Source
Kahl (2026), LIT-1033, read in NOTE-723: §3 (the presupposition argument), §4 (continuity pressure), §5 (the test and its cases, with the group-agency reply in §5.3 and the borderline cases in §5.5).
What was actually shown
An argument, with cases. Nothing is measured.
Two senses of agency. Operational autonomy is task execution with little prompting, and Kahl accepts the word “agentic” for it. Attributive agency is agency “in the sense relevant to responsibility, authority, and governance”. The paper is about the move from the first to the second.
The presupposition argument. Decision theory, planning theory and hierarchical endorsement all take a unified evaluator as given. Dung’s capability dimensions measure “what a system can do”, not “what is at stake for the system in doing it”. So none of them can supply the evaluator, and using them as sufficient for agency “assumes what it ought to explain”.
The condition. Continuity pressure “begins where an inherited evaluative frame no longer settles how future operation should proceed.” Authorship attaches when the system settles a revised frame and later assessment proceeds through it. It is absent “where revision is externally imposed, where rollback avoids carrying forward the consequences of revision, or where the system merely selects among revision paths already ranked by a fixed higher-order criterion.”
The cases. Tool-using agents fail by resetability, reinforcement learners by an “antecedent admissibility structure” over every update, multi-agent workflows because “unity of action is not unity of authorship”, and institutional systems because their standards are “authored elsewhere”. A skill-library agent and a continually learning clinical system come closer and still fail. The paper says what would count: a system that treats failure “as evidence that the governing evaluative frame no longer settles what should count as success”, settles a revision under breakdown, and carries it forward.
Group agents. A workflow may be a group agent in List and Pettit’s sense. What separates it from a court is “whether revision of that rule under breakdown is internal or external to the candidate group agent”, not whether the rule was designed.
What this does not say
- It is a necessary condition, not a theory of agency. Kahl says so. A system meeting it is not thereby responsible.
- It is not minimal agency. THEORY-071’s conditions make a bacterium an agent. This account is about who can bear responsibility and authority, and it drops the organisational grounding: continuity here is “not numerical identity, personhood, or selfhood”. The two answer different questions. Both deny that optimising an externally fixed function makes an agent, and they reach that by different routes.
- It does not say current systems cannot meet it in principle. The negative verdicts are about typical deployments, and the paper declines to rule artificial agency out.
- It does not make “settled under breakdown” precise. Every revision has some prior structure. How much antecedent ranking disqualifies a revision is not said, and the paper gives no positive case.
- It is not an instruction for building systems. The governance sections recommend tracing responsibility to the people who set standards. That is advice to institutions about where accountability lies, and it is not part of this account. Nothing here says how to build a system that would pass.
Connections
- QUESTION-092 (endogenous governance). A partial answer, and not the one the question expected. Internal allocation of authority among components does not count: a workflow’s “voting rules, ranking functions, confidence thresholds” coordinate execution and “do not generate authorship of the governing frame”. What counts is whether the system revises its own integration rule and standards under breakdown. So the scheduler’s location is beside the point, and the location of revision is the point. Two behaviourally matched systems differ on this account when only one can revise its own rule, which is the question’s “stronger sense”, located in revision of standards rather than in allocation of authority.
- CLAIM-288 (fixed objectives cannot show self-constitution). Corroborates it from a different direction; the claim does not stand on it. Kahl’s argument is about responsibility, not self-constitution, and cites no enactive work. It reaches the same verdict on systems whose objectives are fixed from outside, and on self-modification bounded by “an antecedent meta-objective”. It is stricter about the claim’s second answer: a modeller’s objectives that the population later revises would count only if the revision is not ranked by a rule the modeller wrote.
- CASE-081 (learned self-maintenance). Kahl’s account settles the case’s question of origin against present organisation in favour of present organisation, as the group-agency reply shows. The second stage counts only if the system settles revised standards under breakdown. Running on a learned policy after the reward stops is not that.
- CLAIM-285 and its aggregation rule. An endorsement scheme whose aggregation rule no operator controls is, on this account, an orchestration regime, unless the population can revise the rule under breakdown.
- CLAIM-116. Consistent with it. Adaptation, meta-learning and self-modification are regulation within a frame; on this account none establishes an author of the frame.
- CLAIM-025. Continuity pressure resembles the capacity to transform the governing organization. The paper does not show the two coincide, and nothing here rests on it.
- Benthall and Salge (LIT-1031, not read) argue from enactive constitutive autonomy to a condition on legal responsibility. The two papers reach neighbouring verdicts on current systems by unrelated criteria.
Note of 2026-10-28: bearing on QUESTION-092
At the owner’s ruling on the fourth status review, this account bears on QUESTION-092 and does not answer it. Its agency is the kind that bears on responsibility, and the question fixes TERM-045’s sense 5, self-governing. The bridge between the two is the record’s own claim, filed the same day, which answers the question and is grounded in this account.
Note of 2026-10-29: what holds it back
No positive case. The paper gives none, and the line every verdict turns on, between settling a revised frame and selecting among revisions an antecedent criterion already ranks, is not drawn precisely. The record holds no system shown to settle a revision not ranked in advance, together with an independent reason to hold it responsible. What Kahl proposes, as he states it, is held in THEORY-1165.