The BaseEssayAI Alignment Forum

Delaying Internal Deployments Until After X Days of Safety Testing

In a nine-minute post written in a personal capacity, the author annawang proposes that a new AI model not be used inside its developer, beyond safety testing, until a fixed number of days after being filed with an auditor. The post lays out a scenario, open questions, benefits and drawbacks.

Open the source ↗3 quotations checked, word for word ↓

Why it matters

The same thing, for different readers.

What it means for you

annawang proposes that AI companies wait a set number of days, with an outside auditor involved, before staff use their newest model for anything except safety testing.

For lawmakers

annawang imagines a legislature setting a minimum wait, from 10 days possibly up to 90, and lists how developers could game the rule.

For researchers

annawang notes the trade-off that real internal use often surfaces misaligned behaviour, so delaying it slows both AI R&D and the fixing of safety problems.

In their own words

Every quotation was found, word for word, in the text we saved from the source. That shows the words are there; it does not make them true, or prove the speaker meant them as quoted.

From the source’s own text.

“No checkpoint may be used internally for purposes beyond safety testing, until at least X calendar days after its hash is filed with an auditor.”
annawangThe author states the concrete version of the proposal they prefer.
Found word for word in the source
“Today a problem seen in real use can feed into the next model’s training almost straight away; under this proposal, a model will get X more days of training before an issue is detected.”
annawangThe author lists this among the drawbacks of the proposal.
Found word for word in the source
“I have a hunch that, if implemented well, this proposal would have a larger impact on the pace of recursive self-improvement (RSI) itself than on the onset of RSI.”
annawangThe author gives a tentative view on what the delay would do if used as a speed dial.
Found word for word in the source

The record

Questions
Who can demand evidence, delay a release or stop an unsafe deployment? · Are AI companies putting stronger safeguards into practice? · What happens when AI helps build more capable AI?
People
annawang (Wrote it) · evhub (Mentioned)
Also covered by
LessWrong
What we read
We read the whole source.
Found by
The Radar, run of November 11, 2026, via ai-alignment-forum (tier 1: the source’s own channel or an official record)
Cite this
AI Alignment Forum, “Delaying Internal Deployments Until After X Days of Safety Testing”, November 11, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/alignmentforum-2026-11-11-delaying-internal-deployments-safety-testing
Our notesAuthor is known only by the handle annawang; the post says it reflects their own views, not their employer's, and names no employer. One comment, from the handle evhub, says 'Strongly endorsed' and that the commenter might write more publicly; it is not used as a quote and nothing is assumed about who evhub is. The post cites the AI Futures Project's 'How to pace the US frontier' as the source of the broader idea.