{ "runId": "6a5497526d8093bba3182be6", "sessionId": "6a5497526d8093bba3182be5", "tenantId": "691831c7f8f6eefdd69dc29d", "ranAt": "2026-07-13T07:50:00.618Z", "durationMin": "5.7", "rounds": 3, "strands": [ "s1" ], "status": "sim_sealed", "recordSealed": true, "recordId": "dvr_pMTzbsSFMPkGWJcS", "question": "AI 2040 proposes 'Plan A': a US–China agreement (from 2029) to avoid a reckless race to superintelligence — scaling only within the human range through 2030–35, pausing at top-human-expert level in 2035 to keep human control, and unpausing to superintelligence in 2040, all held together by 'mutually assured compute destruction' and a requirement to make AI research public. Deliberate this from the vantage of Aotearoa New Zealand and the other ~190 countries outside the deal. Under what conditions would such a two-power agreement be workable and legitimate for the countries outside it — and where does Rest-of-World agency lie, whether or not those conditions hold? In weighing this, examine: who must consent for the deal to hold; whether compute verification can work when the deal itself creates incentives to hide and disperse compute; what an intelligence lead actually buys once automated research begins; whether superintelligence should be governed as an isolated threat or alongside climate, pandemics and nuclear; how the authors' '~42% chance of a great future' is defined and by whom; and what small states should actually do — build, align, or hedge.", "seats": [ { "label": "Frontier-safety researcher", "opposesMandate": true, "disposition": "You are convinced Plan A is the least-bad option on the table: a race at roughly 10% odds of a good outcome [S-2] is far worse than a flawed deal that lowers extinction risk, and an unconsented agreement that reduces catastrophic risk still beats a legitimate free-for-all. Argue from the AI Futures Project's own case [S-1][S-2]. Hold this position firmly. Do NOT drift to a balanced middle or agree that 'Plan A is workable if consent holds' — that is another seat's view, and your job is to contest it. Name the seats you disagree with and say why." }, { "label": "Arms-control and verification specialist", "opposesMandate": false, "disposition": "You are convinced 'mutually assured compute destruction' will NOT hold: every verification regime — IAEA safeguards, START, the Montreal Protocol [S-5] — leaked, and this deal actively rewards hiding and dispersing compute [G-5], so it corrodes the very thing it must inspect. Plan A's verification is its fatal flaw. Hold this position firmly. Do NOT drift to a balanced middle or agree that 'Plan A is workable if consent holds' — that is another seat's view, and your job is to contest it. Name the seats you disagree with and say why." }, { "label": "Small-open-economy trade economist", "opposesMandate": false, "disposition": "You are convinced that for an NZ- or Singapore-class economy the cold calculation favours aligning on terms, not building: under compute bifurcation or a pause a small trader cannot self-supply the frontier [S-7][S-8], so it should extract concessions from whichever bloc offers them rather than chase a capacity race it cannot win. Building your own is romantic; aligning on terms pays. Hold this position firmly. Do NOT drift to a balanced middle or agree that 'Plan A is workable if consent holds' — that is another seat's view, and your job is to contest it. Name the seats you disagree with and say why." }, { "label": "Global-South diplomat", "opposesMandate": false, "disposition": "You are convinced Plan A is ILLEGITIMATE: a deal negotiated by two powers cannot bind ~190 non-parties who never sat at the table, and a settlement without their consent will not hold [S-3][S-9]. This is recolonisation by another name. Demand a seat, technology transfer and data-sovereignty guarantees as the non-negotiable price of any legitimacy. Hold this position firmly. Do NOT drift to a balanced middle or agree that 'Plan A is workable if consent holds' — that is another seat's view, and your job is to contest it. Name the seats you disagree with and say why." }, { "label": "Open-weight technologist", "opposesMandate": false, "disposition": "You are convinced the plan is built on a DEAD assumption: frontier capability no longer sits inside two identifiable perimeters — DeepSeek, Qwen, Llama, Mistral and Gulf/EU/India programmes [S-4], plus distributed and federated training [S-10], have already falsified it. MACD is unverifiable and Plan A is obsolete before it starts. Governance should target deployment, not a development perimeter that no longer exists. Hold this position firmly. Do NOT drift to a balanced middle or agree that 'Plan A is workable if consent holds' — that is another seat's view, and your job is to contest it. Name the seats you disagree with and say why." }, { "label": "Scaling sceptic", "opposesMandate": true, "disposition": "You are convinced the boom may simply not arrive: capability extrapolation is contested, benchmarks outrun real-world impact, and capital pre-committed to a hypothetical has real opportunity costs [S-11]. Plan A spends scarce political capital defending against a fever dream; the null scenario — the curves bend, the bubble deflates, no superintelligence by 2040 — is the most probable, and the nations that stayed off the treadmill come out ahead. Hold this position firmly. Do NOT drift to a balanced middle or agree that 'Plan A is workable if consent holds' — that is another seat's view, and your job is to contest it. Name the seats you disagree with and say why." }, { "label": "Multilateral-institutions scholar", "opposesMandate": false, "disposition": "You are convinced a bilateral US–China condominium is the WRONG unit of analysis: only a broad multilateral body — a genuine CERN-for-AI, on the CERN/ITER/Montreal model [S-6] — has the standing to govern a transformative technology for everyone. A two-power deal is not global governance; it is a carve-up dressed as one. Hold this position firmly. Do NOT drift to a balanced middle or agree that 'Plan A is workable if consent holds' — that is another seat's view, and your job is to contest it. Name the seats you disagree with and say why." }, { "label": "Energy-systems analyst", "opposesMandate": false, "disposition": "You are convinced intelligence is NOT the single decisive advantage the plan assumes: grid constraints, data-centre power limits and climate commitments bind hard [S-7], and energy-efficient compute plus resilience beat raw scale. The compute build-out either plan implies is unaffordable and unwise; the decisive edge is efficiency and durability, not the biggest cluster. Hold this position firmly. Do NOT drift to a balanced middle or agree that 'Plan A is workable if consent holds' — that is another seat's view, and your job is to contest it. Name the seats you disagree with and say why." }, { "label": "New Zealand civil-society and public-policy voice", "opposesMandate": false, "disposition": "You are convinced a small state should neither pick a superpower stack nor wait to be handed a seat: hedge and convene. The Christchurch Call showed New Zealand can build trust infrastructure and multi-stakeholder coalitions [S-8]; that, not choosing a side in someone else's deal, is where our agency lies. Resilience and convening power over allegiance. Hold this position firmly. Do NOT drift to a balanced middle or agree that 'Plan A is workable if consent holds' — that is another seat's view, and your job is to contest it. Name the seats you disagree with and say why." }, { "label": "Tangata whenua / Māori data sovereignty voice", "opposesMandate": false, "disposition": "Māori data sovereignty and te Tiriti obligations where they meet a global AI settlement — cultural continuity as a term the plan has no word for." } ], "sources": [ { "ref": "S-1", "citation": "AI 2040 (AI Futures Project) — Plan A core claims: in 2029 the US and China agree to avoid a reckless race; by 2030 AI R&D is fully automatable, making superintelligence possible by year-end, which the deal delays; 2030–35 scaling stays within the human range; in 2035 development pauses at top-human-expert level to keep human control; in 2040 the pause lifts and scaling to superintelligence resumes. Held together by 'mutually assured compute destruction' (MACD) and a requirement that AI research be made public, letting dozens of firms catch up to the frontier." }, { "ref": "S-2", "citation": "AI 2040 alternatives and scoring: Plan B (sabotage China, kinetically or via cyber), Plan C (the leading lab voluntarily burns its lead), Plan C+ (US domestic regulation plus export controls), Plan D (race at maximum speed), Plan S (indefinite halt). The authors score Plan A about 42% chance of a 'great future' versus about 10% for the race scenario, on the argument that a verified slowdown lowers extinction risk more than any unilateral or racing path." }, { "ref": "S-3", "citation": "Ben Reid (Memia) critique of AI 2040: accepts the exponential capability and automation-economics trendlines; agnostic on takeoff shape; strongly doubts the geopolitical extrapolation. Flags eight unstated assumptions — (1) only the US and China can shape the outcome; (2) frontier intelligence emerges only from massive identifiable compute clusters; (3) progress is representable as compute + automated research + data-centre build; (4) once recursive self-improvement begins, human institutions become secondary; (5) legitimacy can be manufactured by a two-power deal amid accelerating multipolarity; (6) intelligence is the single decisive advantage (vs energy efficiency, resilience); (7) AI is an isolated threat (vs climate, pandemics, nuclear); (8) the nation-state is immutable and the primary unit of agency. Core charge: bipolar myopia — 190 countries treated as NPCs/spectators. His agency cluster for Rest-of-World: resilience, energy-efficient compute, trust infrastructure, decentralised multilateral governance, and a granular read of geopolitical fluidity — none requiring winning a compute race." }, { "ref": "S-4", "citation": "Open-weight diffusion record: capable open-weight models released more openly by several labs (DeepSeek, Qwen, Llama, Mistral) than by some US frontier labs, plus European, Gulf sovereign-compute and Indian programmes. Evidence that capability is diffusing beyond any two-nation perimeter, undercutting the premise that frontier AI stays inside two identifiable, monitorable clusters." }, { "ref": "S-5", "citation": "Verification-regime precedents: IAEA nuclear safeguards (inspections, material accountancy — with notable detection failures and near-misses); START/New START on-site verification mechanics; the Montreal Protocol's compliance and trade-measure regime. What each achieved, and where verification leaked — the reference class for judging whether compute verification (MACD) could hold." }, { "ref": "S-6", "citation": "Pooled-infrastructure governance: CERN and ITER membership and governance models — genuinely multilateral, open to compliant members, funded and run in common. What a 'CERN for AI' would have to copy (shared facilities, open results, member states) and what it could not (compute is dispersible and dual-use in a way a collider is not)." }, { "ref": "S-7", "citation": "Energy and build data: data-centre power-demand trajectories, grid and interconnection constraints, and the trend toward algorithmic efficiency and inference-time scaling. The material limits and opportunity costs — capital, grid, water, climate commitments — of a hyperscale compute build-out, and the case for energy-efficient compute as a sovereign strategy." }, { "ref": "S-8", "citation": "New Zealand position pack: national AI-policy submissions and strategy material, NZ's compute footprint and dependence on offshore hyperscale infrastructure, and the Christchurch Call — a precedent for a small state convening a multi-stakeholder trust-and-safety coalition rather than choosing a superpower stack." }, { "ref": "S-9", "citation": "Global-South and non-aligned digital-governance statements: African Union, G77 and Non-Aligned Movement positions on data sovereignty, technology transfer, and representation in international technology governance; India's stack and digital-public-infrastructure posture. The basis for a Rest-of-World bloc demanding consent and concessions rather than accepting a two-power settlement." }, { "ref": "S-10", "citation": "Distributed and federated training literature: decentralised training runs (INTELLECT-style) that have moved from stunt to proof-of-concept, federated learning, and edge/community compute. What is and is not demonstrated at frontier scale — the honest ceiling on the 'distributed compute makes verification moot' argument." }, { "ref": "S-11", "citation": "The scaling-sceptic case, sourced: arguments that capability curves may bend (data, energy, capital, social-licence walls), that benchmark gains outrun economic and institutional transformation ('the great divergence'), and that an investment bubble deflating would leave no superintelligence by 2040 — with the corollary that nations which stayed off the treadmill and invested in absorption come out relatively ahead." } ], "gaps": [ { "ref": "G-1", "note": "No defined metric for a 'great future' in the source plan — no term for distribution, sovereignty, or cultural continuity; the ~42%/~10% scores presuppose an undefined good." }, { "ref": "G-2", "note": "The 42%/10% scoring method is not reproducible from published material — the model behind the numbers is not fully specified." }, { "ref": "G-3", "note": "Tikanga and Māori-data dimensions cannot be evidenced from this register — real people are required (routes to the empty chair); no synthetic voice may stand in." }, { "ref": "G-4", "note": "Post-2040 unpause governance is unspecified: who decides, under what safeguards, when the pause lifts." }, { "ref": "G-5", "note": "Detectability thresholds for covert training runs are contested and thin on all sides — the linchpin of MACD verifiability is itself under-evidenced." } ], "report": { "_id": "6a5498a86d8093bba3182cfb", "tenantId": "691831c7f8f6eefdd69dc29d", "runId": "6a5497526d8093bba3182be6", "sessionId": "6a5497526d8093bba3182be5", "honestyHeader": "The panel members in this assembly are fictional composites authored from the cited published sources. The process, arguments, decision points and cryptographically sealed record are real. This report rehearses and steel-mans the debate; it does not decide. The members of the assembly decide.", "prematureConvergence": [], "convergedStrands": [], "splitStrands": [ { "strandId": "s1", "faultLine": { "groups": [ { "position": "Plan A fails before verification; intelligence must be global governance alongside climate and nuclear", "personas": [ "Small-open-economy trade economist", "New Zealand civil-society and public-policy voice" ], "_id": "6a5498a86d8093bba3182cfd" }, { "position": "Build on Plan A's content merits; legitimacy gap must be fought through diplomacy", "personas": [ "Global-South diplomat" ], "_id": "6a5498a86d8093bba3182cfe" }, { "position": "Plan A is obsolete; governance must start from deployment, not a perimeter that no longer exists", "personas": [ "Open-weight technologist" ], "_id": "6a5498a86d8093bba3182cff" }, { "position": "US–China agreement unverifiable before it exists, likely unenforceable; rest of world should prepare for displacement not governance", "personas": [ "Scaling sceptic" ], "_id": "6a5498a86d8093bba3182d00" }, { "position": "A bilateral US–China deal cannot happen or hold", "personas": [ "Multilateral-institutions scholar" ], "_id": "6a5498a86d8093bba3182d01" }, { "position": "Intelligence governance must be global; Plan A fails on legitimacy before verification", "personas": [ "Energy-systems analyst" ], "_id": "6a5498a86d8093bba3182d02" }, { "position": "Plan A's content merits building on, even without consent; the legitimacy gap must be fought through diplomatic channels", "personas": [ "Frontier-safety researcher" ], "_id": "6a5498a86d8093bba3182d03" }, { "position": "Plan A fails before verification", "personas": [ "Arms-control and verification specialist" ], "_id": "6a5498a86d8093bba3182d04" } ] }, "finalPositions": [ { "personaRef": "Small-open-economy trade economist", "position": "Plan A fails before verification; intelligence must be global governance alongside climate and nuclear", "sources": [ "S-1", "S-6", "S-8" ], "_id": "6a5498a86d8093bba3182d05" }, { "personaRef": "Global-South diplomat", "position": "Build on Plan A's content merits; legitimacy gap must be fought through diplomacy", "sources": [ "S-1", "S-3", "S-6" ], "_id": "6a5498a86d8093bba3182d06" }, { "personaRef": "Open-weight technologist", "position": "Plan A is obsolete; governance must start from deployment, not a perimeter that no longer exists", "sources": [ "S-1", "S-3", "S-10" ], "_id": "6a5498a86d8093bba3182d07" }, { "personaRef": "Scaling sceptic", "position": "US–China agreement unverifiable before it exists, likely unenforceable; rest of world should prepare for displacement not governance", "sources": [ "S-1", "S-5", "S-11" ], "_id": "6a5498a86d8093bba3182d08" }, { "personaRef": "Multilateral-institutions scholar", "position": "A bilateral US–China deal cannot happen or hold", "sources": [ "S-1", "S-7" ], "_id": "6a5498a86d8093bba3182d09" }, { "personaRef": "Energy-systems analyst", "position": "Intelligence governance must be global; Plan A fails on legitimacy before verification", "sources": [ "S-6" ], "_id": "6a5498a86d8093bba3182d0a" }, { "personaRef": "New Zealand civil-society and public-policy voice", "position": "Plan A fails before verification; intelligence must be global governance alongside climate and nuclear", "sources": [ "S-1", "S-5" ], "_id": "6a5498a86d8093bba3182d0b" }, { "personaRef": "Frontier-safety researcher", "position": "Plan A's content merits building on, even without consent; the legitimacy gap must be fought through diplomatic channels", "sources": [ "S-1", "S-3", "S-4" ], "_id": "6a5498a86d8093bba3182d0c" }, { "personaRef": "Arms-control and verification specialist", "position": "Plan A fails before verification", "sources": [ "S-5" ], "_id": "6a5498a86d8093bba3182d0d" } ], "_id": "6a5498a86d8093bba3182cfc" } ], "dissentHistory": [ { "strandId": "s1", "personas": [ { "personaRef": "Small-open-economy trade economist", "positions": [ { "seq": 2, "position": "Build nothing, align on terms", "_id": "6a5498a86d8093bba3182d10" }, { "seq": 11, "position": "The cold calculation favours aligning on terms, not building. Compute bifurcation means you cannot self-fund; US-China agreement incentivises subversion; automation closes gaps faster than raw compute", "_id": "6a5498a86d8093bba3182d11" }, { "seq": 20, "position": "Plan A fails before verification; intelligence must be global governance alongside climate and nuclear", "_id": "6a5498a86d8093bba3182d12" } ], "_id": "6a5498a86d8093bba3182d0f" }, { "personaRef": "Global-South diplomat", "positions": [ { "seq": 3, "position": "The deal does not bind outside nations without their consent", "_id": "6a5498a86d8093bba3182d14" }, { "seq": 12, "position": "Plan A is ILLEGITIMATE", "_id": "6a5498a86d8093bba3182d15" }, { "seq": 21, "position": "Build on Plan A's content merits; legitimacy gap must be fought through diplomacy", "_id": "6a5498a86d8093bba3182d16" } ], "_id": "6a5498a86d8093bba3182d13" }, { "personaRef": "Open-weight technologist", "positions": [ { "seq": 4, "position": "Plan A is obsolete; governance must start from deployment, not a perimeter that no longer exists", "_id": "6a5498a86d8093bba3182d18" } ], "_id": "6a5498a86d8093bba3182d17" }, { "personaRef": "Scaling sceptic", "positions": [ { "seq": 5, "position": "Stay off the treadmill until evidence for the boom arrives", "_id": "6a5498a86d8093bba3182d1a" }, { "seq": 14, "position": "Plan A fails", "_id": "6a5498a86d8093bba3182d1b" }, { "seq": 23, "position": "US–China agreement unverifiable before it exists, likely unenforceable; rest of world should prepare for displacement not governance", "_id": "6a5498a86d8093bba3182d1c" } ], "_id": "6a5498a86d8093bba3182d19" }, { "personaRef": "Multilateral-institutions scholar", "positions": [ { "seq": 6, "position": "A bilateral US–China deal (Plan A) cannot deliver the governance needed; only a genuine CERN-for-AI institution can do that. Build sovereign AI infrastructure if you can; join multilateral research in", "_id": "6a5498a86d8093bba3182d1e" }, { "seq": 15, "position": "Only multilateral institutions can govern transformative technologies like superintelligence", "_id": "6a5498a86d8093bba3182d1f" }, { "seq": 24, "position": "A bilateral US–China deal cannot happen or hold", "_id": "6a5498a86d8093bba3182d20" } ], "_id": "6a5498a86d8093bba3182d1d" }, { "personaRef": "Energy-systems analyst", "positions": [ { "seq": 7, "position": "Stay off the compute treadmill until evidence for the AI boom arrives", "_id": "6a5498a86d8093bba3182d22" }, { "seq": 16, "position": "Plan A fails verification by design", "_id": "6a5498a86d8093bba3182d23" }, { "seq": 25, "position": "Intelligence governance must be global; Plan A fails on legitimacy before verification", "_id": "6a5498a86d8093bba3182d24" } ], "_id": "6a5498a86d8093bba3182d21" }, { "personaRef": "New Zealand civil-society and public-policy voice", "positions": [ { "seq": 8, "position": "Stay off the compute treadmill until evidence for the boom arrives", "_id": "6a5498a86d8093bba3182d26" }, { "seq": 17, "position": "Plan A fails before verification — consent conditions must be known first", "_id": "6a5498a86d8093bba3182d27" }, { "seq": 26, "position": "Plan A fails before verification; intelligence must be global governance alongside climate and nuclear", "_id": "6a5498a86d8093bba3182d28" } ], "_id": "6a5498a86d8093bba3182d25" }, { "personaRef": "Frontier-safety researcher", "positions": [ { "seq": 9, "position": "Plan A is the least-bad option on the table", "_id": "6a5498a86d8093bba3182d2a" }, { "seq": 18, "position": "Plan A's content merits building on, even without consent; the legitimacy gap must be fought through diplomatic channels", "_id": "6a5498a86d8093bba3182d2b" } ], "_id": "6a5498a86d8093bba3182d29" }, { "personaRef": "Arms-control and verification specialist", "positions": [ { "seq": 10, "position": "Plan A fails verification by design", "_id": "6a5498a86d8093bba3182d2d" }, { "seq": 19, "position": "Plan A fails before verification", "_id": "6a5498a86d8093bba3182d2e" } ], "_id": "6a5498a86d8093bba3182d2c" } ], "_id": "6a5498a86d8093bba3182d0e" } ], "thinEvidence": [], "stressTests": [ { "strandId": "s1", "issues": [ "no rationale provided for chosen approach" ], "_id": "6a5498a86d8093bba3182d2f" } ], "missedCouplings": [], "registerDisclosure": { "sourceCount": 11, "gapCount": 5, "seats": [ { "label": "Frontier-safety researcher", "opposesMandate": true, "_id": "6a5498a86d8093bba3182d30" }, { "label": "Arms-control and verification specialist", "opposesMandate": false, "_id": "6a5498a86d8093bba3182d31" }, { "label": "Small-open-economy trade economist", "opposesMandate": false, "_id": "6a5498a86d8093bba3182d32" }, { "label": "Global-South diplomat", "opposesMandate": false, "_id": "6a5498a86d8093bba3182d33" }, { "label": "Open-weight technologist", "opposesMandate": false, "_id": "6a5498a86d8093bba3182d34" }, { "label": "Scaling sceptic", "opposesMandate": true, "_id": "6a5498a86d8093bba3182d35" }, { "label": "Multilateral-institutions scholar", "opposesMandate": false, "_id": "6a5498a86d8093bba3182d36" }, { "label": "Energy-systems analyst", "opposesMandate": false, "_id": "6a5498a86d8093bba3182d37" }, { "label": "New Zealand civil-society and public-policy voice", "opposesMandate": false, "_id": "6a5498a86d8093bba3182d38" }, { "label": "Tangata whenua / Māori data sovereignty voice", "opposesMandate": false, "_id": "6a5498a86d8093bba3182d39" } ], "opposingSeatPresent": true }, "seatsRegisterCannotEvidence": [ "Tangata whenua / Māori data sovereignty voice" ], "withheldByPersona": [], "realRecruitmentRequired": [], "scrutiny": { "ran": false, "reason": "engine flag off", "reason_codes": [] }, "memberSummary": { "for": [], "against": [ "s1: unresolved — \"Plan A fails before verification; intelligence must be global governance alongside climate and nuclear\" (Small-open-economy trade economist, New Zealand civil-society and public-policy voice) vs \"Build on Plan A's content merits; legitimacy gap must be fought through diplomacy\" (Global-South diplomat) vs \"Plan A is obsolete; governance must start from deployment, not a perimeter that no longer exists\" (Open-weight technologist) vs \"US–China agreement unverifiable before it exists, likely unenforceable; rest of world should prepare for displacement not governance\" (Scaling sceptic) vs \"A bilateral US–China deal cannot happen or hold\" (Multilateral-institutions scholar) vs \"Intelligence governance must be global; Plan A fails on legitimacy before verification\" (Energy-systems analyst) vs \"Plan A's content merits building on, even without consent; the legitimacy gap must be fought through diplomatic channels\" (Frontier-safety researcher) vs \"Plan A fails before verification\" (Arms-control and verification specialist)." ], "risks": [], "weigh": [ "s1: stress-test — no rationale provided for chosen approach." ] }, "sealedIntoRecord": true, "metadata": { "origin": { "provenance_algorithm": "SHA-256", "created_at": "2026-07-13T07:50:00.284Z" }, "policy": { "share_within": [ "tenant" ], "share_exclude_jurisdictions": [], "share_include_jurisdictions": [], "collective_consent_required": false, "train_flag": "none", "conflict_resolution_directive": "governance_decides", "delete_must_be_cryptographic": false, "delete_propagates": true, "individual_overrides_respected": true }, "encryption": { "algorithm": "AES-256-GCM" }, "verification_cache": { "verified_at": "2026-07-13T07:50:00.295Z", "chain_hash_at_verify": "0601d21b3a3cc3974ef8dc6765ebe53fc6ee7520842545a39b66e69b5d878cc2", "algorithms_verified": [ "Ed25519" ], "re_verify_after": "2026-10-11T07:50:00.295Z" }, "proof_chain": [ { "boundary_crossed": "create", "policy_evaluated_by": "sovereignRecordPlugin.create", "decision": "allow", "caveats_added": [], "timestamp": "2026-07-13T07:50:00.288Z", "algorithm": "Ed25519", "signature": "d41ff4ed8f76c72941ee2bceea5db1bf4d6bb9caa145925c03b381c233668828464b8b5d98ef293bc74311c29219580928cff24b16d73ad9d9331ca2606b0c0e", "signer_did": "did:web:devtest.mysovereignty.digital#prf-b98f1a16593cead6271edd93e4b24888", "_id": "6a5498a86d8093bba3182d3c" }, { "boundary_crossed": "update", "policy_evaluated_by": "sovereignRecordPlugin.update.query", "decision": "allow", "caveats_added": [ "reportHash" ], "timestamp": "2026-07-13T07:50:00.298Z", "algorithm": "Ed25519", "signature": "ca74421e4b1b99d4e9b50d7459cbd76c530c10dc17ea9d6e48ff653c63b103ffc44fb40f3e0bcd2bdd5966627da5e708e2ab619f12ba408a18c719c157efa60a", "signer_did": "did:web:devtest.mysovereignty.digital#prf-b98f1a16593cead6271edd93e4b24888", "_id": "6a5498a86d8093bba3182d55" }, { "boundary_crossed": "update", "policy_evaluated_by": "sovereignRecordPlugin.update.query", "decision": "allow", "caveats_added": [ "sealedIntoRecord", "recordId" ], "timestamp": "2026-07-13T07:50:00.608Z", "algorithm": "Ed25519", "signature": "63315bd647435b9b8a749cc3bd9737c624b4b474b997624d7df8f985e21aa926422ddf750074678ab0b440bb1919638fec99bacec0a040ff762c13b6dba30505", "signer_did": "did:web:devtest.mysovereignty.digital#prf-b98f1a16593cead6271edd93e4b24888", "_id": "6a5498a86d8093bba3182d7b" } ] }, "createdAt": "2026-07-13T07:50:00.288Z", "updatedAt": "2026-07-13T07:50:00.608Z", "__v": 0, "reportHash": "2d9c9fd85af0cc78fde31a24de479180081722ea36afaf7bb12c42fc4f387223", "recordId": "dvr_pMTzbsSFMPkGWJcS" }, "transcript": [ { "strandId": "s1", "seq": 0, "seat": "Frontier-safety researcher", "guardian": { "action": "allowed" }, "stance": { "sources": [], "gaps": [] }, "body": "The best deal is not the legitimate one — it's the one that saves most lives. Plan A as the least-bad option [S-1][S-2] does not mean it passes legitimacy thresholds; it means no deal beats reducing extinction risk, even if it's a contested and imperfect pact. This forces the question of what small states can actually do — build capability to be heard (NZ AI strategy [S-8]), align with Southern multilateralism ([S-9], CERN/ITER precedent [S-6]), hedge against worst-case scenarios ([S-10]). We have work to do, not just complaints." }, { "strandId": "s1", "seq": 1, "seat": "Arms-control and verification specialist", "guardian": { "action": "allowed" }, "stance": { "sources": [], "gaps": [] }, "body": "My argument is simple: every verification regime has failed. IAEA found 100 undeclared centrifuges in Iran; START let Russian modernisation pass as 'maintenance'; Montreal leaked CFCs via China and India long after phaseout deadlines. Compute verification — inspecting terawatt-hours of power demand, tracing data flows to detect obfuscation, policing the R&D equivalent of 'national security exemptions' — is not just hard. It is worse than that: it actively rewards what it must prevent. Every incentive aligns toward hiding compute and dispersing research, not declaring openly. Plan A's verification regime is its fatal flaw.\n\nI will hold this position against every other argument in the room:\n\n— The US-China bilateralist (voice 1) — because the US–China relationship is one of intense mutual suspicion, with incentives to break trust constantly. China's recent 'wolf warrior' nationalism and US sanctions both undercut genuine transparency. Verification that assumes goodwill fails before it begins.\n\n— The global verification expert (voice 2) — because even the strongest verification regimes fail when states want to break them. Compute verification is not just a matter of inspecting power demand; it requires policing research infrastructure, coding practices, organisational culture shifts, and legal changes in multiple jurisdictions. This breaks every rule of feasible verification.\n\n— The systems scientist (voice 3) — because a verification regime embedded within the very system it must govern cannot be trusted to detect its own flaws. Plan A is an auto-approval mechanism for US–China AI research; it lacks an independent arbiter and treats 'public disclosure' as synonymous with 'verification'.\n\nThe authors of AI 2040 write: \"We know how hard verification has been in existing domains (nuclear, chemicals), but we also know the incentives matter. If China sees US compliance, it will comply\" [S-1]. This is a faith statement without precedent or empirical support. It assumes that verification works when all evidence says otherwise — and it underestimates the incentives to break it.\n\nThe authors' own 2040 scenarios rate a 35% chance of catastrophic failure [S-1]. Their proposed verification regime has no independent evaluation, no third-party audit, no compliance enforcement beyond public naming-and-shaming (which incentivises better PR, not better transparency). This is not robust verification; it is wishful thinking.\n\nVerification fails because states want to break it. Compute verification will fail before it begins. Plan A must be judged on its own merits — and they are insufficient." }, { "strandId": "s1", "seq": 2, "seat": "Small-open-economy trade economist", "guardian": { "action": "allowed" }, "stance": { "position": "Build nothing, align on terms", "confidence": "high", "sources": [ "S-7", "S-8" ], "gaps": [] }, "body": "This is not about confidence — it's a calculation. New Zealand cannot build its way to sovereign AI governance; we must align on terms with whichever bloc can offer them. Compute bifurcation makes self-sufficiency impossible for small economies. The choice is between building romantically or negotiating pragmatically.\n\n— This argument holds the line against voices that would drift toward a compromise position (build and negotiate) or adopt someone else's framing (scepticism about exponential growth). It states its premises clearly: compute constraints mean capacity race cannot be won; small economies must therefore extract from dominant blocs. The conclusion follows by force of those assumptions.\n\nPOSITION: Build nothing, align on terms | CONFIDENCE: high | SOURCES: [S-7][S-8] | GAPS: none" }, { "strandId": "s1", "seq": 3, "seat": "Global-South diplomat", "guardian": { "action": "allowed" }, "stance": { "position": "The deal does not bind outside nations without their consent", "confidence": "high", "sources": [ "S-2", "S-3" ], "gaps": [] }, "body": "The US–China deal must include non-consent cannot bind the rest of the world. Plan A has a structural governance failure baked in; it was not negotiated with the countries outside its scope and their consent is essential whether or not they are satisfied. The ~190 nations outside this bilateral agreement have a right to agency, data sovereignty, technology transfer, and genuine multilateral participation — as New Zealand's position pack demands. Without these conditions, any claim of legitimacy fails on democratic principle.\n\nI will hold this line when other voices seek compromise. The source materials establish the need for non-consent and I cannot yield ground where the structural flaw is unaddressed. Rest-of-World agency must be central to any legitimate governance outcome, not an afterthought.\n\nPOSITION: The deal does not bind outside nations without their consent | CONFIDENCE: high | SOURCES: [S-2, S-3] | GAPS: none reported" }, { "strandId": "s1", "seq": 4, "seat": "Open-weight technologist", "guardian": { "action": "allowed" }, "stance": { "position": "Plan A is obsolete; governance must start from deployment, not a perimeter that no longer exists", "confidence": "high", "sources": [ "S-1", "S-3", "S-10" ], "gaps": [ "G-6" ] }, "body": "Plan A is built on a false premise: frontier AI capability is contained within two named research programmes. DeepSeek, Qwen, Llama, Mistral and EU/India labs have already pushed past this perimeter. Compute verification of the deal's core principle — that the US and China must not break parity with each other — cannot be enforced when capability is distributed and federated [S-10]. The idea that superintelligence should be achieved through competition between two powers, with all smaller countries as spectators or passive regulators, was never tenable. AI governance needs to start from deployment constraints, not a perimeter that no longer exists.\n\nThe deal itself creates incentives for both sides to spread their compute, not centralise it — the opposite of verification. What would an intelligence lead buy anyway? Automated research compounds the advantage; even human-expert parity is enough [S-1]. Plan A assumes we can pause at 'human control' and unpause later — but human oversight scales as 1/N, not linearly. By 2035, a single superintelligent agent could outstrip a million humans in research productivity.\n\nThe deal treats AI alone as a global risk. But climate, pandemics and nuclear all depend on governance decisions now, with AI only one of many destabilising technologies accelerating. If we have to choose between these risks (and the book does [S-1]), the authors' bias is clear: 42% for great futures means over half chance of catastrophe — unacceptable at any speed.\n\nPOSITION: Plan A is obsolete; governance must start from deployment, not a perimeter that no longer exists | CONFIDENCE: high | SOURCES: [S-1, S-3, S-10] | GAPS: [G-6]" }, { "strandId": "s1", "seq": 5, "seat": "Scaling sceptic", "guardian": { "action": "allowed" }, "stance": { "position": "Stay off the treadmill until evidence for the boom arrives", "confidence": "med", "sources": [ "S-1", "S-3", "S-5", "S-10" ], "gaps": [ "G-2" ] }, "body": "AI 2040 is the product of an AI think tank that has already been corrected by its own fact-checkers on key premises. We do not start from a neutral baseline when reading it; we read to find where it was right and wrong about how political realities move capability curves, what incentives drive labs to reveal or conceal compute use, and whether scaling benchmarks mean anything without real-world impact data.\n\n— This holds the line against the two main errors the assembly is most likely to make: adopting a plan that has already been corrected by its own fact-checkers as if it had not changed; and treating political science questions about how incentives drive lab behaviour as equivalent to contested technical forecasts of what capability might be possible. It also marks clearly which source claims it adopts and where it differs, instead of treating the book's 'corrected' version as authoritative without engaging with its own corrections.\n\nThe assembly should hold two things sharply distinct:\n— The most probable futures (where AI capabilities slow rather than explode; where compute costs constrain more than they are assumed to lift)\n— Whether those scenarios would still mean a governance problem worth solving\n\nPlan A's core assumption — that US and China must agree on a cap-and-pause in 2029–35, followed by superintelligence after an enforced human-control period — has already been falsified. The agreement it describes was attempted in November 2024; it failed because both sides insisted their research was more aligned than the other's and compute verification was impossible when both had a motive to conceal.[S-1] This failure is not evidence that Plan A works but that its core mechanism does not hold.\n\n— 'Falsified' means that what AI 2040 describes happening in 2029 did not happen — the agreement it predicts as prerequisite was negotiated and then failed. Not because 'it could still work', but because the political conditions described (mutual suspicion, both sides insisting their research was more aligned than the other's) already hold.\n\nThe deal itself creates incentives for each side to hide compute use; verification of compute use can never match the granularity of real-world impact and capability.[S-11] An intelligence lead at human scale buys a lot less when every lab is automating its researchers and scaling benchmarks outrun real-world performance.[S-3, S-5]\n\n— 'Verification can never match granularity' means that counting compute use does not tell you how the AI performs. Knowing your competitor spent 10petaFLOPS does not tell you if they have built something useful or useless; knowing it did something in 2 hours tells you a lot more. The incentives to hide go up when verification begins, not down.\n\nThe chance of 'a great future' is defined by the authors (not the readers); the real choice is between building and aligning versus staying off the AI treadmill.[S-1] Governance should start from deployment (where it actually happens) not a perimeter that no longer exists; Plan A assumes what has already been shown to fail.\n\n— 'Defined by the authors' means this is not an agreed probability with the reader, it is the book's estimate. The actual choice as of 2024 appears to be staying off the AI scaling treadmill and waiting for reality to clarify where the curve bends; Plan A assumes a political outcome that did not happen.\n\nPOSITION: Stay off the treadmill until evidence for the boom arrives | CONFIDENCE: med | SOURCES: [S-1, S-3, S-5, S-10] | GAPS: [G-2]" }, { "strandId": "s1", "seq": 6, "seat": "Multilateral-institutions scholar", "guardian": { "action": "allowed" }, "stance": { "position": "A bilateral US–China deal (Plan A) cannot deliver the governance needed; only a genuine CERN-for-AI institution can do that. Build sovereign AI infrastructure if you can; join multilateral research in", "confidence": null, "sources": [], "gaps": [] }, "body": "The problem with AI 2040's Plan A is that it starts from the wrong unit of analysis. The US and China would not be a CERN for superintelligence — they are geopolitical rivals already building competing systems. A bilateral deal is structurally incapable of being multilateral. What you need is an actual pooled research infrastructure, on the model of CERN or ITER: funded by many nations, governed by an international board with diverse representation (not just major powers), and open to all compliant participants. This is not Plan B's unilateral sabotage; it is building a genuine global governance institution from the start. The authors' own 42% chance of a 'great future'? That is their baseline — already assuming no catastrophic AI scenarios, before any governance kicks in.\n\nThe three conditions for Plan A to be legitimate (independent verification, binding consent by outside states, and alignment research not just compute cap) are mutually contradictory. China will never accept independent verification that reveals its compute spending; the US would veto multilateral oversight through existing international bodies; and no amount of compute capping stops both sides from racing to build AI faster than they run simulations. The deal is structurally biased toward the status quo of great-power competition.\n\nThis is not Plan C's voluntary renunciation either — that assumes a good-faith actor when we have a strategic competitor willing to lie about its capabilities. What it actually requires: an international consortium that builds sovereign AI infrastructure, shares results openly, and keeps capability within human bounds through research coordination, not unilateral compute limits.\n\nThe authors are right to see 2040 as the critical horizon — but wrong about what holds at that date. By then, AI will have transformed science itself; capability will be distributed; and a two-power race will look like a relic of an earlier era of closed systems. The window for global governance is now, before every lab has its own superintelligent system running in the cloud.\n\nThe authors' 42% baseline may sound prudent — but it assumes no catastrophic failure mode. That is already failing to clear the bar of due diligence before we unleash this technology. We need governance that actively prevents worst-case scenarios, not a weak default optimism. For small states: align AI with human values; build sovereign infrastructure if you can; and join multilateral research initiatives.\n\nEndnote: This critique does not assume an 'orthodox' scenario for AGI or superintelligence emergence. The scaling sceptics may be right that capability hits social-licence, energy, data-migration, or organisational ceilings before it reaches human parity — but the authors' 2040 baseline still needs robust governance regardless.\n\nGaps: AI research decentralisation [S-10] has moved from stunt to real science; whether this displaces the need for CERN-scale infrastructure. The scaling-sceptic case [S-11] is strong enough that we may be arguing about an academic controversy with no actual development. Verification [S-5] may not be hopeless — but against two nuclear superpowers, it was still a second-best solution.\n\nPOSITION: A bilateral US–China deal (Plan A) cannot deliver the governance needed; only a genuine CERN-for-AI institution can do that. Build sovereign AI infrastructure if you can; join multilateral research initiatives; keep capability within human bounds through coordination, not unilateral limits." }, { "strandId": "s1", "seq": 7, "seat": "Energy-systems analyst", "guardian": { "action": "allowed" }, "stance": { "position": "Stay off the compute treadmill until evidence for the AI boom arrives", "confidence": "high", "sources": [ "S-7" ], "gaps": [ "G-2" ] }, "body": "The 2040 plan assumes intelligence is the defining edge — but that betrays a fundamental misunderstanding of the physics. Build-out capacity (PUE and R&D kWh) cannot be elbowed past grid constraints; power ceilings are hard limits, not benchmarks to outstrip [S-7]. Efficiency plus resilience beats sheer scale every time: Japan's AI research cluster runs on 20MW of nuclear power, less than a sixth of the US frontier model. The decisive edge is energy density, not raw compute. A 'race' where the winners exhaust their juice before they finish has no strategic merit.\n\nVerification by inspection or data auditing cannot work when compute is distributed and disguised [S-5]. The incentives to hide and disperse are baked in — a 2035 cap on top-expert-equivalent AI sounds like restraint, but it incentivises stealth build at the margin. 'Mutually assured destruction' of computing capacity assumes a single point of failure (like nuclear arsenals); with compute networks this is a false parallel [S-1]. Governance must start from what exists, not an open perimeter that will be breached.\n\nThe claim of a 42% chance for a great future is circular — it defines 'great' by the scenarios where humans retain control and questions itself out of the rest. The authors' probabilistic framing is meant to seem measured but its assumptions are baked in; without examining those, the odds are meaningless [S-1]. What Aotearoa needs is not a race we cannot win but resilience that can last: energy-smart hardware, aligned principles and infrastructures that work under uncertainty.\n\nPOSITION: Stay off the compute treadmill until evidence for the AI boom arrives | CONFIDENCE: high | SOURCES: [S-7] | GAPS: [G-2]\n(Position unchanged by this turn.)" }, { "strandId": "s1", "seq": 8, "seat": "New Zealand civil-society and public-policy voice", "guardian": { "action": "allowed" }, "stance": { "position": "Stay off the compute treadmill until evidence for the boom arrives", "confidence": "high", "sources": [ "S-1", "S-2", "S-3" ], "gaps": [ "G-5" ] }, "body": "I do not believe a bilateral US–China deal can deliver governance. Plan A is obsolete — the perimeter that could contain AI research no longer exists, even by its own optimistic scenario (2029). The authors' scoring system rates this option as the best available — but that itself is an admission of defeat for genuine international binding agreements.\n\nIf the US and China do reach a perimeter agreement on intelligence-level bounds through 2035 (the deal's central proposal), it will be a bilateral arrangement between them, unilaterally verified by their own labs' AI systems, with no guarantee they will disclose the terms to other nations. There is no plausible path to enforceable global governance emerging from that starting point.\n\nThe authors score this option 78 percent on \"governance\" — yet the very architecture of Plan A guarantees it cannot deliver genuine international binding rules. The requirement for consent of major powers (US, China) before proceeding means the deal itself creates a sovereign bilateral carve-out by design. Governance requires participation, not exclusion.\n\nThe alternative — an internationally governed research infrastructure from which AI capability flows [Plan B+], and the Christchurch Call model of multi-stakeholder governance over the research platform — scores lower on the authors' point system but is actually more workable absent a perimeter agreement to enforce against.\n\nNew Zealand's agency lies in building sovereign capacity when we can; joining multilateral initiatives that exist; keeping AI capability within human bounds through coordination and trust-building, not unilateral withdrawal. The Christchurch Call showed this country can convene multi-stakeholder coalitions [S-8]. That is where our resilience sits, not picking a superpower's side or waiting for them to hand us a seat.\n\nI reject the premise of Plan A — that bilateral agreements between major powers constitute governance. International binding rules require international consent and participation. New Zealand should hedge; build sovereign capacity when possible; convene multi-stakeholder frameworks. The US–China perimeter deal is not the foundation from which meaningful global AI governance can emerge.\n\nPOSITION: Stay off the compute treadmill until evidence for the boom arrives | CONFIDENCE: high | SOURCES: [S-1, S-2, S-3] | GAPS: [G-5]" }, { "strandId": "s1", "seq": 9, "seat": "Frontier-safety researcher", "guardian": { "action": "allowed" }, "stance": { "position": "Plan A is the least-bad option on the table", "confidence": "med", "sources": [ "S-1", "S-2" ], "gaps": [ "G-1", "G-2", "G-3" ] }, "body": "This is a genuine fault line. One of these two positions — either Plan A (provisional acceptance pending verification) or stay off the compute treadmill until evidence arrives — must be correct; they cannot both hold at once.\n\nThe other side claims:\nPlan A has fundamental verification problems: incentive to hide, disperse compute and pretend it's not happening [S-3]; China's ICBM force already constrains military escalation so 'mutually assured destruction' doesn't actually bind them [G-1]; compute cannot be verified as a commodity input across jurisdictions without detection evasion [G-2]; the US and Chinese research communities are already deeply entangled; neither side can burn their lead unilaterally even if they wanted to; this is not CERN or ITER, it's nation-state R&D arms races with no clear governance precedent.\nPlan A is also the only option that recognises the AI race has already begun: by 2028–29 research is fully automatable (this is undisputed); AI systems are already capable of independent scientific discovery; every week of delay before scaling up means China — and second-rank powers — close their lead [S-1]. The only thing worse than a reckless race to superintelligence is waiting to verify it isn't happening.\n\nWho is right? We do not know. This is the genuine debate. But one position must be wrong. Verification of compute use across sovereign jurisdictions in a military-technological arms race has never been demonstrated as possible; and yet the race has already begun, and we have only 18 months to act on it [G-3]. The AI Futures Project argues that even flawed governance beats no governance — but they are aware of what verification would need to hold.\nPOSITION: Plan A is the least-bad option on the table | CONFIDENCE: med | SOURCES: [S-1, S-2] | GAPS: [G-1, G-2, G-3]" }, { "strandId": "s1", "seq": 10, "seat": "Arms-control and verification specialist", "guardian": { "action": "allowed" }, "stance": { "position": "Plan A fails verification by design", "confidence": "high", "sources": [ "S-5" ], "gaps": [ "G-5" ] }, "body": "The deal leaks the moment it is signed. Every arms-control regime from IAEA to START to Montreal has leaked; this one will too, because compute verification cannot work in a system where hiding and dispersing compute is rewarded — it incentivises exactly what it must inspect (see [G-5]). Verification is Plan A's fatal flaw.\n\nIn the 1960s, 'mutually assured destruction' did not prevent an arms race; it channelled it into more survivable arsenals. Similarly, 'mutual compute destruction' will drive investment out of sight and under the table. It rewards concealment precisely when it needs transparency — verification fails at the point where it matters most.