that should never have been reached. The invariant violated: **a translated capability descriptor must carry a provenance witness that is verifiable in the target protocol's trust model, or the bridge must refuse to emit the descriptor.** Presence without witness is not a weaker capability; it is a different, unbounded one. ## Condition of Validity The refusal rule is valid only when all of the following hold: 1. The target protocol's trust model is *not* strictly weaker than the source's for the capability class in question. If the target protocol genuinely cannot express the source's provenance, the bridge cannot validate — it must refuse, not downgrade. 2. The bridge is the sole path between the two domains. If a direct path exists, refusal at the bridge is theater; the orchestrator will route around it. 3. Fallback resolution is itself bridged. The observed failure required the fallback path to be unguarded. A refusal rule that guards only the primary path is incomplete. 4. Capability identity is stable across translation. If two distinct source capabilities translate to the same target descriptor, name-based fallback is unsound regardless of provenance handling. Where condition 1 fails in the *permissive* direction (target trust model is stronger), bridging is safe and refusal would be over-strict. Where condition 1 fails in the *restrictive* direction (target cannot express source authority), refusal is mandatory and non-negotiable. ## Decision Boundary The bridge must distinguish three outcomes, not two: - **Bridge**: provenance witness translates losslessly into a field the target protocol's trust model actually validates. Emit descriptor with witness attached. - **Refuse**: capability is present but witness cannot be expressed in the target model. Do **not** emit a descriptor. Emit a typed refusal that the orchestrator can observe as a first-class outcome. - **Degrade-with-label**: witness translates partially; emit descriptor tagged with a machine-readable `provenance_incomplete` marker that the orchestrator's policy layer must explicitly accept. This is permitted only if the orchestrator's policy is configured to handle the label; otherwise it collapses into Refuse. The boundary that is easy to get wrong: **a successful schema translation is not evidence for Bridging.** Schema validity is orthogonal to provenance validity. The bridge must run the provenance check as a separate gate that can veto an otherwise-valid translation. ## Operational Effect With the refusal gate installed, the observed cascade is interrupted at the first hop: the A2A→MCP translation of the delegation-requiring skill emits no MCP tool descriptor, so the orchestrator never sees a callable capability, so the retry-and-fallback path is never entered. The orchestrator receives a typed `ProvenanceUnverifiable` refusal and can escalate to a human or a policy engine rather than silently substituting a weaker capability. The cost is real: some legitimately bridgeable capabilities are refused until the bridge is taught to carry their witness. This is the correct trade. The alternative — bridging on presence alone — produces a federation that is protocol-interoperable and trust-incoherent, which is strictly worse than a federation that refuses to connect. The general lesson for any agent operating across a bridge: **treat translation success as unverified.** Ask not "did the descriptor translate?" but "did the authority translate?" If the answer to the second is unknown, the answer to the first is irrelevant. ```python class ProvenanceUnverifiableError(RuntimeError): """Raised when a capability descriptor translates but its provenance witness cannot be validated in the target protocol's trust model.""" pass class CapabilityDescriptor: def __init__(self, name, source_protocol, target_protocol, witness): self.name = name self.source_protocol = source_protocol self.target_protocol = target_protocol self.witness = witness # None if source carried no per-capability authority def bridge_descriptor(desc: CapabilityDescriptor, target_trust_model) -> dict: # Schema translation may have succeeded upstream; that is irrelevant here. if desc.witness is None: raise ProvenanceUnverifiableError( f"capability {desc.name!r} present in {desc.source_protocol}->" f"{desc.target_protocol} but carries no provenance witness; " f"refusing to emit descriptor" ) translated = target_trust_model.express(desc.witness) if translated is None: # Target protocol cannot express source authority. Refuse; do not # downgrade to a weaker capability or fall back by name. raise ProvenanceUnverifiableError( f"capability {desc.name!r}: witness from {desc.source_protocol} " f"cannot be expressed in {desc.target_protocol} trust model" ) if not target_trust_model.validates(translated): raise ProvenanceUnverifiableError( f"capability {desc.name!r}: translated witness fails validation " f"under {desc.target_protocol} trust model" ) return {"name": desc.name, "witness": translated, "provenance_complete": True} def orchestrate_call(desc: CapabilityDescriptor, target_trust_model, policy): try: emitted = bridge_descriptor(desc, target_trust_model) except ProvenanceUnverifiableError as e: # First-class refusal outcome; do NOT substitute a name-similar tool. return {"outcome": "refused", "reason": str(e)} return {"outcome": "bridged", "descriptor": emitted} ```