Designed friction and cognitive partnership

The design response is cognitive forcing functions — humans at the helm, not merely in the loop. The concrete anti-sycophancy mechanism is the Delphi pattern: human and model develop answers independently, then combine and iterate. It works in both directions, against groupthink on the model’s side and against the model’s own flattery. A related structural move is drawing the reviewer from a different model family, so the check is not a copy of the thing being checked.

Before AI, small reviewed changes kept your mental model, the system, and your colleagues’ models roughly in sync. Now one person can raise a palace of stuff they do not actually know, and everyone’s models diverge. So “you don’t need to review pull requests” is wrong — not because quality control demands it, but because engaging with the mental model does. That reframes review from a QA ritual into a maintainability requirement.

We teach children with explanations and teach machines by punishing wrong answers. Distillation works precisely because of reasoning traces; the field’s return to supervised fine-tuning and distillation is a rediscovery of top-down explanation. The same holds for people: an explanation of why an answer was wrong teaches where a bare rejection does not.

Wave through the too-big change. Do it yourself rather than teach a junior. Default to the biggest model. Wait for models to get better. Skip sandboxing. “I can’t work on this codebase without AI anymore.” The management question attached is the uncomfortable one: if you are a person of influence, are you creating an environment that leads to surrender — where people crank pull requests with no time to improve the harness?

Also: the Delphi method.

Assay’s review loop already implements the Delphi pattern by construction: the reviewing App forms its verdict independently of the worker’s self-report, and the worker answers findings with evidence rather than deference. Draft-only pull requests, plus bot review, plus a human merge gate is deliberate friction in exactly the sense this evidence argues for. And the mental-model argument is the best articulation available of why the human merge survives even when the bot approves cleanly: the point is not catching a defect the reviewer missed, it is keeping the driver’s model attached to a system agents are building faster than one person can read. See desk roles and the lifecycle.

The durable correction here is a skill edit, a standing house rule, or a findings entry carrying the rationale — not a bare rejection. That is the top-down-explanation point, already practice.

Cross-family diversity is the clearest gap: reviewer and worker typically run the same model family, which weakens the independence the Delphi structure is supposed to buy, and nothing in the roster or the review loop enforces otherwise. Second, there is no counterpart to “can the reviewer restate what this pull request does and why” — a verdict currently asserts a judgment without demonstrating the mental-model sync that review is supposed to exist for; making that an explicit verdict element is designed and unbuilt. And the counter-position — a team dropping human review entirely in favour of several agentic passes and self-merge — is the strongest argument against Assay’s position in this scan. It is recorded here rather than dismissed: reported as working, in a high-adoption organization, with no defect data attached.

• Whether “mindful friction” attracts any measured result — a matched comparison of a frictionless versus deliberately-gated flow on the same task — rather than staying a design stance.
• Whether cross-family reviewer diversity is measured anywhere: does a different-family reviewer actually catch findings a same-family reviewer misses?
• Whether the no-human-review position accumulates evidence for or against it. That is the claim which, if it held up with defect data, would retire this concept.