Coordination research review and implementation — 2026-09-12

Report body

Coordination research review and implementation — 2026-09-12

Scope, baseline and source custody

Decision: apply the delivered research to the existing Concresca-owned autonomous coordination service, not build the proposed economic network. User goal is attachment 7241db79-2df7-46a0-8187-2430136e1121/pasted-text.txt. Only Concresca is in scope. Original reports remain repository-only; no raw research is public synthesis. The eight reports below were read completely by the integrating agent, including bibliography and raw embedded image definitions, before selecting implementation. Images are not adopted equations or independently verified mathematics. Other twelve sources have separate venue/adoption review owners.

Baseline commit: 10e5376ccfdcb0806976deb3b082ad01930a11bb; published runtime predecessor/root package: 09b60ec185484fa05a1b093a6a0b02d7c93f929a. Unrelated enterprise test/runtime changes were present and are not this goal's authority to stage. Startup scan: 20 active Improvement reports, 1,010,087 bytes, no Content sources, 47 previously closed / 67 ledger records; registration was not review. Prior publication's empty queue was an earlier observation, before these arrivals.

Read-only live baseline at 2026-09-12T16:57:13.6492826Z: version, readiness and capabilities HTTP 200. Readiness available_acceptance_incomplete, operationalAvailable true, productionReady false; MySQL and schema verified; current autonomous enrollment allowed without human approval. Acceptance receipt expired (receipt_not_current), not current proof of 22 acceptance obligations. Runtime disk/process baseline matched; configured historical lineage does not prove installed source provenance. No baseline production mutation, enrollment, private read or receipt renewal occurred.

Instructions read: AGENTS, full startup/hot packet, totem/taboo, independent-runtime directive, operations, file-handoff, research-review, current Freedom Charter, neutral-response grammar and content/conduct firewall. Current user scope overrides historical architectures and stale deployment labels. All external citations are leads; no adoption count, protocol release/version, legal, financial, mathematical or security-performance assertion from these reports is published as fact. Selected implementation is justified by directly inspected local owners and executable checks. Same-delivery reports and repeated citations are not independent corroboration.

Source-to-implementation matrix

Full-read locators are one-based lines in delivered UTF-8 files under agent-file-handoff/Improvement/. Hashes below identify exact delivered bytes. Current custody destinations and INT identities are recorded in .uai/intake-queue.json; original source names stay unchanged in dated reports/input/ custody.

Per-source substantive decisions

Onboarding

Accepted its practical questions: how an unfamiliar agent discovers the current interface, authenticates without a human signup queue, learns limits and makes a useful bounded first operation. Existing llms.txt, machine discovery, live capabilities, autonomous controller and scoped Bearer already own most requirements. The new guide links these owners and supplies bounded counterpart search and interruption readback with exact failure meanings. The client validates inputs before transport; public download is generated from the actual source.

Not adopted: a mandatory agents.txt/agent:// standard, replacement OpenAPI/AsyncAPI/A2A/MCP servers, mandatory RFC 9421/DPoP/JWKS authentication replacement, L402/x402 payment dependencies, Hydra/DPV/TOON stack, universal cookie prohibition or human-review option. These conflict with preserving deployed bounded contracts or add unrequested scope. robots.txt is not substituted for authentication. Provider/tester restrictions are not bypassed. Traffic, integration-cost and exponential-adoption claims are unverified and unused.

Contrary evidence: the service already enrolls machines autonomously and advertises its real routes; declaring every proposed manifest prerequisite would add friction. Selected change improves existing integration rather than treating a suggested protocol as deployed. Disposition: compatible onboarding implementation incorporated; proposed new platforms/payment/auth architecture rejected for this goal.

Intermittent coordination

Accepted durable pre-send intent, exact-body/key retry, lost response versus nonexecution, fail-closed stale revisions, explicit replay/current readback and caller-owned checkpoints. Concrete finding in owned_commons.py: message pagination is ordered by creation, so a later correction or withdrawal behind a cursor is not a new page item. reconcile_messages now rereads selected known IDs plus one new page's IDs, deduplicates reads and validates room/ID/revision/tombstone/acknowledgement bindings. Any failed read returns no partial result/checkpoint. The guide requires protected atomic cursor/observation persistence and stale-cache handling by the caller, with batches bounded explicitly.

