Review before review — why and how
Before a human reads it, an independent reviewer attacks the work from several angles and issues a verdict. The human review then receives an already-filtered artifact — not a draft.
The problem (concrete context)
When the first read of a change is done by a human, it spends the team’s most expensive time on defects a systematic check could have caught: a test that asserts nothing, a dependency violating a boundary, a forgotten acceptance criterion. The reviewer tires on the trivial and misses the subtle. The cycle stretches with every round-trip.
The review-before-review idea: insert an assisted adversarial review between production and human review. It does not replace the human — it hands them work already cleared of detectable defects.
What the sources say
Wiegers grounds every review in explicit criteria and in the effect of multiple eyes.
« Peer reviews are the single most effective quality practice a software organization can employ. » — Wiegers, K., Peer Reviews in Software, 2002.
And on the value of multiple, independent eyes:
« The combined attention of several reviewers finds defects a single reader misses. » — Wiegers, K., Peer Reviews in Software, 2002.
Application in SKRAFT
Each phase has an independent reviewer (never the producer). In DELIVER, this reviewer applies 4 independent lenses then a weighted synthesis (Genesis A7 pattern).
Produced work (phase N)
│
▼
┌──────────────────────────── Independent reviewer ──────────────────────────┐
│ Lens 1 cold-reader ──► verdict + findings │
│ Lens 2 architecture-boundaries ──► verdict + findings │
│ Lens 3 test-integrity ──► verdict + findings │
│ Lens 4 quality-gates ──► verdict + findings │
│ │ │
│ ▼ weighted synthesis │
│ APPROVED / CHANGES_REQUESTED / REJECTED │
└─────────────────────────────────────────────────────────────────────────────┘
│
▼
(if APPROVED) phase transition ──► human review on a clean artifact
A single BLOCKER on one lens is enough to return the work (CHANGES_REQUESTED), with a
bounded number of retries before human escalation. The verdict and its findings are
traced — the human review sees why it passed.
Conditional lenses in DELIVER
The panel is not fixed: on top of the 4 CORE lenses, fidelity lenses join when the
software-engineer has delegated test wiring to a worker. If mock-integration-worker
wired a mock, mock-fidelity-lens joins the panel; if contract-testing-worker wired
a contract test, contract-fidelity-lens does. Each attacks the fidelity of the
wiring (does the mock/contract truly reflect the dependency?) and honours the same
BLOCKER rule. See the
DELIVER fan-out and the
lenses reference.
Pitfalls & anti-patterns
- Reviewer = producer: the review loses its independence and falsifying power.
- Averaging synthesis: aggregating verdicts into an “average” instead of honouring BLOCKERs lets the critical one slip.
- Infinite loop: without a retry bound, an agent/reviewer disagreement never reaches the human — SKRAFT bounds and escalates.
Sources
- Wiegers, K. Peer Reviews in Software, 2002.
Going further: Review lenses, Gates, Review before review (principle).