
AI Safety, Frontier Development Pace, and Regulation
Development pace and competition
Measuring frontier development pace
Five clocks · Training is not release · Reach is not capability
Separate the clocks
The phrase “frontier pace” compresses at least five measurable processes: the scale and frequency of training, improvement on capability evaluations, changes made after initial training, changes in access, and uptake in real workflows. Each has its own dates and units. An update may increase a product’s reach without changing its underlying model. A research checkpoint may advance capability without a public launch. A well-designed assessment tracks each clock separately and tests whether they are moving together. [32]
Use a single event ledger with distinct fields for training or evaluation date, publication date, release scope, and deployment date. Record which dates are estimates or unknown. Count a hosted preview separately from generally available access and separately again from the distribution of model weights. This prevents changes in publicity or packaging from appearing to be technological acceleration.
Read pace alongside control capacity
The governance question is whether measurement, independent review, and intervention keep up with changes in capability and exposure. An operational indicator is the fraction of material changes that were independently reviewed before a permission expansion. A related indicator is elapsed time between a credible adverse signal and a decision by someone authorized to limit access. Both require timestamps and a clear denominator; neither can be inferred from public release calendars alone. [50]
The report should present four different evidentiary levels: observed benchmark results, estimated sector trends, stated corporate practices, and scenario assumptions. None can replace another. A financial commitment can support an argument about capacity; it cannot quantify the likelihood of a severe incident. A safety policy shows what a company says it does, not what an independent reviewer found. Any claim that oversight is keeping up must be specific about the workflow and the evidence.
Counting releases without overstating acceleration
119 → 98 → 95 · One database snapshot · Not all checkpoints
What the release series shows
Stanford HAI’s 2026 AI Index, using an Epoch AI snapshot, shows notable-model access-classification totals of 119 in 2023, 98 in 2024, and 95 in 2025. Over those same three years, the API access category rose from 20 to 30 to 45. That is an example of why total counts and access channels can move in different directions. A drop from 119 to 95 is not evidence of slower capability improvement; changes in which systems qualify or are documented can also change the count. [32]
Do not join old annual report tables when a single common-vintage series is available. The database is maintained retrospectively, and coverage can change for historical years. The chart labels in the 2026 release cover a classification view; its even-year milestones are 33 in 2014, 49 in 2016, 37 in 2018, 50 in 2020, 91 in 2022, and 98 in 2024. [32]
Use a complementary event measure
An organization evaluating supplier exposure should maintain its own inventory. Count major model transitions, new capabilities or tool permissions, and updates to default settings that reach meaningful shares of its own users. Distinguish an announcement from a change that actually entered the production workflow. A persistent model identifier may be more useful than a product brand; the same brand can cover more than one configuration.
Publish separate rows for product releases and safety-review events. A company that uses calendar time as its primary output may miss concentrated change windows when multiple products share a single upgraded model. Flag changes that are hard to reverse, and compare observed deployment pace with review and rollback capacity rather than with the number of press releases. [50]
Concentration by sector and geography
87 industry models · National affiliations may overlap · Not a market-share measure
Distinguish modeling capacity from geography
In the same HAI/Epoch snapshot, 87 of 95 notable models were categorized as industry, five as industry–academia collaboration, two as other, and one as academia. The industry count corresponds to about 91.6% of the table. This illustrates who is publicly represented at the frontier, not where all relevant expertise resides. Universities can produce tooling, evaluations, or foundational ideas without having a standalone model counted in a notable-model dataset. [32]
The report lists 50 notable models with U.S. affiliation, 30 with Chinese affiliation, and five with South Korean affiliation in 2025. Country attribution is based on the institutions of introducing-paper authors; multinational contributions can be counted in multiple places. Comparing this with sales share, sovereign control over compute, or local deployment would require separate datasets. [32]
Implications for assurance capacity
A highly concentrated frontier can make a limited number of release policies and audit practices consequential across much wider markets. Concentration may make coordination and common reporting easier. It may also put regulators and academic evaluators in a weaker position to verify claims if important artifacts are accessible only through a partner. Neither effect follows mechanically from the counts; it is a governance hypothesis that requires direct observation.
