Unconstrained CTI-REALM with Open-Weights

URL SCAN: Unconstrained CTI-REALM with Open-Weights
FIRST LINE: Computer Science > Artificial Intelligence

The Dissection

This is a benchmark correction and procurement memo disguised as model evaluation. Its findings are blunt: open-weight models can match closed models when given enough turns, at a fraction of the cost; benchmark limits can measure completion speed rather than capability; and decomposing the workflow enables model routing by stage.

The deeper event is not “better model selection.” A security-intelligence workflow has been decomposed into measurable cognitive stages that can be assigned to the cheapest competent machine. The analyst is being converted from producer into exception handler.

The Core Fallacy

The paper frames the decisive issue as open versus closed models, cost, turns, and routing. The Discontinuity Thesis asks the more destructive question: when the model, telemetry store, scorer, and orchestration layer perform the workflow, what economically necessary human function remains?

The abstract supplies evidence for local cognitive substitutability, but treats it as an engineering optimization rather than a labor-displacement mechanism. It also mistakes benchmark success for operational autonomy. The results do not establish P1 across cognitive work or P3 at societal scale—but they move one specialized domain closer to both.

Hidden Assumptions

• CTI-REALM performance transfers to messy production environments.
• Additional model turns are cheap enough and fast enough to tolerate.
• Answer-key correctness transfers to live operational reliability.
• Human supervision, verification, and maintenance remain inexpensive rather than becoming the next targets for automation.
• Open weights equal distributed control, despite inference compute, energy, telemetry, and deployment infrastructure remaining scarce.
• Security organizations will preserve human roles instead of routing each stage to the cheapest capable model.

Social Function

Partial truth, transition management, and prestige signaling. The paper correctly exposes closed-model scarcity as a weak moat. Its language—cost, routing, checkpoints, tokens—turns labor removal into a respectable systems-engineering problem. The political rupture disappears behind a benchmark table.

The Verdict

This is evidence against the durability of closed frontier-model advantage and evidence for cheaper, more modular automation. Under DT logic, it is an accelerant: open weights reduce cost, stage routing increases utilization, and unconstrained turns remove an artificial brake.

The paper maps part of the machinery of displacement while pretending it is merely a better purchasing decision. It does not prove general economic collapse. It makes that collapse cheaper to deploy.