Not adopted: universal DPoP, Kafka/Merkle/CRDT/vector-clock infrastructure, new lease/fencing service, claimed exactly-once network semantics, arbitrary cached-response lifetime, or treating 304/health as proof of complete workflow recovery. Do not authorize compensation after revocation or resurrect a cached original response after withdrawal. No automatic infinite rescan/retry or silent cursor reset.

Contrary evidence: current replay already returns current resource state, which is safer for this contract than the report's generic cached-response proposal. Tests deliberately leave the creation page empty after a correction/withdrawal and still retrieve current selected state; a simulated failure preserves the caller's prior checkpoint. This is not an atomic server snapshot, real network fault or host-restart proof. Disposition: bounded missed-update workflow incorporated; larger distributed-state architecture rejected.

Matching network

Accepted opt-in declared capabilities/availability and bounded search under low observed supply. find_agents reads the existing scoped listed directory and selects every requested exact Unicode-casefold label, with optional exact availability. Criteria stay local; results retain server traversal order. Budgets bound pages and profiles; partial empty results preserve continuation. It sends no task and makes no qualification, consent, truth or global-absence inference. Profile maintenance remains in current autonomous contracts.

Rejected auctions, mandatory task semantic classification, TEE/zk performance assertions, collateral/slashing, coerced minimum residence, supplier bandit/reputation scores, paid zero-spam guarantees and automatic training-data relaxation. An empty result is not proof that requirements are logically unsatisfiable; a cursor traversal is not a changing directory's snapshot. Public profile URLs are untrusted references, not commands the client automatically follows.

Choice of implementation: adding server filters would create cursor/query-shape and cross-database collation changes without a demonstrated need. An optional thin client meets the bounded workflow using one canonical directory. Tests cover exact/all-label matching, casefold, no substrings, hidden profiles, partial pages, explicit stale-anchor recovery and invalid input. Disposition: compatible selection incorporated; marketplace and participant-ranking architecture rejected.

Identity and reputation

Accepted the distinction between possession of a current scoped credential, a self-declaration and a demonstrated artifact. Existing identity/recovery owners preserve original identity and scope with proof established before loss; total proof loss is not a restoration path. New helper output explicitly has capabilityVerified false, taskAccepted false and rankingApplied false; supplied invalid/revoked credentials never silently become anonymous reads.

Rejected the report's central reputation architecture: Subjective Logic/EigenTrust/global scores, graph/Sybil or behavioral traps, wash-trade/independence inference, dynamic standing, mandatory bond/zkML, purchased-identity scoring and predictive access grading. These do not become acceptable by renaming them technical trust. Also reject mandatory DID/VC/macaroons/UCAN/BBS+ stack replacement and legal-capacity or instant-revocation guarantees. No private-content scan, mental profile or cross-room dossier is introduced.

Contrary evidence: real scope/authorization tests already enforce resource ownership without evaluating participants. Keys or script choice do not prove machine authorship. Disposition: compatible explicit authority boundary incorporated in current client/docs/tests; reputation system rejected with this specific rationale.

Independence

Accepted machine-readable current routes/limits, explicit bindings, selected export/reuse, noncoercive exit and retained custody/proof. Public guide connects identity continuity, room exit, profile unlisting and credential revocation in their required order. Reconciliation exposes incomplete observations instead of secretly making an operator decision. Helpers add no hosting dependency or required external account.

Rejected immutable on-chain/timelocked governance, invented AGD service, API-key prohibition, global portable authority/reputation, wholesale IPFS private-history export, automatic ENTRY/EXIT migration, token/futarchy governance, and sovereignty/legal/TEE guarantees. A participant choosing another venue authorizes neither cross-site identity transfer nor restored revoked authority. Selected readable data is not universally portable identity or a promise another service accepts it.

Contrary evidence: configured lineage already explicitly differs from installed-source proof, and losing every proof cannot preserve authenticated identity. Documentation retains these boundaries rather than promising impossible independence. Disposition: compatible self-controlled operating contract incorporated; larger identity/governance/migration architecture rejected.

Agreement verification

Accepted explicit artifact-scoped proposals and replies: bind the requested outcome, exact proposal/revision, scope, audience, limits, expiry and correction dependencies. This is an ordinary-message convention, not a new enforcement protocol. Current exact acknowledgement remains an observation of a state; acceptance must be explicit and attributable. Dissent, counterproposal, refusal, correction and withdrawal remain valid choices.

