Methodology

This page states, once, what came from the rubric, what came from the framework package's own back matter, and what was written for this site rather than for the framework. Every figure below is read from the same files the assessment reads at run time — none is typed into this page, so a rubric change moves the numbers here without anyone having to remember to.

Weights

The corporate variant scored on this site divides its full weight across 9 pillars and 38 sub-indicators. The table below is generated from loadRubric() — the same function every score on this site reads — and must always agree with the framework's own copy at §2.3, Weights by variant.

Changelog

Every released version, newest first, verbatim from the package's own back matter.

Version 0.7.1 — August 2026

A revision in three layers, responding to the August 2026 evidence review.

Evidence and consistency (what a 0.5.8 would have carried). Evidence Base updated to August 2026 with eleven added or extended entries: the Fable 5 / Mythos 5 restoration (30 June – 1 July), US chip-export posture codification (January–July 2026), DeepSeek-V4 with day-0 Ascend adaptation, the Nexperia standoff (a pre-snapshot gap), the June–July rare-earth enforcement wave, the AI Gigafactories call, the first EU sovereignty-criteria cloud procurement, the Digital Omnibus deferral with on-schedule GPAI enforcement, the CADA and Chips Act 2.0 proposals, the New Delhi Declaration (promoted from Part J), and the Bartz final settlement approval. The Fable/Mythos arc is reconciled everywhere it is cited. The v0.5.7 L0–L4 relabelling is completed in the BgGPT case study and the corporate guide (INSAIT profile Models 12 → 11; totals adjusted). MacroPolo tracker corrected to 2.0 throughout; CAISI naming standardised; a stale reference in the Evidence Base verification note removed.

Methodology. Every sub-indicator now carries an R (resilience) or A (autonomy) tag, and assessments report the two subtotals alongside the headline (Framework §1.1) — the two-axes definition now reaches the arithmetic. AI-SAF-C Models: portfolio diversification is rescored as portfolio-and-jurisdiction diversification (a single-jurisdiction portfolio caps at 2 of 5), reflecting the June 2026 lesson that vendor count is not the binding risk. Version Deadlock taxonomy gains the restoration-is-not-recovery rule (§3.5). The Infrastructure anchors add the proposed CADA Cloud Sovereignty Framework alongside SecNumCloud 3.2 and GAIA-X L3. §4.3 now states plainly which rigour checks are design commitments rather than completed exercises, and Part V owns the limits of additive aggregation. Modifier discipline is enforced: the national worked examples apply all modifiers in a fixed order, and the ad-hoc bonuses in earlier revisions (France “+5 rule-setting,” UAE “+5 energy”) are removed — France recomputes to ~70 (from ~71), the UAE to ~57 (from ~62), Bulgaria is unchanged at ~46.

Structure and figures. The package is reorganised: the corporate variant (AI-SAF-C) now leads as Document 2; dated market material moves from the field guides into a new Document 4, State of Play, refreshed each revision; the Evidence Base becomes Document 5; the three glossaries merge into one master glossary in Document 1; and repeated passages (the “same pillar, opposite meaning” argument, the CrowdStrike rationale) are stated once. The package is roughly a third shorter than v0.5.7 with no scoring content removed. Both figures are redrawn: the architecture diagram with the correct weights, fixed modifier order, and R/A tags; the trajectory chart in the package palette with the stale pre-rename axis label corrected.

Pillar weights are unchanged. Toward 1.0: publication of the Monte Carlo sensitivity runs and PCA cross-checks (§4.3), and at least two independent assessments landing within one band of the author’s.

Version 0.5.7 — July 2026

Responded to the first external review: Models ladder rescaled to the 16-point weight (L0–L4 = 0/4/11/13/16); pooled EU compute credited in Local compute (≤2 of 4 points); front matter consolidated; residual inconsistencies corrected.

Version 0.5.6 — June 2026

Added the data-egress dimension to the Data pillar in both field guides; pillar totals held by redistribution.

Version 0.5.5 — June 2026

Added the Anthropic Fable 5 / Mythos 5 suspension (12 June 2026) as a flagship case; no scores or methodology changed.

Version 0.5.4 — May 2026

First public release of the document package.

Citation

To cite the framework package itself, in the package's own words:

Alexandrov, G. (2026). AI Strategic Autonomy Framework: Assessment Framework (Version 0.7.1) [Document package]. https://doi.org/10.5281/zenodo.22178006

To cite one completed assessment rather than the framework — the score, both axis readings, confidence level and date — use the rubric's own format instead:

AI-SAF-C v0.7.1 · {score} (R {resilience} / A {autonomy}) · self-assessed, {confidenceLabel} · {ISO8601 date}

What is uncertain

The framework documents its own limits in Part V. Areas of genuine uncertainty. What follows is this tool's list: the parts of this site that go beyond that document, said plainly here rather than only as a caveat on the report.

• 29 of 38 sub-indicator anchors were written for this tool and are still awaiting the framework author's sign-off.
• The Version Deadlock thresholds are authored for this tool. The four states themselves are the framework's own taxonomy.
• The rubric holds capability data only. Framework §1.5 derives its three-year projection from the commitment half of each sub-score, which the rubric does not carry — so the projection cannot be computed, and the assessment report says exactly that rather than estimating around it.

Each of those is an invitation rather than a disclaimer: how to review this framework sets out what would settle them, and where to send it.