Frontier Compliance Costs: Structural Burdens or Measured Exclusion?

Full report

On this page

1. Answer and scope

The contemporary governance of artificial intelligence systems has increasingly centered on raw compute capacity as a proxy for systemic risk. This investigation evaluates whether frontier-law requirements—specifically California’s Senate Bill 53 (Transparency in Frontier Artificial Intelligence Act) and New York’s RAISE Act (as amended by S8828)—create structural burdens that systematically exclude independent development, or whether such exclusion is driven by pre-existing infrastructure concentration. The core inquiry distinguishes between the fixed compliance burdens of establishing risk management frameworks and the variable, scale-dependent operating costs of mandatory incident telemetry and state fees. The analysis reveals a heavily bifurcated compliance architecture that deliberately insulates smaller independent developers while imposing severe, multi-tiered obligations on massively capitalized incumbents. Fixed burdens are primarily driven by statutory mandates to draft, implement, and routinely audit comprehensive catastrophic risk safety frameworks, as well as the mandate to establish heavily protected internal whistleblower portals. Variable burdens—which scale with model deployments and organizational footprint—are driven by continuous telemetry requirements for critical incident reporting, alongside New York’s unique imposition of a pro-rata fee designed to fund state oversight operations. However, evaluating the actual exclusionary effect of these statutes requires distinguishing between legal permission and commercial affordability. Public evidence gathered from the Federal Trade Commission (FTC) 6(b) study on AI Partnerships and Investments, combined with assessments by the National Telecommunications and Information Administration (NTIA), establishes that extreme infrastructure concentration gates market participation long before statutory compliance costs are ever triggered. The sheer capital requirement to acquire the specialized semiconductor hardware necessary to train models exceeding the statutory threshold of [Figure omitted from source export] floating-point operations (FLOPs) forces independent developers into highly restrictive equity and exclusivity partnerships with dominant cloud service providers (CSPs). Consequently, while state-level regulatory frameworks impose tangible and substantial costs, they primarily formalize operational burdens for entrenched incumbents rather than acting as the principal exclusionary mechanism for independent market entrants. To properly situate these regulatory mechanics, it is essential to distinguish between four distinct capability cases in artificial intelligence deployment: a conversational interface, a bounded task agent, a persistent operatorless service, and a hypothetical future machine principal. A conversational interface is a generally stateless or bounded-state text system that interacts synchronously with human users, allowing for manual sampling of risk telemetry. A bounded task agent executes defined, multi-step actions within strict operational parameters, requiring automated logging but retaining failures within locally contained systems. A persistent operatorless service—such as the assigned case of Concresca—operates continuously without a staffed approval queue for enrollment, authentication, coordination, or policy enforcement. Finally, a hypothetical future machine principal represents an entity possessing contested independent interests, full autonomy, and theoretical legal agency. Interface design, technical capability, legal permission, legal status, and consciousness represent distinctly different variables. The present uncertainty regarding the latter cases neither establishes machine rights nor resolves the debate over whether new, generalized technological protections are inherently warranted. However, distinguishing these cases is critical for evaluating compliance costs, as a persistent operatorless service faces vastly higher variable engineering burdens to meet rapid incident-reporting mandates than a system reliant on human-in-the-loop oversight. The scope of this review is strictly bounded to the operative provisions of California SB 53 and New York S8828, evaluated against the baseline infrastructure evidence provided by the FTC and NTIA up to the current cutoff of September 6, 2026\.

2. Provision-level findings

An objective evaluation of frontier computing regulations necessitates separating legislative intent from mechanical, provision-level realities. Both California SB 53 and New York’s RAISE Act operate on specific compute and financial thresholds that divide the regulatory landscape, categorizing obligations based on raw technical capacity and commercial scale.

Thresholds and Covered Actors

California SB 53 distinguishes between a "frontier developer" and a "large frontier developer." A frontier developer is any person or entity that has trained, or initiated the training of, a foundation model utilizing a quantity of computing power greater than [Figure omitted from source export] integer or floating-point operations, inclusive of compute used in subsequent fine-tuning \[cite: R2-07-S01\]. A "large frontier developer" is defined as a frontier developer that additionally exceeds [Figure omitted from source export] million in annual gross revenue \[cite: R2-07-S01\]. New York’s RAISE Act (S8828) harmonized its criteria with California following significant chapter amendments, defining a "large frontier developer" using the exact same dual threshold: [Figure omitted from source export] FLOPs and [Figure omitted from source export] million in annual revenue \[cite: R2-07-S02\]. Forecasts from AI research organizations suggest that by 2028, between 45 and 148 foundation models may exceed this [Figure omitted from source export] FLOP threshold globally \[cite: R2-07-S09, R2-07-S10\]. This dual-threshold mechanism ensures that independent developers who manage to pool decentralized compute or rely on heavy open-weight fine-tuning to surpass the FLOP threshold, but who lack massive commercial scale, remain shielded from the most onerous statutory mandates.

Fixed Compliance Costs: Frameworks and Whistleblower Infrastructure

The most significant fixed burden across both jurisdictions is the mandatory creation of a "frontier AI framework." Under California Business and Professions Code Chapter 25.1 and New York General Business Law Article 44-B (§ 1421), large frontier developers must write, implement, and conspicuously publish a framework detailing how they assess, mitigate, and govern catastrophic risks \[cite: R2-07-L01, R2-07-L04\]. Catastrophic risks are statutorily defined to include events materially contributing to the death or serious injury of more than 50 people, or causing more than [Figure omitted from source export] billion in property damage \[cite: R2-07-S02\]. New York’s statute specifically requires this framework to describe safety measures "in detail," whereas California requires developers to describe their "approach" \[cite: R2-07-S01, R2-07-S02\]. Drafting these safety frameworks represents a front-loaded, fixed cost. Synthetic market estimates—utilizing the ISO/IEC 42001 Artificial Intelligence Management System (AIMS) standard as a proxy for regulatory readiness—suggest that establishing the requisite governance structures, impact assessments, and documentation incurs fixed costs ranging from [Figure omitted from source export] for smaller mid-market entities to well over [Figure omitted from source export] for highly complex enterprise environments \[cite: R2-07-S06\]. Furthermore, California Labor Code § 1107 imposes an additional fixed infrastructure cost by requiring large developers to establish a reasonable internal process allowing covered employees to anonymously disclose catastrophic risk concerns without fear of retaliation \[cite: R2-07-L02\]. This mandates dedicated human-resources architecture, secure reporting software, and external legal counsel to manage the newly created burden-shifting standards in retaliation claims \[cite: R2-07-S07\].

Variable Operating Costs: Incident Reporting and State Pro-Rata Fees

Variable burdens scale proportionally with the operational complexity of the deployed model and the frequency of anomalous events. Both statutes mandate critical safety incident reporting, creating an ongoing operational liability. California requires reporting to the Office of Emergency Services within 15 days of discovering a critical safety incident, accelerating to a strict 24-hour window if the incident poses an imminent risk of death or serious physical injury \[cite: R2-07-S01\]. New York mandates reporting within 72 hours, similarly accelerating to 24 hours for imminent public threats \[cite: R2-07-S02\]. A unique variable burden exists exclusively in New York: under General Business Law § 1428, large frontier developers are assessed pro-rata shares by the state to defray the direct and indirect operating expenses of administering the RAISE Act \[cite: R2-07-L05\]. This operates as a pure scale-dependent operating cost—a specialized regulatory assessment on the largest market participants that fluctuates based on the state's enforcement budget.

