Adversarial Review
Reviewing work with an explicit mandate to refute it, from a perspective that did not produce it.
Adversarial review is critique with an assignment to disprove rather than to approve. The framing matters because a reviewer asked "does this look right?" is disposed to agree, while a reviewer asked "find where this is wrong, default to rejecting if unsure" applies real pressure.
Two structural requirements make it work.
The reviewer must be cold. A critic that inherits the author's context inherits the author's assumptions and blind spots. Self-review reliably declares work satisfactory and misses the things that most needed catching — dead exports, inconsistent naming, an assertion that never runs. Where a system delegates implementation, review has to come from a fresh perspective rather than from the implementer's own continuation.
Independent verdicts must be aggregated, not chained. Several critics reviewing the same finding through different lenses — correctness, security, does-it-actually-reproduce — catch failure modes that three identical passes cannot. Requiring a majority to survive filters plausible-but-wrong findings, which is the dominant failure mode of automated review.
The same discipline applies to humans reviewing machine output, and to the reviewer's own reasoning: a lint rule set to error plus a clean run feels like execution, but it is configuration plus an inference across a gap nobody tested. See Plausible Mechanism and Code Review.
See also8
Subagent
A model instance spawned by another to handle a scoped task with its own context.
Agents & Language Models11 connections
Multi-Agent Orchestration
Coordinating several model instances on one body of work, each with its own context.
Agents & Language Models11 connections
LLM-as-Judge
Using a language model to score another model's output against a rubric.
Agents & Language Models12 connections
Falsifiability
A claim is only worth something if you know what observation would refute it.
Method47 connections
Provenance
The recorded origin of a change or a claim — who made it, when, and on what evidence.
Method44 connections
Ground Truth
The measurement that arbitrates, as opposed to the artifact that merely reports.
Method99 connections
Exhaustive Claim
A statement of the form "the last one" or "nothing else does this" — load-bearing, and only as good as the search behind it.
Method23 connections
Model Card
A short standard document published alongside a machine learning model describing what it's for, how it was trained and evaluated, where it performs worse, and what it shouldn't be used for, now routine on model hubs and for frontier releases.
Agents & Language Models5 connections
Linked from9
- Automation BiasMethod
Over-trusting an automated recommendation, including against available contrary evidence.
- Code ReviewVersion Control & Delivery
A second person reading a change before it lands, and the practices that make it worth the time.
- LLM-as-JudgeAgents & Language Models
Using a language model to score another model's output against a rubric.
- Model CardAgents & Language Models
A short standard document published alongside a machine learning model describing what it's for, how it was trained and evaluated, where it performs worse, and what it shouldn't be used for, now routine on model hubs and for frontier releases.
- Multi-Agent OrchestrationAgents & Language Models
Coordinating several model instances on one body of work, each with its own context.
- Plausible MechanismMethod
A causal explanation that was inferred rather than tested, and reads as more rigorous for being specific.
- Segmented SleepBody & Medicine
The claim that pre-industrial night sleep came in two blocks separated by a waking interval, and the dispute over how general that was.
- SubagentAgents & Language Models
A model instance spawned by another to handle a scoped task with its own context.
- Threat ModelingTesting & Verification
A structured way of asking what could go wrong with a system's security before an attacker answers it, by drawing where data flows, marking where trust changes, and walking each crossing through a list of threats.