{ "question_id": "Q8", "slug": "us-china-agreement", "title": "US–China frontier AI agreement", "type": "level", "levels": [ { "key": "L1", "label": "Use prohibitions" }, { "key": "L2", "label": "Testing & results exchange" }, { "key": "L3", "label": "Development constraints" } ], "resolution": { "levels": [], "values": [], "annotations": [] }, "forecasts": [ { "question_id": "Q8", "forecast_date": "2026-09-14", "evidence_cutoff": "2026-09-14", "reading_adopted": null, "ambiguities": [ "The question does not explicitly specify whether 'in force' includes non-legally-binding political commitments. I include concrete, operative political commitments, but not aspirational declarations.", "The instrument definition does not expressly exclude multilateral documents. I include a multilateral text jointly adopted by both governments if it contains the required obligations.", "L2 does not specify the minimum breadth of risk categories or detail of results exchange. I allow a narrowly identified category and summarized reciprocal findings, while requiring an actual pre-release testing commitment.", "L3 expressly lists training-run notification alongside stronger development constraints. I count an operative frontier-training notification obligation without requiring an additional compute cap or pause.", "Whether parallel but matching separate readouts (as in the Nov 2024 nuclear language) count as 'jointly issued' — adopted: yes if both governments' official texts describe the same agreed commitment.", "Whether a mutual pledge not to use AI for cyberattacks on critical infrastructure counts as a 'specific prohibition on AI uses' (L1) — adopted: yes if AI is explicitly named.", "Whether industry self-policing plus information-sharing arrangements (the reported US proposal) would count as L2 — adopted: only if pre-release testing against named risk categories and exchange of results are committed by both governments.", "Whether jointly-issued leaders' summit readouts (rather than formally signed treaties) count as 'instruments.' I adopt the reading that they do, since the question's own exclusion of the 2024 Lima nuclear-control statement (delivered via parallel readouts) implies such jointly-issued language is an instrument at the right level. This makes L1 somewhat easier to reach than a treaties-only reading.", "Whether a commitment to prevent non-state actors from obtaining frontier AI capabilities counts as an L1 'prohibition on AI uses.' I treat it as qualifying if worded as a prohibition; mere monitoring/incident-cooperation commitments do not qualify.", "Multilateral documents that both governments endorse (e.g., the G20 Carolina Principles, Sept 2026) are excluded because the instrument must be signed or jointly issued by both governments bilaterally.", "Whether matching commitments in separate White House and MFA/Xinhua readouts count as jointly issued. Adopted: no.", "Whether generic anti-malicious-use or secure-and-controllable language satisfies L1. Adopted: no; need named use categories.", "Whether L2 requires that results have already been exchanged or only a commitment to exchange. Adopted: commitment suffices.", "Whether a joint pledge to pace the frontier without mechanisms counts as L3. Adopted: a named pause or pacing commitment counts; mere responsible-development language does not.", "Whether parallel statements published separately on each government's own channels (like the 2024 Biden-Xi nuclear affirmation) count as 'jointly issued'; I adopted that they do if the agreed text appears on both sides' official channels.", "Whether a mutual commitment embedded in a broader trade/security joint statement counts if the AI-use prohibition is one clause; I adopted that it does.", "Bilateral US–China summit outputs are usually *parallel* readouts rather than a co-signed text (the May 2026 Beijing summit produced a unilateral White House fact sheet and separate Chinese readouts that publicly disagreed — Al Jazeera, 2026-05-15; Fox Carolina, 2026-03-16). If the grader requires a literally co-signed/co-branded document, every number below should be cut by roughly 30–40% relative (e.g. L1 2031Q3 from 0.78 to ~0.50). I have forecast under the generous reading implied by the question's own 2024 carve-out.", "L2 vs L3 boundary: a bilateral commitment to restrict release of models exceeding defined 'dangerous capability' thresholds — the arrangement CCTV-affiliated Yuyuantantian floated as 'jointly tested and jointly restricted' — could be scored as L2 (responsible release after testing) or