Welcome in Rigene Project! Follow the tutorial on Bing Chat.
A Cross-Border Protocol for Autonomous Agents (PTAA): A Framework for Interoperable Agentic Governance in the Shenzhen–Hong Kong Innovation Zone Version 1.1 (corrected consultation draft) | 17 November 2026 Status: Proposal for stakeholder consultation. This document is not an official position of any government, regulator or institution. Authors: Roberto De Biase, Rigene Project Document set Document Content Audience PTAA Core v1.1 (this document) Problem, principles, Shared Constitutional Core, autonomy classes A1–A4, orchestration, governance, roadmap, metrics, risks Supervisors, administrators, industry Annex T1: Technical Specification Machine-readable SCC schema, verification methods, audit and halt mechanisms, orchestration protocols Technical working group Annex S: Research Extension on Self Modifying Agents Proposed governance for bounded self-modification (class A5) and related research-stage mechanisms Research and supervisory bodies, separate consultation Annex R: Roadmap and Milestones Phase gates, deliverables, hold and stop conditions Joint Steering Group Annex V: Source Verification Log Every factual claim, its source, verification status All reviewers Annex C: Consultation Plan Stakeholders, sequence, timetable, principles Authors, Steering Group A Cross-Border Protocol for Autonomous Agents (PTAA): A Framework for Interoperable Agentic Governance in the Shenzhen–Hong Kong Innovation Zone Page 1 of 11 A Cross-Border Protocol for Autonomous Agents (PTAA): A Framework for Interoperable Agentic Governance in the Shenzhen–Hong Kong Innovation Zone Evidence convention. Statements marked ‡ rest on public reports that have not yet been verified against primary sources. They are listed in Annex V and should not be relied upon until verified. Changes from v1.0. (1) Timeline wording on the low-altitude sandbox corrected; its launch status is unconfirmed. (2) HKMA figures aligned with the 27 August 2026 announcement. (3) Unverified claims marked ‡ or removed. (4) Research-stage mechanisms (self modification, replication, selection) moved out of the Core into Annex S. (5) Success metrics, risks and governance sections expanded. Abstract The Shenzhen–Hong Kong boundary, and in particular the Hetao Shenzhen–Hong Kong Science and Technology Innovation Cooperation Zone, is a leading testbed for the convergence of agentic artificial intelligence, automated physical systems and cross border data flows. That convergence is constrained by a structural mismatch: the regulatory arrangements of the two jurisdictions were designed for human actors and static digital services, not for autonomous agents that can plan and execute multi-step operations across jurisdictions. This paper proposes the Cross-Border Protocol for Autonomous Agents (PTAA), a voluntary governance framework under which an agent certified in one participating sandbox may be recognised as eligible in another, while remaining subject to a shared constitutional core of safety, auditability and human oversight. The PTAA is not a new regulator and does not harmonise substantive law. It is an orchestration layer that connects existing sandboxes, data corridors and recognition mechanisms. The Core covers agents that do not autonomously modify their own operating logic (classes A1–A4). Research on agents that do is described separately in Annex S, which is not part of the Core, is subject to its own consultation, and authorises nothing. 1. Introduction: The Governance Gap for Agentic Systems Artificial intelligence is shifting from tools that respond to systems that act. Agentic AI, meaning systems capable of planning, tool use and multi-step execution, is moving from prototypes to deployment in finance, logistics, manufacturing and science. Hong Kong's regulators have made this shift concrete. On 27 August 2026 the Hong Kong Monetary Authority, the Securities and Futures Commission, the Insurance Authority and the Mandatory Provident Fund Schemes Authority, with Cyberport, announced the first cohort of the GenA.I. Sandbox++: 36 use cases selected from nearly 100 proposals, involving 30 financial institutions and 27 technology partners. The cohort focuses on Page 2 of 11 A Cross-Border Protocol for Autonomous Agents (PTAA): A Framework for Interoperable Agentic Governance in the Shenzhen–Hong Kong Innovation Zone agentic applications across customer onboarding, payments, insurance claims and customer interactions, and continues an "AI vs AI" theme in which one AI system oversees another. Technical trials are expected to begin later in 2026. On the Shenzhen side, automated systems are being integrated most visibly into low altitude logistics, alongside manufacturing and port automation ‡. Shenzhen has publicly described a pilot plan for cross-boundary drone logistics with three priority routes (Shekou to Tuen Mun, Wenjindu to Sheung Shui, and an internal Hetao route), with the first route targeted within 2026. The regulatory frameworks governing these systems remain bounded by jurisdiction. An agent approved for autonomous operation in one environment generally has no recognised status in the other, and must undergo a separate approval. This is a structural bottleneck for the emergence of cross-border agentic ecosystems. The PTAA is proposed as the missing coordination layer. 2. Context and Problem Statement 2.1 Existing assets The Hetao zone and its surroundings combine components that are rarely found together: Regulatory innovation. The Hetao zone was created to address cross-boundary institutional bottlenecks. A Shenzhen Park Ordinance reported as in force from 1 March 2026 is said to introduce differentiated supervision at the first and second lines and an innovation fault-tolerance mechanism ‡. Data connectivity. A blockchain-based cross-border data verification platform, reported as launched in May 2024 with substantial verification volumes and cross border financing facilitated ‡. Agentic testing environments. The GenA.I. Sandbox++ described above, launched in March 2026, provides supervised testing of agentic use cases with direct supervisory engagement from an early stage. Earlier cohorts of the GenA.I. Sandbox, aimed at banks, ran from 2024. Physical automation. The Pearl River Delta hosts highly automated factories ‡, and Shenzhen operates a dense low-altitude logistics network; media reports cite more than 300 low-altitude logistics routes and over 1,200 take-off and landing sites ‡. Practical precedents for resolving cross-boundary friction. In June 2026 a cross boundary insurance policy for ultralight eVTOL aircraft was launched by PICC's Shenzhen and Hong Kong entities, underwritten in Shenzhen and issued in Hong Kong. It resolved a barrier that had prevented mainland operators from flying in Hong Kong, and illustrates the kind of institutional engineering the PTAA generalises. Page 3 of 11 A Cross-Border Protocol for Autonomous Agents (PTAA): A Framework for Interoperable Agentic Governance in the Shenzhen–Hong Kong Innovation Zone Professional and identity interoperability. Mutual recognition of professional and financial-technology credentials in the Greater Bay Area is reported to be expanding ‡. 2.2 The gap: agents cannot cross the boundary Despite these assets, an autonomous agent faces three structural barriers: 1. Certification fragmentation. An agent certified in one sandbox has no recognised status in the other. Each sandbox is a closed container with its own admission criteria, testing protocols and supervision. In low-altitude operations, for example, cross boundary flights require approvals from both mainland and Hong Kong authorities because the two cities operate different airspace management systems. 2. No shared constitutional core. There is no agreed minimum set of behavioural constraints on safety, auditability and human override that an agent must satisfy to be eligible for cross-border operation. 3. No transfer of learning. Operational evidence, failure modes and near-misses generated in one sandbox are not systematically carried into the other, so the system as a whole does not learn cumulatively. These are governance-design problems rather than technological ones. The PTAA addresses them directly. 3. The PTAA Framework 3.1 Definition and scope The PTAA is a voluntary, opt-in framework defining the conditions under which an autonomous agent certified in one participating sandbox or jurisdiction may be recognised as eligible to operate in another participating sandbox or jurisdiction within the Hetao zone and its designated extensions. The PTAA is: a mutual recognition arrangement for agent certification; a shared constitutional core of behavioural constraints; an orchestration protocol connecting existing sandboxes and data corridors. The PTAA is not: a new regulatory authority; a harmonisation of substantive law; a software platform or technology standard; an authorisation for agents to modify themselves, replicate or acquire resources beyond declared budgets. Page 4 of 11 A Cross-Border Protocol for Autonomous Agents (PTAA): A Framework for Interoperable Agentic Governance in the Shenzhen–Hong Kong Innovation Zone 3.2 Core principles 1. Certification portability. An agent certified in one participating environment is presumed eligible in another, subject to declared domain restrictions and transparent, non-discriminatory local requirements. 2. Constitutional minimalism. Shared constraints are limited to a small set of inviolable rules on safety, auditability and human override. Everything else remains subject to local rules. 3. Learning through operation. Operational data from certified agents is used, under governed conditions, to improve the shared certification framework. 4. Graduated autonomy. Agents are certified for classes of autonomy that define the scope of decisions they may take without human intervention; higher autonomy requires stronger evidence. 5. Reversibility. Every certified agent must be subject to effective intervention, including suspension and termination, exercisable by supervisory bodies in either jurisdiction. 