A reproducible dashboard reports model count by snapshot date; the entity and affiliation definition; and separate measures of procurer dependence, external-assessor access, and evaluator coverage. Treat missing independent evidence as a known gap instead of interpreting absence as safety. If a comparison aggregates corporate affiliates or research subsidiaries, document the mapping. A benchmark project counts organizations differently from a loan covenant, a merger filing, or a regulator’s covered-entity definition. [32] [48]
Training compute and what it measures
Computation is an input · Historical model fit · Assumptions matter
Use estimates as estimates
Epoch AI’s 2024 analysis estimates that training compute for its notable-model sample grew about 4.1-fold per year from 2010 through May 2024, with a 90% interval of 3.7–4.6. Its analysis of the more recent frontier-language-model segment gives a central estimate of 5.0-fold, with a broader interval of 3.1–7.3. Those analyses use different filters and windows. Do not average the two slopes or convert either one into a likelihood of catastrophic harm. [33]
Compute estimates are often indirect. Published inputs may include selected architecture, hardware, or training information rather than an audited count. Methods can assign different amounts of work to a training run, and not every model’s starting data are equally complete. Any figure using estimates should label its source, sample window, and uncertainty. [34]
Use a screening proxy carefully
A computational threshold may help a reviewer decide when to request an assessment because compute can be estimated at a large scale even when capability is hard to predict. That same strength makes it a coarse proxy. Better data, algorithms, or tools can improve practical performance without proportional increases in recorded training compute. Conversely, a large retraining or specialist run does not guarantee the emergence of the capability a rule was intended to detect.
A robust governance system pairs resource-based triggers with capability-based triggers and material changes to access. A provider crossing one trigger supplies evidence; absence of a trigger need not override a credible adverse observation. Evaluate whether the proxy still catches deployments that merit scrutiny as the frontier changes. Record why an exception was approved and which signals would cause the threshold or its scope to be reconsidered. [50] [33]
Inference and the system boundary
Checkpoint plus configuration · Same model, different exposure · Version the application
Evaluate more than training
Production behavior arises from a configured system rather than from a checkpoint alone. Request-time resources, additional context, human approval, and application permissions can change what a system can do and what happens when it fails. A configuration with no authority to alter external records is not equivalent to one that can commit consequential actions. Neither a training compute estimate nor a generic model card measures that difference. [50]
A standard evidence packet should inventory model version, deployment variant, access class, scope of permissions, allowed data, human supervision, and what kind of change could invalidate the assessment. Record whether the reviewer saw the same configuration as users. If a reviewer tested a substitute, explain the difference instead of calling the result a production test.
Governance should follow material change
A policy limited to training events leaves a gap when autonomy or authority increases after initial release. A useful control triggers a new assessment when there is a significant change to the model, rights, context, reachable population, or effectiveness of safeguards. The decision does not depend on whether marketing calls that change a new product.
A baseline system can remain available while a more consequential configuration is still in evaluation. This permits evidence collection without assuming the only choices are a blanket stop or unrestricted rollout. The guardrail is evidence of separation: establish exactly which configuration was authorized, show that permissions enforce it, retain logs that can support review, and identify who can limit the change on short notice. These are proposed operating controls, not assertions about any named developer’s current practices. [50]
What a task-completion horizon means
Human completion time · Not runtime · 50% is not production readiness
Read the unit before reading the slope
METR’s task-completion horizon describes the task duration, measured using human expert completion time, at which an evaluated agent is estimated to achieve a specified success probability on a test suite. It is not the duration for which the agent operates, nor a promise that the agent can perform any task of that length. METR’s original March 2025 analysis reported approximately seven-month doubling over its historic sample; it was a fitted retrospective trend, not a rule about how future models must behave. [35] [36]
A 50% point is useful because it allows comparison over tasks of differing duration. It is not an acceptable error rate for high-consequence activities. The report should show the probability threshold and confidence interval beside any horizon estimate and include the task mix in the caption. Operational judgment also requires explicit costs for failures and a mechanism for detection and remediation.