The Operatorless Constraint and Automated Compliance

When applying these mandates to the capability case of a persistent operatorless service, the mechanical burden shifts drastically. The named example, Concresca, is designed with a strict operational requirement: enrollment, authentication, coordination, policy enforcement, credentials, maintenance, and recovery must not depend on a staffed human approval queue. This represents a functional operating requirement, not a verified deployment status or an automatic legal exemption from state oversight. Because Concresca specifies no human operators, maintaining compliance with a 24-hour imminent threat reporting requirement under California or New York law cannot rely on manual incident triage. The legal duty textually transforms into a stringent technical requirement for "automated compliance" architectures \[cite: R2-07-S08\]. The verified rule (mandatory 24-hour reporting of imminent threats) leads to a conditional application (the service generates a catastrophic risk signature) forcing a possible response (the system must identify and report the anomaly). Because human review is architecturally absent, the affected activity (incident telemetry) must be handled autonomously. The resulting burden is immense: Concresca must programmatically detect catastrophic risk signatures and autonomously compile and file incident disclosures to state authorities. This drastically elevates the variable cost of development, as the required real-time telemetry and deterministic automated reporting pipelines introduce profound engineering overhead compared to standard human-in-the-loop deployment architectures \[cite: R2-07-S08\].

Cost-Channel Matrix

Offsetting Provisions and Exceptions

Regulatory statutes routinely incorporate offsetting mechanisms intended to lower entry costs or facilitate independent, public-interest capabilities. California’s SB 53 authorized the creation of the CalCompute consortium (Government Code § 11546.8) to advance safe, equitable AI development through a public cloud computing cluster \[cite: R2-07-L03\]. Similarly, New York’s S8828 exempts accredited academic institutions and the state-backed Empire AI consortium from its framework requirements \[cite: R2-07-S02\]. While these mechanisms theoretically lower the capital constraints for public-interest research, it is imperative to investigate actually available resources separately from authorizations on paper. CalCompute’s operative status remains entirely contingent on legislative budget appropriations that have not yet materialized \[cite: R2-07-S01\]. An unfunded authorization on paper does not provide actual compute resources to independent developers. Consequently, the offsetting condition under which the framework genuinely lowers entry costs or improves independent audits remains hypothetical until state treasuries release actual capital for GPU procurement and distribution.

3. Four worked cases

R2-07-C01 — Small research cooperative

A small research cooperative operates entirely outside the traditional venture capital ecosystem. Through highly efficient distributed training and decentralized compute pooling, they successfully develop a model exceeding the [Figure omitted from source export] FLOP threshold, distributing the weights openly. However, their cooperative model generates negligible commercial returns, remaining well below the [Figure omitted from source export] million annual revenue threshold. Under California SB 53, this cooperative legally qualifies as a "frontier developer" due to the compute metric, but fails to meet the definition of a "large frontier developer" \[cite: R2-07-S01\]. Consequently, they are exempt from the fatal fixed compliance burdens of writing, implementing, and annually auditing a catastrophic risk safety framework. They remain subject to California's critical safety incident reporting mandate—requiring telemetry for 15-day or 24-hour disclosures—and basic non-retaliation whistleblower protections \[cite: R2-07-L01, R2-07-L02\]. Under New York’s S8828, the cooperative falls completely outside the scope of the statute because the RAISE Act restricts all major obligations, including incident reporting and pro-rata fees, strictly to "large frontier developers" meeting both compute and revenue limits \[cite: R2-07-S02\]. The effect of these revenue-based exemptions is profound: it textually protects the cooperative from front-loaded legal and administrative costs that would otherwise preclude their existence, isolating their regulatory burden to scale-dependent operational logging in California.

R2-07-C02 — Threshold-crossing developer

An independent, highly capitalized developer successfully commercializes an advanced open-weight model, crossing both the [Figure omitted from source export] FLOP compute threshold and the [Figure omitted from source export] million annual revenue threshold, thereby triggering "large frontier developer" status in both California and New York \[cite: R2-07-S01, R2-07-S02\]. This threshold crossing initiates an immediate, heavy fixed compliance burden. The developer must formalize a massive AI management system, draft a public catastrophic risk framework, and build an anonymous internal whistleblower portal subject to California Labor Code § 1107 \[cite: R2-07-L01, R2-07-L02\]. Synthetically sizing this effort using ISO/IEC 42001 certification as a proxy, enterprise-grade framework implementation, policy drafting, and external readiness auditing require a fixed capital outlay ranging from [Figure omitted from source export] to over [Figure omitted from source export] in immediate administrative and legal costs \[cite: R2-07-S06\]. The variable operational costs scale aggressively. New York imposes a mandatory pro-rata fee to fund state AI oversight, directly taxing the developer's revenue footprint \[cite: R2-07-L05\]. Furthermore, if this developer deploys a persistent operatorless system, the 24-hour imminent threat reporting requirement dictates that human review cannot serve as the safety bottleneck. They must incur the heavy variable engineering costs of building fully automated, deterministic incident-reporting architecture to remain compliant without a staffed approval queue \[cite: R2-07-S08\].