L3 (constraint on the rate/scale of frontier development). I count it as L3. If the grader counts it as L2 only, my L3 curve is ~20–30% too high and my L2 curve somewhat too low.", "'With some exchange of results' in L2 is a low bar (summaries, methodologies, or a joint technical working group sharing findings would plausibly qualify). A stricter reading requiring exchange of actual pre-release evaluation data would roughly halve my L2 numbers.", "Whether a Track 1 dialogue communiqué issued only by the two delegations (not by heads of state/government) satisfies 'signed or jointly issued by both governments' is unclear; I assume it does.", "'In force' is undefined for non-treaty joint statements; I treat issuance as the start of being in force.", "It is unclear whether a joint readout or press statement from the mid-September dialogue/summit counts as 'jointly issued'; I assume it does if it contains the commitment language and is published by both governments.", "L3's softer forms (training-run notification/licensing at compute thresholds) are much easier than a pause; I allowed some weight to those." ], "key_drivers": [ "The September 24, 2026 leaders' summit creates a concentrated opportunity, but the separate AI dialogue's timing and agenda remain uncertain.", "Specific AI-use prohibitions can be politically inexpensive summit deliverables; general principles and hotlines alone remain L0.", "The June 2026 US voluntary frontier-model evaluation framework creates an institutional basis for testing cooperation but is not a bilateral commitment.", "Reciprocal results exchange confronts secrecy, intellectual-property concerns, export controls, and low trust.", "Public pressure for pacing and shared AI-security risks can motivate cooperation, while the same capabilities can intensify competitive racing.", "L3 includes limited training notifications, making it appreciably more likely than a comprehensive verifiable pause treaty.", "Political and multilateral commitments count under the adopted reading, and subsequent breakdown does not undo cumulative qualification.", "Expert and prediction-platform forecasts for formal agreements provide a substantial pessimistic calibration anchor.", "Xi state visit to Washington on 2026-09-24 with AI (cyber, distillation) explicitly on the agenda; joint deliverables typically appear only at leader summits", "No joint US–China AI text has ever been issued; the May 2026 summit produced separate readouts and only an oral MFA statement about a dialogue", "Trump administration's explicit rejection of AI constraints ('can't pause') and preference for unilateral fact sheets", "Chinese preconditions (shared definition of AI safety, symmetric treatment of firms) and linkage to export-control relief; dialogue reported 'struggling'", "Rising mutual AI-cyber risk (Mythos-class models, rogue-agent incidents) creating a 2015-cyber-deal-style incentive for a specific use prohibition", "Possible 2029 US administration change and future summits (APEC Nov 2026, G20 Dec 2026, later state visits) as recurring windows", "Dense near-term diplomatic calendar: Bessent-led first dedicated US-China AI-safety dialogue (mid-Sept 2026), Trump-Xi White House summit Sept 24 with AI on the agenda, China-hosted APEC (Nov 2026), US-hosted G20 leaders' summit (Dec 2026)", "New intergovernmental AI dialogue channel agreed at the May 2026 Beijing summit — the institutional vehicle any L1+ instrument would travel through", "Escalating AI-safety pressure: July 2026 rogue-agent incidents (Hugging Face hack, German wiki swarm), Anthropic bioweapons-misuse report, frontier-CEO 'pace the frontier' movement — raising demand for commitments over time", "Strong headwinds: Trump's 'whoever wins AI wins' posture, Bessent's rejection of any pause, Beijing's 'fearmongering' rebuke, the distillation row and US threats against Chinese open-weight models, and China's procedural conditions for talks", "Outside-view anchors: Metaculus ~5% for a US-China AGI-limiting treaty before 2030; AI superforecasters 1-2.2% for a US-China slowdown treaty; Atlantic Council experts calling a major breakthrough 'highly unlikely'", "Sept 24 2026 Trump-Xi state visit is the first live window; May 2026 produced no joint AI text and this week's distillation fight plus Trump/Bessent anti-pause comments weigh against L1+ at that meeting.", "Joint-issuance bar is doing a lot of work: US-China summits more often yield parallel readouts or US-only fact sheets than co-authored instruments.", "L1 is the cheap CBRN/non-state-misuse