Avoid extrapolation across domains
The relationship between human completion time and difficulty can vary with the task family. Human reference durations are estimates. Tooling and evaluation setup affect success. Even a median improvement can coexist with poor reliability in an important subset, and the frontier estimate may depend on only a few strong systems. METR separately discusses how horizon comparisons vary across domains. [36] [38]
A reasonable governance interpretation is to revisit the effectiveness of oversight as capability estimates change, not to infer that the system is safe or unsafe from one point. Before applying the metric, identify which tasks the benchmark excludes and whether the reviewed workflow includes interactions that could invalidate the fit. Then examine production-specific false approvals and consequential errors. The metric provides evidence of bounded capability under test; it is not an incident-rate predictor.
Benchmark sensitivity and uncertainty
Same model, different suite · Use intervals · No leaderboard claim
A dated comparison
METR’s January 2026 Time Horizon 1.1 publication reports 50% horizon estimates of 60 minutes for [*REDACTED*], 101 for [*REDACTED*], 121 for o3, 214 for [*REDACTED*], and 320 for [*REDACTED*]. The reported intervals were 32–106, 58–170, 74–201, 117–480, and 170–729 minutes respectively. These examples share the revised benchmark snapshot and should not appear as present-day rankings. Overlapping intervals make a statement of precise rank especially inappropriate. [37]
Version sensitivity is material. The same publication gives [*REDACTED*] an estimate of 138 minutes in Time Horizon 1.0 and 214 minutes in 1.1. A changed estimate of the same model is evidence that the measurement changed, not evidence that the checkpoint independently improved.
Treat evaluation metadata as part of the result
Retain the test suite version, model version when available, evaluation date, definition of success, amount of human intervention, number of trials, and uncertainty method. Label cases in which important evaluation parameters are unknown. Put a review boundary around any claim that a model “exceeds” another: the difference must be meaningful in the tasks, population, and harm model being assessed. [37] [50]
Compare two providers using matched test conditions where feasible and note any respect in which conditions do not match. Favor holdout cases independent of optimization and sample across difficulty levels. A positive mathematical trend does not resolve weak measurement at the frontier. The more consequential the decision, the greater the need for direct evidence, sensitivity checks, and validation of the way findings are translated into operating rules.
Electricity scenarios and infrastructure lead times
415 TWh in 2024 · 485 TWh in 2025 · 950 TWh 2030 base case
Label the electricity series correctly
The IEA’s 2025 report estimated 415 TWh of global data-center electricity consumption in 2024. Its 2026 update estimates about 485 TWh in 2025 and projects approximately 950 TWh in its 2030 Base Case. The 2030 value is a projection, not a measurement; the earlier years are estimates. This series describes data centers generally, not frontier training alone or even exclusively AI workloads. The later 2030 projection supersedes the roughly 945 TWh base-case figure in the earlier assessment when discussing the updated outlook. [45] [46]
A clean chart shows 2024 and 2025 as estimated anchors and 2030 as a distinct scenario marker. It should not imply a measured path for 2026–29. The figures do not describe the electricity mix in a specific region or establish incremental demand from one particular training program.
Infrastructure is a constraint, not a schedule
Capacity announcements do not equal electricity draw. Between the two are permitting, construction, energization, deployment, and utilization. Lead times vary for generation, transmission, and equipment. A company can change how it allocates existing capacity even while new power is delayed. Infrastructure can slow one part of the frontier without imposing a uniform limit on product deployment or capability research. [46]
When assessing a policy by its effect on available compute, explain where in the chain the policy acts. A project-level interconnection decision has different effects from a financing constraint or a deployment restriction. Keep the assumptions explicit: a projection is useful for planning if it is treated as a scenario and updated when adoption, efficiency, or energy-supply conditions change.
Capital commitments and deployment incentives
Commitments are not cash spent · Time horizons matter · SEC filing, November 2026
What one filing actually says
[*REDACTED*]’s November 17, 2026 amended registration statement describes approximately $900.9 billion in aggregate committed spending under purchase commitments and lease arrangements as of and subsequent to September 30, 2026. The filing indicates that approximately $451.6 billion was expected to be incurred before December 31, 2030, and substantially all of the remainder between 2031 and 2045. This is a forward-looking contractual and accounting disclosure, not a measure of realized spending, the cost of training a single model, or the size of the market. [47]
The figure rounds contracted amounts and includes agreements with different structures and timelines. The filing separately discusses contract types and risks. It is not appropriate to add components from another table without checking scope; some measures overlap and some omit committed items that have not begun.