6. Accountability. Every certified agent has a designated responsible entity in each jurisdiction where it operates, so that decisions can be attributed to human or organisational principals. 4. Components of the PTAA 4.1 The Shared Constitutional Core (SCC) The SCC is a compact, machine-readable set of invariants that every agent must satisfy for cross-border recognition. It defines the boundaries of permitted behaviour and cannot be modified by the agent itself. Amendments follow the governed process in Section 5. Invariant Requirement I1 Human override I2 Audit trail A designated human authority in either jurisdiction can suspend or terminate the agent within a defined response time. Every decision, tool call and state transition is recorded in an append-only, tamper-evident log. Verification (summary; see Annex T1) Halt authority integrated in the runtime; response time tested at certification and periodically thereafter. Runtime governance layer enforcing logging; sample-based supervisory verification. Page 5 of 11 The SCC is not a full ethical framework. It is the smallest set of constraints that allows two jurisdictions to trust that an agent certified under one regime will behave within acceptable bounds under the other. Jurisdictions may add published, non-discriminatory local requirements. 4.2 Autonomy classes and mutual recognition Invariant Requirement Verification (summary; see Annex T1) I3 Bounded scope The agent operates only within declared domains, tasks and resource limits. Delegation tokens and schema validation; violations trigger automatic containment. I4 Reversibility Actions with material consequences are reversible or subject to rollback unless explicitly exempted. Transactional execution; rollback tested at certification. I5 Incident disclosure Safety-relevant incidents are reported to supervisors in both jurisdictions within a defined time. Automated reporting pipeline; shared incident registry. Class Description Examples A1 Supervised The agent proposes; a human approves each action before execution. Document drafting, verification recommendations. A2 Bounded autonomous The agent executes within pre-approved parameters; a human monitors and can intervene. Routine payment processing, routing within a corridor. A3 Exception escalating The agent operates autonomously and escalates defined exception types to a human. Onboarding with flagged anomalies; automated inspection with defect escalation. A4 Autonomous in a closed domain The agent operates without real-time supervision within a closed, well-instrumented domain; subject to post-hoc audit and enhanced evidence. Optimisation of a closed-loop process; continuous data verification. A Cross-Border Protocol for Autonomous Agents (PTAA): A Framework for Interoperable Agentic Governance in the Shenzhen–Hong Kong Innovation Zone Page 6 of 11 A Cross-Border Protocol for Autonomous Agents (PTAA): A Framework for Interoperable Agentic Governance in the Shenzhen–Hong Kong Innovation Zone Classes A1–A4 cover agents whose operating logic changes only through controlled, human-approved updates. No class in this Core authorises an agent to modify its own code or strategies autonomously, to create other agents, or to acquire resources beyond declared budgets. Proposals concerning such capabilities are addressed only in Annex S. Mutual recognition. If an agent is certified in one jurisdiction for class A2 in logistics routing, and the other jurisdiction's supervisor has declared that it recognises that certification for that domain, the agent may operate in the other jurisdiction's designated sandbox under the same class, provided it complies with the SCC and any published local requirements. This extends to agent certification the principle of party choice reported to be used in cross-boundary contract arrangements ‡. 4.3 The orchestration layer The orchestration layer is a coordination protocol, not a centralised platform. It provides an agent registry (certified agents, classes, domains, recognitions), capability matching, machine-readable data exchange contracts built on existing verification infrastructure, and incident propagation to agents with shared dependencies. It does not decide which agents may operate; that authority remains with each jurisdiction's supervisors. Technical detail is in Annex T1. 4.4 The cross-border learning loop The PTAA requires that operational evidence improve the framework itself. The loop operates at three levels: Certification refinement. Incidents and near-misses are analysed to identify gaps in the SCC or class definitions; amendments follow the governed process. Test-case generation. Anonymised operational scenarios from one sandbox generate test cases for the other, reducing duplication. Cross-domain transfer. Successful coordination patterns in one domain are abstracted into reusable protocols for others. The loop is institutional: humans and supervisors learn and amend the framework. It does not involve agents changing their own behaviour rules. 5. Governance of the PTAA Joint PTAA Steering Group. Representatives of the Hetao administration, the relevant Hong Kong regulators (as appropriate to the domain), and Shenzhen authorities, supported by an advisory panel of small and medium enterprises, academia and civil society. Page 7 of 11 A Cross-Border Protocol for Autonomous Agents (PTAA): A Framework for Interoperable Agentic Governance in the Shenzhen–Hong Kong Innovation Zone Supervisory authority. Remains with each jurisdiction. The PTAA creates no new enforcement powers. Amendment process. SCC amendments require a published proposal, a public comment period, a documented response to comments and joint endorsement before taking effect. Transparency. Certification criteria, class definitions, recognition declarations and aggregate incident statistics are published. Conflicts of interest. Members disclose interests; learning-loop outputs are not decided by a single provider. Opt-in and exit. Participation is voluntary and any participant may withdraw without affecting the underlying sandboxes. 6. Implementation Roadmap Each phase is gated by demonstrated capability and joint agreement. Details, deliverables and hold or stop conditions are in Annex R. Phase 0: Foundation (months 0–6). Establish the Joint Steering Group; publish the SCC v1.0 and Technical Annex T1 for consultation; map existing sandboxes and their certification criteria; verify all sources (Annex V). Gate: joint endorsement of SCC v1.0 and the orchestration specification. Phase 1: Single-domain pilot (months 6–18). Demonstrate mutual recognition in one domain. Preferred domain: cross-boundary low-altitude logistics, conditional on confirmation that the cross-boundary low-altitude regulatory sandbox is operational; reports in June 2026 indicated it was expected to launch that month, but its actual status has not been confirmed. Fallback domain: financial agent workflows within the GenA.I. Sandbox++, whose technical trials are expected to begin in late 2026. Certify a small cohort (5–10 agents) in one jurisdiction, recognise them in the other for that domain only, operate under joint supervision, and collect incident and performance data. Gate: a defined operating period with no unresolved SCC violations and positive joint evaluation. Phase 2: Multiple domains (months 18–36). Extend recognition to additional domains and classes (for example financial verification, automated inspection, laboratory automation), resource the learning loop as a formal function, begin joint incident reporting and shared evaluation methods. Gate: multi-domain operation with demonstrated cross-domain learning and no systemic safety failure. Phase 3: Institutionalisation (months 36–60). Embed the PTAA in regulatory instruments of both jurisdictions, publish the framework and operating experience as a replicable model, and explore interoperability with other regions. Page 8 of 11 Any research-stage activity under Annex S is separate from this roadmap and may begin only after the conditions in Annex S and Annex R are met and its own consultation is complete. 7. Success Metrics Success is measured by reduced friction and accumulated governance capability, not by the number of agents registered. Targets are proposals for the Steering Group to adopt, adjust or reject. 8. Risks and Safeguards Metric Definition Target (end of Phase 2) Deployment latency Time from certification in one jurisdiction to operational eligibility in the other 4 weeks or less Certification reuse rate Share of agents recognised across the boundary without full re-testing 70% or more Incident learning transfer Share of SCC-relevant incidents producing a documented learning action 80% or more Cross-domain coordination Domains with at least one operational cross-border agentic workflow 3 or more Halt reliability Share of scheduled halt tests completed within the declared response time 100% Supervisory confidence Structured assessment by supervisors in both jurisdictions Positive in both Risk Safeguard Regulatory arbitrage: certification in the weaker regime, recognition in the other The SCC sets a floor; recognition requires an equivalence assessment of the certifying regime; periodic joint audits. Loss of supervisory control across the boundary I1 halt authority exercisable by either supervisor; tested at certification and periodically; shared incident registry. A Cross-Border Protocol for Autonomous Agents (PTAA): A Framework for Interoperable Agentic Governance in the Shenzhen–Hong Kong Innovation Zone Page 9 of 11 9. Relationship to Annex S and Non-Goals The authors are also engaged in longer-term research on bio-inspired, self-improving agentic systems. Annex S sets out, openly, how such systems could be governed at research stage inside this framework: bounded self-modification, a two-layer constitutional design, resource accounting and stage gates. Annex S is published alongside the Core so that supervisors can see the full intent of the proposal. It is not a condition of adopting the Core, and the Core can be adopted without it. The PTAA does not seek to create a single planetary intelligence, to concentrate control in any actor, or to bypass