Follow-on research synthesis — owners, evidence, and non-equivalent boundaries

This document closes WI-028. It compares the eight completed extractions in their three tracks, then reconciles the overloaded vocabulary that appears across products, evaluators, optimizers, and protocols. It is a synthesis of pinned evidence, not a ranking, benchmark, or implementation plan. The canonical pattern catalogue is changed only when a reconciliation row supports an existing definition; no new catalogue entry is admitted by analogy alone.

Method and provenance

Documented — The synthesis applies WI-019’s shared spine, product template, evaluation template, and protocol template. Track comparisons come before cross-track claims. Every source claim below is routed through one of these extraction records, each of which freezes its own authority and read boundary:

Boundary — “Confirmed” below means that the candidate matches an existing catalogue meaning under a comparable evidence boundary. “Track-specific” means useful vocabulary whose owner or guarantee cannot be promoted to the shared catalogue from this corpus. “Contradicted” rejects the candidate as a safe positive rule. No score, adoption claim, or cross-source performance claim is introduced.

Within-track comparison: durable/local products

Within-track conclusion — Opinion — The product subjects converge on explicit ownership and durable boundaries, but not on one “session” or “memory” primitive. Letta’s agent-owned memory, Aider’s repository candidate, goose’s session isolation, and ADK’s event/confirmation model are commensurable as boundary declarations; their state stores and recovery guarantees are not interchangeable.

Within-track comparison: evaluation and optimization

Within-track conclusion — Opinion — Inspect AI is primarily an evidence-producing evaluator; DSPy is an optimizer whose metric changes a program artifact. Their typed scores, logs, and limits are comparable as evaluation-contract fields, but a score from one cannot prove the other’s optimization claim. Both reject the shortcut “a passing score is a trustworthy general improvement.”

Within-track comparison: interoperability protocols

Within-track conclusion — Opinion — MCP and A2A are commensurable at the level of negotiated capability, explicit lifecycle, error, and trust boundaries. They are not substitutes: MCP centers host-mediated capabilities and context; A2A centers independent-agent task ownership and artifacts. Normative text establishes requirements, not conformance of an arbitrary implementation.

Cross-track boundary table

Contradiction — Opinion from equivalent boundaries — None is claimed between equivalent normative/implemented statements. The apparent contradictions (“session is durable” versus “session is in-memory”, or “approval is authentication” versus “approval is a pending state”) disappear once the owner and evidence type are restored; treating them as one claim would be a category error.

Candidate adjudication

Each row below is one candidate from exactly one extraction. The evidence column links to the extraction’s pinned catalogue-consequence section; the provenance register above supplies the full SHA and authority type.

WI-020 — Letta Code

WI-021 — Inspect AI

WI-022 — Aider

WI-023 — goose

WI-024 — Google ADK

WI-025 — MCP

WI-026 — A2A

WI-027 — DSPy

Catalogue reconciliation and downstream boundary

Opinion — No edit to pattern-catalog.md is warranted by these rows. Confirmed candidates map to existing definitions (bounded-agent-loop, event-stream-not-audit-log, agent-facing-interface, external-authority-precedence, ownership-changing-handoff, prompt-writes-artifact, protocol-with-escape-hatch, structural-gates, typed-output-gate, explicit-output-designation, and the resumable-approval-state/approval-bound-to-request pair). The track-specific candidates are deliberately not promoted: doing so would turn one product’s storage or one protocol’s lifecycle into a guarantee for another track. The single contradicted candidate, test-substitution, is retained as a warning in the DSPy extraction and is not added to the catalogue.

Boundary — No module or CLI change follows from this synthesis. Existing WI-029 remains the only implementation-facing follow-up and is not silently expanded; any future adoption of follow-on evidence must be a separately planned work item with its own catalogue/ADR decision.

Findings and final conclusions

1. Confirmed across tracks — Opinion. Explicit ownership, bounded loops, structured outputs, and evidence retention recur, but only when the owner and guarantee are carried with the noun.
2. Track-specific — Implemented/Normative. Durable memory/session stores, evaluator scores and compiled demos, MCP capabilities, and A2A tasks/artifacts cannot be merged into one state model.
3. Contradicted shortcut — Opinion. A metric or event stream is not a held-out test or audit trail merely because it is durable or typed.
4. Strongest counter-evidence — Implemented/Normative. Letta memory commits, Inspect crash recovery, DSPy trainset reuse, MCP host authority, and A2A duplicate delivery each leave a different unrecovered boundary. No subject supplies a universal exactly-once, authenticated, human-accountable agent loop.

The synthesis therefore confirms the catalogue’s boundary-first vocabulary, keeps new nouns track-specific unless a current definition already covers them, and closes without changing modules/, CLI code, or the canonical catalogue.