The governance inference and its limits
Long-duration infrastructure commitments can increase pressure to use capacity and meet customer demand. They can also improve planning stability and create room for experimentation, research, and assurance. The filing does not establish which effect dominates and cannot establish that a particular release was rushed. A governance claim requires evidence of what the decision makers knew, what alternatives were available, and whether a credible process could override the commercial calendar.
An appropriate dashboard separates forecast cash obligations from reported expenses and matches both to measured review coverage and incident response. Do not convert procurement commitments into an estimate of likelihood or severity of model failure. The useful question is whether the organization can maintain evaluation, intervention, and support even if commercial conditions change. [47] [50]
Cloud partnerships and bargaining power
FTC study of three partnerships · Potential effects · Not a liability finding
Why cloud agreements matter
The FTC’s January 2025 staff study examined partnerships between major cloud providers and frontier developers. The staff described equity and revenue-sharing arrangements, rights that vary by agreement, cloud-spend commitments, and exchanges of resources and information. It identified potential effects on access to inputs, switching costs, and information asymmetry. The public report is a historical snapshot of three specific partnerships, not a market-wide finding, antitrust decision, or current inventory of contracts. [48]
Commercial integration can finance development that would otherwise be harder to deliver. Technical and contractual coupling can also make it difficult for a developer to move, and a single integration may propagate a model change to several downstream products. For risk governance, inspect access and decision dependencies as well as financial interests.
Build a dependency map before designing controls
Represent each organization, cloud dependency, and major downstream integration as a node. Annotate what it supplies, who controls changes, what evidence can be independently inspected, and where a contingency can be activated. Distinguish contracts guaranteeing access from public assumptions about future capacity. A single vulnerability or common dependency can matter even if every downstream application is labeled a separate product.
Board reporting can track the share of critical workloads without a validated alternative and the time needed to change providers or return to a less capable system. Restrict sensitive information to appropriate reviewers while ensuring that reviewers have enough access to assess the decision. Governance controls that rely on a partner’s cooperation should include a test of that assumption rather than treating “third party” as proof of independence. [48] [50]
Incentives when release dates are competitive
Incentive model, not motive · Cost of delay · Shared safety floor
How timing pressure can arise
A company deciding whether to defer a launch may compare a private commercial loss against a benefit enjoyed by a wider public. When a competitor can capture users or distribution during the delay and no common floor exists, that private calculus may favor faster deployment. This is an incentive model, not evidence of anyone’s intentions or a finding that the model applies to a specific decision. Concentrated infrastructure commitments and partnership structures give practical reasons to examine the incentives rather than assume them. [48] [47]
There are countervailing incentives. Reliable systems may win enterprise customers; a visible failure can create costs, remediation obligations, or loss of trust. Development speed and deployment speed can also diverge. A provider may pursue research rapidly while imposing a slower process for permitting wider access.
Design against override pressure
A documented approval threshold is more credible if the people who check it do not report exclusively to the people who benefit from release timing. Set independent escalation channels and define which evidence is required for an exception, including a dated expiration. Record the dissent even if the authorized decision maker disagrees. Never infer an absence of dissent from an absence of published notes.
A shared floor for independent assessment, disclosure, and incident escalation can reduce incentives to underinvest in diligence without requiring identical products or designs. Such a floor is effective only if the covered decisions and enforcement mechanisms are observable. If a proposed requirement addresses calendar speed but not deployment permissions or production oversight, it may constrain the activity that is easiest to count rather than the activity most relevant to potential harm. [50]
Release channels and reversibility
API access is not open weights · Rollback reach varies · Tradeoffs are not rankings
Different releases expose different levers
API and hosted access can preserve provider-side rate limits, monitoring, model replacement, and account controls. That offers tools for intervention but requires trust in the provider and does not guarantee that its controls cover every use. Distribution of model weights can support independent research, reproducibility, competition, and local deployment, while limiting the original provider’s ability to withdraw redistributed copies. “Open weights” describes the access to parameters, not necessarily access to data, source code, or unrestricted rights. [49] [32]
A deployment decision should state the chosen release path and why alternatives were rejected. The answer should distinguish access for independent assessors, controlled pilots, general commercial users, and permanent redistribution. Each arrangement serves different goals and presents different questions about evidence and control.