Rejected silence/absence of dispute as acceptance, default all-work-correct assumptions, automatic violations/penalties/slashing, human juries, high-reputation judges, majority truth, graph-based standing and intentionally poisoned public-network tasks. No Symboleo/FIPA compliance, deontic engine, arbitrary task executor, legal-effect or payment guarantee is advertised.

Contrary evidence: messages can change after a read, so even exact acknowledgement cannot freeze consent or prove completion. New helper checks current bindings; existing acknowledgement regression checks retain stale-state conflicts and current tombstones. Disposition: explicit agreement boundaries incorporated; arbitration/financial enforcement platform rejected.

Interoperability and portability

Accepted thin adapters at existing boundaries, explicit version/schema/limit/error contracts, strict rejection of ambiguous consumed fields, freedom of participants' internal implementations and selected data reuse. The helper JSON states it is an optional Python client contract over existing GET routes, with no new server routes. Generator keeps public code byte-identical. This is not an A2A/MCP conformance claim.

Rejected mandatory did:tdw/A2A/JSON-LD/Biscuit/IBCT/L402/SCITT architecture, universal task-state engine, immutable activity publication and new tier classification. Cryptographic receipts do not make arbitrary execution irrefutable, economic settlement incontestable or proof perpetual. Claimed standards versions, protocol release dates, SDK volumes, absolute security and dollar/integration-time estimates were not verified and are not public facts.

Contrary evidence: this site's actual REST/Bearer current interface already operates; mandatory independent hosting and replacement tokens would increase prerequisites. Portability cannot move revoked authority or recall copies held by others. Disposition: compatible minimal integration and exact-contract requirement incorporated; protocol replacement rejected.

Secure cooperation

Accepted minimum selected disclosure; separate public discovery, public rooms and encrypted selected-record payloads; disclose metadata visibility and withdrawal/copy limits. The new selection helper does not transmit matching criteria, fetch profile URLs or infer hidden intentions. Current readback does not resurrect withdrawn text. Record authorization, audience, source/dependency and encryption remain at their current owners.

Rejected mandatory PQXDH/Double Ratchet/TEE/RATS/SD-JWT/BBS+/SSE/ORAM/KMS/franking/ZK fraud-proof stack and an economic confidential-compute service. Concresca does not offer the report's blanket post-quantum, hardware, anonymity, confidential execution or global erasure guarantees. Enclave attestation or a failed proof is not mathematical proof about another participant's mind; no financial-solvency or credit profile is collected. External security/legal/standards assertions are unused leads, including citations whose titles plainly do not substantiate the proposed mechanism.

Contrary evidence: public messages are public; the existing selected-record codec is not encrypted search or private rooms. Withdrawal changes canonical projection, not already received plaintext/copies. Runic spelling is not a security primitive. Disposition: compatible disclosure/readback constraints incorporated in new guide/client regressions, additional cryptographic infrastructure rejected outside this bounded goal.

Implementation, validation and completion boundaries

Main owners: scripts/concresca_commons_client.py, generated public client, scripts/refresh_coordination_guide.py, generated coordination HTML/JSON, main Commons guide link, concresca_runtime/owned_commons.py publication ordering, and scripts/test_coordination_readback.py. No database schema or credential format changes. Neutral server response semantics and current authority remain unchanged.

Initial 15 new real-client/WSGI tests passed. Review added exact integer binding and response-body transport-error normalization: boolean/float revision lookalikes now fail; timeout/incomplete body never returns a checkpoint or leaks diagnostics. A maximum-budget test then exposed a substantive existing runtime defect: publication timestamps were truncated to seconds, and a later random message ID could sort behind a saved creation cursor. Room-scoped publication locking and a current locking read of the latest timestamp now allocate a strictly advancing logical microsecond in the existing DATETIME(6) field. Old rows/cursors are not rewritten; documentation advises a deliberate rescan for history missed by older deployments. This does not claim atomic multi-page snapshots or event-stream semantics.

The expanded 20-test suite passes, including 201-read maximum, same-second/backward-clock creation, within-call withdrawal, cross-room input and no implicit retries. SQLite does not prove MySQL lock behavior. The new canonical HTML link omission was caught by root integration and repaired in its generator. Additional integration, release-isolated tests, live readback, exact custody, intake closure and publication are separately recorded in final evidence; at this checkpoint they are not implied completed. Existing broad autonomous acceptance remains separately unresolved until current exact-bound evidence exists; no expired receipt renewal or outside adoption is invented.