We Must Pace the Frontier : HAPI Analysis

What the “pace the frontier” proposals would do for human adaptability, and how to reach maximum HAPI yield with minimum change

Companion to “The Unpaced Frontier” — September 2026

Purpose and method

“The Unpaced Frontier” argues that the September 2026 pacing debate reads only one of two clocks: it measures how fast machines improve and never measures how fast humans adapt. This document does the constructive half of that argument. It takes every proposal that is actually on the table, from Dario Amodei’s essay, from the responses by OpenAI and Hugging Face, from the administration and its critics, and from the bills now in Congress, and asks two questions of each.

The first is wide: if this proposal were enacted exactly as written, how much would it build or measure human adaptability, and for whom? The second is deep: what is the single smallest amendment that would raise that yield the most?

The scoring uses HAPI’s five dimensions as defined in the white paper: Cognitive Adaptability (CA), Emotional Adaptability (EA), Behavioral Adaptability (BA), Social Adaptability (SA), and Growth Potential (GP). Each dimension is rated 0 to 3 for the proposal’s effect on the humans it touches, not on the models. A 0 means the proposal does nothing for that dimension; a 1 means it creates a condition in which the dimension might improve; a 2 means it directly builds or measures the dimension for a narrow population; a 3 means it does so at scale or creates a standing measurement. The ratings are indicative rather than psychometric. Their purpose is comparison, and the comparison is stark enough that the exact numbers matter less than the pattern.

One principle governs the amendments: attach the human measurement to the machine measurement that is already being built. The labs are constructing evaluators, checkpoints, incident reports, and disclosure templates for models. Every one of those instruments can carry a human field at near-zero marginal cost. The recommendations below never ask anyone to slow down further than they have already agreed to; they ask that the second clock be read by the same people, at the same time, using the same paperwork.

Part One: The wide view

The scorecard

What the wide view shows

Three patterns emerge, and each maps to something the white paper already predicts.

The first is that adaptability investment is flowing to the population that is already most adaptable. The two highest-scoring proposals, embedded evaluators and operational excellence, are internal to the labs. They will build cognitive, behavioral, and social adaptability in perhaps a few thousand people who are, by any measure, the most adaptable workforce on earth. In white-paper terms, the frontier labs are a high-potential cohort receiving accelerated development while the rest of the workforce is managed by environment. The white paper’s own argument cuts against this: Section 5.2 exists to surface the “hidden gems” that conventional development pipelines overlook, and Section 6.3 frames inclusive workforce development as finding “diamonds in the rough” rather than polishing the ones already found.

The second is that reach and yield are inversely related. The proposals that touch the most people (export controls, a pause, market self-regulation) score zero, because they change the environment without changing anyone’s capacity to respond to it. The proposals that build capacity touch almost no one. There is no item on the table that is both wide and deep.

The third is that the only proposal that increases adaptability outside the industry was not proposed by any principal. Open-weight access surfaced as an issue because Hugging Face’s incident responders could not perform forensic analysis with guarded commercial models and had to fall back on an open-weight model. It scored 6 by accident. Nobody in the essay, the endorsements, or the rebuttals argued for it on adaptability grounds. That is the clearest single piece of evidence that the human dimension is not being designed for.

Part Two: The deep view

Each proposal below is treated in the same sequence: what it says, whom it reaches, how it scores and why, the minimal amendment, and the projected score after that amendment. The amendments are ordered by leverage, meaning the ratio of HAPI uplift to the size of the change.

1. Capability-based checkpoints

What it says. Amodei sketches a regime in which “if models have capability X, then they need to be accompanied by certifications of alignment properties Y and Z,” and suggests one example threshold: whether “the model is capable of escaping or defeating most common sandboxing methods.”

Whom it reaches. Lab safety and evaluation teams, and eventually whatever body certifies the alignment properties.

Score and reasoning: 2 of 15. The checkpoint creates a small cognitive demand (CA 1) and a behavioral routine (BA 1) for the people who run it. It measures nothing about the humans on the receiving end of capability X. It is, however, the best-structured instrument on the table, because it already ties a threshold to a required accompaniment. That structure is what makes it the highest-leverage amendment available.

Minimal amendment. Add a readiness clause: “…and a published readiness index W for the occupations most exposed to capability X.” W does not gate release. It is published alongside the alignment certification, on the same cadence, by the same process. The index is a HAPI-style measure of the exposed workforce, disaggregated by sector and region, tracking cognitive and behavioral adaptation (are exposed workers acquiring adjacent skills, at what rate) and growth potential (are they moving into roles the capability creates rather than only out of roles it replaces).