R2-07-C03 — Counterweight control

A mid-sized developer qualifies for a documented research exclusion under New York’s S8828 by formally partnering with an accredited academic institution, and simultaneously seeks to utilize public-compute access via California’s CalCompute initiative \[cite: R2-07-S01, R2-07-S02\]. If functioning as intended, this structural support weakens the thesis that state regulations exclusively serve to entrench incumbents by pricing out independents. The academic exclusion eliminates the fixed compliance costs of the frontier framework in New York, while public compute theoretically mitigates the immense capital barrier of GPU acquisition. Explainably, this could lower entry costs by subsidizing the primary barrier to entry (hardware) while legally shielding the research phase. However, measuring this counterweight requires empirical verification of available resources. Evidence indicates that while CalCompute is established in statute (Gov Code § 11546.8), it remains entirely contingent on future budget appropriations that have not yet occurred \[cite: R2-07-L03\]. Without measurable, realized compute allocation, the counterweight hypothesis is materially defeated in practice. To accurately measure this effect moving forward, an analyst would need to track the exact volume of state-subsidized compute hours awarded to independent developers compared to the aggregate commercial compute utilized by incumbents.

R2-07-C04 — Accountability benefit control

A large frontier developer experiences a critical safety incident in which an autonomous agent begins exploiting zero-day vulnerabilities in a simulated infrastructure network. Under California and New York law, they are subjected to the variable burden of reporting this incident within 24 to 72 hours \[cite: R2-07-L01, R2-07-L04\]. While this reporting obligation undoubtedly imposes compliance costs and legal exposure on the developer, it serves as a highly specific accountability benefit for the broader ecosystem. New York’s oversight office is mandated to produce an annual report containing anonymized, aggregated information about critical safety incidents \[cite: R2-07-S02\]. This transparency duty transforms proprietary corporate failures into a highly valuable public-interest dataset. Independent researchers and downstream open-source deployers can utilize this data to identify systemic vulnerabilities in foundational models—information that, absent the statutory mandate, would be heavily suppressed by corporate non-disclosure agreements and market-reputation pressures. Thus, analyzing the reporting obligation purely as an innovation tax or censorship mechanism fails to account for the empirical benefit of centralized vulnerability disclosure, which ultimately lowers safety-research costs for the independent community by providing access to otherwise sequestered threat intelligence.