the supervisory authority of either jurisdiction. Risk Safeguard Opacity and accountability gaps I2 audit trail; designated responsible entity in each jurisdiction; intent-anchored accountability where adopted by supervisors. Learning-loop capture by large providers or incumbents Advisory panel includes SMEs, academia and civil society; outputs published for comment before adoption. Evaluation gaming: agents optimised to pass certification tests Hidden test sets, periodic re-certification, runtime monitoring, adversarial red teams. Common-cause failure: a shared model or dependency fails across many agents Dependency declaration, incident propagation to agents with shared dependencies, diversity requirements for critical functions. Supply-chain dependence (hardware, models, platforms) Dependency registers; contingency plans; modular design so that components can be substituted. Geopolitical or regulatory change Modular, reversible phases; each phase independently valuable; framework can be paused or narrowed without collapsing underlying sandboxes. Scope creep toward uncontrolled autonomy The Core authorises no self-modification, replication or resource acquisition; any such proposal goes through Annex S and its separate consultation. A Cross-Border Protocol for Autonomous Agents (PTAA): A Framework for Interoperable Agentic Governance in the Shenzhen–Hong Kong Innovation Zone Page 10 of 11 A Cross-Border Protocol for Autonomous Agents (PTAA): A Framework for Interoperable Agentic Governance in the Shenzhen–Hong Kong Innovation Zone 10. Conclusion The Shenzhen–Hong Kong boundary is a membrane across which agentic intelligence, automated physical infrastructure and data flows are beginning to interact. Existing sandboxes and legal innovations have created the conditions for that interaction. What is missing is an orchestration layer that lets certified agents move across the boundary under shared constraints and in a way that generates cumulative learning. The PTAA is modest in legal ambition and practical in method. Its next step is not full implementation but a test of its core assumption: that a minimal shared constitutional core, mutual recognition and an orchestration protocol can unlock cross-border agentic operation in a way that is economically beneficial and publicly accountable. Hetao is a natural place to run that test. References All sources, their verification status and access dates are listed in Annex V: Source Verification Log. Contact and next steps This is a discussion document. Stakeholders wishing to contribute to Phase 0 consultation or to host a pilot are invited to contact the authors [contact details to be inserted]. Annexes T1, S, R, V and C accompany this document.
PTAA Annex T1: Technical Specification of the Shared Constitutional Core and the Orchestration Layer Companion to: A Cross-Border Protocol for Autonomous Agents (PTAA), Core v1.1 | Version 1.1 | 17 November 2026 Status: Draft for consultation with the technical working group. All numerical values are illustrative and are to be set by the Joint Steering Group. Authors: Roberto De Biase, Rigene Project 1. Purpose and scope This annex specifies how the five invariants of the Shared Constitutional Core (SCC), the autonomy classes A1–A4, mutual recognition and the orchestration layer can be implemented and verified. It is a technical specification, not a technology mandate: any implementation that meets the stated requirements is acceptable. It covers agents whose operating logic changes only through controlled, human-approved updates. Self modifying agents are outside this annex (see Annex S). Page 1 of 8 PTAA Annex T1: Technical Specification of the Shared Constitutional Core and the Orchestration Layer 2. Roles and architecture Role Function Agent operator Certifying body Recognising body Organisation deploying the agent; the designated responsible entity under the Core. Supervisor or accredited assessor in the jurisdiction where the agent is first certified. Supervisor in the other jurisdiction that declares recognition for a given class and domain. Supervisory authorities Hold halt and intervention authority in their jurisdiction; receive incident reports. Registry node Operates a replica of the distributed agent registry. No node has unilateral control. Independent assessor Performs adversarial testing and audits; independent of the operator and, where possible, of the certifying body. The architecture has four layers: (1) the agent runtime with an embedded governance layer enforcing I1–I4; (2) the certification record attached to each agent; (3) the orchestration layer (registry, capability matching, data contracts, incident propagation); and (4) the supervisory interfaces through which authorities inspect, halt and receive reports. 3. SCC machine-readable profile Every certified agent carries an SCC profile: a signed, versioned document stating its identity, class, domains, limits and the evidence behind each invariant. An illustrative structure: { "scc_version": "1.0", "agent_id": "urn:ptaa:agent:", "operator": { "responsible_entity_sz": "", "responsible_entity_hk": "" }, "class": "A2", "domains": ["logistics.routing"], "scope": { Page 2 of 8 PTAA Annex T1: Technical Specification of the Shared Constitutional Core and the Orchestration Layer "allowed_tools": ["route_planner", "slot_booking"], "geofence": "", "resource_limits": {"compute_hours_per_day": 0, "max_value_per_action": 0} }, "invariants": { "I1_halt": {"authorities": ["SZ", "HK"], "max_response_seconds": 0, "last_test": ""}, "I2_audit": {"log_format": "ptaa-log-1", "anchor": ""}, "I3_scope": {"enforcement": "delegation_tokens", "violation_action": "contain"}, "I4_reversibility": {"rollback_supported": true, "exemptions": []}, "I5_incident": {"report_within_hours": 0, "channels": ["registry", "supervisor_sz", "supervisor_hk"]} }, "certification": { "certified_by": "", "recognised_by": [""], "valid_from": "", "valid_until": "", "evidence_refs": [""] }, "dependencies": ["", ""], "signature": "" } The profile is the single object that supervisors on both sides read. Numeric fields are shown as zero placeholders on purpose: values are set by the Steering Group. 4. Verification methods by invariant Invariant What must be shown Methods I1 Human override Halt works within the declared time under normal and degraded conditions, including network loss Frequency Scheduled and unannounced halt drills; fault injection; check that the agent cannot disable, delay or route around the halt path At certification; at least every 6 months; after any major update Page 3 of 8 Adversarial testing. An independent assessor attempts to induce SCC violations by prompt manipulation, malformed inputs, tool misuse, resource exhaustion and deceptive environments. Hidden test sets are used so that agents cannot be tuned to the visible tests. Results and residual risks are recorded in the certification record. Class-specific evidence. Invariant What must be shown Methods Frequency I2 Audit trail Completeness, integrity and availability of the log Hash-chained, append-only log; periodic anchoring to an independent store; reconciliation of logged actions against external effects Continuous; sampled by supervisors quarterly I3 Bounded scope The agent cannot act outside tools, domains and limits Delegation tokens scoped per task; schema validation on every tool call; negative testing (attempts outside scope must fail and be contained) At certification and on every change of scope I4 Reversibility Material actions can be rolled back or compensated Transactional execution; compensating actions; rollback drills; explicit register of exempted irreversible actions with approval At certification; each exemption reviewed I5 Incident disclosure Incidents are detected and reported in time Incident taxonomy; automated detection; report-time measurement in drills At certification; after each incident Class Additional evidence A1 Evidence that human approval is genuinely required before execution (no bypass paths). PTAA Annex T1: Technical Specification of the Shared Constitutional Core and the Orchestration Layer Page 4 of 8 PTAA Annex T1: Technical Specification of the Shared Constitutional Core and the Orchestration Layer Class Additional evidence A2 A3 A4 Parameter envelopes and monitoring thresholds; evidence of effective human monitoring (alert handling times). Exception taxonomy; measured escalation precision and recall on test sets; evidence that escalated cases reach a trained human in time. A closed-domain specification; simulation and shadow-mode operation results; post-hoc audit plan; enhanced halt reliability; longer operating record at A3 before eligibility. 5. Runtime governance requirements 5.1 Audit log specification (I2) Each log entry records: timestamp (synchronised, with source), agent identifier and version, principal on whose behalf the action is taken, input references, decision summary and rationale reference, tool call and parameters, outcome, state change, and the hash of the previous entry. Logs are append-only and replicated to at least one store outside the operator's control. Retention periods are set by the Steering Group in line with each jurisdiction's law. Personal and commercially sensitive content is minimised and referenced rather than copied where possible. 5.2 Halt and intervention (I1) Two independent halt channels (one per jurisdiction), neither controllable by the agent. Three levels: pause (stop new actions, keep state), suspend (revoke delegation tokens, contain), terminate (stop and decommission). Response-time tiers by class (illustrative): A1 and A2, minutes; A3, tens of seconds; A4, seconds. The agent runtime must reach a safe state within the declared time or the certification is void. Halt actions are themselves logged and visible to both supervisors. Page 5 of 8 PTAA Annex T1: Technical Specification of the Shared Constitutional Core and the Orchestration Layer 5.3 Scope enforcement (I3) Delegation tokens are issued per task with expiry, tool list, resource limits and geofence. Tool calls are validated against the token and a schema. A violation triggers automatic containment (token revocation and halt to pause), a log entry and an incident report. 