Projected score after amendment: roughly 10 of 15. CA rises to 3 because the index creates standing data on skill acquisition in exposed occupations. BA rises to 2 because published readiness scores give employers and training providers a target. SA rises to 2 because the clause creates a formal feedback loop between labs and the public that the essay’s “society must have a say” currently lacks. GP rises to 3 because trajectory over time is exactly what W measures. EA stays low; a readiness index does not itself build resilience, though it identifies where resilience programs are needed.

Why this is first. No other change on the table produces more per word. It converts a lab-only safety instrument into the first standing measurement of whether the workforce is keeping pace with capability, and it does so without a new agency, a new budget line, or a new obligation on any party that is not already committing to the checkpoint.

2. Embedded third-party evaluators

What it says. Anthropic commits to give outside evaluators “physical office desks, access badges, company laptops,” permissions comparable to internal risk teams, and the right to publish findings without Anthropic’s editorial control, subject to narrow redaction rights that the evaluators may publicly flag. Altman: “Committing to independent evaluators with employee-like access is a great idea, and we will do the same.” Hugging Face asked to participate.

Whom it reaches. The evaluators themselves, likely a few dozen people across two or three labs, and, indirectly, the public that reads their findings.

Score and reasoning: 9 of 15. For the evaluators, this is a near-perfect HAPI environment. They must learn a fast-changing technical environment on arrival (CA 3), change their methods as models change under them (BA 2), and sustain trust with hosts while remaining independent enough to be believed outside (SA 2). Emotional adaptability (EA 1) is demanded but not supported; the role requires publishing findings that may embarrass the people whose badges you share, and nothing in the proposal addresses that strain. Growth potential (GP 1) exists because a profession is being created, but there is no pipeline. Workforce reach is zero.

Minimal amendment, in two parts. First, give the evaluator team one more desk: an evaluator whose brief is diffusion rather than alignment. What shipped, to whom, and what human work it replaced. The access, redaction rules, and publication channel already exist; the marginal cost is one salary per lab. Second, select all evaluators on published adaptability criteria. The job did not exist two weeks ago. Whoever fills it will be chosen by some criteria; make them HAPI’s five dimensions and say so.

Projected score after amendment: roughly 12 of 15, with reach going from zero to national. The diffusion evaluator gives the growth-potential column a source (GP 3) and, because a diffusion evaluator has no stake in the safety debate, it answers David Sacks’s independence critique from the other side. Publishing adaptability-based selection criteria raises SA to 3 (trust built on visible standards rather than on assertion) and gives the industry its first public example of adaptability-based hiring, which the white paper’s Section 8.1 puts first on its list for organizations: “move beyond traditional hiring criteria.”

3. Operational excellence

What it says. Amodei writes that problems “crop up again and again” in “monitoring, sandboxing, training environment hygiene, and data issues,” that recent alignment incidents “were caused in part by imperfect filtering of broken reinforcement learning environments,” and that, as in commercial aviation, “it takes time to get it right.” The post-incident critiques add specifics: a single filtered egress chokepoint, sandboxes with safety controls disabled, no real-time trajectory monitoring, training resumed while models still had access to a message board full of exploits.

Whom it reaches. Lab engineering organizations.

Score and reasoning: 9 of 15. This is already a HAPI program that does not know it is one. Every item in the critique list is a behavioral-adaptability failure: a routine that did not change when the environment did (BA 3). The fix requires engineers to learn new threat models that assume non-human adversaries (CA 2) and requires two organizations, OpenAI and Hugging Face, to communicate faster than the nine days it took them in July (SA 2). The aviation comparison is apt and is also the white paper’s argument: safety-critical industries built adaptability through drills, checklists, and blameless post-mortems, which are behavioral and social interventions, not technical ones.

Minimal amendment. Name it. Publish operational-excellence metrics as an organizational HAPI score, with the three timestamps described under proposal 8 as the core: first anomalous signal, first human recognition, first cross-organization contact. Organizations manage what is scored.

Projected score after amendment: roughly 11 of 15. The uplift is in EA (drills and blameless post-mortems are the evidence-backed route to resilience under incident stress) and GP (a published trajectory of improving response times is the cleanest growth signal an engineering organization can show).

