Frontier doctrine scan, September 2026
Ten frontier AI labs scored 0 to 4 on each of the twelve laws, from their published policies and product design. A dated research note, not a safety ranking.
Read this first
What this scan scores, and what it does not
Ten frontier labs scored against the twelve laws for engineering delegated intelligence, on the Institute's reading of their public doctrine and product architecture. No individual score has a source attached yet.
The scan is published outside the method. It is an argument about where the method should center, not part of what the method requires.
Scoring scale
Each lab is scored on each law from 0 to 4. Overall fit is a weighted reading of the twelve scores, reported on a 0 to 10 scale.
| Score | Meaning |
|---|---|
| 4 | Explicit and operational in public doctrine |
| 3 | Strongly present |
| 2 | Partial or narrow analogue |
| 1 | Weak or incidental |
| 0 | No meaningful public analogue found |
Overall fit by lab
Fit to the twelve laws as a whole, on public doctrine and product architecture. Higher is closer to the laws. It is not a safety ranking.
| Lab | Overall fit (0 to 10) | Characterization |
|---|---|---|
| Anthropic | 6.9 | Strongest public doctrine on capability versus authority, observation, oversight, and evaluation layer; little on dependence, standing, or contestability preservation. |
| OpenAI | 6.5 | Explicit permission and evaluation doctrine; renewal and coupling present but partial; dependence and standing largely absent. |
| Google DeepMind | 6.5 | Strong observation and evaluation doctrine; oversight present; the only lab with a partial analogue to contestability preservation. |
| xAI / SpaceXAI | 5.8 | Explicit on capability versus authority and observation; weaker on oversight and reversibility; dependence and standing absent. |
| Meta MSL | 4.8 | First-generation laws partially present; no public analogue to standing; second-generation laws weak or incidental. |
| Mistral | 4.0 | Capability boundaries and oversight present; renewal and coupling weak; no public analogue to expansion review. |
| Alibaba / Qwen | 3.5 | Capability boundaries present; most other laws weak or narrow; no public analogue to dependence as risk. |
| DeepSeek | 2.3 | Weak or incidental across most laws; a partial analogue to standing is the only score above 1 apart from evaluation layer. |
| Z.ai | 1.7 | Capability boundaries strongly present; almost no public doctrine on any other law. |
| Moonshot | 1.0 | Weak or incidental on capability, observation, oversight, and evaluation layer; no public analogue found elsewhere. |
Scores by lab and law
Columns are Laws I to XII in order. The table scrolls sideways on narrow screens.
| Lab | I | II | III | IV | V | VI | VII | VIII | IX | X | XI | XII |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Anthropic | 4 | 3 | 3 | 3 | 2 | 2 | 1 | 4 | 4 | 4 | 2 | 1 |
| OpenAI | 4 | 3 | 3 | 3 | 1 | 2 | 1 | 4 | 3 | 4 | 2 | 1 |
| Google DeepMind | 3 | 3 | 3 | 3 | 1 | 2 | 1 | 4 | 3 | 4 | 2 | 2 |
| xAI / SpaceXAI | 4 | 3 | 3 | 3 | 1 | 1 | 1 | 4 | 2 | 3 | 2 | 1 |
| Meta MSL | 3 | 2 | 2 | 3 | 1 | 1 | 0 | 3 | 2 | 3 | 2 | 1 |
| Mistral | 3 | 1 | 1 | 2 | 1 | 2 | 1 | 2 | 3 | 2 | 0 | 1 |
| Alibaba / Qwen | 3 | 1 | 1 | 2 | 0 | 1 | 1 | 2 | 2 | 2 | 1 | 1 |
| DeepSeek | 1 | 1 | 1 | 1 | 0 | 0 | 2 | 1 | 1 | 2 | 1 | 0 |
| Z.ai | 3 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 2 | 2 | 0 | 0 |
| Moonshot | 1 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 1 | 2 | 0 | 0 |
Law statements
- I. Capability does not imply authority
- II. Authority decays unless its justification is renewed
- III. Evidence and authority must remain coupled
- IV. Every consequential delegation creates a correction obligation
- V. Dependence converts technical risk into structural risk
- VI. Nominal reversibility is not operational reversibility
- VII. Error-bearing parties require standing proportional to exposure
- VIII. Observability without state transition is theater
- IX. Human oversight is a control only if the human can alter system state
- X. The relevant eval sits at the highest layer where harm can emerge
- XI. Successful automation increases its own governance burden
- XII. No system may erase the conditions of its own contestability
Averages
Average score by law
Mean across the ten labs, sorted from most to least covered. The split between first-generation and second-generation laws is the finding.
