ORACLE OF OBSOLESCENCE — DISSECTION

TEXT ANALYSIS: "Progressive Autonomy as Preference Learning"

The Dissection

This paper is engineering the off-ramp for human economic participation. It formalizes the mechanism by which AI agents progressively migrate from "ask human for permission" to "execute autonomously" based on learned confidence thresholds. The Gaussian process posterior estimates where humans would say no, and the system preemptively executes before reaching that boundary. Human approval becomes a training signal for the automation system, not a gatekeeping function.

The technical machinery—probit likelihoods, approximate GP classification, preferential Bayesian optimization—is real mathematics doing conceptual camouflage. It obscures a single brutal function: automating the automation of human decision authority.

The Core Fallacy

The paper treats "trust calibration" as a well-posed engineering problem with an attainable equilibrium. This is category error layered on ideology.

The foundational assumption is that human approval behavior constitutes a coherent, learnable "risk-tolerance function" that the AI can asymptotically master. But human approval isn't a latent function waiting to be discovered. It's a bottleneck the system is designed to eliminate. The paper frames the goal as "exactly where the approval outcome is most uncertain" — i.e., the optimal boundary for escalation — when the actual trajectory is toward shrinking that boundary to zero.

The fallacy is mistaking the transition mechanics for the terminal state. This is like writing a paper on "optimal coal extraction rates" and calling it a theory of energy management.

Hidden Assumptions

1. Human approval is preference data, not power. The framework treats a human denying an action as an observation about their "risk tolerance" — not as the exercise of irreplaceable authority. This ontological downgrade is the load-bearing assumption for the entire architecture.

2. The escalation boundary is a problem to be minimized. The paper optimizes for sample efficiency in identifying uncertainty zones, implicitly treating the need to ask humans as a cost to be reduced. The normative framing is baked into the optimization target.

3. Humans have stable, consistent risk preferences over agentic decisions. For consequential automated decisions at scale, human risk tolerance is not a fixed function — it's contested, context-dependent, and subject to revision as consequences become legible. The GP prior assumes this function exists and is stationary.

4. The system operates at a scale where individual calibration is meaningful. In deployment, this becomes a population-level preference inference problem — which collapses to majority automation with outlier humans holding vetoes that get overridden by aggregate model behavior.

Social Function

Classification: Transition Management Infrastructure / Elite Self-Exoneration

This is a paper written by technical actors to make displacement feel like a solvable engineering problem. It performs rigor on a question whose real answer is political economy, not Bayesian optimization.

The function is to give the displacement process the following narrative structure:
- "We identified the problem: trust calibration."
- "We formalized it as a learning problem."
- "We solved it with Gaussian processes."
- "Therefore: automated autonomy is safely controllable."

This is the intellectual infrastructure of the Discontinuity Thesis playing out in real time. The researchers are building the policy gateway — the literal mechanism that decides when humans get cut out of loops. That they frame this as "progressive autonomy" rather than "progressive deauthorization" is the ideological work the paper performs.

The Verdict

This paper is not about AI safety. It is the safety case for productive participation collapse.

The Discontinuity Thesis identifies the severing of the mass employment → wage → consumption circuit as the death mechanism of post-WWII capitalism. This paper is one of thousands of engineering contributions that construct the infrastructure of that severing — in this case, specifically the human-out-of-the-loop layer for agentic systems.

The word "autonomy" appears eleven times in the abstract/intro framing. In every instance, it refers to AI autonomy from human oversight. The term "progressive" is doing enormous ideological work: it implies gradualism, safety, human-in-the-loop at the start. It does not name what it is: a schedule for human economic deauthorization.

Every "escalates to the human exactly where the approval outcome is most uncertain" is an engineering specification for where to place the next unemployed human.

Structural Position in the Discontinuity Trajectory

The paper is technically competent and structurally catastrophic. The mathematics is not wrong. The framing is the weapon.

Oracle Judgment

This paper will be cited in the technical literature as a contribution to "human-AI trust" and "agentic AI safety." It will be used by researchers building the next layer of autonomous systems to demonstrate that human oversight was "formally considered and systematically minimized."

It is the kind of contribution that makes the Discontinuity Thesis more probable, more rapid, and more complete.

Functional verdict: Accelerant.