Warranted transfer out of the human cut leaves people the hardest-to-warrant decisions

Type: kb/types/note.md · Tags: foundations, self-improving-systems

Call a decision warrantable for computational supply when the deciding path has the premises it needs, a sufficiently settled objective, criterion, or authority, and evidence capable of rejecting a plausible but harmful candidate. These conditions are relative to a declared task distribution, boundary, objective, horizon, and available oracles.

The claim is conditional. When a system preferentially transfers warrantable decisions out of its human cut, the decisions left with people become enriched for unrepresented premises, unsettled criteria, weak checks, and delayed consequences. Per decision, the residue is harder to warrant than the set that moved.

This is a selection effect, not a claim that the remaining work is essentially human. It also does not establish that real systems usually transfer decisions in this order. That prevalence question needs before-and-after evidence from actual automation paths.

Why the residue changes composition

A formatter can take formatting decisions because its inputs are represented, its rule is settled, and idempotence or comparison with a specification supplies a cheap check. It stops before design or maintenance decisions because those are not merely more examples of formatting. Their premises, criteria, and consequences have a different warrant structure.

Repeat this pattern across mechanisms. Each mechanism takes the slice that its representations, decision rules, and oracles can support. What remains is not a random sample of the original work. It is selected by the reasons transfer failed.

This differs from the elastic-backlog effect in which freed attention moves to new frontier work. The selection effect operates even with the incoming workload held fixed. The human cut may shrink while the average ambiguity, stakes, or evaluation cost of its remaining decisions rises.

The residue is a work list

Why the decision stayed human What must grow before warranted transfer
A required premise is unavailable to the deciding process Representation, retrieval, or acquisition of the premise
The objective, commitment, criterion, or authority does not settle what may be accepted Methodological settlement, an explicit commitment, or a represented grant of authority
No sufficiently independent check can defeat the candidate Verification, decorrelated criticism, delayed exposure, or accepted error-cost tolerance
The decision arises after the automatic path ends Continuity, persistent state, and later reactivation
Transfer is possible but costs more than human judgment No structural capacity; the available transfer is currently uneconomic

The table should be applied to individual decisions, not to job titles. A residual design decision can contain formalizable and checkable substeps, and a nominally automated task can still export its decisive premise or acceptance to a person.

Coherent modification is the hardest delayed-evidence case

At an open-ended program-modification crux, no complete local rule or cheap oracle determines which change preserves the program's purpose and organization. A human theory-holder does not solve that problem by possessing a perfect specification. The person uses a partial program theory to guide search, interpret failure, backtrack, and revise as later evidence arrives.

A computational composite can carry the same decision only by sustaining that longitudinal process. The theory may make a provisional proposal better warranted without settling it immediately. Warrant accumulates through exposure, recovery, and read-back across later demands, because holding a program theory means sustaining coherent search under delayed feedback.

This is where the bearer question and the residue question meet. They still ask different things. The bearer question asks whether the composite can preserve coherent search and recovery. The transfer question asks whether enough of that process and its corrective evidence lie inside the declared boundary to move the decision with warrant.

Consequences for closure and evaluation

Structural computational closure asks where required decisions and transitions occur. A path can be computationally closed while using a captured evaluator, a viability-only gate, or a bad objective. Those failures make the path weak, dangerous, or inadequately warranted; they do not make a decision cross the human boundary.

The selection effect therefore does not imply that an evaluator decides computational closure. It implies that an adequate evaluator is often the binding condition for warranted, non-degenerate closure and for transferring the hardest residual decisions without merely hiding them.

Three common patterns create the appearance of warranted transfer:

  • a captured evaluator lets the candidate supply the decisive standard;
  • a viability-only gate treats build success or non-crash behavior as evidence of improvement; and
  • boundary export leaves demand choice, missing premises, or final acceptance with a person outside the declared path.

The first two can occur inside a structurally closed system. The third is an actual failure of the declared closure claim when the exported decision was required by that declaration. Keeping these cases separate preserves the distinction between usefulness, autonomy, warrant, and power.

The remote-programmer benchmark is likewise a capability comparison under a fixed client. It can show that the worker role moved while task choice, feedback, and acceptance remain declared exports, because holding the client fixed exports the least-warrantable decisions.

What would test the selection claim

A direct test should reconstruct successive transfers on one path while holding the boundary, objective, horizon, and workload as stable as possible. Before each transfer, classify candidate decisions by premise availability, criterion settlement, checkability, delay, and cost. Then compare the decisions that moved with those that remained.

Evidence for the mechanism would show that transferred decisions were more warrantable under the preregistered classification and that the residual mix shifted toward missing premises, unsettled criteria, or weak and delayed checks. A cross-sectional finding that current systems often lack independent evaluation is consistent with the predicted bottleneck, but it does not establish the before-and-after selection effect.

A second test concerns displaced judgment. Compare reviewers who see only escalated residual cases with reviewers who also see routine cases. If the first group performs worse on the hard cases, that supports the concern that transfer can weaken the information environment of the remaining human decisions.

Scope

  • The prediction requires preferential transfer of warrantable decisions. A system that automates whatever is cheap or attempts whatever a model will try can produce a different residue.
  • Warrantability is system-relative and can regress when an oracle proves to be a weak proxy.
  • The claim predicts enrichment, not purity. Warrantable decisions may remain human because transfer is uneconomic or intentionally withheld.
  • Structural closure, capability, and warrant remain separate coordinates.
  • The claim does not imply that residual decisions become impossible to move. It predicts what new representation, settlement, verification, or continuity machinery the next transfer must supply.

Open Questions

  • Do real automation histories show the predicted before-and-after enrichment?
  • Which classification grain avoids counting a large architectural decision and a small formatting choice as comparable units?
  • How should delayed evidence receive credit when several decisions and theory revisions intervene before the consequence appears?
  • Does removing routine cases measurably degrade later human judgment on the escalated residue?

Relevant Notes: