Examples
Examples from published documents
- A developer cannot delay or block an external evaluator's publication for more than 30 business days after learning of the findings, unless a longer timeline is exceptionally necessary (EU GPAI Code of Practice, Measure 3.5).
- Anthropic's external reviewers are bound by confidentiality but 'will not be restricted in what they can publish, including concerns about the Risk Report or Anthropic's conduct' (Anthropic RSP v3.4, §3.6.3).
- In METR's cross-developer report, participants gave written approval for the final text of the appendices containing their attributed claims (METR Frontier Risk Report, Feb-Mar 2026).
Counter-examples
Not to be confused with
- A confidentiality agreement is not a publication rights clause: it says what must stay private, not what the evaluator is free to publish or who may delay it.
Related terms
Related AI safety terms
Terms defined under this element that are not NIKOLAI elements themselves, each as the cited document uses it. Every term is also listed in the AI safety glossary.
- Editorial control (evaluation)
- Editorial control is an evaluator's right to decide how it presents the results of its evaluation, including how it characterises what was tested and the system's performance, without the developer rewriting its conclusions.
- Source: AEF-1: Minimum Operating Conditions for Independent Third Party AI Evaluations (v1, 4 Dec 2025)
- No contingent release
- No contingent release is the principle that a developer must not narrow who sees an evaluation's results because the results are unfavourable, such as showing poor performance or unmitigated risks.
- Source: AEF-1: Minimum Operating Conditions for Independent Third Party AI Evaluations (v1, 4 Dec 2025)
- Silent exit option
- A silent exit option is an agreed right for either a developer or an evaluator to end a joint assessment before publication, after which the evaluator treats the developer as never having participated.
- Source: METR Frontier Risk Report (February-March 2026)
Source of record
Where this definition comes from
We Must Pace the Frontier (essay by Anthropic's CEO), Embedded Evaluators
“External reviewers should have the right to publish key findings about risk levels, incidents, practices, and the access they received or didn't receive — without editorial control by Anthropic.”
https://darioamodei.com/post/we-must-pace-the-frontierAnthropic Advanced AI Framework, p.8
“Evaluators should be bound by obligations not to copy, retain, or disclose confidential information (e.g., confidential intellectual property, matters of national security, or proprietary details such as model architecture, cost, or size), but beyond that should generally not be restricted in what they can publish, including concerns about the risk report or the developer's conduct in connection with the review process.”
https://www-cdn.anthropic.com/files/4zrzovbb/website/0a58d567024a8b448ff15158ebc3625328dfcc1f.pdfEU GPAI Code of Practice, Safety and Security Chapter, Measure 3.5
“Additionally, Signatories will not take any legal or technical retaliation against the independent external evaluators as a consequence of their testing and/or publication of findings as long as the evaluators: ... (5) adhere to the Signatory's publicly available procedure for responsible vulnerability disclosure, which will specify at least that the Signatory cannot delay or block publication for more than 30 business days from the date that the Signatory is made aware of the findings, unless a longer timeline is exceptionally necessary”
https://ec.europa.eu/newsroom/dae/redirection/document/118119
Crosswalk
How named organisations use this concept
| Organisation | Their term, as published | Match & verification | Source |
|---|---|---|---|
| AnthropicShadow mapping We Must Pace the Frontier (essay by Anthropic's CEO) / Advanced AI Framework / RSP v3.4 | “Essay: "External reviewers should have the right to publish key findings about risk levels, incidents, practices, and the access they received or didn't receive — without editorial control by Anthropic." AAF: "Evaluators should be bound by obligations not to copy, retain, or disclose confidential information ..., but beyond that should generally not be restricted in what they can publish, including concerns about the risk report or the developer's conduct in connection with the review process." (p.8). RSP v3.4 §3.6.3: external reviewers "will not be restricted in what they can publish, including concerns about the Risk Report or Anthropic's conduct in connection with the external review."” Also cites the Advanced AI Framework and Anthropic's Responsible Scaling Policy v3.4. | exactEQ confidence: high | We Must Pace the Frontier (essay by Anthropic's CEO) |
| EUShadow mapping EU GPAI Code of Practice, Safety and Security Chapter | “Measure 3.5 (post-market monitoring): Signatories "will not take any legal or technical retaliation against the independent external evaluators as a consequence of their testing and/or publication of findings" as long as the evaluator meets five conditions, including adherence to "the Signatory's publicly available procedure for responsible vulnerability disclosure, which will specify at least that the Signatory cannot delay or block publication for more than 30 business days from the date that the Signatory is made aware of the findings, unless a longer timeline is exceptionally necessary." Commitment 7/Measure 7.4: evaluators "maintain control over the publication of their findings, without implicit endorsement by the Signatories of the content of such reports."” | exactEQ confidence: high | EU GPAI Code of Practice, Safety and Security Chapter |
| METRShadow mapping METR (site) / OpenAI-Hugging Face incident investigation | “Frontier Risk Report: "participants gave written approval for the final text of these appendices. Participants likewise gave written approval for the use of transcripts that relied on privileged access." Incident investigation: "OpenAI was able to redact any non-public information from this post. We worked with OpenAI to find mutually agreeable language to describe redactions where they occurred".” Also cites METR's OpenAI/Hugging Face incident investigation. Shows a negotiated model: the developer approves or redacts, the evaluator keeps editorial control of the rest and states what was redacted. | closeCL confidence: medium | METR Frontier Risk Report (February-March 2026) |
| AI Evaluator Forum (AEF-1)Shadow mapping AEF-1: Minimum Operating Conditions for Independent Third Party AI Evaluations | “AEF-1 "3.4 Editorial control Requirement: The evaluator retained editorial control over how they present the results of their evaluation"; "4.2 Disclosure rights Recommendation: The system provider granted upfront any necessary rights to disclose results to the evaluation's intended audiences"; "4.3 No contingent release Recommendation: The intended audiences for an evaluation were not narrowed based on the results."” Verified against the AEF-1 PDF (v1, 4 Dec 2025). | exactEQ confidence: high | AEF-1: Minimum Operating Conditions for Independent Third Party AI Evaluations (v1, 4 Dec 2025) |
| AnthropicShadow mapping Anthropic Responsible Scaling Policy v3.4 | “"External reviewers will be bound by obligations not to disclose confidential information (e.g., confidential intellectual property, matters of national security, or proprietary details such as model architecture, cost, or size), but beyond that will not be restricted in what they can publish, including concerns about the Risk Report or Anthropic's conduct in connection with the external review." (§3.6.3)” Policy text in force from 8 Jul 2026, distinct from the legislative proposal in the AAF. | exactEQ confidence: high | Anthropic Responsible Scaling Policy v3.4 |
What do these codes mean?
- exact
- The source term is equivalent to this element
- close
- The source term is close but not equivalent to this element
- broad
- The source term is broader than this element
- narrow
- The source term is narrower than this element
- none
- No mapping claim — used for false-friend and declared-but-undefined rows
- EQ
- Equivalent
- CL
- Close
- BR
- Source is broader than the element
- NR
- Source is narrower than the element
- FF
- False friend — same or similar label, different meaning
- DU
- Declared but undefined by the source
- UV
- Unverified
Related, not mapped
Pointers that are not crosswalk claims
These sources mention this concept but do not define or map it clearly enough to count as a crosswalk row — noted here so the research is visible without overstating it as a mapping.
- OpenAI
No publication clause was found in the Preparedness Framework or the Frontier Governance Framework, and the CEO's post gives no publication terms. In the METR investigation: "OpenAI was able to redact any non-public information from this post. We worked with OpenAI to find mutually agreeable language to describe redactions where they occurred" -- a pointer, not a mapping.
Post on X by OpenAI's CEO; METR OpenAI-Hugging Face incident investigation - Google DeepMind
UK AISI partnership includes "Joint reports and publications sharing findings with the research community"; whether AISI has independent publication rights is unstated -- a pointer, not a mapping.
DeepMind / UK AI Security Institute partnership post
Gap
The EU Code of Practice's Measure 3.5 is the one text CASRAI found that puts a number on how long a developer may delay an evaluator's publication (30 business days). The Frontier Model Forum report sets out no publication-rights regime and lists 'protocols for sharing findings' as future work. California SB 53 and the RAISE Act do not address evaluator publication.