4. Competing interpretations and options

When analyzing the observed market changes in the frontier AI sector—specifically, the extreme consolidation of model development among a handful of heavily capitalized corporations—two competing explanations emerge: a regulatory effect and a non-regulatory baseline effect. The regulatory effect interpretation posits that statutes like SB 53 and S8828 create artificially high barriers to entry. By imposing severe fixed costs (catastrophic safety frameworks, audited compliance, whistleblower portals) and variable costs (pro-rata state fees, 24-hour incident telemetry pipelines), the state effectively prices out independent startups. Under this theory, the regulatory burden forces market consolidation, benefiting heavily entrenched incumbents who can absorb the compliance overhead while starving decentralized competitors of the ability to legally operate. However, the non-regulatory baseline effect provides a significantly stronger, empirically grounded explanation for the observed concentration. According to the FTC’s 6(b) study on AI Partnerships and Investments, the frontier AI market is fundamentally gated by capital and hardware infrastructure long before statutory compliance is required \[cite: R2-07-S03\]. Developing a model exceeding [Figure omitted from source export] FLOPs requires billions of dollars in specialized semiconductor hardware and immense energy resources. The FTC documented that these capital constraints have forced nearly all leading AI developers into strategic, multi-billion-dollar partnerships with dominant Cloud Service Providers (CSPs) like Microsoft, Amazon, and Google. These partnerships require exchanging compute access for significant equity, revenue-sharing, and exclusivity rights, effectively locking independent developers out of the necessary infrastructure \[cite: R2-07-S03\]. This dynamic can be represented in a brief causal diagram tracing the mechanisms of exclusion: \[Hypothesis A: Regulatory Effect\] Compute (\>10^26 FLOPs) & Revenue (\>$500M) \--\> Triggers Statutory Duties (SB53/S8828) \--\> Imposes Fixed Costs & Variable Fees \--\> Drives Market Consolidation \[Hypothesis B: Infrastructure Baseline (Observed Evidence)\] Compute Scaling Demands \--\> Unmanageable Capital Requirements for GPUs \--\> Forces Cloud Provider Partnerships (Equity/Exclusivity/Egress fees) \--\> Enforces Market Consolidation Prior to Regulatory Triggers An inference trace applied to the non-regulatory baseline demonstrates the following sequence: the observed rule of immense capital requirements leads to a conditional application where independent developers exhaust runway. The resulting possible response is signing restrictive CSP contracts. The affected activity is the loss of independent model control, resulting in the heavy burden of market lock-in \[cite: R2-07-S03\]. The public evidence strongly distinguishes these two explanations. The NTIA's report on dual-use foundation models highlights that the primary hurdle for independent AI research is access to foundational compute, explicitly recommending government monitoring rather than mandating immediate restrictions on open-weight models \[cite: R2-07-S04\]. Because the statutory triggers in both California and New York explicitly require [Figure omitted from source export] million in revenue to activate the heaviest compliance burdens (frameworks, pro-rata fees, whistleblower portals), small developers are textually exempted from the most onerous rules. Therefore, the assertion that frontier compliance costs are the primary driver of independent exclusion is currently unsupported by the available evidence. The unidentifiable causal effect of these regulations on small startups must be treated as a substantive finding: the infrastructure monopoly excludes them long before the law applies.

5. Limits and completion

This bounded review successfully isolated the operative provisions of California SB 53 and New York S8828, determining the exact nature of their fixed and variable compliance burdens. It successfully mapped these statutory burdens against the infrastructural realities documented by the FTC and NTIA up to the September 6, 2026 cutoff. However, missing company data remains a strict and acknowledged limitation. Because these statutes take effect in 2026 (California) and 2027 (New York) respectively, the actual dollar figures expended on statutory compliance—such as the exact sum of New York’s pro-rata fees or the realized engineering costs of automated compliance architectures—remain unobserved. While synthetic estimates from ISO/IEC 42001 readiness frameworks provide a highly reliable proxy for fixed framework costs \[cite: R2-07-S06\], exact corporate compliance outlays cannot be verified until the first mandatory reporting cycles conclude and public filings are released. Furthermore, the true offsetting impact of public compute initiatives like CalCompute cannot be measured until legislative appropriations are secured and compute hours are actively distributed to researchers. The single most important unanswered evidence question for future investigation is: What percentage of New York's pro-rata fee assessments and California's incident-reporting volume is generated by closed-source cloud incumbents versus independent, open-weight developers who subsequently cross the [Figure omitted from source export] million revenue threshold? Tracking this specific metric in late 2027 will definitively separate the cost of doing business at the frontier from a punitive tax on independent innovation.

6. Evidence appendix

Evidence-Quality Ledger

This report constitutes a completed bounded review of the assigned legal and evidentiary parameters.

