AI-SAF

Review this framework

AI Strategic Autonomy Framework is published at v0.7.1 — before it is finished, on purpose. It is a scoring instrument making falsifiable claims, and the useful thing to do with it is not to adopt it. It is to break it.

What stands between this and v1.0

Two things, both stated in the framework's own Part V rather than discovered by a reader:

  1. Published weight-sensitivity analysis. Monte Carlo weight-sensitivity and PCA cross-checks are described in §4.3 as design commitments. They have not been run. They are a precondition for 1.0, not a decoration on 0.7.
  2. At least two assessments performed by someone who is not the author, landing within a band of the author's own reading. This is the harder of the two, and it is the reason this page exists.

What is already known to be uncertain

Part V of the framework document lists it in full, and reading it first saves reporting something the author already concedes. The three largest:

  • The aggregation partly contradicts the thesis. Pillars sum linearly, so a strong regulatory position can compensate for near-zero model capability — while the framework's own argument is that sovereignty fails at the choke point. The dependency-type labels are the non-compensatory reading; the honest fix is open.
  • 9 weights set by judgement. They are argued from the record on the methodology page, not derived from a model, and the framework says so.
  • Whether export controls now reach model access itself is scored as a live uncertainty rather than a settled fact. One event is one event.

Read Part V in the framework document →

The standing questions

Answering any one of these, with evidence, is a complete review. None of them is rhetorical — each names a place the author expects to be wrong.

  1. Which pillar weight would you move, and on what dated evidence?
  2. Which sub-indicator is tagged to the wrong axis — and which way would you flip it?
  3. Where does the national/corporate inversion break down? Which pillars carry it weakly?
  4. Name a Version Deadlock case the four states misclassify — any technology, any decade.
  5. Is the single-jurisdiction portfolio cap too harsh, too soft, or the wrong mechanism?
  6. What moved AI sovereignty between 2024 and 2026 that the Evidence Base is missing?
  7. Does the additive aggregation contradict the weakest-link thesis badly enough to replace it?

Where to send it

A public issue, so the argument stays citable and the next reviewer can read it. There is no form on this site and there will not be one: nothing here leaves your browser, and a form would need a server to receive it.

Read what others have raised →

What happens to it

Criticism that survives is credited in the changelog with the version it changed, and the framework's release history is public — v0.7.1 shipped August 2026. Criticism that kills a claim is worth more than agreement that confirms one, and it is treated that way: two worked examples in the current version exist because earlier ones were wrong and said so.

How it was made

AI assistance disclosure: Drafting, editing, and evidence-base compilation were assisted by Claude (Anthropic). All analysis, judgments, and final content are the sole responsibility of the author.