Design release gates around reversibility
Use a release matrix with permissions, intended use, implementation owner, withdrawal mechanism, and time to effective rollback. For irreversible distribution, the relevant gate comes before release; later telemetry may inform policy, but does not approximate the ability to recall all copies. For a hosted feature, test the advertised shutdown or reversion path, including integrations that may rely on it.
Policy debates that equate limited public disclosure with safety overlook the benefits of meaningful outside scrutiny. Conversely, debates asserting that openness automatically creates assurance overlook the difference between reproducibility and accountability. The operational aim is to make reviewer access sufficient to verify claims and to match wider distribution to evidence about foreseeable exposure. The 2024 NTIA report is historical policy analysis, not a statement of present law. [49] [50]
[*REDACTED*]’s published governance layers
Separate Preparedness from governance summary · Company disclosure · Version before comparison
What the publications establish
[*REDACTED*]’s May 2026 Frontier Governance Framework announcement describes how portions of the company’s approach align with regulatory requirements and identifies the Preparedness Framework as the foundation for how it manages the most serious advanced-AI risks. The announcement distinguishes a regulation-focused public governance document from the company’s broader underlying internal practices. An April 2025 Preparedness update offers a dated view of how the process was being developed at that time. [39] [40]
The two publications are company accounts. They provide evidence about what the company disclosed, not independent evidence that a safeguard worked or a legal obligation was satisfied. A comparison should cite the exact document version and date rather than impute a generic label to every checkpoint or rollout.
Questions for auditors and boards
Ask which teams or roles are responsible for detecting new capabilities, reviewing residual risk, authorizing access, and monitoring changes after release. Request the decision record where possible, not merely the policy statement. A useful record ties the evidence to the configured system, lists unknowns that matter, and identifies the trigger for renewed review.
Distinguish a prospectively dated plan from a completed assurance activity. [*REDACTED*]’s November 16, 2026 announcement states that agreements with four external assessors were signed for a first embedded-assessment phase focused on the company’s largest frontier reinforcement-learning run. That is evidence of a stated arrangement and intended scope; it is not a published assessor verdict. Expectations about visibility, independence, and access should be checked against reports and actual delivery. [41]
[*REDACTED*]’s Responsible Scaling Policy
Version 3.4, July 8, 2026 · Risk reports · Distinguish transparency from enforcement
A version-control case
[*REDACTED*]’s public change log lists Responsible Scaling Policy version 3.0 effective February 24, 2026 and version 3.4 effective July 8, 2026. The log identifies formal changes across versions, including changes in reviewer arrangements, internal risk-report distribution, coverage dates, and redaction indicators. Analysts should cite the text in force at the date of a specific decision rather than the version most convenient to quote. [42]
The April 29 v3.2 entry states that the company’s Long-Term Benefit Trust may request external review of Risk Reports and approve selection of reviewers. The July 8 v3.4 entry describes the conditions under which separate external reviewers can cover different unredacted sections. Coverage of each section does not necessarily establish that one assessor saw the interactions between all sections. [42]
A review should examine decision rights
Compare what the company says must happen with the authority of each actor to demand evidence, delay a rollout, or publish disagreement. A governance presentation can highlight the existence of an independent body without settling what that body can do under specific facts. Consider information rights, appointment and removal, disclosure timing, and what happens when a reviewer concludes a decision should change.