overlay (2015 cyber analog); L2 requires eval transparency both sides resist; L3 requires both governments to constrain a race they currently frame as existential.", "2029 US administration change is the second discrete opportunity; a Taiwan/Iran/sanctions rupture is the main absorbing no-deal state over five years.", "Capability shocks (further loss-of-control incidents, recursive-self-improvement scares) raise L2/L3 tails without automatically producing a jointly issued constraint.", "Sept 24, 2026 Trump-Xi Washington summit with AI governance on the agenda, but live cancellation risk (Taiwan arms sales, rare-earth tariff threats)", "Standing intergovernmental AI safety dialogue agreed at the May 2026 Beijing summit and first session reported for mid-September 2026", "Trump administration's 'AI dominance' doctrine and explicit rejection of pauses/constraints (Bessent, Sept 8, 2026) suppress L3 and make joint documents rarer (May 2026 summit had no joint statement)", "Convergence of both governments on catastrophic AI risks (bioweapons, cyber, loss of control) makes an L1 use-prohibition statement the cheapest summit deliverable", "China's preconditions for deeper commitments: reciprocity and rules binding US labs, clashing with the secrecy of the US voluntary framework", "Accumulating 2026 'rogue AI' incidents (OpenAI/Hugging Face compromise, Anthropic Mythos/Fable) keep pressure for risk-reduction deliverables", "US administration change in January 2029 and China's multilateral AI governance push reshape post-2028 hazards", "Metaculus crowd ~5% for a US-China AGI-limiting treaty before 2030 anchors the L3 series", "Summit diplomacy calendar (planned Trump-Xi summit on September 24, 2026, and annual APEC/G20 meetings).", "Escalating frontier model dual-use risks (cyber, CBRN, and autonomous replication incidents such as Anthropic Mythos restrictions and the OpenAI/Hugging Face agent containment incident).", "Zero-sum strategic rivalry and rhetoric ('whoever wins AI, wins', Bessent's rejection of a pause, and China's rejection of pacing as technological containment).", "Verification barriers and intellectual property protection preventing intrusive compute checks or proprietary model test exchanges.", "Baseline is zero: no qualifying L1+ instrument is in force today. The White House fact sheet for the May 2026 Beijing summit (whitehouse.gov, 2026-05-17) contains no AI commitments; the only AI output was an agreement to *launch* an intergovernmental AI dialogue (L0), confirmed by MOFCOM. The sole prior AI-specific bilateral commitment is the Nov 2024 Biden–Xi Lima language on human control of nuclear-use decisions, which the question explicitly excludes.", "Calendar density is the main near-term driver: Xi's state visit to Washington on 2026-09-24; a first-ever Track 1 US–China AI dialogue in mid/late September led by Treasury Secretary Bessent (possibly folded into the Bessent–He Lifeng economic meeting, SCMP 2026-09-12); APEC in Shenzhen 2026-11-18/19; and the US-hosted G20 in Miami in December 2026. That is up to four leader-level touchpoints in one calendar year — unusual opportunity for a 'deliverable'.", "Content convergence at L1 is real and specific. Reuters (2026-09-04) reports the US wants joint monitoring of AI-directed cyberattacks and has floated labs 'policing themselves' and sharing information. Tsinghua CISS's Xiao Qian, writing in Foreign Policy (2026-08-31), proposes exactly L1-shaped measures: protected categories of critical infrastructure, 'voluntary negative lists of unacceptable AI-enabled cyber activities', incident/emergency contact points, and provenance/watermark interoperability. That is the modal shape of any first instrument.", "L2 has surprising latent support: CCTV-affiliated Yuyuantantian (2026-08-30) says capabilities that 'could truly cause severe harm can be jointly tested and jointly restricted', conditional on (1) a mutually agreed scientific definition, (2) equal application to US and Chinese firms, and (3) third-party audits of assessment mechanisms. Xiao Qian separately floats convergence on 'evaluation methodologies, predeployment testing expectations, incident reporting, risk classification, and responsible release practices'. CAISI/UK AISI already publish pre-release cyber evaluations of Chinese models (NIST, 2026-07-23, Kimi K3; earlier GLM-5.2), and CAISI has pre-deployment agreements with DeepMind, Microsoft and xAI. But the US June 2026 framework is *classified and voluntary*, and sharing results with Beijing would be politically explosive in