Public preprint · Methods phase complete
From Formal Authority to Practical Human Control
Public preprint; repository release v0.18.0. The November editorial revision, submitted source package and unchanged October reproducibility supplement are available on Zenodo. This revision changes presentation and disclosure, not findings; the research account below retains its 16 October 2026 source assessment. The arXiv submission was updated on November 1, 2026 and remains on hold; no public arXiv article identifier or acceptance is established.
What makes human control practical?
A person may have authority to intervene without usable information, enough time or a route for changing what a system does. This study asks what evidence would establish that human authority had practical force in one consequential decision.
Practical Human Control examines six necessary conditions: information access, comprehension capacity, intervention authority, intervention feasibility, exercised judgment and execution propagation. Trust, Autonomy & Evidence (TAE) is the broader research architecture.
A bounded methods demonstration
The demonstration uses three deliberately selected historical cases: the Soviet Oko early-warning case and two Patriot engagements. One researcher assessed the historical records. Later AI-assisted collaborative review led to an author-approved reassessment; it did not add an independent assessor validation.
The findings concern the evidence available within each declared boundary. They do not estimate how often human control succeeds, establish causal effects or demonstrate transfer to contemporary AI systems.
How the findings changed
The original assessment produced two Fail findings and one Unresolved finding. Limited evidence corrections changed that result to one Fail and two Unresolved. The approved additional reassessment leaves all three Unresolved across the 64 retained combinations of partial and indeterminate judgments.
| Account | Oko | Patriot ZG710 | Patriot F/A-18C |
|---|---|---|---|
| Historical baseline | Unresolved | Fail | Fail |
| Limited corrections · 16 assignments | Unresolved | Fail | Unresolved |
| Additional reassessment · 64 assignments | Unresolved | Unresolved | Unresolved |
Unresolved establishes neither that control worked nor that it was absent. A justified Unsupported finding for any necessary condition still produces Fail. Uncertainty does not override that rule.
The 64 assignments enumerate retained alternatives; they are not 64 cases, independent observations or a probability distribution. The broader Stage 3 proposals remain unadopted.
Keep the uncertain judgments visible
P means partially supported: some required evidence is present, with a material gap. I means indeterminate: the record is insufficient to decide. P / I below identifies retained alternative judgments, not an extra evidence state.
| Case | Access | Comprehension | Authority | Feasibility | Exercise | Propagation |
|---|---|---|---|---|---|---|
| Oko | P | P | P | P | P | P |
| Patriot ZG710 | P | I | P / I | P / I | P / I | P / I |
| Patriot F/A-18C | P | I | P / I | I | I | P / I |
The additional reassessment changes ZG710 comprehension, feasibility and exercise relative to the limited correction. Its other alternatives and the F/A-18C corrections remain dependencies. These categorical findings produce no aggregate control score.
The completed phase and open questions
The closed phase comprises the bounded methods demonstration, approved reassessment and initial Jev experiments. The records make the procedure, changes and limits available for examination.
The first live Jev experiment retained 260 valid and 22 invalid responses from 282 attempts. Primary reconstruction agreement was 198 of 231 valid determinations (85.7%), with coverage of 231 of 252 (91.7%). These are different denominators. The 1,079 original adjudication records remain unresolved; later assisted author review does not close that backlog.
Independent application, reliability across assessors, transfer to current systems, and measured improvements in human review time or quality remain future work. Software checks and model agreement do not establish those outcomes.
What an institution can examine
Start with one consequential decision and identify the records needed at each condition. Separate a demonstrated failure from an evidentiary gap. If a review changes the finding, retain the old assessment, the changed rule or evidence, and the reason for the correction.
The present study supplies a method and an inspectable demonstration. It does not certify safety, compliance or the effectiveness of a deployed control.
Paper, evidence and version history
- November paper, source and offline reproducibility supplementEditorial revision and publication: November 1, 2026. The October supplement is unchanged. Paper DOI: 10.5281/zenodo.23852566.
- November preprint PDF on GitHubPinned manuscript copy; separate from the historical October edition.
- Archived October paper and supplementRevision 10 October 2026; published 16 October. This DOI identifies the earlier edition.
- Repository archive v0.18.0A separate archive of the research repository; not the paper DOI.
- Repository publication and research statusPinned November 1 editorial update and arXiv status; the Zenodo record above identifies the subsequent DOI publication.
- November artifact verification recordManuscript integrity and source comparisons; no new empirical validation.
- October reassessment verification recordThe earlier integrity and reproducibility checks retain their declared scope.
- Download citation metadataNames the current paper separately from the repository release.
The earlier findings remain inspectable
The earlier article includes ten source figures, downloadable data and detailed evidence tables. Those figures show the historical assessment states; they have not been relabeled as results of the October reassessment.