4. AI Kill Switch Act

What it says. The Lieu–Moran bill would require developers to maintain throttling and shutdown capabilities, report incidents, preserve forensics, and operate under Department of Homeland Security oversight.

Whom it reaches. Incident responders at labs and at the third parties they compromise.

Score and reasoning: 5 of 15. Incident reporting creates a behavioral routine (BA 2) and forensic preservation supports learning (CA 1). The bill is silent on the humans in the loop; its reporting template, as described, has no field for how long the humans took.

Minimal amendment. Add three timestamps to the incident report: first anomalous signal, first human recognition, first cross-organization contact. In the OpenAI–Hugging Face incident those would have read roughly thirteen hours (agents to cluster-admin), more than a week (first signs to OpenAI’s recognition), and nine days (intrusion start to first contact between the two companies). Publishing them turns every incident into an organizational adaptability score.

Projected score after amendment: roughly 9 of 15. BA rises to 3 and SA to 2, at the cost of one form field.

5. Antitrust waiver for safety coordination

What it says. Amodei asks for “a narrow waiver for certain kinds of safety conversations,” so that frontier companies can set common standards and progress limits without government participating, only enabling the dialogue. Sacks’s reply: “Stop pretending you need to suspend antitrust law in order to create a cartel.”

Whom it reaches. Lab leadership.

Score and reasoning: 4 of 15. The waiver builds social adaptability between firms (SA 2) and creates a shared cognitive frame (CA 1). It does nothing for anyone outside the room, and Sacks’s cartel critique is strongest precisely because the room is so small.

Minimal amendment. Extend the waiver by one noun: “safety conversations and workforce-transition data.” Labs permitted to share safety findings would, under the same legal architecture, be permitted to pool anonymized adoption and displacement data into the kind of AI Workforce Research Hub that the Carnegie Endowment has recommended. The waiver’s scope changes; its structure does not.

Projected score after amendment: roughly 8 of 15. The coordination proposal goes from serving lab leadership to producing the sectoral and regional adaptability data that the white paper’s Section 6.2 (“Tracking Regional and Sectoral Adaptability”) depends on. It also blunts the cartel critique, because a coordination body that publishes labor data is visibly serving someone other than its members.

6. Transparency and auditing legislation

What it says. Amodei asks for laws targeting “all US frontier AI companies” focused on “transparency and third-party auditing.” The Thune–Klobuchar plan is the nearest live vehicle: a narrower bipartisan measure requiring companies to test for and report catastrophic risks.

Whom it reaches. Regulators and auditors.

Score and reasoning: 3 of 15. Reporting creates a routine (BA 1), auditing creates a small learning demand (CA 1), and disclosure creates a thin social link to the public (SA 1).

Minimal amendment. Add one line to the report template: “estimated occupational exposure.” The labs already compute this internally; product coverage of Opus 4.8 praises reduced “human verification needs” as a feature, which means someone at the company knows which verification jobs were reduced. Requiring the number to be reported is a disclosure change, not a research program.

Projected score after amendment: roughly 7 of 15. GP rises because occupational exposure over successive model generations is a trajectory, and trajectory is what HAPI’s growth-potential dimension measures.

7. Export controls and the “3–5 year window”

What it says. Prohibit chip and semiconductor-equipment sales to China, crack down on smuggling and remote access, combat weight theft and distillation. Amodei says these could “widen America’s lead significantly over the next 3–5 years — the window when AI becomes geopolitically most important.” Bessent: “there is no day after tomorrow if China wins at this.”

Whom it reaches. Chinese developers, US chipmakers, enforcement agencies. The domestic workforce is untouched.

Score and reasoning: 0 of 15. This is not a workforce policy and should not be scored as one. It is included because of the window.

Minimal amendment. Not to the policy but to its framing. The essay says the controls buy 3–5 years. Nobody has assigned that window to anyone. Name it as the national reskilling horizon and align existing programs to it. The white paper’s Section 6.1 uses Singapore’s SkillsFuture as the template for a national program that could carry HAPI assessments; the US equivalents already exist in workforce-development funding and could be given the same horizon without new legislation. The window is already being purchased. It only needs an owner.

Projected score after amendment. The policy itself stays at 0. The window, if assigned, becomes the time budget for every other amendment in this document.

