adversarial-review-lenses

Procedure for executing 4 independent lenses and weighted synthesis to produce a defensible adversarial verdict on a phase artefact.

When to use

Every SKRAFT reviewer agent invokes this skill once per review pass, after reading the upstream phase artefact(s)
The number of lenses executed is governed by the depthTier recorded in state.json

Depth tier	Lenses required
`basic`	1 (Completeness only)
`standard`	2 (Completeness + Business Fit)
`comprehensive` (default)	4 (all lenses)
`custom`	as configured in `userPreferences.customDepth.reviewerLenses`, minimum 1

Entry contract

Artefact(s) of the phase under review (relative path)
The phase’s *-review-criteria skill (read before executing lenses)
state.json::userPreferences.depthTier to determine the number of lenses to execute

Exit contract

Review file under reviews/{YYYY-MM-DD}/{phase}-{slug}-review.md
Verdict: APPROVED, NEEDS_REWORK, or REJECTED
Weighted score computed from the 4 lenses (fixed weights)
List of required actions when verdict is NEEDS_REWORK or REJECTED

Invariants

No cross-lens contamination — findings from one lens must not influence another
A single INVARIANT_VIOLATION in Lens 4 forces REJECTED regardless of weighted sum
The reviewer never modifies upstream artefacts — read-only
Fixed lens weights: Completeness 0.30 — Business Fit 0.30 — Quality 0.15 — Risk 0.25

Lens	Weight	Finding tags
1 — Completeness	0.30	`MISSING`, `THIN`, `OK`
2 — Business Fit	0.30	`MISALIGNED`, `AMBIGUOUS`, `OK`
3 — Quality	0.15	`BROKEN`, `INCONSISTENT`, `OK`
4 — Risk	0.25	`INVARIANT_VIOLATION`, `HIDDEN_COUPLING`, `AMBIGUOUS_ASSUMPTION`, `OK`

Verdict table:

Weighted sum	Verdict
>= 0.85 and no lens scored 0.0	`APPROVED`
>= 0.55	`NEEDS_REWORK`
< 0.55, or any lens scored 0.0 on an invariant	`REJECTED`

Why this shape

Adversarial review draws on the Genesis Step 7 pattern: independent judge panels produce more reliable verdicts than collegial reviews where conformity bias suppresses dissent. The weighted synthesis preserves the dominance of business lenses (Completeness + Business Fit = 60%) over structural ones.

« Peer reviews consistently find more defects per hour than any other technique. » — Wiegers, K., Peer Reviews in Software, 2002.

Allowed customisation

Depth tier (via state.json — L1)
Minimum lenses for custom (minimum 1 — L2)
Output file format (L1)