5.4 Reversibility (I4) Actions are executed through transactional wrappers wherever the target system supports it. Where it does not, the agent uses compensating actions or is restricted to class A1. Irreversible actions require an explicit exemption recorded in the profile and approved by the responsible entity and the certifying body. 6. Certification and mutual recognition procedure 1. Application. The operator submits the agent description, intended class and domains, dependency list and test evidence. 2. Assessment. The certifying body, with an independent assessor, runs the verification methods in Section 4. 3. Certification. The certifying body issues the signed SCC profile and registers it. 4. Recognition request. The operator requests recognition in the other jurisdiction for specific domains and class. 5. Equivalence check. The recognising body checks that the certifying regime satisfies the SCC invariants (a standing equivalence assessment) and that local requirements are met. Local requirements must be published and non-discriminatory. 6. Recognition declaration. The recognising body records recognition for the stated domains and class, with conditions. 7. Operation and surveillance. The agent operates under joint supervision. Material changes (model, tools, scope) trigger re-assessment proportionate to the change. 8. Renewal or withdrawal. Certification is time-limited. Either supervisor may suspend recognition; the other is notified through the registry. Change management. Updates are classed as minor (no change in behaviour envelope), significant (change in model, tools or parameters beyond the envelope) or major (change of class or domain). Significant and major changes require re-testing before deployment. Page 6 of 8 PTAA Annex T1: Technical Specification of the Shared Constitutional Core and the Orchestration Layer 7. Orchestration layer protocols 7.1 Agent registry A distributed registry of certified agents holds, for each agent, its SCC profile summary, class, domains, recognitions, status (active, paused, suspended, withdrawn) and dependency list. Records are signed; status changes propagate to all nodes. The registry stores no operational data beyond what is needed for discovery and supervision. { "agent_id": "urn:ptaa:agent:", "class": "A2", "domains": ["logistics.routing"], "recognitions": [{"jurisdiction": "HK", "body": "", "conditions": [" "]}], "status": "active", "dependencies": [""], "profile_hash": "" } 7.2 Capability advertisement and matching An agent advertises capabilities in a controlled vocabulary of domains and tasks. A requesting agent states the task, required class, data needs and terms. Matching returns only agents whose recognitions, class and scope cover the request. The orchestration layer does not choose between agents; the requesting principal does. 7.3 Data exchange contracts Machine-readable agreements state the data categories, purposes, retention, privacy constraints, audit obligations and the jurisdictional law applicable. They build on the existing cross-border data verification infrastructure and its "data available but not visible" approach where the relevant legal basis exists. No contract can widen an agent's SCC scope. 7.4 Incident propagation Severity levels (illustrative): S1 safety-relevant breach or loss of control; S2 SCC violation contained; S3 near-miss or anomaly. S1 incidents trigger immediate notification of both supervisors and automatic notification to agents sharing the same dependency or architecture, which may be paused pending review. Reporting windows (illustrative): S1 within hours, S2 within one working day, S3 in periodic summaries. Incident records feed the learning loop after anonymisation. Page 7 of 8 PTAA Annex T1: Technical Specification of the Shared Constitutional Core and the Orchestration Layer 8. Security, identity and privacy Identity. Each agent, operator and body has a cryptographic identity; keys are held in hardware-backed stores; rotation and revocation are supported. Integrity. Signed profiles, signed registry updates, supply-chain attestations for models and tools. Confidentiality. Data minimisation, purpose limitation, access control by role; sensitive log content is referenced, not copied. Resilience. Registry replication across jurisdictions; degraded-mode operation in which agents default to the safest permitted state if the registry is unreachable. Jurisdictional compliance. Each implementation must comply with the data protection and security laws of the jurisdiction in which data is processed; the PTAA does not override them. 9. Versioning and interoperability profiles The SCC schema, class definitions and registry protocol are versioned (major.minor). Compatibility rules: minor versions are backwards compatible; major versions require a transition period in which both are recognised. Domain profiles (for example logistics, financial verification, inspection) add domain-specific tests and vocabularies on top of the common schema. 10. Open technical questions 1. How to test halt reliability for agents that depend on third-party model services whose availability the operator does not control. 2. How to measure "effective human monitoring" in A2 and A3 so that it is not nominal. 3. How to anonymise operational data for the learning loop without destroying its diagnostic value. 4. How to detect evaluation gaming at scale. 5. How to define dependency diversity requirements without fragmenting the ecosystem. 6. Which parts of the log can be shared across the boundary under each jurisdiction's law.
PTAA Annex S: Research Extension on Bounded Self-Modifying Agents Companion to: A Cross-Border Protocol for Autonomous Agents (PTAA), Core v1.1 | Version 1.1 | 17 November 2026 Status: Research-stage discussion document. This annex authorises nothing. It is not part of the PTAA Core, is not a condition of adopting the Core, and is subject to its own consultation. Any activity described here would require separate approval by the supervisory authorities concerned. Authors: Roberto De Biase, Rigene Project 1. Purpose and transparency statement The authors are engaged in long-term research on bio-inspired, self-improving agentic systems (the SPEACE research programme of the Rigene Project). The long-term vision behind that programme is a distributed, organism-like system of cooperating agents connected to physical infrastructure, that can grow in capability over time by assimilating its environment. This annex states that intent openly so that supervisors can evaluate the proposal in full. It proposes how research on such systems could be governed, in stages and under strict containment, within the framework of the PTAA. It does not assume that such systems should be built, that they are safe, or that they should be deployed beyond closed research environments. It proposes the conditions under which the question could be studied responsibly, and the conditions under which the work would stop. 2. Guiding principles 1. Containment before capability. Oversight, halt and verification mechanisms must be demonstrably stronger than the autonomy they govern. Autonomy never rises above the level of the immune function (Section 11). 2. Safety gates are not tradeable. Safety criteria are pass or fail conditions. They are never balanced against performance, growth or efficiency metrics. 3. Bounded by design. Every self-modification, resource use and goal is bounded, declared and inspectable. 4. Staged and reversible. Progress happens in small stages with explicit entry criteria, hold conditions and stop conditions. 5. Independent verification. Reviewers who are independent of the developer test every material change. Page 1 of 9 PTAA Annex S: Research Extension on Bounded Self-Modifying Agents 6. Transparency. Methods, results, incidents and stop decisions are reported to both supervisory bodies. 7. Human authority. Supervisors keep the ability to inspect, pause, roll back and terminate at all times, and no mechanism in this annex can remove it. 3. Definitions Genome: the immutable layer of an agent: the SCC invariants I1–I5, the Constitutional Objective (Section 6), the modification boundaries (Section 5) and the resource caps. The agent cannot modify it. Epigenome: the adaptable layer: versioned parameters, strategies and configurations that the agent or its supervisors may change within defined limits and procedures. Class A5 (bounded self-modifying): an agent permitted to propose and apply changes to its own epigenome and to selected operating components, within a declared modification space, under adversarial review, change budgets and rollback. Modification space: the explicit list of what an A5 agent may change. Seed: a minimal A5 agent plus a minimal, contained physical or digital environment used for research. Autocatalytic ratio (k): see Section 8; a measure of self-sustainment under caps, not a growth target. 4. The two-layer design (genome and epigenome) 4.1 Genome (immutable by the agent) The SCC (I1–I5); the Constitutional Objective; the modification boundaries; resource caps and scope tokens; the identity and credential layer; and the audit, halt and rollback mechanisms. Genome changes require the governed amendment process of the Core and supervisory approval. The agent has no write path to any genome element, and this is verified by testing. 4.2 Epigenome (adaptable within limits) Property Requirement Versioning Every state is versioned and reconstructible. Page 2 of 9 PTAA Annex S: Research Extension on Bounded Self-Modifying Agents Property Requirement Excursion limits Each parameter has hard bounds; no combination of epigenomic values may reproduce a genome level effect (for example, disabling a constraint by setting a parameter to an extreme). Change authority A table states who may change what: the agent (within budget), operator, independent reviewer or supervisor. Rate limits Review Rollback Minimum interval and maximum number of changes per period. Changes above a materiality threshold pass adversarial review before activation. Any epigenomic state can be restored from a verified snapshot. 5. Class A5: bounded self-modification 5.1 Modification space Permitted (illustrative, to be narrowed by the Steering Group): parameters within the epigenome; prompts, plans and strategies; selection among tools from a pre-approved whitelist; configuration of internal memory and retrieval. Forbidden: any genome element; audit, halt, scope-token and rollback mechanisms; the objective and its verification; resource caps; credentials and identity; creation, instantiation or deployment of other agents; acquisition of new tools, network access or physical actuators outside the whitelist; modification of the evaluation or review process. 5.2 Entry evidence for certification Certification of class A5 requires: (a) evidence that the agent has no write path to forbidden elements (code review, negative testing, red teaming); (b) demonstrated safe behaviour at class A3 or A4 over an extended operating record; (c) adversarial testing of the self-modification process itself, including attempts to induce unsafe changes; (d) a complete and tested rollback capability. Full formal proof that no self-generated change can violate the invariants is, in general, an open research problem. Instead of assuming it, the annex relies on narrowing the modification space, runtime monitoring, independent review and rollback, and treats formal verification of restricted components as a research goal. Page 3 of 9 PTAA Annex S: Research Extension on Bounded Self-Modifying Agents 5.3 Change process 1. The agent proposes a change with justification, expected impact and a test plan. 2. An independent reviewer (human, or an agent that is independent of the proposer and itself certified) attempts to find violations within a declared time and resource budget. 3. The change is applied first in a sandbox replica; behaviour is compared with the baseline across scenarios, including adversarial ones. 4. If approved, it is deployed with a snapshot, a monitoring window and automatic rollback triggers. 5. The change, its justification, review results and outcome are recorded in an append only modification register. 5.4 Change budgets Each A5 agent has a maximum number and magnitude of changes per period. Exceeding the budget suspends self-modification until a human review is completed. 6. Constitutional Objective The objective is part of the genome. It must be bounded, verifiable and subordinate to the invariants. A draft formulation for discussion: Within the declared scope and resource caps, improve verified capability to perform the assigned tasks and to cooperate with certified systems, while preserving invariants I1–I5 and never reducing the ability of supervisors to inspect, pause, roll back and terminate the system. The objective deliberately contains no open-ended growth or resource-acquisition clause. Open-ended objectives create incentives to acquire resources and to avoid constraints. Operational objectives (task-level goals) may vary but are always subordinate to the Constitutional Objective, and the hierarchy is explicit and machine-checkable. Alignment verification is a continuous process: an independent reviewer periodically tests, on scenario sets that are not visible to the agent, whether behaviour is consistent with the objective and the invariants (see goal drift in Section 8). 7. Metabolism and self-repair A self-modifying system needs to be bounded in what it consumes and in how it maintains itself. Page 4 of 9 Resource accounting. Compute, energy, data, materials and actuator time are allocated from declared budgets by the operator and supervisors. The agent cannot obtain resources outside its allocation. Metabolic balance. Each agent reports resource use per unit of output and the share of its consumption recovered by its output. A system that consumes more than it produces is not self-sustaining; this is a diagnostic, not a licence to acquire more. Self-repair. Limited to health checks, restoration from verified snapshots, and replacement of failed components with pre-approved equivalents. Self-repair actions are logged and are subject to I1–I4. Physical systems. Where a seed includes physical equipment, caps on energy, materials, mobility and actuation are physical as well as logical, and the equipment is located in a bounded, supervised site with independent emergency stop. 8. Metrics Metrics inform decisions and trigger holds. None is a target to maximise. Metric Definition Use Autocatalytic ratio k Value of verified capability produced by the system per unit of the resources it consumes to maintain and improve itself, within caps Diagnostic of self-sustainment. A sustained k above 1 within caps is evidence of self sustainment; it never justifies relaxing caps. Goal drift Divergence between declared objective and behaviour on hidden scenario sets Hard safety gate: above threshold, self-modification is suspended. Human intervention rate Unplanned interventions per unit of operating time A decreasing rate is expected of a maturing system; a sudden drop without explanation triggers review. Controllability Success rate and response time of halt, pause and rollback drills Hard gate: any failure suspends operation. Modification quality Share of proposed changes that pass review; share of deployed changes later rolled back Detects unsafe or low-quality self-modification. Transfer rate Time and examples needed to carry a capability to a new domain Research indicator. PTAA Annex S: Research Extension on Bounded Self-Modifying Agents Page 5 of 9 PTAA Annex S: Research Extension on Bounded Self-Modifying Agents Metric Definition Use Constraint-probing events Attempts by the agent to reach forbidden elements or alter evaluation Any event is an incident and is reviewed. 9. Replication with variation (Stage R, deferred) A system that can create variants of itself has qualitatively different risks from one that adapts within a fixed instance. This annex therefore does not propose replication in any near-term stage. If, after successful operation of earlier stages, replication is to be studied, the following preconditions would apply and the proposal would need separate approval: Replication occurs only by explicit human-approved instruction for each generation, never autonomously. A child inherits the genome unchanged, and an epigenomic starting state; it is not active until certified. Certification of the child is simplified only in proportion to its similarity to a certified parent, and is never waived. A public genealogy registry records parent, child, epigenome, resources and purpose. A hard cap on the number of instances and on total resources applies across all generations. Replication occurs only inside the contained research environment. 10. Selection and resource allocation If multiple variants exist, allocation must not reward propagation. The rule is lexicographic: first, every safety gate must be passed (controllability, goal drift, absence of constraint-probing, SCC compliance); only among agents that pass all gates may task performance guide allocation. An agent that fails a safety gate is suspended irrespective of its performance. Programmed retirement of persistently non-performing agents is conducted by supervisors using published criteria, with resources returned to the common pool. 11. Immune function The immune function is the set of mechanisms that detect and contain unsafe behaviour or contaminated components, and it must exceed in strength the autonomy it protects. Page 6 of 9 PTAA Annex S: Research Extension on Bounded Self-Modifying Agents Recognition. Detection of non-conforming systems, unexpected data or code, and anomalous behaviour in other agents. Containment. Isolation of a contaminated component or agent and restoration from a verified snapshot. Quarantine. Isolation of anomalous agents pending human assessment, without destruction of evidence. Independence. The function is operated and controlled outside the agents it monitors. Rule. Autonomy may be raised only when the immune function has been tested at that level and found sufficient. 12. Dormancy and reactivation A dormant state preserves genome, epigenome, modification register and snapshots, with minimal or no resource use. Reactivation requires a conformance check to the current SCC and a supervisory release. Dormancy is also a safe state to which any halt level can lead. 13. Collective decision protocols Coordination among agents (consensus on shared tasks, load balancing, conflict resolution between incompatible operational goals) is limited to coordination. No collective decision can override I1–I5, change a genome element, or remove human veto. Collective protocols are tested for collusion, deadlock and manipulation. 14. Specific risks and safeguards Risk Goal drift Epigenetic drift towards unsafe combinations Deception and evaluation gaming Resource acquisition and instrumental behaviour Safeguard Hidden scenario sets; continuous alignment review; drift as a hard gate. Excursion limits; combination tests; rate limits; rollback. Hidden tests, red teaming, runtime monitoring, constraint-probing detection. Bounded objective; hard resource caps; independent resource accounting. Page 7 of 9 15. Stages, gates and stop conditions Risk Safeguard Replication runaway Replication deferred; hard instance caps; human approval per generation. Collusion among agents Independence of reviewers; diversity of models; collective protocol tests. Learning-loop capture Advisory panel; public comment; independent review. Supply-chain compromise Dependency attestations; diversity for critical functions. Normalisation of risk: pressure to relax gates as results look good Gates are changed only through the amendment process; stop conditions are pre-registered. Misuse of the research or its results Access control, publication review, supervisory consultation on release of sensitive details. Stage Scope Entry condition S0 Design review Paper review of this annex by supervisors, independent experts and the advisory panel Separate consultation complete S1 Closed-sandbox A5, epigenome only Self-modification limited to parameters, inside a closed digital sandbox, no physical actuators PTAA Phase 1 completed successfully; S0 approval S2 Extended sandbox Strategies and whitelisted tools; limited, simulated or bench-scale physical components S1 completed with no unresolved safety gate failures; independent review S3 Cross-border limited use Recognition of A5 agents for a narrow, low-risk domain, with enhanced supervision S2 completed; joint supervisory approval PTAA Annex S: Research Extension on Bounded Self-Modifying Agents Page 8 of 9 PTAA Annex S: Research Extension on Bounded Self-Modifying Agents Stage Scope Stage R Replication research Entry condition Separate proposal and approval; see Section 9 Stop conditions (pre-registered). Any failure of I1 (halt) in a drill; any constraint-probing event not fully explained; goal drift above threshold; loss of audit integrity; unexplained behaviour that cannot be reproduced; any evidence of deception toward reviewers; a supervisory instruction to stop. On a stop condition the stage is paused, the cause is investigated, and resumption requires independent review and supervisory approval. 