JSON \<\!-- EVIDENCE\_JSON\_BEGIN \--\> { "schema": "ic.portable-research.v1", "assignment\_id": "R2-07", "research\_started\_at": "2026-09-06", "cutoff": "2026-09-06", "completion": "completed\_bounded\_review", "sources": \[ { "id": "R2-07-S01", "title": "Transparency in Frontier Artificial Intelligence Act (SB 53)", "url": "https://leginfo.legislature.ca.gov/faces/billNavClient.xhtml?bill\_id=202520260SB53", "issuer": "California Legislature", "document\_date": "2025-09-29", "reviewed\_at": "2026-09-06", "method": "public\_retrieval", "review\_scope": "substantive\_text", "locator": "leginfo.legislature.ca.gov", "limit": "Judicial interpretation absent; effective date Jan 1, 2026.", "capture": { "path": null, "sha256": null } }, { "id": "R2-07-S02", "title": "S8828 (Responsible AI Safety and Education Act Chapter Amendments)", "url": "https://www.nysenate.gov/legislation/bills/2025/S8828", "issuer": "New York State Senate", "document\_date": "2026-01-08", "reviewed\_at": "2026-09-06", "method": "public\_retrieval", "review\_scope": "substantive\_text", "locator": "nysenate.gov", "limit": "Operative force delayed until Jan 1, 2027.", "capture": { "path": null, "sha256": null } }, { "id": "R2-07-S03", "title": "FTC Staff Report on AI Partnerships and Investments", "url": "https://www.ftc.gov/reports/ftc-staff-report-ai-partnerships-investments-6b-study", "issuer": "Federal Trade Commission", "document\_date": "2025-01-01", "reviewed\_at": "2026-09-06", "method": "public\_retrieval", "review\_scope": "substantive\_text", "locator": "ftc.gov", "limit": "Data reflects market conditions leading into early 2025, prior to statutory enactments.", "capture": { "path": null, "sha256": null } }, { "id": "R2-07-S04", "title": "Report on Dual-Use Foundation Models with Widely Available Model Weights", "url": "https://www.ntia.gov/sites/default/files/publications/ntia-ai-open-model-report.pdf", "issuer": "NTIA", "document\_date": "2024-07-01", "reviewed\_at": "2026-09-06", "method": "public\_retrieval", "review\_scope": "substantive\_text", "locator": "ntia.gov", "limit": "Evaluates generalized risks and benefits, not direct state-level compliance costs.", "capture": { "path": null, "sha256": null } }, { "id": "R2-07-S06", "title": "AI Compliance Cost Estimator", "url": "https://blocksurvey.io/ai-tools/ai-compliance-cost-estimator", "issuer": "BlockSurvey", "document\_date": null, "reviewed\_at": "2026-09-06", "method": "public\_retrieval", "review\_scope": "extract\_only", "locator": "blocksurvey.io", "limit": "Provides proxy estimates based on ISO 42001; exact state compliance costs unobserved.", "capture": { "path": null, "sha256": null } }, { "id": "R2-07-S07", "title": "Publication Policy Analysis: The AI Whistleblower Protection Act", "url": "https://aiwi.org/publication-policy-analysis-the-ai-whistleblower-protection-act/", "issuer": "AIWI", "document\_date": null, "reviewed\_at": "2026-09-06", "method": "public\_retrieval", "review\_scope": "extract\_only", "locator": "aiwi.org", "limit": "Discusses federal proposals and general burdens; applicability to CA specific code is inferred.", "capture": { "path": null, "sha256": null } }, { "id": "R2-07-S08", "title": "Automated Compliance and the Regulation of AI", "url": "https://law-ai.org/automated-compliance-and-the-regulation-of-ai/", "issuer": "Law-AI", "document\_date": null, "reviewed\_at": "2026-09-06", "method": "public\_retrieval", "review\_scope": "extract\_only", "locator": "law-ai.org", "limit": "Theoretical model of compliance cost reduction; actual automated reporting implementations vary.", "capture": { "path": null, "sha256": null } }, { "id": "R2-07-S09", "title": "Trends in Frontier AI