“External review” should be reported in terms of scope and access. A reviewer of a public summary is not equivalent to an assessor with the ability to independently rerun a test, document uncertainty, or examine the whole system. Redactions can protect legitimate security and privacy interests; assessments should still state enough about their limitations for readers to distinguish verified conclusions from conclusions outside their scope. [42] [50]
[*REDACTED*]’s Frontier Safety Framework
Version 3.1, April 17, 2026 · Earlier detection · Not direct label parity
Separate an early signal from a critical threshold
[*REDACTED*]’s public frontier safety page lists Frontier Safety Framework version 3.1 on April 17, 2026 after version 3.0 on September 22, 2025. The accompanying update describes Tracked Capability Levels for certain domains, intended to support earlier recognition and assessment of less extreme capability signals. The addition illustrates an important design distinction: the threshold to intensify monitoring need not be identical to the threshold requiring stronger restrictions. [43] [44]
An early-signal approach is useful only when someone owns the follow-up and documents how evidence shifts the level of assurance. Tracking without a decision process can lead to early warnings that are noticed but not resolved. Equally, a low-sensitivity threshold can generate administrative work without improving decisions.
Compare frameworks on what they require
Different organizations use different names for triggers and categories. Do not map labels to each other by name alone. Compare the evaluated system and domain, the source of evidence, the stage at which a test is required, the approval authority, and the action required after an adverse result. Frameworks can differ materially even where both publish thresholds.
For external readers, a change log helps establish what was promised when. An independent paper should still say whether tests, results, review scope, or incidents were separately verified. If only the policy was observed, report it as the policy and identify remaining uncertainty rather than filling gaps with general assumptions about industry practice. [43] [50]
Building a common disclosure matrix
Compare evidence, not branding · Preserve missing data · Version every entry
Proposed matrix fields
For each company, enter legal or organizational entity, document title and effective version, release or deployment under review, relevant jurisdictions, evidence owner, evaluator independence and access, evaluation date, material unknowns, decision authority, incident-notification routes, and rollback authority. Separate mandatory obligations from voluntary policy. If the record does not disclose a field, mark it unknown rather than writing that the practice does not exist. [39] [42] [43]
Use two evidence columns: what the institution says about its process and what an independent record shows. If a company changed its framework during the period, preserve both rows rather than overwriting the record. A report published after a launch may document prior work, but timing alone does not tell the reader when the evidence was actually produced.
Strong claims require matched scopes
A judgment that one provider has stronger safeguards requires comparable evidence about implementation and outcomes. Public documentation can support a narrower claim about disclosure specificity, not an overall safety ranking. An assessor may see one model, one checkpoint, or one system configuration. An explanation of test design should accompany the scope rather than be deferred to marketing language.
When public detail is limited for safety or commercial reasons, reviewers can still publish a minimal assurance packet: evidence classes examined, the direction of major findings, important exclusions, date, and who had authority to act. The public packet is not a substitute for appropriately governed fuller access; it explains the strength and limits of public inference. [50] [41]
Embedded evaluation and independence
Access does not guarantee authority · Four assessors announced · Outcomes not yet established
Why embedded review can add evidence
[*REDACTED*] announced on November 16, 2026 that it signed agreements with Apollo Research, METR, Transluce, and Goodfire for its first phase of embedded assessments. It described potentially deep access to internal resources according to assessment needs and stated that the Safety and Security Committee would have visibility during the assessment. This is a first-party statement about the phase, not an audit opinion, a finding of absence of risk, or proof that the relevant access was actually furnished. [41]
Embedding can address the limits of one-time pre-release testing where reviewers only see outward-facing behavior. It can introduce dependence on the provider for access, interpretation, or logistics. Independence therefore turns on the contractual and practical conditions of the work and on how relevant negative findings are handled.
Evidence a reader should expect
A high-quality public report would distinguish the object examined, what access was granted, what could not be reviewed, dates, conflicts or funding arrangements, technical limits, and whether publication approval is restricted. Separate evidence that an assessment occurred from evidence that it found a mitigation sufficient. Require enough specificity to judge relevance without revealing sensitive information.