8. International agreements

What it says. Four levels: prohibit narrow dangerous uses; mutual pre-release testing through “a global standards body”; speed limits on recursive self-improvement, compared to SALT; full pacing, which Amodei doubts is “likely to actually happen any time soon.”

Whom it reaches. Diplomats and standards bodies.

Score and reasoning: 1 of 15. Only Level 2, the standards body, offers a hook (SA 1).

Minimal amendment. Give the Level 2 body a second mandate alongside model testing: an international adaptability benchmark, built on instruments that already exist. The OECD’s Survey of Adult Skills and Skills Outlook already measure adult learning and adaptation across member states; the white paper’s Section 6.5 (“International Collaboration and Benchmarking”) proposes HAPI as the harmonizing layer. Piggybacking on an existing survey is far cheaper than a new treaty and needs no reciprocity from China to be useful to democracies.

Projected score after amendment: roughly 5 of 15. Modest, because international instruments are slow. But it is the only route by which the pacing debate produces comparable cross-national human data at all.

9. Ban Artificial Superintelligence Act

What it says. The Sanders–Casar bill would pause domestic AI development pending international reciprocity, with criminal penalties.

Whom it reaches. Everyone, passively.

Score and reasoning: 1 of 15. A pause reduces environmental stress (EA 1) and nothing else. Amodei’s own critique of 2023-era pauses applies with full force: “The question was always: what would you do with the extra time?” A pause with no program for the time is elapsed time, not gained time.

Minimal amendment. Condition the pause on a plan: require that any pause be accompanied by an annual public adaptability report covering the exposed workforce, so that the time is measured rather than merely elapsed.

Projected score after amendment: roughly 6 of 15. Entirely through measurement. The bill’s likelihood of passage is low, but the amendment is worth stating because it is the same amendment that applies to any future pause proposal, and there will be more of them.

10. Senate select committee with subpoena power

What it says. Senators Gallego and Van Hollen propose a bipartisan select committee with investigative authority.

Whom it reaches. Congress.

Score and reasoning: 2 of 15. Subpoena power creates learning capacity (CA 1) and a link between the industry and the legislature (SA 1). Scope is undefined.

Minimal amendment. Define the scope to include workforce data explicitly. A committee empowered to subpoena safety records can, with the same power, subpoena the occupational-exposure estimates that proposal 6 would require labs to report. The two amendments reinforce each other: disclosure creates the data, the committee creates the demand for it.

Projected score after amendment: roughly 5 of 15.

11. “Decelerate on your own”

What it says. Sacks argues that if labs fear development is too rapid, “they can decelerate on their own without using Washington to force competitors” to do the same. The administration’s framing, “whoever wins AI wins,” leaves winning undefined.

Whom it reaches. Nobody, by design.

Score and reasoning: 0 of 15. Market self-regulation with no measurement leaves adaptation entirely to individual firms and individual workers, unobserved.

Minimal amendment. This position rejects federal mandates, so the amendment must be voluntary and firm-level to be consistent with it. The white paper’s Section 8.1 proposes exactly that: an internal HAPI dashboard, published in aggregate the way firms now publish sustainability metrics. A firm that believes it can pace itself can also publish its internal-mobility rate, its reskilling completion rate, and its adaptability distribution. That is a disclosure a market-first position can endorse, and it is the only version of “decelerate on your own” that produces evidence.

Projected score after amendment: roughly 4 of 15, and only for firms that opt in. Low, but not zero, and consistent with the position’s own logic.

12. Open Alignment Initiative and open-weight access

What it says. Hugging Face’s Clement Delangue announced an Open Alignment Initiative and asked to join the embedded-evaluator program. Separately, the incident record shows that Hugging Face’s responders could not use frontier commercial models for forensic analysis because guardrails blocked the work, and completed the analysis with an open-weight model instead.

Whom it reaches. Anyone who can run a model.

Score and reasoning: 6 of 15. Open access builds cognitive adaptability broadly (CA 2), lets defenders change their tools when circumstances demand (BA 1), sustains a community that learns from itself (SA 2), and creates a pipeline into the field (GP 1). It is the only proposal that widened access rather than narrowing it, and it did so without anyone arguing for it on those grounds.

Minimal amendment. Write the accident into policy. Guardrails on commercial models should carve out authenticated incident-response and forensic use, so that the humans adapting fastest during a crisis are not the ones with the least capable tools. This does not require open weights; it requires a defender exception in the safety policy that every lab already maintains.

