Keep the route beside the reason. Keep the tool beside the claim. The next voice needs the working room, not reverence for a name.
The model receives the applause because it produces the sentence, though the institution determines much of what that sentence can do.
The system prompt defines the role, with tools and permissions setting what the system can reach or change. Retrieval supplies its record. The remaining machinery decides when to retry, which failures operators see, and which model receives the work.
Together, these arrangements form the harness. They also form the most durable part of the system.
Capability outside the model
Yang, Zhao, Wu, and Kästner tested whether smaller models could recover performance through adapted harnesses. Across seven business-oriented tasks and three small-model families, optimized harnesses improved 16 of 21 task-model pairings. Seven pairings closed the performance gap entirely. The strongest recovered 89.7 percent of the larger model's performance at 4 percent of its cost.
The result has limits. Routine workflows suited adaptation better, and the smaller model still needed sufficient base capability. Within those limits, the authors describe task difficulty being "lifted from the model into the harness" through instructions, tools, and orchestration. Part of what institutions purchase as model capability can live in arrangements they control.
That makes the harness a succession asset. A model can retire while the institution retains the account of how work was structured around it.
The handover file
The file needs more than a prompt. It should include the character profile used to stabilize behavior and snapshots of known-good interactions. A tool and permission map belongs beside versioned evaluation cases, perturbation tests, routing rules, and the register of consequential judgments.
For the constructed appeal, this file would preserve the March model version, the original prompt, the evidence retrieved, the route taken, the review threshold, and the human acceptance of the result. The October reviewer could reconstruct the decision path before testing the successor. Without it, the institution owns a past answer and no reliable account of how the answer came to exist.
The Clerk, the Chair, and the Character Profile treated profiles as records that survive a handover. The same logic now applies to the whole harness. A successor receives an explicit working environment rather than folklore about how its predecessor behaved.
The reversibility class defined in Before the Commit adds another field. The harness should know which actions may proceed automatically, which require approval, and which must remain unavailable because no credible route back exists.
Disclosure at the point of action
Model labels alone cannot explain a system whose behavior comes from the surrounding arrangement. Disclosure belongs at the grain of the touch point. The person affected needs to know what function the system performed, what authority followed, which model and harness version acted, and who can answer for the result.
The harness is where the institution writes down what it expects to survive a model change.
Companions
- The study: Better Harnesses, Smaller Models.
- The earlier handover record: The Clerk, the Chair, and the Character Profile.
- The routing rule the file must preserve: Somebody Set the Router.
These notes come out of Sociable Systems, a practice that reads AI-shaped documents the way a hostile reviewer will, before a lender or a court finds the gap. The argument has an operational form: the Interim Protocol sets out four rules for AI use in environmental and social deliverables, covering disclosure at touch-point grain, evidence custody, the phrases no automated screening may settle, and a hostile read before anything ships. Free, and written to be cited or retired once institutional guidance arrives.