Skip to content

Clarify the trust scope of MCP gateway policy receipts #142

Open

Description

@RobLe3

Question, not an implementation defect

Clarify the intended trust scope of the MCP gateway's policy_receipt. This is informational review, deferred until after pre-1.0 qualification. No signing change is currently required or authorized.

The Correctover / CCS report dated 2026-09-28 tested Python iicp-client / iicp-node 0.7.109 from PyPI, not the initially preferred Rust CLI. The unmodified gateway exposed format_json,summarize_text with a local Directory stub/local MCP server and an external verifier 1.3.0. This is bounded execution-assurance composition, not complete IICP discovery/selection/dispatch-ticket interoperability or qualification credit.

Source baseline

Release v0.7.109 resolves to 7718dd35904568f3a10777a11ee3a3ffeeeaffab.

  • McpToolPolicy.receipt emits structured, content-minimized policy metadata; it has no detached signature, signer, expiry or nonce.
  • Gateway response places task_id beside the successful result and policy receipt. Correlation is not itself authenticated attestation of execution.
  • Rust and TypeScript expose related redacted policy receipts. They are parity consumers, not three independent owners of a new shared semantic contract.
  • The IICP specification owns binding/trust semantics; this repository owns the implementation used in the experiment. Generic composition is tracked in Track composable external execution-assurance evidence across IICP bindings IICP#256.

Review questions

  • Is the current receipt local/transient diagnostic evidence, application-visible policy-decision evidence, externally verifiable authorization evidence, or an intermediate input to a later artifact?
  • Which claims can a caller reasonably rely on today? What is outside the receipt's intended trust scope? The report's broader suggestion that both receipt models were detached-signed contradicts the tested policy receipt and must not become documentation.
  • Do existing task/attempt identifiers or signed IICP artifacts already provide the needed correlation/evidence? Would a later completion/provenance artifact be a better location than changing this object?
  • Should receipts remain unsigned? Only an independently justified trust requirement can lead to a separate signing proposal; external CCS signatures are not a reason by themselves.
  • If signing is later justified, review signer/key authority, lifecycle/rotation, canonicalization, nonce, freshness/expiry, replay resistance, immutable correlation, privacy, verification scope and shared SDK behavior. Determine Profile versus binding-local or implementation-local scope before implementation.
  • Preserve exclusion of tool arguments, prompts, credentials and responses. Avoid treating digests or a signed issuer assertion as proof of confidential execution or the truth of the assertion.

Acceptance and revisit trigger

After qualification, revisit on an accepted trust-model review or a concrete external relying-party need. Publish an as-is trust-scope decision, reuse existing architecture, and document any remaining requirement separately. A reviewed decision to keep the receipt unsigned is a valid outcome. Any future cross-SDK requirement belongs in the specification and needs separately reviewed owning-repository changes.

Do not change the current candidate, support matrix, wire contract, qualification state/credit, gateway code, vocabulary or signatures in this task. Do not add CCS packages, import ELv2 source, schedule a new experiment, or infer certification from the externally reported 14/14 case outcomes and 415/426 checker checks.

Evidence record

Sanitized dated record is available in draft documentation PR #257, pending merge and normal review/validation closure. It retains report-only results and unavailable reproduction artifacts; no qualification credit or implementation requirement is introduced.

Activity

  1. added
    documentationImprovements or additions to documentation
    questionFurther information is requested
    on Sep 29, 2026
  2. RobLe3 commented on Nov 6, 2026

    @RobLe3
    OwnerAuthor

    6 November 2026 — as-is source review, no signing change

    The McpToolPolicy.receipt methods at reported release 7718dd35904568f3a10777a11ee3a3ffeeeaffab and qualified candidate 9c5b3d894ee89ef9fcde52ea0ad99d388ed62183 have identical ASTs. The qualified implementation emits selected policy-decision metadata, not a detached signature, signer, expiry or nonce. Selected strings have character limits; this is not a proven total serialized-size bound.

    The qualified gateway caller includes the receipt on policy denial and alongside task_id and the successful result. Adjacency supplies application correlation, not authenticated execution attestation by itself.

    Independent bounded source review confirms this interpretation. No gateway or CCS experiment was executed, and no signing behavior changed. Enclosing channel authentication or a separate signed artifact could supply distinct assurances; those compositions have not been established by this review.

    This advances the as-is evidence question but does not decide future signing requirements, create a cross-SDK contract, resolve RobLe3/IICP#256, or close this issue. The remaining step is a Protocol-owned relying-party/trust-scope decision and a source-linked review of any existing correlated signed artifacts. No accepted qualification evidence or credit changed.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    documentationImprovements or additions to documentationquestionFurther information is requested

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions