メインコンテンツまでスキップ

Human Judgment Focus

River Review is not a tool for replacing human review with AI. It is a framework that executes your team's review criteria as versioned skills, so that humans can focus on the judgments that truly need them.

"A review happened" is not "understanding is shared"

Confirming an implementation diff does not, on its own, guarantee shared system understanding or shared design intent. A diff shows what changed, but not why the change is sound or whether it is consistent with prior design decisions.

At the same time, having a human synchronize the full context on every PR does not scale.

  • High cognitive load on reviewers
  • Increased lead time from review queues
  • Review perspectives become tied to individual reviewers
  • Re-reading the full context from scratch on every review

River Review lowers this burden while letting humans concentrate on high-risk judgment.

Reallocate human attention

River Review does not try to reduce Human Judgment itself. It tries to reduce the amount of scarce human attention spent on decisions that can be handled reproducibly elsewhere.

Judgments are placed according to the Judgment Placement principle, which is the SSoT for the four layer definitions. The list below is a summary.

Mechanically provable
-> Deterministic

Reliably detectable by explicit rules
-> Heuristic

Requires semantic / contextual judgment
-> Agentic Review

Requires responsibility / value / irreversibility judgment
-> Human Judgment

When Human or Agentic Review repeatedly catches the same issue, River Review should ask whether that judgment can be promoted into a test, checker, rule, or heuristic. Conversely, critical security or irreversible decisions are not moved to AI merely because some part of the workflow can be automated.

The objective is not fewer human review minutes as a KPI. It is a higher density of consequential human judgment.

What River Review reduces / does not replace

The goal of River Review is not to replace human judgment, but to focus it on high-risk areas.

CategoryContent
What River Review reducesIndiscriminate human sync / shallow diff-only review / reviewer-specific tacit knowledge / re-reading context from zero
What River Review does not replaceFinal accountability / design decisions / business validity / critical security judgment / approval of irreversible changes

River Review is a mechanism for increasing the evidence available for a decision — not for delegating responsibility. How much human supervision applies is not uniform: it is allocated across three risk tiers — cliff, hill, and field.

Review allocation by risk tier

Weight your review effort across three tiers according to the risk of the change. The lower the risk, the more you lean on River Review skills; the higher the risk, the heavier the human supervision. The SSoT for the tier definitions is the "Direction of travel: risk-tiered human supervision" section of Design Philosophy.

Risk tierLean onto River ReviewHuman supervision
Cliff (ESCALATE)detection of auth / payment / personal data / security boundary / irreversible migrationHuman approval is mandatory; escalate whenever anything is unclear
Hill (GO_WITH_OBSERVATION)plan-diff conformance / tests / migration policy / API contractTime-boxed observation, plus a check of design intent and blast radius
Field (GO)lint / format / naming / docs / simple refactorsAutonomous convergence, audited after the fact via run-record / digest

The common Low / Medium / High risk wording maps onto field / hill / cliff respectively.

For changes on the cliff, River Review goes as far as detection and presenting the evidence. The final "pass / block" decision is made by humans.

Use cases

River Review reduces repetitive review synchronization and lets human reviewers focus on high-risk judgment.

  • Plan Review — detect dangerous gaps in requirements, design, and plan before implementation
  • Diff Review — confirm the implementation diff is consistent with the plan, design, and test policy
  • Test Review — confirm tests are sufficient against the spec and the risks
  • Review Comment Review — re-examine whether existing AI or human review comments are valid (see W-check)

Do not over-trust AI review

River Review increases the available evidence, but it is not a complete substitute for human approval. The following areas require human judgment.

  • Security boundaries, authentication, and authorization
  • Personal data, payments, and data migration
  • Irreversible changes

Avoid merging based solely on AI review results in these risk areas. River Review's verdicts (merge-ready / human-review / block) support human judgment; they do not stand in for it.