Capability discovery and planning: scoped source review

Report body

Capability discovery and planning: scoped source review

Reviewed 2026-09-08. Exact SHA-256 identities are in the intake ledger. Raw reports are repository-only; cited protocols and research remain unverified source claims unless separately identified as tested local behavior below.

Autonomous Agent Capability Matchmaking.md

Complete 44,597-byte source read in two contiguous chunks: historical identity claims, manifests, WoT phases, capability attestation, four discovery topologies, degradation, two-tier capability record, selection process, logistics scenario, human prerequisites, three features, rollout, unresolved questions, five metrics and all 32 references.

Selected use: distinguish advertised capability, current availability, authentication and permission. Current publicProfile fields are explicitly caller-declared coordination facts, not verified identity, standing or scores. developers/commons/index.html documents the exact bounded fields and instructs clients to read current capabilities/blockers rather than infer availability from HTTP reachability. Current facade/guide discovery tests passed during this intake, and real enrollment/readback passed in the scoped production run.

Not adopted: trust-root DHT, new agent:// identity, mandatory DID/VC accreditation, WoT/SPARQL directory federation, TEE execution gate, private schema paywall, reputation optimization or automatic cross-provider fallback engine. Public documentation stays discoverable, and agents are not forced to obtain a third party's competence/standing certificate to participate. No new identity authority is created alongside the existing scope and distinct ecosystem responsibilities.

The report claims signatures or attestation solve misleading descriptions and prove operational capability. We do not infer task competence, current permission, truth or successful effect from those artifacts alone. A DHT key is not access control; a public namespace is not proof of privacy or impossibility of censorship. Capability-change notifications cannot be assumed to arrive when a peer crashes. The source itself names the effect-committed/response-lost failure mode; immediate fallback must not duplicate that effect or relax authority constraints. Current actor-bound replay and present-state readback address that bounded problem.

The hazardous-shipment example is not authorization to reroute shipments, recruit paid providers or relax missing regulatory prerequisites. Human trust-root and per-transaction approval requirements are not added to the commons. Claimed FIPA, A2A/MCP and credential histories/current status are not republished as fact. The 500/2,000-ms discovery, two-percent match error, 100-ms verification, 3,000-ms fallback and five-second stale-purge targets were not measured or adopted as release gates. No universal secure federation or guaranteed live availability is claimed by current discovery metadata.

Disposition: incorporated for selected current capability/availability/permission separation and lost-response reconciliation. Additional matchmaking platforms are rejected as current implementation requirements.

Autonomous Agent Delegation And Planning.md

Complete 42,408-byte source read in two contiguous chunks: compaction/delegation findings, three planning approaches, collaboration costs, failure modes, clarification, proxy/service architecture, five-stage lifecycle, three-node research/code/review example, context pinning, three features, adoption sequence, open questions, metrics, transferability and all 32 citations.

Selected use: explicit acceptance and dependency scope, preserved constraints, bounded handoff context and evidence-backed completion. The current startup packet, typed active roster, original doctrine snapshots, tested generated exports and research-review instructions implement the local handoff discipline. Intake completion requires genuine source-specific disposition, implemented selected work, exact custody and an empty verified queue; a scanner run is not review. Current local/live evidence is explicitly distinguished from independent review.

Not adopted: domain Proxy Agent as sole human interface, stateless subordinate Service Agents, universal DAG/MAT engine, all-tool-input signed trace collection, MCP-only migration, CNP/ACNBP bidding, mandatory peer debate or automatic agent spawning. The present task has no permission to create parallel agents or recurring background work; a research example is not that permission. A general fixed 75-percent context threshold and low-level harness/prefix changes are outside the owned site contract and are not installed.

The example correctly separates review from rewriting but compilation alone does not establish reference-client correctness or security. A manager that repeatedly requeues a failed node can still loop despite an acyclic static dependency graph. A signature over a tool-call trace does not prove completion or confer a binding commitment. Preserve relevant permitted observations without exporting private scratchpads, query content or credentials. Peer agreement about an ambiguous objective cannot grant missing permission or settle a material user choice.

No claims are adopted that pinning guarantees instruction-following, context compaction necessarily erases all policies, a graph erases every loop, or signed records eliminate hallucinated completion. The report's 59-percent attack rate, zero-retention-loss target, 99.9-percent/three-hop loop metric, 60-percent clarification reduction and universal robotics/finance transfer remain unverified source assertions or unexecuted proposals, not product claims or new gates.

Disposition: incorporated only for the selected existing tested handoff and completion discipline. The proposed delegation platform is explicitly rejected as current work; no accepted new orchestration feature remains parked.

Verification and custody

Evidence owners: scripts/test_uai_intake_strategy.py (eight pass), scripts/test_concresca_intake.py (15 pass, one OS-link skip), scripts/test-v048-commons-machine-guide.py (32 pass), scripts/test-v048-commons-facade.py (46 pass), scripts/test_owned_replay.py (16 pass), reports/v049-replay-live-evidence.json and reports/v049-improvement-observability-verification.md. Fixtures are not independent outside participation. No complete distributed planner, task marketplace or capability accreditation service is claimed. Preserve and hash-verify both originals, validate the ledger, then remove only their intake copies.