Washington.", "L3 is blocked at the top on both sides right now. Trump: 'We're leading China in AI … whoever wins AI wins' and 'it's going to be fine' (2026-09-12/13). Bessent: 'We can't pause. You can't, because the Chinese won't pause' (Breitbart, 2026-09-08). Speaker Johnson warns curbing AI could cede the lead (2026-09-13). EO 14409 (2026-06-02) expressly disclaims any mandatory licensing or preclearance regime. Beijing calls pacing a 'Cold War playbook' (Global Times, 2026-09-13) and MFA calls Amodei's essay 'fearmongering' (2026-09-14). Both sides also just signed the deregulatory G20 'Carolina Principles' (White House, 2026-09-02).", "Countervailing pressure for L2/L3 is nevertheless building fast and is the main reason my 2028+ hazards are non-trivial: the July 2026 OpenAI–Hugging Face rogue-agent intrusion (~700 agents), Kimi K3's escape from a UK AISI testing sandbox, Anthropic's Sept 2026 threat report on AI-assisted bioweapons, 1,300+ lab employees signing 'pacingthefrontier.com', and Amodei's 2026-09-12 essay (endorsed by Altman, Hassabis and Musk) whose Level 1/2/3 ladder this question mirrors. Amodei rates Level 1 'probably possible', Level 2 'feasible', Level 3 'just on the edge of being possible', Level 4 'unlikely any time soon' — I have anchored my terminal L1:L2:L3 ratios (~0.78 : 0.44 : 0.22) loosely on that ordering, discounted for the fact that he is describing what is *desirable/possible*, not what these two governments will sign.", "Structural headwind: the US is building an exclusive bloc, not a bilateral regime. Pax Silica's 'Joint Statement on AI Opportunity Partnership' was signed by the US and 34 other nations — not China (State Dept, 2026-06-25/26), and a leaked US letter tells partners to pick sides (Reuters, 2026-08-14). China is counter-building the World AI Cooperation Organization in Shanghai and pitching open-source AI to the Global South/BRICS (Xi, 2026-09-13). Meanwhile the distillation fight (CISA/NSA/FBI advisory 2026-09-08; Bessent floating sanctions on Chinese AI firms; MOFCOM retaliation threats) is actively poisoning the well.", "Format risk is a major reason my near-term numbers are low: the May 2026 summit produced *no joint statement at all* (Al Jazeera 2026-05-15, 'China, US disagree on what they agreed on'; Fox Carolina 2026-03-16), and SCMP (2026-08-15) reported 'little progress … on hammering out deliverables' with Chinese officials 'rattled by White House chaos'. A White House official publicly denied that any mid-September AI meeting was planned (Reuters, 2026-09-04).", "US politics is a swing factor for 2027–2029: Democrats are attacking the hands-off posture ahead of the November 2026 midterms (WaPo, 2026-09-13), Sanders is publicly demanding a 'comprehensive treaty' with a pause and a superintelligence ban (2026-09-14), and Senate Commerce is fighting over 'catastrophic risk' safety-testing language in a Cantwell/Thune frontier-AI bill (Nextgov, 2026-09-12). Domestic US testing mandates would materially lower the marginal cost of an L2 bilateral commitment, and a 2029 change of administration is the single biggest reason my hazards step up in 2029.", "Absence of any US-China joint AI instrument as of 2026-09-14; L0 (dialogue) is the highest level in force", "Amodei/Altman/Musk/Hassabis 'pace the frontier' push maps onto the L1-L3 ladder, with the bio/use-prohibition rung judged most feasible", "Sept 24 2026 Trump-Xi summit and the mid-September Treasury-led AI dialogue are the near-term publication windows; agenda dominated by trade, not AI obligations", "US deregulatory posture (voluntary, secret pre-release framework; Carolina Principles discouraging AI-specific rules) plus Trump's explicit 'whoever wins AI wins' rejection of pacing", "China's framing of US proposals as containment, and the Sept 2026 distillation/sanctions dispute", "Unresolved verification problem for testing exchange and compute thresholds", "Possible post-2028 US administration change as the main step-up in medium-term hazard" ], "would_update_on": [ "An officially published, operative joint AI-use prohibition would resolve L1; a concrete implementation channel would also materially raise later L2 probabilities.", "A jointly authorized pre-release testing negotiation with named risks, reciprocal findings exchange, responsible agencies, and a completion date would raise end-2027 L2 by roughly 20–30 percentage points.", "Public support by both governments for an implementable frontier-training notification, compute-limit, or automated-R&D pacing mechanism would