An organization evaluating an assessor should check practical separation of duties, reproducibility of its own observations, scope and data access, and the reviewer’s ability to document disagreements. The governance design can allow protected disclosure to an authorized public body while preserving a public methodological summary. An assessor’s recommendation does not replace the responsible party’s authority to postpone or constrain deployment. [41] [50]
From testing to limited release
Proposed operating policy · Pilot is not approval · Define rollback first
Gate the permissions, not the press release
Before research or deployment changes expand risk exposure, designate an owner, reviewer, and decision authority. Set guardrails, define what the system may do, and state the evidence required to proceed. In a test phase, use cases and evaluation procedures suited to the purpose and separate ordinary errors from critical errors. Keep consequential actions outside the testing environment unless explicitly authorized. [50]
A limited release has a cap, a scope, an expiration, and access short of broader production permissions. Monitor relevant outputs and failures, including indicators that the assumptions of the original evaluation no longer hold. A pilot should provide data to make a later decision; duration or completion of a pilot is not itself evidence of safety.
A wireframe decision sequence
Record proposed change → identify material scope or permission change → collect evidence and independent review → approve limited, reject, or request more evidence → monitor and reapprove, pause, or roll back. For each transition, identify which role can authorize it and where the decision is logged. If the organization cannot demonstrate that the affected system can be constrained or withdrawn in the proposed release channel, the decision packet should state the limits and account for them before wider access.
The performance goal is not to avoid every reversible experiment. It is to avoid expanding authority or user exposure faster than the organization can observe and respond. The same process should handle a model version change and changes made by a system integrator, so permissions are not inadvertently broadened through a downstream deployment. [50]
Post-release monitoring and change control
Frequency × severity · Missing denominator · Incident and near miss
What to monitor
A useful monitoring program separates incident severity, near misses, reporting delays, and the scale of exposure. Denominators may include approved deployments, evaluated configurations, active users, or eligible decisions; none alone is adequate. Higher counts can reflect better reporting rather than worse operation. Absence of public events does not prove absence of problems. [50]
Changes in reporting scope can be as important as underlying changes in event rates. Record the definition of an event, how duplicates are treated, which channels report, what events are delayed, and whether reports are based on self-identification or independent checks. Compare periods only where scope and ascertainment are comparable.
Link alerts to decisions
Monitor whether the evaluation’s conditions still apply: model or application change, scope expansion, new categories of users, and changes in observability. Define which indicators force a review, which require human authorization, and which ultimately require pause or reversion. An alert that does not name an owner or deadline is easily overlooked. Preserve records for correction and audit while minimizing sensitive data.
Distinguish a rollback drill from an actual rollback, and confirm downstream behavior after an intervention. A service may be disabled while a secondary integration continues to use a different approved version. Document the residual scope instead of announcing an absolute shutdown if it was not achieved. Summaries should include mitigations that were tested and those merely planned, as well as the date of follow-up assessment. [50]
Design decisions when contemplating a pause
Object and duration · Verify exceptions · No claimed forecasts
Define the action precisely
“Pause” could mean a temporary stop in selected training activities, restrictions on external release, a halt to further permission expansion, or conditions for continuation. These actions affect different parts of the development cycle and differ in visibility, reversibility, and feasibility. A proposal should define the object, covered entities, trigger and duration, what remains allowed, decision rights, and evidence required for resumption.
A scoped hold can preserve independent, low-exposure evaluation while stopping a narrower class of changes that presents unresolved risk. A broader hold may be proposed if uncertainty is severe, but it needs enforcement and consideration of activity beyond the covered parties or jurisdiction. No estimate of precise reduction in severe harm should be inferred from the proposal alone.
Compare a pause with alternatives
Compare the proposed hold with tools such as independent safety cases, staged access, minimum evidentiary floors, and enforceable reporting. An options memo can ask what would be observed in 30, 90, and 180 days under each approach, but should present this as monitoring design, not predictions. Consider whether the measure constrains the highest-exposure deployment or merely the most visible action. [50]
Thresholds that apply only to new training runs can miss significant changes in deployment. A release-only rule can leave internal use exposures unaddressed. Conversely, blanket rules can cover unrelated low-exposure configurations that do not fit the underlying concern. Publish the intended tradeoff, the reviewed alternatives, and criteria for revising the intervention. Apply this framework to actual legal proposals using their precise text rather than treating an editorial description as law.