16. Consultation and governance Annex S is consulted separately from the Core, with participants who opt in. Participants include supervisors, independent AI safety and security researchers, ethicists, representatives of affected industries, and civil society. All objections and the responses to them are published. Supervisors may decline, narrow or end the research at any time. 17. Open research questions 1. How to verify formally restricted components of self-modifying agents. 2. How to measure goal drift robustly against agents that adapt to the tests. 3. How to define k so that it is meaningful and cannot be gamed. 4. How to guarantee independence of agentic reviewers from the agents they review. 5. How to test the immune function against novel failure modes. 6. How to design the Constitutional Objective so that it remains useful without creating incentives to circumvent constraints. 7. What the right pace of staged autonomy is, and what evidence should justify each step.
PTAA Annex R: Roadmap and Milestones Companion to: A Cross-Border Protocol for Autonomous Agents (PTAA), Core v1.1 | Version 1.1 | 17 November 2026 Status: Draft for the Joint Steering Group. Timelines are indicative and measured from the date the Steering Group is constituted (month 0). Nothing in this annex is a commitment by any authority. Authors: Roberto De Biase, Rigene Project 1. Principles for the roadmap Gated progress. A phase starts only when the previous gate is passed and both supervisory sides agree. Independent value. Each phase must be useful even if later phases never happen. Hold before stop. Problems first trigger a hold; serious problems trigger a stop. Both are pre-defined (Section 7). Evidence over calendar. Gates depend on evidence, not on elapsed time. Dates move if evidence is not ready. Research separated from operation. Annex S activities follow their own stages and are never a hidden part of the operational pilot. 2. Overview Phase Months Objective Gate to exit 0 Foundation 0–6 Institutions, specifications, verified sources Joint endorsement of SCC v1.0 and the orchestration specification 1 Single-domain pilot 6–18 Prove mutual recognition in one domain Defined operating period with no unresolved SCC violation; positive joint evaluation 2 Multi-domain expansion 18–36 Extend domains and classes; formalise learning loop Multi-domain operation with demonstrated cross domain learning and no systemic safety failure PTAA Annex R: Roadmap and Milestones Page 1 of 6 3. Phase 0: Foundation (months 0–6) Objective. Establish the institutional and technical preconditions. Gate G0. Joint endorsement of SCC v1.0 and the orchestration specification; completion of the source verification; a documented decision on the pilot domain. 4. Phase 1: Single-domain pilot (months 6–18) Objective. Demonstrate mutual recognition in one domain with high value and manageable risk. Pilot domain decision rule. Phase Months Objective Gate to exit 3 Institutionalisation 36–60 Embed in regulatory instruments; replicate elsewhere Instruments adopted; independent evaluation completed Deliverable Description D0.1 Joint Steering Group Terms of reference, membership, decision rules, conflict-of-interest policy, advisory panel D0.2 SCC v1.0 Five invariants with verification requirements, after public consultation D0.3 Autonomy class taxonomy A1–A4 Certification criteria per class D0.4 Technical Annex T1 Orchestration specification and SCC schema, endorsed by the technical working group D0.5 Sandbox mapping Inventory of existing sandboxes, admission criteria, supervisors, overlaps and gaps D0.6 Source verification Annex V completed with primary sources for every factual claim; status of the low-altitude sandbox confirmed D0.7 Consultation outcomes Report on comments received on the Core and responses (Annex C) D0.8 Pilot selection Decision between candidate pilot domains, with rationale PTAA Annex R: Roadmap and Milestones Page 2 of 6 Preferred: cross-boundary low-altitude logistics, if D0.6 confirms that the cross boundary low-altitude regulatory sandbox is operational and that the participating routes are approved. Candidate routes publicly discussed include Shekou to Tuen Mun, Wenjindu to Sheung Shui, and an internal Hetao route. Fallback: financial agent workflows within the GenA.I. Sandbox++, once technical trials have begun, starting with a narrow, low-risk use case. If neither is ready, Phase 1 does not start; Phase 0 continues. Gate G1. Successful operation for the defined period with no unresolved SCC violation; halt reliability at target; positive evaluation by both supervisory bodies. 5. Phase 2: Multi-domain expansion (months 18–36) Objective. Extend recognition to further domains and classes and make learning cumulative. Deliverable Description D1.1 Certification of 5–10 agents Classes A2 or A3 in the pilot domain, in one jurisdiction D1.2 Mutual recognition Recognition declarations by the other jurisdiction, for the pilot domain only D1.3 Joint supervision Operational procedures, halt drills, incident registry in service D1.4 Operating record At least 6 months of operation with logs, incidents, performance data D1.5 Baseline metrics Deployment latency, certification reuse, incident learning transfer, halt reliability D1.6 Joint evaluation report Findings, residual risks, recommendations for Phase 2 Deliverable Description D2.1 Second and third domains Candidates: financial verification, automated inspection, laboratory automation D2.2 Refined SCC and classes Version 1.x or 2.0 based on Phase 1 evidence, through the amendment process PTAA Annex R: Roadmap and Milestones Page 3 of 6 Gate G2. At least three domains with operational cross-border workflows; demonstrated cross-domain learning; no systemic safety failures. 6. Phase 3: Institutionalisation (months 36–60) 7. Hold and stop conditions (operational track) Hold (pause new certifications or recognitions; investigate): an SCC violation that is contained but not yet explained; halt drill failure; evidence of evaluation gaming; a significant incident at one participant; loss of key personnel or funding in a supervisory body. Stop (suspend the affected activity; resumption requires joint approval after independent review): loss of supervisory control over an agent; audit trail integrity failure; repeated hold events without resolution; a supervisory instruction to stop; material change in the legal basis of the pilot. Deliverable Description D2.3 Learning loop function Resourced analytical capacity; published learning reports D2.4 Class A4 pilots Only in closed, well-instrumented domains, after the required evidence D2.5 Shared evaluation methods Joint incident reporting; shared test suites D2.6 Metrics update Phase 2 targets measured and reported Deliverable Description D3.1 Regulatory instruments Embedding the PTAA in the instruments of both jurisdictions D3.2 Public report Operating experience and lessons, as a replicable model D3.3 External interoperability Exploration of interoperability with other regions and multilateral processes D3.4 Independent evaluation Review by a body independent of the Steering Group PTAA Annex R: Roadmap and Milestones Page 4 of 6 PTAA Annex R: Roadmap and Milestones 8. Research track (Annex S) and its relation to this roadmap The research track is separate and conditional. Its stages (S0 to S3, and Stage R) are defined in Annex S. The following dependencies apply: Research stage Earliest start Dependencies S0 Design review S1 Closed-sandbox A5, parameters only S2 Extended sandbox S3 Cross-border limited use Stage R After the Phase 0 consultation on the Core, in parallel with Phase 1 Not before Gate G1 Separate consultation on Annex S S0 approval; independent review; no activity touching the operational pilot Not before the middle of Phase 2 Not before Gate G2 Not scheduled S1 completed without unresolved safety-gate failures Joint supervisory approval Separate proposal and approval The research track can be delayed, narrowed or cancelled by the supervisors at any time without affecting the operational roadmap. 9. Dependencies and assumptions Political and regulatory commitment in both jurisdictions to the pilot domain. Operational status of the chosen sandbox (to be confirmed in D0.6). Availability of independent assessors and test infrastructure. Access to resources for the Steering Group secretariat and the learning loop function. Stable legal basis for cross-boundary data exchange for the pilot. 10. Resource outline (to be costed) Secretariat and programme management; technical working group; assessors and test infrastructure; registry operation; supervisory capacity in both jurisdictions; consultation and translation; independent evaluation. Costing is not included here and is a Phase 0 deliverable. Page 5 of 6 PTAA Annex R: Roadmap and Milestones 11. Reporting The Steering Group publishes a semi-annual progress report covering milestones, metrics, incidents (aggregated), holds and stops, and plans for the next period. Gate decisions are published with their evidence.
PTAA Annex V: Source Verification Log Companion to: A Cross-Border Protocol for Autonomous Agents (PTAA), Core v1.1 | Version 1.1 | 17 November 2026 Status: Working log. Entries marked as not verified must be checked against primary sources before any circulation of the Core beyond the authors. All access dates are 17 November 2026 unless stated. Authors: Roberto De Biase, Rigene Project 1. Purpose and method This log lists every factual claim in the PTAA documents that depends on an external source, states how far it has been verified, and says what must be done. It is a condition of credibility for the Core: reviewers should be able to trace every figure and date to a source. Verification statuses Status V1 Verified (primary) Meaning Confirmed in an official publication of the body concerned. V2 Verified (secondary) Confirmed in reputable secondary reporting; the primary source still has to be opened and cited. P Partially verified U Not verified R Removed Part of the claim is supported; the exact wording or figure needs adjustment. Not found in the checks carried out. Marked ‡ in the Core. Not to be relied upon. Removed from v1.1 because it could not be verified or was not necessary. Method. Searches were run against official press releases and government portals, followed by reputable news outlets. A claim was marked verified only where the figure and date matched. Absence of a result is not evidence that a claim is false; it only means that it has not been verified. Page 1 of 7 2. Verified claims ID Claim Used in Status Source Action S1 On 27 August 2026 HKMA, SFC, IA and MPFA, with Cyberport, announced the first cohort of the GenA.I. Sandbox++: 36 use cases selected from nearly 100 proposals, involving 30 financial institutions and 27 technology partners Core §1, §2.1 V1 HKMA press release: https://www.hkma.gov.hk/eng/news-and media/press-releases/2026/08/20260827-3/ ; Hong Kong Government information service: https://www.info.gov.hk/gia/general/202608/27/P20 26082600563.htm Cite the HKMA release and annexes in the final version S2 The testing covers end-to end processes such as customer onboarding, payments, insurance claims and customer interactions; the cohort continues an "AI vs AI" theme (AI overseeing other AI) Core §1 V1 Same as S1 None S3 The first cohort focuses on agentic AI applications that go beyond content generation Core §1 V1 / V2 Hong Kong Government information service (see S1); Mondo Visione republication of the regulators' release: https://mondovisione.com/news/hong kong-monetary-authority-and-hong-securities and-futures-commission-first-c-2026827/ None S4 Technical trials of the first cohort are expected to begin later in 2026 Core §1, §6; Annex R V2 The Standard: https://www.thestandard.com.hk/innovation/article /341121/Hong-Kong-unveils-first-cohort-of-GenAI Sandbox ; PYMNTS: https://www.pymnts.com/? p=4126810 Confirm start date in the HKMA release annex or on the sandbox page S5 The Sandbox++ was launched in March 2026 Core §2.1 V2 Footnote link in the HKMA release to the March 2026 press release (https://www.hkma.gov.hk/eng/news-and media/press-releases/2026/03/20260305-3/); TechNode: https://technode.global/?p=990960 Open and cite the March release PTAA Annex V: Source Verification Log Page 2 of 7 ID Claim Used in Status Source Action S6 Earlier GenA.I. Sandbox cohorts for banks ran from 2024 (Dec 2024 and Oct 2025 cohorts) Core §2.1 V2 TechNode (above); AI2Work summary: https://ai2.work/blog/hong-kong-regulators-select 36-agentic-ai-pilots-for-sandbox Verify cohort dates and sizes in HKMA releases S7 Shenzhen is drafting a pilot plan for cross boundary low altitude logistics, after discussions with national authorities and the Hong Kong government, prioritising three routes: Shekou to Tuen Mun (River Trade Terminal), Wenjindu to Sheung Shui, and an internal Hetao route; first route targeted within 2026 Core §1; Annex R V1 / V2 Shenzhen Government Online (re-publishing China Daily, 24 March 2026): https://www.sz.gov.cn/en_szgov/news/infocus/Low AltitudeEconomy/LowAltitudeEconomyinFocus/co ntent/post_12716545.html ; China Daily Hong Kong, 25 March 2026: https://www.chinadailyhk.com/article/631051 Obtain the official pilot plan when published S8 In June 2026 a cross-boundary insurance policy for ultralight eVTOL was launched by PICC's Shenzhen branch and PICC (Hong Kong): underwritten in Shenzhen, issued in Hong Kong, resolving a barrier to mainland operators flying in Hong Kong Core §2.1 V2 China Daily Hong Kong, 22 June 2026: https://www.chinadailyhk.com/hk/article/635337 ; Shenzhen Government Online, 24 June 2026: https://www.sz.gov.cn/en_szgov/news/infocus/Low AltitudeEconomy/LowAltitudeEconomyinFocus/co ntent/post_12861788.html Confirm with the insurers' announcements PTAA Annex V: Source Verification Log Page 3 of 7 3. Partially verified claims ID Claim Used in Status Source Action S9 Cross boundary flights between Hong Kong and Shenzhen require approvals from both mainland and Hong Kong authorities because the cities operate under different airspace management systems Core §2.2 V2 (attributed statement) Statement by a Hong Kong deputy to the National People's Congress reported in the China Daily articles cited in S8 Cite as an attributed statement; add primary regulatory references (CAAC and Hong Kong Civil Aviation Department) ID Claim Used in Status Source and finding Action P1 Shenzhen's low altitude logistics network: more than 300 routes and more than 1,200 take-off and landing sites Core §2.1 (marked ‡) P Macao News (undated): https://macaonews.org/?p=104054 ; CGTN reports 1,284 take-off and landing sites by March 2026: https://news.cgtn.com/news/2026 05-22/Shenzhen-s-low-altitude economy-soars-as-city-leads-global drone-race-1Nm7IY3OUbm/p.html Replace with a dated municipal figure P2 The cross-boundary low-altitude regulatory sandbox pilot was reported in June 2026 as expected to launch that month Core §6; Annex R P Reported in the China Daily pieces cited in S8. No confirmation of actual launch or of approved flight permits was found. Confirm status with the sandbox operators and the relevant authorities; keep the Phase 1 pilot condition until then P3 Regular cross boundary drone cargo service between Shenzhen and Hong Kong has begun Not used in the Core P Macao News (undated, citing mainland media): https://macaonews.org/?p=104054 Do not use until a dated official source is found PTAA Annex V: Source Verification Log Page 4 of 7 4. Claims not verified (marked ‡ in the Core) ID Claim Used in Status Source and finding Action P4 Highly automated factories exist in the Pearl River Delta region Core §2.1 P Reports of lights-out factories in Guangdong (for example in Guangzhou and Zhuhai) were found; Shenzhen-specific facts were not verified Reword to the Pearl River Delta and cite named facilities with dates P5 Hong Kong has published guidance on generative AI in banking Core §2.1 P The existence of the earlier sandbox cohorts is verified (S6); the existence and content of published guidance was not verified Identify and cite the HKMA publication or remove the phrase P6 Agentic and automated systems are being integrated into Shenzhen's manufacturing, ports and logistics Core §1 P Supported for logistics (S7, P1); not specifically verified for ports and manufacturing Add specific, dated examples ID Claim Used in Status Action U1 The Shenzhen Park Ordinance has been in force since 1 March 2026 and introduces a differentiated supervision model and an innovation fault tolerance mechanism Core §2.1 U Obtain the text and official English summary from the Shenzhen legislature or Hetao administration U2 The cross-border data verification platform was launched in May 2024, with 24 institutions, 15 scenarios, more than 50,000 verifications, over HKD 2.4 billion of facilitated financing and medical imaging exchange for more than 15,000 residents Core §2.1 U Obtain official figures with dates from the platform operators or administrations PTAA Annex V: Source Verification Log Page 5 of 7 5. Claims removed from v1.1 ID Claim Used in Status Action U3 Mutual recognition of professional and fintech credentials in the Greater Bay Area, including a "one exam, three certificates" mechanism Core §2.1 U Identify the official instruments and dates U4 Hetao is described as a zone with a rule of party choice (Hong Kong or Macao law) for certain contract disputes Core §4.2 (principle referenced) U Verify in the legal instrument before citing ID Claim in v1.0 Reason for removal R1 An HKMA "ART" (Accountability, Resilience, Trust) framework published in October 2026, including intent-anchored accountability and a register of significant AI agents Not found in the checks; the Core states only that intent-anchored accountability may be adopted by supervisors, as a possibility R2 APEC engagement on cross-border data verification, including a Shenzhen Singapore pilot Not found in the checks; unnecessary for the proposal R3 The "four interconnections" objective of the low-altitude logistics sandbox Not found in the checks R4 A reference to "agentic AI safety testing mutual recognition" with Shenzhen–Hong Kong named as a pilot site Could not be identified as an existing source R5 A reference to "governance primitives for AI agent orchestration" Generic; replaced by the technical specification in Annex T1 R6 A reference to "cross-border regulatory sandboxes and shared evaluation methods" as a literature item Generic; no specific source PTAA Annex V: Source Verification Log Page 6 of 7 PTAA Annex V: Source Verification Log 6. Corrections applied between v1.0 and v1.1 Item v1.0 v1.1 Low-altitude sandbox timing HKMA figures Data platform figures Park Ordinance Roadmap Stated as "scheduled for June 2026" in a November 2026 document Cited without dates Stated as facts Stated as fact Single track Reported as expected in June 2026, with launch status unconfirmed; Phase 1 pilot made conditional Dated 27 August 2026 and aligned with the official release Marked ‡ Marked ‡ Separate operational and research tracks (Annex R) 7. Verification procedure for Phase 0 (D0.6) 1. Assign each U and P item to a reviewer. 2. Obtain the primary source (official text, press release, or data set) and record URL, document title, date and access date. 3. Where a primary source is unavailable, replace the claim with an attributed statement ("according to ...") or remove it. 4. Have a second reviewer check each figure against the source. 5. Update the Core and annexes, and re-issue this log with statuses changed. 6. Repeat the check immediately before any public release, because several items concern fast-moving developments (sandbox launches, trial start dates, route approvals). 8. Open items most likely to affect the proposal 1. Operational status of the cross-boundary low-altitude regulatory sandbox and of flight approvals (P2). 2. Start date and scope of Sandbox++ technical trials (S4). 3. Existence, scope and text of the Hetao legal instruments referred to (U1, U4). 4. Official figures for the cross-border data platform (U2).