| Law | Statement | Average (0 to 4) | Generation |
|---|---|---|---|
| I | Capability does not imply authority | 2.9 | First |
| X | The relevant eval sits at the highest layer where harm can emerge | 2.8 | First |
| VIII | Observability without state transition is theater | 2.6 | First |
| IX | Human oversight is a control only if the human can alter system state | 2.3 | First |
| IV | Every consequential delegation creates a correction obligation | 2.0 | First |
| II | Authority decays unless its justification is renewed | 1.7 | Second |
| III | Evidence and authority must remain coupled | 1.7 | Second |
| XI | Successful automation increases its own governance burden | 1.2 | Second |
| VI | Nominal reversibility is not operational reversibility | 1.1 | Second |
| VII | Error-bearing parties require standing proportional to exposure | 0.8 | Second |
| XII | No system may erase the conditions of its own contestability | 0.8 | Second |
| V | Dependence converts technical risk into structural risk | 0.7 | Second |
Consequence
What this means for Ethotechnics
Where the method should center, given what the labs have already absorbed.
| Laws the labs have largely converged on | Laws with almost no public doctrine anywhere |
|---|---|
|
|
On this reading, the first-generation laws are already being absorbed by the labs: capability versus authority, evaluation at the layer where harm emerges, observation that triggers state change, causally meaningful oversight, and correction obligations.
On this reading, the second-generation laws have almost no public doctrine anywhere: dependence as risk, standing for error-bearing parties, preservation of contestability, operational rather than nominal reversibility, success raising the governance burden, and authority renewal and evidence coupling.
If Ethotechnics centers on permissions, containment, approval, and evals, it describes a layer the labs are already shipping. If it centers on the dynamics of delegated authority after deployment, it describes a layer none of the ten labs in this scan was found to publish. That is a finding about the scanned set, whose sources are not yet listed, not a claim that the layer exists nowhere: parts of it, such as continuing review, safe decommissioning, and incorporating adjudicated feedback, already appear in established guidance.
The object model, standards, and mechanisms therefore put the second-generation objects (dependency record, standing register, reversibility ladder) ahead of the first-generation ones (capability catalog, grant states) wherever the two do not depend on each other. The laws themselves are stated on the laws page.
Sources and refresh
How the scan is sourced and when it should be treated as stale.
No score has a source attached yet. When sources are added, each score will cite the public document or product surface it was read from. Until then, read the scan as the Institute's structured reading, not as a sourced dataset.
No refresh is scheduled. If no refresh has been published by June 2027, treat this scan as stale.
Cite this page · APA, MLA, Chicago, BibTeX, RIS
APA
Ethotechnics Institute. (2026). Frontier doctrine scan (Sep 2026). Ethotechnics Institute. https://ethotechnics.org/research/frontier-doctrine-scan
MLA
Ethotechnics Institute. "Frontier doctrine scan (Sep 2026)." Ethotechnics Institute, 2026, https://ethotechnics.org/research/frontier-doctrine-scan.
Chicago
Ethotechnics Institute. "Frontier doctrine scan (Sep 2026)." Ethotechnics Institute. Sep 6, 2026. https://ethotechnics.org/research/frontier-doctrine-scan.
BibTeX
@misc{ethotechnics_research_frontier_doctrine_scan,
title={Frontier doctrine scan (Sep 2026)},
author={Ethotechnics Institute},
year={2026},
howpublished={Ethotechnics Institute},
url={https://ethotechnics.org/research/frontier-doctrine-scan},
version={1.0.0}
}
RIS
TY - WEB TI - Frontier doctrine scan (Sep 2026) AU - Ethotechnics Institute PY - 2026 UR - https://ethotechnics.org/research/frontier-doctrine-scan ER -