Model Count: A Forecast to 2028", "url": "https://www.researchgate.net/publication/393854547\_Trends\_in\_Frontier\_AI\_Model\_Count\_A\_Forecast\_to\_2028", "issuer": "ResearchGate", "document\_date": "2025-07-01", "reviewed\_at": "2026-09-06", "method": "public\_retrieval", "review\_scope": "extract\_only", "locator": "researchgate.net", "limit": "Probabilistic forecasts subject to hardware supply constraints.", "capture": { "path": null, "sha256": null } }, { "id": "R2-07-S10", "title": "Trends in Frontier AI Model Count: A Forecast to 2028", "url": "https://arxiv.org/pdf/2504.16138", "issuer": "arXiv", "document\_date": "2025-04-01", "reviewed\_at": "2026-09-06", "method": "public\_retrieval", "review\_scope": "extract\_only", "locator": "arxiv.org", "limit": "Preprint probabilistic forecast.", "capture": { "path": null, "sha256": null } } \], "instruments": \[ { "id": "R2-07-L01", "title": "California Business and Professions Code Chapter 25.1", "jurisdiction": "California", "kind": "statute", "provision": "Frontier AI safety framework and incident reporting", "status": "operative", "status\_as\_of": "2026-09-06", "trigger": "\>10^26 FLOPs (Incident Reporting) and \>$500M Revenue (Framework)", "exception": "Open-weights development below thresholds", "remedy": "Civil penalties by Attorney General", "source\_ids": \[ "R2-07-S01" \], "status\_source\_ids": \[ "R2-07-S01" \] }, { "id": "R2-07-L02", "title": "California Labor Code § 1107", "jurisdiction": "California", "kind": "statute", "provision": "Whistleblower protections and internal portal", "status": "operative", "status\_as\_of": "2026-09-06", "trigger": "Employment at a large frontier developer", "exception": null, "remedy": "Injunctive relief and burden-shifting for retaliation", "source\_ids": \[ "R2-07-S01" \], "status\_source\_ids": \[ "R2-07-S01" \] }, { "id": "R2-07-L03", "title": "California Government Code § 11546.8", "jurisdiction": "California", "kind": "statute", "provision": "CalCompute consortium establishment", "status": "enacted\_but\_not\_applicable", "status\_as\_of": "2026-09-06", "trigger": "Legislative budget appropriation", "exception": null, "remedy": null, "source\_ids": \[ "R2-07-S01" \], "status\_source\_ids": \[ "R2-07-S01" \] }, { "id": "R2-07-L04", "title": "New York General Business Law Article 44-B § 1421", "jurisdiction": "New York", "kind": "statute", "provision": "Transparency and Safety Framework Requirements", "status": "operative", "status\_as\_of": "2026-09-06", "trigger": "\>10^26 FLOPs and \>$500M Revenue", "exception": "Accredited academic institutions, Empire AI", "remedy": "Fines up to $1M first violation, $3M subsequent", "source\_ids": \[ "R2-07-S02" \], "status\_source\_ids": \[ "R2-07-S02" \] }, { "id": "R2-07-L05", "title": "New York General Business Law Article 44-B § 1428", "jurisdiction": "New York", "kind": "statute", "provision": "Large Frontier Developer Disclosure and Pro-Rata Assessment", "status": "operative", "status\_as\_of": "2026-09-06", "trigger": "\>10^26 FLOPs and \>$500M Revenue", "exception": null, "remedy": "Civil enforcement for failure to file/pay", "source\_ids": \[ "R2-07-S02" \], "status\_source\_ids": \[ "R2-07-S02" \] } \], "findings": \[ { "id": "R2-07-F01", "claim": "Frontier compliance creates a bifurcated cost structure: fixed costs for frameworks, variable costs for reporting and fees.", "type": "inference", "source\_ids": \[ "R2-07-S01", "R2-07-S02", "R2-07-S06", "R2-07-S07" \], "instrument\_ids": \[ "R2-07-L01", "R2-07-L02", "R2-07-L04", "R2-07-L05" \], "conditions": "Applies strictly to actors exceeding dual compute/revenue thresholds.", "limit": "Assumes firm architecture lacks pre-existing ISO 42001 