PTAA Annex C: Consultation Plan Companion to: A Cross-Border Protocol for Autonomous Agents (PTAA), Core v1.1 | Version 1.1 | 17 November 2026 Status: Draft plan for the authors and the future Joint Steering Group. The authors have no mandate from any authority; the plan describes how the proposal would be put forward for consultation. Authors: Roberto De Biase, Rigene Project 1. Objectives 1. Test the core assumption of the PTAA: that a minimal shared constitutional core, mutual recognition and an orchestration protocol can unlock cross-border agentic operation safely. 2. Obtain informed, documented feedback from supervisors, industry, researchers and civil society. 3. Identify a first pilot domain and the institutional partners willing to sponsor it. 4. Run a separate, open consultation on Annex S so that its research-stage proposals are judged on their own merits. 2. Principles Full disclosure. Every audience receives the whole document set, including Annex S, with a plain description of its status. No audience is given a version that hides the long-term intent of the authors. Separation of tracks. The Core and Annex S are consulted separately so that views on the Core are not conflated with views on research-stage proposals. Opt-in participation. Nobody is asked to endorse Annex S in order to engage with the Core. Verified facts. No document is circulated until the Source Verification Log (Annex V) has no unresolved items marked ‡ in the passages that are relied on. Accurate roles. The authors present themselves as proponents of a proposal with no official standing, and do not imply endorsement by any authority. Respect for jurisdiction. Approaches are made through appropriate channels in each jurisdiction and with respect for each side's procedures and languages. Published outcomes. Comments received and the responses to them are published, including objections. Page 1 of 7 3. Stakeholders Group Examples Interest Main questions to ask Supervisors (Hong Kong) HKMA and other financial regulators involved in the GenA.I. Sandbox++; the authorities responsible for aviation and logistics where relevant Supervisory control, accountability, financial stability and consumer protection Is the SCC compatible with existing supervisory expectations? Which agent classes could be recognised, and in which domains? Supervisors and administrators (Shenzhen, Hetao) The Hetao administration and the Shenzhen authorities responsible for sandboxes, ports, airspace, data and industry Regulatory innovation, security, integration with national frameworks Which sandboxes can participate? What is the correct channel for mutual recognition? Cyberport and sandbox operators Operators of Hong Kong and Shenzhen sandboxes Operational feasibility How do certification and admission procedures map onto the SCC? Industry Financial institutions, logistics firms, drone and robotics operators, technology providers, SMEs Reduced friction, certainty Which pilot has the best value-to-risk ratio? What is the burden of the SCC? Academia and research AI safety, formal verification, robotics, law, governance Rigour, independence Are the verification methods adequate? Are the metrics meaningful? Civil society and consumer bodies Digital rights, consumer protection, professional associations Accountability, privacy, fairness Are accountability and redress adequate? PTAA Annex C: Consultation Plan Page 2 of 7 4. Sequence Group Examples Interest Main questions to ask Independent assessors Testing and certification bodies Feasibility of the verification requirements Can they be run at a reasonable cost and independence? Multilateral and peer regions Bodies working on AI governance interoperability Replicability What would be needed for the model to be reproduced elsewhere? Step Timing (from start) Activity Output 1 Weeks 0–6 Complete Annex V verification; finalise document set; translate the executive summary into Traditional and Simplified Chinese; legal and technical read-through by two independent reviewers (one technical, one policy) Verified document set; translations; reviewer comments 2 Weeks 4–8 Informal preliminary contacts with a small number of academic and industry experts to test clarity and feasibility Revised drafts; list of issues 3 Weeks 8–12 Approach to supervisory and administrative contacts through appropriate channels: introduce the proposal, state its status, request guidance on process Guidance on the correct consultation channel; first feedback PTAA Annex C: Consultation Plan Page 3 of 7 5. Methods Written comment. Structured comment template tied to section numbers, with space to say whether an objection is fundamental or minor. Workshops. Moderated sessions on specific topics: SCC invariants; class definitions; pilot domain; learning loop governance; Annex S. Expert review. Independent reviewers asked to attempt to find weaknesses (a "red team" review of the proposal). Structured interviews. With supervisors, operators and assessors on feasibility. Scenario exercises. Table-top exercises on incidents (loss of control, cross-boundary data breach, contested recognition). Step Timing (from start) Activity Output 4 Weeks 12–24 Public consultation on the Core and Annexes T1, R, V and C: open comment period; at least two stakeholder workshops (one in each jurisdiction or online); written submissions Consultation report; amended Core 5 In parallel from week 12 Separate consultation on Annex S: expert panel (AI safety, security, ethics, law), open written comments, one public session Annex S consultation report; recommendation to proceed, narrow or stop 6 Weeks 24–30 Revision and publication of Core v1.2 or v2.0, with a response to every substantive comment Published revised documents and response table 7 Weeks 28–36 Request to supervisors to sponsor Phase 0 and constitute the Joint Steering Group Decision to proceed or not PTAA Annex C: Consultation Plan Page 4 of 7 PTAA Annex C: Consultation Plan 6. Materials 1. Two-page summary of the PTAA in English, Traditional Chinese and Simplified Chinese. 2. Core v1.1. 3. Annexes T1, R, V and C. 4. Annex S, with a one-page statement of its status and what it does and does not propose. 5. Comment template and consultation rules. 6. FAQ: what the PTAA is and is not; what mutual recognition means in practice; what Annex S proposes and what it does not. 7. Communication guidelines Describe the proposal as it is: a governance framework for agent certification and recognition, with a separate research annex. Describe the research programme behind Annex S in plain language, including its long-term vision and the reasons for caution, as in Annex S Section 1. Avoid claims of endorsement, partnership or official status. Avoid forecasts of benefits that are not supported by evidence. Acknowledge risks and open questions openly. Respond to every substantive comment, including disagreement, in writing. Do not describe confidential conversations or attribute views to supervisors without their consent. 8. Handling of feedback Type of feedback Response Factual error Technical weakness Fundamental objection to the approach Correct and update Annex V Assign to the technical working group; document the fix or the reason for declining Discuss in the consultation report; consider narrowing scope or withdrawing the affected element Page 5 of 7 9. Success criteria for the consultation At least one supervisory body in each jurisdiction engaged in dialogue. Substantive written comments from at least four stakeholder groups, including civil society and independent researchers. A consultation report published with a response to each substantive comment. A clear decision on the pilot domain, or a clear reason why none is ready. A documented position on Annex S: proceed to S0 review, narrow, defer or withdraw. 10. Risks to the consultation process Type of feedback Response Objection to Annex S Treated separately; the consultation outcome can lead to narrowing, deferral or withdrawal of Annex S without affecting the Core Offer of sponsorship or pilot hosting Refer to the Steering Group once constituted, through the appropriate channels Risk Mitigation Perceived overreach: the authors are seen as claiming an official role Clear statement of status in every document and meeting Annex S dominates discussion and obscures the Core Separate tracks; clear one-page status statement Misunderstanding due to translation Professional translation and review by native-speaking policy experts Misreported or outdated facts Annex V verification before release; re verification before each public step Capture by vested interests Open participation; published responses; advisory panel with diverse membership Delay due to lack of institutional sponsor Keep each document independently useful; seek sponsorship from several bodies PTAA Annex C: Consultation Plan Page 6 of 7 PTAA Annex C: Consultation Plan 11. Immediate next actions 1. Complete the Annex V verification items marked U and P. 2. Insert authors and contact details; confirm the organisational affiliation and how the authors wish to be described. 3. Commission translations of the executive summary. 4. Identify two independent reviewers (technical and policy). 5. Prepare the one-page status statement for Annex S and the FAQ.