Grounds, Frames, and Standing: The Absorbable and Irreducible Classes of Human Intervention in Agentic Work
Task-level studies of AI exposure record a human-AI interaction as one automation-versus-augmentation bit, and an index built on that bit can report stability across exactly the period in which the character of human work inverts. The bit has a composition, and the parts of the composition have different fates under capability growth. This paper derives those fates.
The premise is a thirteen-mechanism inventory of human interventions in the work of LLM agents, mined in a companion paper from seven months of one practitioner's records. The inventory is the only import: the discovery corpus, the survey, and the reliability measurements stay with the companion, and no derivation here rests on them. If the inventory is approximately right, whoever assembled it and in whatever domain, the results hold. If it is not, the results degrade from derivations to observations of one practice.
Mechanisms sort into three durability classes by what they supply. Grounds, considerations bearing on entitlement to conclusions, live in the world, so retrieval, wider context, and stronger verification reach them, at a rate set by the domain's verifier availability rather than by capability alone. Frames, replacements of the hypothesis space, require a position outside the current framing, so the class yields to a second reasoner with decorrelated priors, and the absorber is a population fact rather than a capability fact. Under monoculture that route closes, and because a frame cannot be requested by name, remaining frame supply turns on a decision to invoke a source, which is threshold-setting and relocates the burden onto the third class. Standing, what the zero-content stop signal, judgment demand, and purpose re-anchor transfer, is the partition's irreducible class, absorbed by neither capability growth nor architecture. A stop's timing and sign reveal the threshold's value, a fact learnable from a record, while the authority to set and revise the threshold does not reduce to a fact, so a stop an agent issues to itself is a decision rather than a stop. Standing is not self-conferrable under the conferral semantics institutions currently enforce, which is a governance fact with a repeal path rather than a theorem, and the requirement the class leaves is for a principal, not for a human, a role architecture can relocate but not eliminate.
Under incomplete contracting, capability improves inference from a fixed information set without enlarging it, so expected loss from a misset delegation threshold converges not to zero but to the conditional variance of the principal's threshold, a strictly positive floor. The floor rests on either of two premises, non-enumerability that persists under learning or preference construction in the threshold itself, so the framework's sharpest falsifier is the failure of both premises together. The mechanism list transfers across substrates while its couplings do not, and a minority of mechanisms run with the sign of their human-literature analog inverted, changing what the move does rather than how much of it is needed. The policy-facing corollary is compositional: the grounds-shaped fraction of human input should decline as capability grows, the standing-shaped fraction should not, and the recorded bit is blind to both movements at once.
PDF · DOI: 10.5281/zenodo.21563154 · v3.33
This paper was one document with its companion through v2.11. The split follows the argument's dependency structure rather than its section numbering: the entailments rest on the mechanism inventory being approximately right, not on the corpus that produced it, so the inventory is taken here as a premise and cited. Readers arriving from an earlier citation will find the corpus, the coding scheme, the reliability measurements, and the novelty claims in the companion.
A fifteen-page treatment of the same argument, carrying the pilot evidence and the research program without the formal appendix or the full methodological record, is available as a research proposal.