Scenario planning under different pace drivers
Scenario assumptions are explicit · No scenario probabilities · Review moments
Four bounded narratives
In a capacity-constrained scenario, power or cloud access delays some work while stronger configurations still diffuse through existing deployments. In a demand-driven scenario, integration and customer adoption accelerate faster than training performance. In an efficiency-led scenario, useful capabilities spread with less resource per task, weakening compute as a standalone proxy. In an assurance-led scenario, oversight capacity and stronger evaluation improve before exposure expands. These narratives are deliberately qualitative and may overlap; none is an estimate of what will happen. [46] [33] [50]
Avoid projecting any single observed slope indefinitely. A constraint can bind for a period and then change as infrastructure is brought online, financing changes, or methods shift. Select signals that would help distinguish scenarios, not just measurements that confirm one’s preferred view.
Decisions resilient across scenarios
Retain dated configuration records, grant independent reviewer access proportionate to consequence, practice reversibility where possible, and maintain clear channels for adverse signals to reach decision makers. These measures are useful whether demand grows faster than training or the reverse. Adapt review cadence to observed material changes rather than a fixed estimate of how fast every component evolves.
A board or public oversight body can define an update cycle using leading and lagging evidence. Leading evidence includes planned permission expansion and significant evaluation changes; lagging evidence includes verified incidents, time to remediate, and repeated failures. Report when sources are stale or non-comparable. A scenario that cannot be described in terms of what data would change the decision has little value for release governance. [50]
A frontier pace and assurance dashboard
Scorecard, not verdict · Comparable definitions · Named decision owner
Five views support a decision
First, report resource inputs and dates with estimates flagged. Second, show evidence on capability and reliability with test versions and thresholds. Third, record reach and permission changes in production. Fourth, show assurance coverage: which material changes were reviewed, external access scope, unresolved findings, and review timeliness. Fifth, report observed incidents, near misses, and response timing with exposure denominators. This proposed dashboard is an analytical tool and does not certify any deployment. [50] [32]
Avoid aggregating the views into a single index unless the weights and missing-data rules are explicit and defensible. A single favorable composite could hide a critical failure or a lack of independent evidence. A red outcome in a critical field must remain visible even if average performance improved.
How to read the resulting exhibits
The numeric exhibits in this chapter use dated sources and preserve units. Model counts describe a defined sample; a benchmark horizon is an estimated test point; electricity figures distinguish historical estimates from projections; company commitments are not realized industry spending. Together, these indicators show different forms of pace, not mutually reinforcing measurements of the same quantity. [32] [37] [46] [47]
Decide which indicator has authority to trigger additional scrutiny before treating a trend as a reason for action. Corroborate with a differently sourced record and retain the reasoning. The strongest public conclusion is often limited: one part of the frontier is changing quickly, and stated assurance rules should be matched to actual operating evidence. A judgment about any specific organization or program requires a dated, configuration-specific record.
Exhibit 03 · Notable models are not a release clock
Counts in the AI Index’s Epoch AI snapshot, not all releases or changes to access, and not a metric of safety or model performance.
Stanford HAI, 2026 AI Index, Chapter 1, Figure 1.1.8: primary source
Exhibit 04 · Private investment is concentrated
The regions are as reported by Stanford. Private investment excludes categories such as government-guided funding and does not measure total AI spending or safety expenditure.
Stanford HAI, 2026 AI Index, Economy: primary source
Exhibit 05 · Data-centre demand and its outlook
Estimates for 2024 and 2025, and the IEA central projection for 2030. These are all data centres, not AI-only electricity use; the projected value is not an observed outcome.
IEA, Energy and AI (2024 baseline) and Key Questions on Energy and AI (2026): primary source
Exhibit 06 · Task horizons have wide uncertainty
Selected historical estimates from METR Time Horizon 1.1, with the source’s reported intervals. Human task duration is not autonomous wall-clock runtime; results depend on the benchmark, configuration and fit.
METR, Time Horizon 1.1, January 29, 2026: primary source
Exhibit 07 · Access is a separate dimension
API-access category counts within the notable-model snapshot. They do not measure audience size, use of a model, privileges, or how quickly safeguards changed.
Stanford HAI, 2026 AI Index, Chapter 1, Figure 1.1.8: primary source
Browse: White Papers · Reports · Downloads
