DISCOVERY — Next Evidence Wave (2026 Q2)¶

• Status: Historical discovery note; preferred candidate is now formalized as PLAN — K1, the first bounded K1-A adapter slice is now public in v3.4.0, and the next bounded MCP evidence line now lives in PLAN — K2 with K2-A — Phase 1 formal freeze plus an implemented first slice on main
• Date: 2026-03-25
• Owner: Evidence / Product
• Scope (this document now): Historical ranking input only. The formal next-wave choice now lives in PLAN — K1.

1. Decision frame¶

This note originally did not formalize the next wave. It recorded the candidate surfaces after G4-A Phase 1 and P2c and applied the same discipline as earlier waves: evidence reality first, productization later.

That choice has now been made: the preferred candidate below is formalized as PLAN — K1. This note remains the discovery record that justified that choice.

Two-question filter¶

Working implication: A new pack is not justified without both evidence reality and a bounded claimset. Yes / yes is plausible for a future P-slice, not a mandate to ship one.

Why no ROADMAP change in this PR¶

This note is discovery-only and does not commit maintainers to the next formal wave; updating ROADMAP here would imply a strategic decision that belongs after explicit maintainer choice.

2. North star¶

The next meaningful step after P2c is not automatically another pack. It is whichever surface has the strongest case under the filter above.

Assay should continue to prefer:

• Adapter-emitted, bounded seams
• Typed / canonical evidence over blob interpretation
• Presence / visibility claims over validity or trust claims
• One surface at a time
• New protocol topic ≠ automatic pack — evidence reality and an honest claimset come first (same seam ≠ same pack discipline)

G4-A-PHASE1-FREEZE.md remains the reference pattern for how a later freeze should look once a surface is chosen and implemented.

3. Repo reality at time of discovery¶

From the current A2A adapter implementation:

• Canonical event mapping is in crates/assay-adapter-a2a/src/adapter_impl/mapping.rs: agent.capabilities, task.requested, task.updated, artifact.shared; unknown upstream types map to a generic path (assay.adapter.a2a.message — see convert.rs).
• At the time of discovery, first-class A2A surfaces in shipped evidence were primarily capabilities, task lifecycle, artifact exchange, and (after G4-A) discovery / card visibility via payload.discovery (discovery.rs).
• At the time of discovery, there was not yet a first-class, bounded handoff / delegation-route seam in adapter output (no dedicated mapping or typed subobject for that route in the codepath then under review).
• G3 applies to assay.tool.decision authorization context (g3_authorization_context.rs) — not MCP authorization-discovery / resource-metadata as a separate bundle seam (see § Candidate 2).

These repo facts were the primary basis for the candidate ranking below. They are now historical context; K1-A Phase 1 has since added the first bounded top-level handoff seam on main.

4. Candidate 1 — A2A handoff / delegation-route visibility¶

Status: Preferred candidate for the next evidence wave (not a pack wave).

Why this surface matters¶

A2A’s public framing emphasizes agent-to-agent collaboration, task management, and interoperability. Recent community discussion also highlights ambiguity around transfer / handoff semantics — a signal that protocol value can exist before clean evidence seams do.

Why this is not “more G4” or “more P2c”¶

This is a different surface from discovery/card visibility. It must not be treated as:

• an automatic extension of payload.discovery
• an automatic extension of P2c
• a reason to append rules onto existing discovery/card packs (“same seam ≠ same pack”)

Current gap¶

At the time of discovery, the adapter’s mapped evidence did not expose a bounded, typed handoff / delegation-route seam. That bottleneck is why the path chosen here was evidence-shape first (adapter-first), not pack productization.

Likely outcome under the filter¶

Discovery questions¶

• Which upstream A2A flows or fields carry handoff/delegation-route information?
• Which are typed and stable vs blob-like or lossy?
• What is the minimum honest seam (one small subobject, one bounded visibility set, etc.)?
• Explicit non-goals: no trust score, no provenance “success”, no correctness of delegation outcome as an observed claim unless evidence supports it.

5. Candidate 2 — MCP authorization-discovery¶

Status: Second candidate.

Why this surface matters¶

The MCP authorization line requires OAuth 2.0 Protected Resource Metadata and authorization-server discovery as part of the protocol story — authorization-discovery is a real protocol surface, not a minor implementation detail.

Product guard: Even if the MCP spec makes this surface important, it is not a next-wave candidate unless canonical Assay evidence already exposes it or a dedicated evidence-wave is explicitly chosen. Spec relevance ≠ bundle reality.

Critical contrast with G3¶

G3 is not this surface.

Current gap¶

The spec can make the surface important; that does not prove Assay bundles already emit it as bounded canonical evidence.

Likely outcome under the filter¶

Discovery questions¶

• Where would PRM / AS-discovery appear in current Assay bundles (event types, payloads)?
• Is it already first-class and typed, or only implied by spec-level expectations?
• Could a bounded claimset exist without inventing new semantics?

6. Candidate 3 — Trust / signed-card provenance¶

Status: Not a candidate for the next wave.

Public discussion and research emphasize signing, verification, and spoofing risk — visibility evidence (as in G4-A / P2c) can be honest; trust / provenance claims without bounded signals are high-risk theater.

Conclusion: Park until there is bounded first-class evidence, a non-theatrical claimset, and alignment with G4 / P2c non-goals (no reopening signature semantics as product claims).

7. External context (non-normative)¶

This section is second-order motivation only. It does not define Assay scope.

• A2A positions discovery, collaborative tasks, and secure collaboration as core protocol surfaces; official samples stress treating external agent data (including Agent Cards and task/artifact content) as untrusted input where applicable.
• A2ASecBench-style work highlights discovery/card spoofing and related protocol attacks as practical risks — supporting careful visibility evidence, not premature trust marketing.
• NIST / ecosystem work (e.g. agent identity, authorization, auditing) favors observable controls over marketing-style “trust” claims without evidence.
• MCP authorization work makes authorization-discovery (Protected Resource Metadata, authorization server locations) a protocol-level requirement in current spec direction.

Hard rule: External sources explain why surfaces matter; bundle evidence reality in this repo determines what Assay can honestly ship.

8. Provisional decision¶

9. Next steps¶

1. Done: Candidate 1 is now formalized as PLAN — K1.
2. Done in this slice: the next execution artifact is now recorded as K1-A — Phase 1 formal freeze, with explicit source mapping, emitted examples, and no implementation in the freeze slice itself.
3. Done: the first bounded adapter-first implementation slice for K1-A is now merged on main as the top-level A2A handoff seam, still without a pack wave.
4. Any further K1 widening remains a separate maintainer decision; no downstream pack follows automatically from this slice.
5. The next formal wave after K1-A stabilization now lives in PLAN — K2, and its active contract now lives in K2-A — Phase 1 formal freeze; the first bounded K2-A slice is now implemented on main, and this note remains the historical discovery input that ranked the earlier post-P2c candidates.

References (repo)¶

• PLAN-G4-A2A-DISCOVERY-CARD-EVIDENCE-2026q2.md
• G4-A-PHASE1-FREEZE.md
• PLAN-P2c-A2A-DISCOVERY-CARD-FOLLOWUP-PACK.md
• RFC-005 §6 — sequencing; review gates for future execution
• mapping.rs (A2A event mapping)
• g3_authorization_context.rs