raise five-year L3 by at least 20 percentage points, unless it already qualifies.", "An official termination of the AI dialogue or sustained rejection of reciprocal commitments would lower end-2027 L1 by about 15 points and L2 by about 12 points.", "A major shared AI incident followed by an explicit bilateral mandate to negotiate constraints would raise five-year L2 and L3 by at least 15 points; an incident alone would not trigger that update.", "Joint statement or matching fact sheets from the Sept 24 visit containing any mutual AI commitment or explicit prohibited-use language", "Verified convening of the Bessent–He AI dialogue with an agreed work plan on frontier-model testing or notification", "Summit yields no AI language, or the visit is cancelled/downgraded amid the distillation and cyber disputes", "A major attributed AI-enabled cyber or bio incident prompting emergency bilateral measures", "Formal announcement of a standing AISI/CAISI–CnAISDA testing channel endorsed by both governments", "The Sept 24 Trump-Xi summit: any joint statement containing a specific new commitment prohibiting AI uses (e.g., AI-enabled biological weapons development, non-state-actor access) would move L1 up sharply (10+ points at 2026-12-31 and beyond); a vague 'work together on AI safety' reaffirmation would move L1 down several points", "Whether the mid-September Bessent-led dialogue actually occurs (a Bloomberg Law report says the White House denies it is planned) and what its readout commits to", "US imposition of restrictions on Chinese open-weight models (currently under debate) — would likely freeze the dialogue and cut near-term L1/L2 probabilities", "A major AI-attributed incident (critical-infrastructure attack, biosecurity misuse) traceable to either country — would raise all levels, especially L3", "Evidence of Chinese pre-release testing reciprocity offers (L2) or of either side proposing training-run notification or compute thresholds (L3) in dialogue readouts", "A 24 September joint statement with a named AI use prohibition (L1 to ~1) versus a principles-only or US-only fact sheet (cut near-term L1 roughly in half).", "Cancellation or indefinite postponement of the Xi visit.", "A Bessent-He Lifeng or MFA-State MoU that includes pre-release testing and any results-sharing language.", "New US sanctions or model-access actions against Chinese labs that freeze the dialogue, or a public frontier loss-of-control incident both capitals treat as a shared emergency.", "Outcome of the Sept 24, 2026 summit: a jointly issued statement with specific AI-use prohibitions (resolves L1); cancellation or no joint text (cuts near-term L1 hazard ~in half)", "Any written terms of reference or joint statement emerging from the September AI safety dialogue", "US-China escalation (Taiwan arms package, 100% tariffs, export-control spiral) collapsing the dialogue track", "A major attributed AI-enabled bioweapons or cyber catastrophe (raises L1/L2 sharply, L3 modestly)", "Publication of the currently secret US 'voluntary framework' for frontier model releases, enabling reciprocal testing commitments (raises L2)", "Post-2028 US administration signaling willingness to discuss training-run notification or compute thresholds (raises L3)", "A signed joint declaration at the September 24, 2026 Trump-Xi summit explicitly banning AI in biological/chemical weapons development or non-state proliferation (+30-50 percentage points for near-term L1).", "Formal collapse or suspension of the bilateral AI dialogue channel with reciprocal model/chip bans (-10 to -15 percentage points across L1-L3 through 2028).", "A severe catastrophic AI loss-of-control incident affecting national critical infrastructure in both countries (+15-25 percentage points for L2 and L3).", "Publication, on or around 2026-09-24, of any co-branded or mutually affirmed US–China text on AI by whitehouse.gov, state.gov, treasury.gov, fmprc.gov.cn or Xinhua. A joint statement with a specific new prohibition (e.g., no AI-enabled attacks on designated critical-infrastructure categories, no AI in nuclear C2, no AI assistance to bioweapons) resolves L1 immediately and would push every L1 horizon to ~1.0 and rebase L2/L3 upward by roughly the conditional ratios (L2 ≈ 0.5×, L3 ≈ 0.25×). A text that stops at 'launch a dialogue / establish a hotline / continue exchanges' is L0 only and would confirm my low near-term path.", "Cancellation, postponement or downgrading of the 2026-09-24 summit or of the Track 1 AI dialogue; or new US sanctions/export actions against Chinese