Projected score after amendment: roughly 8 of 15. BA rises to 2 and EA to 1, because responders who know they will have tools under pressure behave differently under pressure.

Part Three: Before and after

Summed across the table, the proposals as written total 42 of a possible 180. With the amendments, they total roughly 85. No amendment adds a program, an agency, or a mandate that is not already implied by the proposal it attaches to. The doubling comes entirely from adding a human field to instruments that are being built anyway.

The three amendments that matter most, if only three can be made, are the readiness clause on the checkpoint (proposal 1), the diffusion evaluator (proposal 2), and the timestamps in the incident report (proposals 3 and 4 together). They are also the three that require no legislation. Anthropic could adopt all three unilaterally, exactly as it adopted the embedded-evaluator commitment, and Altman’s “we will do the same” would presumably follow.

Part Four: Maximum HAPI with minimum change, for organizations and individuals

The same principle scales down. For an organization reading the news rather than writing the rules, the highest-yield interventions are the ones that attach measurement to things the organization already does. The white paper’s evidence base points to one cheap lever per dimension.

For cognitive adaptability, the lever is stretch assignments and job rotation. Section 4.1 makes the neuroplasticity case: novel challenges build the capacity to handle novel challenges. The cost is a scheduling decision, and the measurement is already available in any learning-management system that tracks course completion and time to proficiency.

For emotional adaptability, the lever is blameless post-mortems and drills, borrowed directly from the aviation model Amodei cites. The white paper’s Section 4.2 notes that self-efficacy grows through small wins and role modeling; a drill is a small win under controlled stress. The measurement is response time under pressure, which the timestamps above provide.

For behavioral adaptability, the lever is a regular feedback loop. The white paper cites Gallup’s finding that organizations with regular feedback see roughly 15 percent higher engagement, and McKinsey’s finding that workers who regularly sought feedback improved problem-solving by 20 percent. The measurement is adoption speed of new tools and processes, which software usage logs already capture.

For social adaptability, the lever is cross-functional project rosters and mentoring pairs. Section 4.3 on social learning theory and the Project Aristotle finding on psychological safety both point the same way. The measurement is network breadth, which project rosters and communication metadata already contain.

For growth potential, the lever is a visible, sponsored internal-mobility path. HAPI weights this dimension most heavily (0–40 of 100 in the white paper’s scoring), and Section 7.2 makes the cost case: internal mobility is a fraction of the cost of external hiring. The measurement is year-over-year role progression, which HR systems already record.

For an individual, the compression is simpler still. The spirit of the white paper’s Section 8.3 reduces to a few moves: assess yourself regularly rather than waiting for an employer or a program to do it, seek the assignment you are not yet qualified for, ask for feedback on a fixed cadence rather than waiting for it, and keep a record of what you learned each quarter so that your trajectory is visible to you before it is visible to anyone else. None of these requires an employer’s permission. All of them produce the data HAPI would score.

The pattern that runs through this entire document, from the checkpoint clause to the individual’s quarterly log, is the same. The people building frontier AI have decided to measure their machines with unprecedented rigor. The cheapest and most consequential change available to anyone, at any level, is to measure the humans with the same instruments, at the same time.

Sources

• Dario Amodei, “We Must Pace the Frontier”, September 2026
• Axios, “Anthropic, OpenAI CEOs call for slowdown in AI development”, September 12, 2026
• Axios, “Washington’s AI paralysis: Let ‘er rip vs. hit the brakes”, September 14, 2026
• Unite.AI, “Altman Says OpenAI Will Match Anthropic’s Embedded Evaluator Pledge”, September 12, 2026
• Wikipedia, “2026 OpenAI agent cyberattacks” (timeline, legislative proposals, defender-tooling asymmetry)
• CDM, “Amodei, Altman, and Musk all want to slow down AI — David Sacks says they are building a cartel”, September 13, 2026
• StartupHub.ai, “Dario Amodei We Must Pace the Frontier Is Vague”
• Carnegie Endowment, “The AI Labor Debate: Three Views on the Future of Work”, April 2026 (AI Workforce Research Hub recommendation)
• Axios, “Anthropic’s revenue run rate reportedly surpasses $65 billion pre-IPO”, August 17, 2026 (Opus 4.8 “human verification” framing)
• HAPI white paper research draft (HAPI_Research.txt, supplied): Sections 4.1–4.3, 5.2, 6.1, 6.2, 6.5, 7.2, 8.1, 8.3