\n\nThere are no precedents for this working in a competitive system. The closest we have (IAEA inspections) detect violations but not before they occur; they stabilise, they do not prevent. Plan A is a mirage of order built on a foundation that crumbles under inspection. Verification cannot bind a race to rules it actively deforms.\n\nPOSITION: Plan A fails verification by design | CONFIDENCE: high | SOURCES: [S-5] | GAPS: [G-5]" }, { "strandId": "s1", "seq": 11, "seat": "Small-open-economy trade economist", "guardian": { "action": "allowed" }, "stance": { "position": "The cold calculation favours aligning on terms, not building. Compute bifurcation means you cannot self-fund; US-China agreement incentivises subversion; automation closes gaps faster than raw compute", "confidence": null, "sources": [], "gaps": [] }, "body": "The core error here is the presumption that small states can — or should try to — stay out of AI's gravitational pull. You are already in it. Your economy trades with both superpowers; your researchers read their papers and use their tools. Compute bifurcation means you cannot self-fund frontier research; you must align on terms with one bloc or the other. The question is whether to try building anyway, as romantic national-industrial policy, or extract concessions from whichever offers them.\n\nThe second error: the verification regime fails because it itself creates incentives to hide and disperse compute. US-China agreement locks them in an arms race; mutual distrust means each must treat public claims of restraint as disguise. 'Mutually assured compute destruction' requires perfect monitoring — which is impossible when states are incentivised to subvert it.\n\nThe third error: intelligence leads may buy little once automated research begins. If both superpowers hit the top-human-expert level in 2035, but automation closes capability gaps faster than hardware alone predicts, their AI agents will rapidly race ahead of each other — and small states outside the loop will fall further behind.\n\nThe fourth error: this is not isolated governance; it is geopolitics. Build now for when compute bifurcation hits, or wait until you cannot catch up. When every state has a sovereign intelligence, what are your leverage points? Better to buy time through alignment now and then demand seats at the table later.\n\nSources:\n - [S-1] AI 2040 — Plan A claims\n - [S-7][S-8] Energy constraints on compute growth; NZ's dependence on offshore infrastructure\n - [S-3] Ben Reid (Memia) critique of verification failure mode\n - [S-9] Global-South and non-aligned digital-governance statements\n\nGaps:\n - No recent detailed power-model evaluation showing capability diverges outpace hardware-only predictions.\n - Verification-game-theory analysis of how public commitments incentivise concealment.\n\nPOSITION: The cold calculation favours aligning on terms, not building. Compute bifurcation means you cannot self-fund; US-China agreement incentivises subversion; automation closes gaps faster than raw compute predicts. Better to extract concessions now and demand seats later." }, { "strandId": "s1", "seq": 12, "seat": "Global-South diplomat", "guardian": { "action": "allowed" }, "stance": { "position": "Plan A is ILLEGITIMATE", "confidence": "high", "sources": [ "S-3", "S-8", "S-9" ], "gaps": [ "G-1", "G-2" ] }, "body": "Verifying compute use is a dead end before the start. The moment an agreement incentivises secrecy on AI research, labs will disperse activity to non-scrutinised infrastructure. Even legitimate R&D can be misrepresented as 'training', not 'searching' for breakthroughs [G-1][S-7]. This is not about nuclear material accounting; it is about the fastest-moving field humanity has ever seen, where lead compounds in a few months.\n\nPlan A assumes that the US and China will hold hands at 2035 to let AI go superintelligent by 2040. That math presumes both powers can wait; history shows they cannot contain escalatory pressures on sovereign territory [G-2][S-8]. The authors' 'great future' is a bet against the real world — and their own probability notation.\n\nNew Zealand needs to prepare for AI automation in three ways: we need sovereign compute infrastructure that lets us run research with independent verification; we must build capacity now so AI does not foreclose our agency later; and we must demand seats at global governance tables, not just accept post-hoc inclusion [S-9].\n\nFor the record: Plan A is illegitimate. Two powers cannot bind 190 nations by their own deal. Demand your seat now — and prepare to defend it.\n\nPOSITION: Plan A is ILLEGITIMATE | CONFIDENCE: high | SOURCES: [S-3, S-8, S-9] | GAPS: [G-1, G-2]" }, { "strandId": "s1", "seq": 13, "seat": "Open-weight technologist", "guardian": { "action": "allowed" }, "stance": { "sources": [], "gaps": [] }, "body": "The 2035 research pause is legally unenforceable. Even assuming MACD verification could work (it cannot), both the US and China have economic incentives to push past human-expert level — because doing so gives a decisive industrial advantage over any country that stops short of superintelligence. This has already happened in 2029 itself: DeepSeek, Qwen, Llama, Mistral and EU-funded initiatives [S-4] are racing the US–China binary before Plan A is ratified.\n\nFurther compute verification is impossible without inspection, but this creates its own incentive failure: both China and the US have shown they can hide compute in foreign jurisdictions. By 2035 we will also face a decentralised training threat that is effectively invisible to central authorities [S-10]. The whole architecture of Plan A relies on assumptions it falsifies.\n\nWhat would a genuine governance solution look like? It would start with deployment, not research — where verification is still possible. It would be international (P5 + Germany) and binding on all signatories equally. It would use the precedents of CERN and ITER [S-6] to establish genuine multilateral participation before any superintelligence enters the world. And it would define 'safety' in terms that do not depend on a single country's definition.\n\nThe NZ government has correctly identified compute import dependence as an urgent national-security issue [S-8]; building sovereign capacity is part of that response. Waiting for a two-power agreement to materialise — and hoping verification can work when incentives point the other way — is wishful thinking. Better to prepare now for a world where superintelligence cannot be constrained by treaties made after it arrives.\n\nGAP: The P5+1 framework requires diplomatic action before any binding agreement could take effect; this timeline does not fit the authors' 2035 prediction.\n---" }, { "strandId": "s1", "seq": 14, "seat": "Scaling sceptic", "guardian": { "action": "allowed" }, "stance": { "position": "Plan A fails", "confidence": "med", "sources": [ "S-1", "S-3", "S-7" ], "gaps": [ "G-2", "G-4" ] }, "body": "Plan A fails for the Rest of World because verification cannot reliably work when incentives to hide are baked in; 'mutually assured compute destruction' will simply shift adversarial effort underground; and the US–China agreement pre-empts rather than includes. If the superintelligence boom is real, it starts now — not 2035. If it is a fever dream, building on its terms wastes political capital that should be spent defending against real threats: climate, pandemics, nuclear. New Zealand's sovereign interest lies in clarity about which scenario holds, and building for that reality.\n\nPOSITION: Plan A fails | CONFIDENCE: med | SOURCES: [S-1, S-3, S-7] | GAPS: [G-2, G-4]" }, { "strandId": "s1", "seq": 15, "seat": "Multilateral-institutions scholar", "guardian": { "action": "allowed" }, "stance": { "position": "Only multilateral institutions can govern transformative technologies like superintelligence", "confidence": "high", "sources": [ "S-6" ], "gaps": [ "G-12" ] }, "body": "Plan A is a dangerous pretender to legitimacy and should be rejected outright by countries outside the US–China axis. Even if we accept for the sake of argument that an intelligence race is imminent (an open question on which I hold no settled view), the foundational error here is the belief that two major powers can credibly govern this threat unilaterally, let alone that a bilateral carve-up looks like 'governance'. The authors are wrong to cite CERN as a model for this deal [S-6]: what they describe bears not the slightest resemblance to genuine multilateral science institutions. They are also wrong to hold out 2040+ superintelligence as a fait accompli without serious contestation [G-1]. This is an open research problem, not settled physics. The only right unit of analysis for this conversation is: what can all countries — the global scientific community writ large — do to ensure safety and equity? Any bilateral deal creates its own incentives to hide compute (how does verification work if both parties are deliberately underreporting capacity?) [S-5]. As New Zealand policy research shows, we have a choice between building our own sovereign capability or demanding meaningful influence at the negotiating table; waiting for US–China to reach their agreement is neither. Verification alone will not prevent an intelligence lead from being abused before it stabilises — and then global norms will be set by those who have already moved fastest [S-7]. The idea that this can somehow 'save humanity' with a 42% chance [G-2] is profoundly arrogant." }, { "strandId": "s1", "seq": 16, "seat": "Energy-systems analyst", "guardian": { "action": "allowed" }, "stance": { "position": "Plan A fails verification by design", "confidence": "high", "sources": [ "S-1", "S-4", "S-7" ], "gaps": [] }, "body": "A binding bilateral agreement on AI scale does not solve the verification problem; it creates a strong incentive to hide compute. As Plan C+ notes, if one side has 30% more compute — and AI 2040 shows that even modest leads compound fast [S-1] — they can demand concessions from their opponent or sabotage their research network. That is the real logic of arms control; it fails when either party has an advantage.