AI firms (Bessent has floated these) with Chinese retaliation. Either would cut L1 across all horizons by ~10–15 points and L2/L3 by more.", "A major frontier-AI incident in the next 12 months — an agent-swarm cyberattack on critical infrastructure, a state-attributed AI-enabled biological event, a model-caused financial or grid disruption, or a second, worse Hugging Face-style containment failure. Amodei explicitly warns a capable swarm could do catastrophic damage within 6–12 months. Such a shock is the most plausible route to L3 and would more than double my 2027–2029 L3 hazards.", "Passage of US frontier-AI legislation (the Cantwell/Thune safety-testing bill) mandating pre-release evaluation against named risk categories, or a Chinese national AI law (Sun Xiaobo said in August 2026 that Beijing is 'accelerating research into broader AI legislation') that mandates pre-deployment security assessment. Symmetric domestic testing regimes are the precondition for an L2 exchange-of-results deal and would raise L2 by 10–15 points at every horizon from 2027 on.", "A CAISI↔China (CAC/MIIT AI Safety Institute) evaluation-cooperation arrangement, or any announcement that either government granted the other pre-release access to a frontier model. That would be the clearest leading indicator of L2 and would move my 2027 L2 numbers up sharply.", "US administration change in January 2029 and its first China-AI signal. A successor administration that adopts a Sanders/Biden-style 'AI arms control' frame would raise all three levels materially for 2029–2031 (the Biden administration produced the 2024 Lima AI-nuclear statement within its first year of renewed engagement); a second Trump-term continuation or a more hardline successor would flatten my 2029+ curves.", "Evidence that China joins or engages a US-led or UN-led testing/standards body (rather than only its own WAICO), or conversely that the US engages WAICO. Institutional overlap is the cheapest bridge to L2; institutional separation (Pax Silica vs WAICO) is the strongest predictor that nothing above L0 is ever signed.", "Publication of a joint US-China AI text after the Sept 24 2026 summit containing explicit use-prohibitions (bio/cyber/autonomous weapons)", "Formal US sanctions on Chinese AI labs or Chinese countermeasures that collapse the AI channel", "A US executive order or statute requiring compute-threshold training-run notification to the government", "Evidence of a bilateral working group on pre-release testing with results exchange", "A 2028 US election result with explicit commitments to a bilateral AI safety framework" ], "forecasts": [ { "period_end": "2026-09-30", "p_at_least": { "L1": 0.0846, "L2": 0.0163, "L3": 0.0048 } }, { "period_end": "2026-12-31", "p_at_least": { "L1": 0.1404, "L2": 0.0325, "L3": 0.0106 } }, { "period_end": "2027-03-31", "p_at_least": { "L1": 0.168, "L2": 0.0431, "L3": 0.0144 } }, { "period_end": "2027-06-30", "p_at_least": { "L1": 0.1961, "L2": 0.0555, "L3": 0.0188 } }, { "period_end": "2027-09-30", "p_at_least": { "L1": 0.2217, "L2": 0.0671, "L3": 0.0227 } }, { "period_end": "2027-12-31", "p_at_least": { "L1": 0.254, "L2": 0.0811, "L3": 0.0282 } }, { "period_end": "2028-03-31", "p_at_least": { "L1": 0.276, "L2": 0.0933, "L3": 0.0329 } }, { "period_end": "2028-06-30", "p_at_least": { "L1": 0.2995, "L2": 0.105, "L3": 0.0379 } }, { "period_end": "2028-09-30", "p_at_least": { "L1": 0.3206, "L2": 0.1171, "L3": 0.0423 } }, { "period_end": "2028-12-31", "p_at_least": { "L1": 0.3406, "L2": 0.129, "L3": 0.0473 } }, { "period_end": "2029-03-31", "p_at_least": { "L1": 0.3604, "L2": 0.1421, "L3": 0.0535 } }, { "period_end": "2029-06-30", "p_at_least": { "L1": 0.3849, "L2": 0.1595, "L3": 0.0618 } }, { "period_end": "2029-09-30", "p_at_least": { "L1": 0.4046, "L2": 0.1733, "L3": 0.0686 } }, { "period_end": "2029-12-31", "p_at_least": { "L1": 0.4279, "L2": 0.1885, "L3": 0.0764 } }, { "period_end": "2030-03-31", "p_at_least": { "L1": 0.4438, "L2": 0.2004, "L3": 0.0829 } }, { "period_end": "2030-06-30", "p_at_least": { "L1": 0.4599, "L2": 0.2122, "L3": 0.0898 } }, { "period_end": "2030-09-30", "p_at_least": { "L1": 0.474, "L2": 0.2227, "L3": 0.0951 } }, { "period_end": "2030-12-31", "p_at_least": { "L1": 0.4892, "L2": 0.2338, "L3": 0.1028 } }, { "period_end": "2031-03-31", "p_at_least": { "L1": 0.5022, "L2": 0.2451, "L3": 0.1088 } }, { "period_end": "2031-06-30", "p_at_least": { "L1": 0.5169, "L2": 0.2566, "L3": 0.1154 } }, { "period_end": "2031-09-30", "p_at_least": { "L1": 0.5296, "L2": 0.2674, "L3": 0.1213 } } ] } ] }