What labs have committed to, and what new laws ask, in their own words.
Each is quoted from its source, beside what the record gives. Most of them ask a lab to watch its models, write down what happened, and let someone outside check.
Four layers of controls and audits.
The accord, signed September 29, 2026, sets out four layers for the companies that signed it: internal controls, an internal team, an independent outside auditor or evaluator, and a board committee. It is voluntary today, and it says it may make sense to codify the steps in law.
“Implement robust internal controls to monitor the capabilities and alignment of its models during training and deployment”
What the monitoring sees can be written as it happens and committed where no one can change it, PacSpace included.
“Partner with an independent external auditor or evaluator to carry out independent assessments of whether the controls, monitoring, and detection are operating as intended.”
The auditor opens one link and checks every entry against what was committed, with no visit and no access to the lab's systems.
“Designate an independent committee of the board of directors to oversee and receive reports from the teams operating the controls”
The committee gets its own link to the same record, showing the fields the lab chooses for it.
California and New York ask for incident reports.
“shall report any critical safety incident pertaining to one or more of its frontier models to the Office of Emergency Services within 15 days”
The record shows when the incident was first written down and what was written then, unchanged since.
“A frontier model that uses deceptive techniques against the frontier developer to subvert the controls or monitoring of its frontier developer…” The law counts this only outside an evaluation built to draw it out, and only where it shows a materially higher catastrophic risk.
Entries committed before a model works against the monitoring stay as they were written, and a changed copy fails the check.
A large frontier developer's framework must cover “Using third parties to assess the potential for catastrophic risks”.
A third party assesses from a record it checks itself, instead of from what the developer hands over.
“shall report any critical safety incident pertaining to one or more of its frontier models to the office within seventy-two hours from a determination”
When the determination was written down is on the record, unchanged, beside everything written before it.
Each law applies to the developers it defines, from the date it sets. Both set a higher bar, over $500 million in yearly revenue, for the large frontier developers that must publish a framework.3
Evaluations, serious incidents, and outside evaluators.
The European Commission's AI Office and national authorities began enforcing the AI Act on 2 August 2026.6
“perform model evaluation in accordance with standardised protocols and tools reflecting the state of the art, including conducting and documenting adversarial testing”
Each test run can be written as it happens, so the documentation includes the record of the run itself.
“keep track of, document, and report, without undue delay, to the AI Office and, as appropriate, to national competent authorities, relevant information about serious incidents”
What was tracked, and when, is on the record unchanged, for the lab and for the office it reports to.
“…Signatories will provide an adequate number of independent external evaluators with adequate free access…”
The access each evaluator got, and when, can be written as it's given, so it's on the record unchanged.
What labs have already written about logs and outside review.
“We will comprehensively gather, centralize, and maintain logs for all critical AI-development activities, and use AI to analyze them for issues”
Centralized logs still live on the lab's own systems. Committed, the same entries become checkable by someone outside the lab without asking.
“Agent actions are all logged in an uneditable database, and asynchronous monitoring routines review those actions for evidence of harm”
A committed record is out of everyone's reach, the lab's and PacSpace's included, and anyone given the link can check that for themselves.
“If we assess that a model has reached a CCL that poses an unmitigated and material risk to overall public safety, we aim to share relevant information with appropriate government authorities where it will facilitate safety of frontier AI.”
What's shared can be a link to entries that check against what was committed, with the fields the lab chooses.
Each framework is its lab's own commitment, quoted as published. PacSpace makes no claim about how any lab applies its framework.
The record answers part of each commitment, never the whole of it.
PacSpace is not a compliance product and doesn't make a lab compliant with any law, code or framework on this page. What a record satisfies is for the lab, its counsel, and whoever applies the rule.
The lab still chooses what to write, the same limit every log has. What it gives up is changing the record afterward. If writing stops, the gap shows: the records on either side put a start and an end on it.
Sources
- The American Presidency Project, “White House Accord on Super Intelligence”, Sept 29, 2026.
- Cal. Bus. & Prof. Code §22757.13, from SB 53, the Transparency in Frontier Artificial Intelligence Act, in effect Jan 1, 2026.
- Cal. Bus. & Prof. Code §22757.11, definitions.
- Cal. Bus. & Prof. Code §22757.12, frontier AI frameworks.
- New York, S8828, amending the RAISE Act (General Business Law Article 44-B), signed Mar 27, 2026.
- European Commission, “Commission starts enforcing AI Act rules and new transparency requirements on 2 August”, July 31, 2026.
- EU AI Act, Article 55, Regulation (EU) 2024/1689.
- General-Purpose AI Code of Practice, Safety and Security chapter, Measure 3.5, July 10, 2025.
- Anthropic, Responsible Scaling Policy, version 3.4, in effect July 8, 2026.
- OpenAI, Preparedness Framework, version 2, Apr 15, 2025.
- Google DeepMind, Frontier Safety Framework, version 3.1, Apr 17, 2026.
Bring the case you think breaks it.
We would rather be evaluated by use than by description. Talk to us and we'll put you in a live environment: commit a record, do your best to change it, then check it yourself, with us out of the loop. The change shows.
The record must exist.