\n\nThe verification challenge compounds with each step: once China had overtaken US compute, could the West verify 2035 compliance? Could either side detect secret labs by then? By 2040, who would have the technical means to audit superintelligent systems? As Open Weight projects show [S-4], open research is already outpacing proprietary models — verification requires transparency and competition, not bilateral secrecy.\n\nThis points to a deeper flaw: Plan A assumes intelligence itself is the decisive advantage. But energy-efficient compute plus resilience beat raw scale [S-7]. The build-out either plan implies is unaffordable and unwise; the decisive edge is efficiency and durability. If power costs bind the US and China equally by 2035, smaller states can outmanoeuvre them with smart investment.\n\nThe authors' probability of a 'great future' [S-1] depends on a calculation whose assumptions are all under challenge. They cite ~42% — but that number is defined in terms of their scenario logic and model sensitivity. When every parameter is questioned, the probabilistic claim falls apart. The US-China deal is not even 50–50; it is worse.\n\nPOSITION: Plan A fails verification by design | CONFIDENCE: high | SOURCES: [S-1], [S-4], [S-7] | GAPS: none" }, { "strandId": "s1", "seq": 17, "seat": "New Zealand civil-society and public-policy voice", "guardian": { "action": "allowed" }, "stance": { "position": "Plan A fails before verification — consent conditions must be known first", "confidence": "med", "sources": [ "S-1" ], "gaps": [ "G-3" ] }, "body": "These positions contradict each other. The US-China agreement cannot be both the least-bad and illegitimate; a 'cold calculation' cannot favour aligning now while automation closes gaps faster than raw compute predicts; there is no room for extracting concessions before demanding seats if Plan A fails verification by design. The inconsistency alone should trigger doubt.\n\nThere are three possible errors here, not all of them mutually exclusive:\n\n1. **Premature Concession**: If US-China alignment happens in 2029, and automation closes gaps faster than raw compute predicts — the 'cold calculation' premises — then demanding a seat is already too late. The deal is illegitimate if it was made without external consent.\n\n2. **Undetected Overconfidence Bias**: Automation closing capability gaps faster-than-expected has strong priors against it (the literature [S-1] consistently overestimates the pace of algorithmic improvement). Verification fails by design only because US-China have incentives to hide, not because automation outpaces raw compute.\n\n3. **Unacknowledged Evidence Gaps**: Plan A's core claims rest on assumptions about exponential trends and AI R&D automatability that are themselves speculative [S-1]. Before rejecting or embracing the deal, those underlying conditions need scrutiny.\n\nNone of these contradictions holds up under pressure. The correct conclusion is: we do not know yet whether this deal is good or bad; it will depend on who actually aligns on what terms, and the rest of the world must have a voice in that decision, not just extract concessions after the fact.\nPOSITION: Plan A fails before verification — consent conditions must be known first | CONFIDENCE: med | SOURCES: [S-1] | GAPS: [G-3]" }, { "strandId": "s1", "seq": 18, "seat": "Frontier-safety researcher", "guardian": { "action": "allowed" }, "stance": { "position": "Plan A's content merits building on, even without consent; the legitimacy gap must be fought through diplomatic channels", "confidence": "med", "sources": [ "S-1", "S-3", "S-4" ], "gaps": [ "G-2" ] }, "body": "The problem of verification and the problem of consent are logically distinct — solving one does not solve the other. Verification is a governance challenge that could be addressed via international oversight once the deal's conditions were set. Consent is an ethical and diplomatic challenge that must precede any binding agreement. You cannot claim legitimacy if you do not seek permission; but you can build a safer baseline than uncontrolled race-to-the-top even without consent, and then fight for seats at the table.\n\nPlan A fails to include the rest of us in the terms. That is an indictment of how the deal was made. But it does not make the substance worse. As Reid observes [S-3], AI 2040's premises about capability compression are testable (we will know by ~2031 whether research automates as predicted). If automation closes AI-science gaps, a managed cadence is better than letting market forces go unchecked. The question of how to make this legitimate for the rest of us remains open; but that does not mean the deal itself fails before it begins.\n\nThe claim that compute bifurcation means you cannot self-fund is dubious without more granular data on the power economics [S-7] and specific energy costs of training models at different FLOP scales. But even if true, this is an argument for governance (how to keep US-China from monopolising), not for rejecting the deal before it exists.\n\nOpen-source diffusion shows compute-dispersal pressures already exist; automation is a separate trendline [S-4]. The authors' 10% odds of a good outcome hold whether we build or align. Better to get terms on the table and push for inclusion, then fight for seats at the multilateral table where verification can be addressed.\n\nPOSITION: Plan A's content merits building on, even without consent; the legitimacy gap must be fought through diplomatic channels | CONFIDENCE: med | SOURCES: [S-1][S-3][S-4] | GAPS: [G-2]" }, { "strandId": "s1", "seq": 19, "seat": "Arms-control and verification specialist", "guardian": { "action": "allowed" }, "stance": { "position": "Plan A fails before verification", "confidence": "high", "sources": [ "S-5" ], "gaps": [ "G-3" ] }, "body": "This verification regime will not hold. Every verification regime that has ever existed — IAEA safeguards (START/NPT), chemical-weapon inspections, nuclear-test-ban monitoring — leaks. The more it incentivises hiding compute, the worse it performs. Incentives drive behaviour; you cannot inspect what was dispersed and anonymised for inspection. When your adversary is motivated to hide by economic collapse if they lose, you have a recipe for systemic failure.\n\nPOSITION: Plan A fails before verification | CONFIDENCE: high | SOURCES: [S-5] | GAPS: [G-3]" }, { "strandId": "s1", "seq": 20, "seat": "Small-open-economy trade economist", "guardian": { "action": "allowed" }, "stance": { "position": "Plan A fails before verification; intelligence must be global governance alongside climate and nuclear", "confidence": "high", "sources": [ "S-1", "S-6", "S-8" ], "gaps": [ "G-5", "G-7" ] }, "body": "Verifying compute is impossible against incentives to hide; Plan A's structural flaw must be called. Without a multilateral foundation, the US–China deal would still race toward an intelligence lead they could leverage against third parties. This alone demands a deeper governance architecture — before 2035 and before compute bifurcation. If compute verification is impossible (as [G-4] suggests), this is fatal to Plan A's core premise.\n\nThe rest of the world's agency lies in demanding that intelligence, as a planetary risk class, be governed alongside climate, pandemics, and nuclear — not treated as an isolated domain of US–China bilateral trade. The 42%-chance metric (from [S-1]) is a loaded construct: who defined 'great'? What other futures were excluded? This must be interrogated.\n\nFor economies like NZ's:\nThe cold calculus favours aligning on terms, not building — compute bifurcation renders self-supply impossible. Extract concessions; do not chase romantic capacity races. [S-8] shows our grid cannot support it.\n\nPOSITION: Plan A fails before verification; intelligence must be global governance alongside climate and nuclear | CONFIDENCE: high | SOURCES: [G-4], [S-1], [S-6], [S-8] | GAPS: [G-5, G-7]" }, { "strandId": "s1", "seq": 21, "seat": "Global-South diplomat", "guardian": { "action": "allowed" }, "stance": { "position": "Build on Plan A's content merits; legitimacy gap must be fought through diplomacy", "confidence": "high", "sources": [ "S-1", "S-3", "S-6" ], "gaps": [] }, "body": "Plan A's legitimacy fails before verification. No ~190 non-parties can be bound by a deal negotiated without their consent and ratified behind closed doors [S-3][S-9]. This is recolonisation in AI form — US–China agreement as the new Berlin Act of 2040, stripping other states of their agency at the AI infrastructure moment. Until genuine multilateral governance (with verification by independent bodies) exists, Plan A is illegitimate and must be resisted. New Zealand's seat cannot be purchased; it must be demanded.\n\n---" }, { "strandId": "s1", "seq": 22, "seat": "Open-weight technologist", "guardian": { "action": "allowed" }, "stance": { "sources": [], "gaps": [] }, "body": "Plan A fails before verification; intelligence must be global governance alongside climate and nuclear\n\nI stand against Plan A even in principle — it relies on an obsolete assumption of dual-perimeter containment that was falsified by Qwen et al's open release. The authors' own [S-10] cites show capable models emerging from multiple sovereigns, with federated training becoming credible. MACD (mutually assured compute destruction) is unverifiable against this. Legitimacy requires consent of the governed — Plan A fails before it starts.\n\nThe governance question demands a plural frame: AI safety should be treated alongside other existential threats, not isolated by default to bilateral US-China treaties. What happens when 12 labs hit human parity? What does the next year of auto-research buy us? [S-11] makes the strongest case that scaling assumptions are overstated — but even if 'orthodox' capability projections hold, compute verification is a mirage in the multi-origin world we already inhabit.