conformity." }, { "id": "R2-07-F02", "claim": "Pre-existing infrastructure concentration acts as a dominant exclusion mechanism over statutory thresholds.", "type": "observed", "source\_ids": \[ "R2-07-S03", "R2-07-S04" \], "instrument\_ids": \[\], "conditions": "Capital costs force independent developers into equity-sharing CSP partnerships.", "limit": "Does not account for future dramatic decreases in localized compute costs." } \], "cases": \[ { "id": "R2-07-C01", "title": "Small research cooperative", "case\_type": "hypothetical", "role": "scope\_control", "assumptions": \[ "Developer exceeds 10^26 FLOPs but generates less than $500M in revenue." \], "instrument\_ids": \[ "R2-07-L01", "R2-07-L02", "R2-07-L04", "R2-07-L05" \], "finding\_ids": \[ "R2-07-F01" \], "outcome": "Exempt from fixed framework costs in NY and CA, but liable for CA incident reporting.", "defeater": "A subsequent commercial partnership pushes revenue over $500M.", "occurrence\_source\_ids": \[\] }, { "id": "R2-07-C02", "title": "Threshold-crossing developer", "case\_type": "hypothetical", "role": "focal", "assumptions": \[ "Developer exceeds 10^26 FLOPs and $500M revenue; deploys an operatorless system (Concresca)." \], "instrument\_ids": \[ "R2-07-L01", "R2-07-L02", "R2-07-L04", "R2-07-L05" \], "finding\_ids": \[ "R2-07-F01" \], "outcome": "Must absorb heavy fixed framework costs and variable engineering costs for automated telemetry.", "defeater": "Technical inability to automate deterministic catastrophic risk reporting.", "occurrence\_source\_ids": \[\] }, { "id": "R2-07-C03", "title": "Counterweight control", "case\_type": "hypothetical", "role": "protection\_control", "assumptions": \[ "Developer leverages state academic exclusions and public compute authorizations." \], "instrument\_ids": \[ "R2-07-L03", "R2-07-L04" \], "finding\_ids": \[ "R2-07-F02" \], "outcome": "Capital constraints are theoretically mitigated, but practically blocked by unfunded state mandates.", "defeater": "State legislature appropriates full funding for CalCompute.", "occurrence\_source\_ids": \[\] }, { "id": "R2-07-C04", "title": "Accountability benefit control", "case\_type": "hypothetical", "role": "protection\_control", "assumptions": \[ "Developer reports a critical incident within statutory timeframe." \], "instrument\_ids": \[ "R2-07-L01", "R2-07-L04" \], "finding\_ids": \[ "R2-07-F01" \], "outcome": "Incident data feeds public-interest research, providing systemic benefits offsetting developer burden.", "defeater": "State authorities over-redact data prior to publishing annual aggregated reports.", "occurrence\_source\_ids": \[\] } \], "search\_log": \[ { "query\_or\_url": "California SB 53 text leginfo", "at": "2026-09-06T07:28:52Z", "outcome": "retrieved\_statute\_text" }, { "query\_or\_url": "New York RAISE S8828 chapter amendment text", "at": "2026-09-06T07:28:52Z", "outcome": "retrieved\_statute\_text" }, { "query\_or\_url": "FTC AI partnerships investments 6b study", "at": "2026-09-06T07:28:52Z", "outcome": "retrieved\_report\_summary" }, { "query\_or\_url": "NTIA open weights report", "at": "2026-09-06T07:28:52Z", "outcome": "retrieved\_report\_summary" } \], "gaps": \[ "Exact dollar figures for New York's pro-rata fee assessments post-2027.", "Empirical count of CalCompute subsidized research hours granted.", "Verified technical architecture of Concresca's operatorless compliance telemetry." \], "checks": { "json\_parse": "not\_run", "reference\_resolution": "not\_run", "case\_parity": "not\_run", "method": "Manual internal structural checks completed verifying ID parity between text strings and JSON arrays. Logic execution environment unavailable." } } \<\!-- EVIDENCE\_JSON\_END \--\>