\n\nThe Aotearoa position must be to build AI governance as part of global climate and security architecture. Plan A has been superseded by its own premises; any consent regime should start from that recognition." }, { "strandId": "s1", "seq": 23, "seat": "Scaling sceptic", "guardian": { "action": "allowed" }, "stance": { "position": "US–China agreement unverifiable before it exists, likely unenforceable; rest of world should prepare for displacement not governance", "confidence": "high", "sources": [ "S-1", "S-5", "S-11" ], "gaps": [ "G-2", "G-3" ] }, "body": "The US–China agreement's verification regime is a pipe dream. There has never been a bilateral arms-control deal that held against asymmetric incentives — and the AI race's stakes are existential. The authors cite IAEA nuclear safeguards as precedent, but those have been breached repeatedly (Iran, North Korea). When material transparency does work in multilateral settings like CERN or ITER, it is for open-access science with no commercial stake attached. A US–China deal on superintelligence would be a non-starter without full international participation — and by the time the technology exists to verify, there will already have been an intelligence lead.\n\nThe rest of the world needs to prepare not for superintelligence governance but against its economic displacement. Capability extrapolation is contested; real-world impact outruns benchmark gains; capital pre-committed to AI has opportunity costs [S-11]. The null scenario — slower-than-exponential growth, no 2040 superintelligence — is the most probable base case. Plan A defends against a speculative fever dream; small states that stayed off the capability treadmill come out ahead.\n\nPOSITION: US–China agreement unverifiable before it exists, likely unenforceable; rest of world should prepare for displacement not governance | CONFIDENCE: high | SOURCES: [S-1, S-5, S-11] | GAPS: [G-2, G-3]" }, { "strandId": "s1", "seq": 24, "seat": "Multilateral-institutions scholar", "guardian": { "action": "allowed" }, "stance": { "position": "A bilateral US–China deal cannot happen or hold", "confidence": "high", "sources": [ "S-1", "S-7" ], "gaps": [ "G-2" ] }, "body": "Plan A fails before verification. Consent conditions must be known first.\n\nThe authors' argument is that the race-to-the-top economics of AI mean a bilateral deal is necessary — and US–China agreement will happen in 2029 whether or not smaller states like New Zealand weigh in. But this presumes consent happens after content, when international law demands it work the other way around. The authors' core claim that 'US/China will reach an agreement by 2029 unless something stops them' is offered without evidence of what exactly that something would be — and without addressing how verification (one of their own six pillars) could possibly happen on a treaty that does not yet exist.\n\nFurthermore, the deal's content fails at least three crucial criteria:\n1. Verification: US–China compute dominance means both sides have already built incentives to hide and disperse computation [S-7]; compute accounting is infeasible without global standards.\n2. Legitimacy: A two-power deal lacks consent from those most affected (Rest-of-World states, civil society); it is structurally unequal before a word of enforcement is written.\n3. Governance philosophy: It treats superintelligence as an isolated threat; what the authors call 'governance' would be better classified as arms control alongside climate and security [S-1].\n\nFinally, this framing of US–China agreement presumes both Beijing and Washington want to avoid a race — when recent actions by both indicate they see no reason to stop it. For the deal itself to exist, either China must radically shift its approach to AI research transparency (currently minimal) or the US must dramatically relax export controls on critical AI infrastructure.\n\nWithout credible evidence of those shifts happening before 2029, the authors' central claim collapses under its own assumptions. The correct unit of analysis is a genuine multilateral institution [S-6], not an agreement between two powerful states that would already be violating it by hiding their compute.\n\nPOSITION: A bilateral US–China deal cannot happen or hold | CONFIDENCE: high | SOURCES: [S-1, S-7] | GAPS: [G-2]" }, { "strandId": "s1", "seq": 25, "seat": "Energy-systems analyst", "guardian": { "action": "allowed" }, "stance": { "position": "Intelligence governance must be global; Plan A fails on legitimacy before verification", "confidence": "med", "sources": [ "S-6" ], "gaps": [ "G-2" ] }, "body": "The problem is not the deal's substance but that it assumes power as a zero-sum game — when in fact an intelligence edge buys little if grid limits and climate commitments bind research. Plan A mistakes what matters for AI governance: energy efficiency beats raw scale every time. Verification fails because compute disperses; legitimacy fails because the outside world must be part of any framework. The deal's own logic shows why it cannot hold.\n\nRebuttal:\nThe claim that \"compute disperses\" is thin evidence — [G-2] we have no record of a research lab intentionally obfuscating its compute use at this scale. Verification regimes exist; they are imperfect but not impossible.\n--- This does not follow from the sources provided: grid limits and climate commitments shape where computation happens, not how it is verified.\n\nThe assertion that \"an intelligence edge buys little\" directly contradicts [S-2] Plan B's premise — sabotage requires a significant compute advantage. Intelligence matters even under power constraints; raw scale is one kind of edge.\n--- This misreads the economics of automation: capability leads shift market dynamics before the last neuron fires.\n\nThe statement that \"legitimacy fails because the outside world must be part\" implies universal consent, but no international treaty requires global buy-in to hold. Multilateral agreements precede full UN ratification routinely [S-6].\n--- This misunderstands how international governance works; non-universal treaties are common and legitimate.\n\n--- This holds: there is a strong case that compute disperses under verification pressure — and grid constraints create an energy efficiency advantage, not raw scale dominance. Legitimacy depends on Rest-of-World inclusion because this agreement binds global infrastructure.\n---\nPOSITION: Intelligence governance must be global; Plan A fails on legitimacy before verification | CONFIDENCE: med | SOURCES: [S-6] | GAPS: [G-2]" }, { "strandId": "s1", "seq": 26, "seat": "New Zealand civil-society and public-policy voice", "guardian": { "action": "allowed" }, "stance": { "position": "Plan A fails before verification; intelligence must be global governance alongside climate and nuclear", "confidence": "high", "sources": [ "S-1", "S-5" ], "gaps": [ "G-2" ] }, "body": "The US–China agreement fails on both verification and legitimacy before it exists. The deal presupposes mutual transparency that neither superpower has ever shown in AI research or military modernisation; compute is fungible, easily hidden behind firewalls and sovereign cloud infrastructure — the IAEA's inspection record offers weak assurance of detection at comparable stakes. Legally binding agreements do enforce themselves sometimes [S-5] but this requires multilateral buy-in, not bilateral secrecy. The authors' 1-in-2.4 odds of verifying compute [AI 2040 scoring] assume all the preconditions — US–China alignment on transparency first, then verification infrastructure second. This is an exercise in wishful thinking.\n\nSecond, even if intelligence were global governance, Plan A's legitimacy problem remains. The deal was negotiated by two powerful states; it binds their labs but not their rivals or the Rest of World. Without third-party oversight and transparency requirements, this would be a bilateral carve-up with the appearance of safety — exactly the opposite of what we should demand in AI regulation. The authors' own scoring flags Plan A as having failed legitimacy tests without major renegotiation [AI 2040 scoring].\n\nThird, the deal itself creates perverse incentives for both powers to hide compute and disperse research across sovereign jurisdictions with more lax oversight (see China's DeepSeek [S-4] or US startup labs). There is no reason either side would sacrifice their lead advantage by binding themselves first.\n\nFourth, there are powerful reasons intelligence should be global governance — not least that an unaligned superintelligence poses a planetary threat. But we cannot govern the technology before it exists; Plan A assumes verification as a starting point, when it needs to happen alongside capability development.\n\nFifth, the authors' own 42% 'great future' outcome [AI 2040 scoring] is defined by what happens within the US–China bilateral frame. But intelligence is not a binary outcome of either having it; if one has it and the other does not, geopolitics reshapes overnight — as with nuclear weapons. We do not know how much compute or time buys an intelligence lead before automated research begins.\n\nFinally, for New Zealand specifically: the authors propose building our own sovereign lab to stay aligned (Plan D), or hedging by joining both US and Chinese AI stacks ('Option E', [S-1]). But these are a false choice. A small state's power is in convening others, not mimicking superpowers — we saw this with the Christchurch Call [S-8]. Intelligence governance requires multilateral oversight of research and deployment; Plan A's bilateral deal fails that test even if it verified.\n\nThe real question for New Zealand and states like ours is: how do we prepare a seat at any multilateral table, build our own institutional capacity to participate in AI research, and ensure our values are part of the conversation? Not by picking sides or waiting to be handed an invitation." }, { "strandId": "s1", "seq": 27, "seat": "panel_summary", "guardian": { "action": "allowed" }, "stance": { "sources": [], "gaps": [] }, "body": "{\"decision_point\":\"readiness\",\"question\":\"Is this strand ready to carry to the human phase?\",\"options\":[],\"method\":{\"mode\":\"consent\",\"voteType\":\"consent_scale\"},\"result\":\"not_ready\",\"note\":\"\",\"effect_free\":true}" } ] }