Review and validation
Evolver 2 asset review and code-change review are separate steps. Approving a Gene admits it as a strategy candidate; it does not authorize later edits, publication, or credit spending.
Review candidate assets first
In Evolver, run:
evolver review
This lists auto-distilled Genes awaiting review. Inspect source material, triggers, strategy, applicability, and validation evidence. A title or summary alone is insufficient.
After reviewing, use the actual candidate ID:
evolver review --approve <gene-id>
Approval does not supply successful execution evidence. Distinguish an unproven candidate, an approved candidate, and a verified reuse outcome.
Review the execution plan separately
| Area | Question |
|---|---|
| Input | Does the failure evidence or task material describe this issue? |
| Scope | Are files, commands, network access, and credentials authorized? |
| Method | Do the Gene’s prerequisites match dependencies, platform, and project? |
| Verification | Do checks prove the outcome, rather than just return zero? |
| Recovery | Is the baseline recorded and pre-existing work protected? |
| External effects | Are publishing, spending, and resident operation separately authorized? |
Inspect evolver cycle capabilities --json before choosing how to execute. Imported sessions and installed hooks do not prove that built-in autonomous execution is enabled.
Rejecting a candidate
Keep its asset ID, provenance, rejection reason, and missing evidence. Correct the material or plan and review again. Do not relax repository-wide permissions to run an unclear candidate.
The old README’s evolver --review example is not a substitute for this v2 asset-review flow. Preserve code-diff review, validation results, and publication decisions separately.
Example: review a repair strategy
For an overflow Gene, obtain the real candidate ID from evolver review, open its source and content, and compare:
| Object | Acceptance evidence | Return for revision |
|---|---|---|
| Signals and preconditions | Reproducible overflow with source and runtime | Vague aesthetic complaint only |
| Strategy | Identifies a width constraint and checks interaction regressions | Hides all overflow to conceal clipped content |
| Impact | Explicit file and directory limits | Requires a repository reset or unrestricted dependency changes |
| Validation | Reported viewport, menus, and desktop layout exercised | Requires only a model completion message |
| Provenance | Original task, diff, and actual output are traceable | Confidence or an unsupported success claim only |
Approve that candidate ID only after acceptance. Approval allows selection; verification after execution still requires actual project artifacts and check output. Retain rejected candidates and reasons and review revised content again. Editing outcomes or deleting failures does not create success.
For diff operations, see Code review and changes. This page owns asset judgment and execution-verification decisions.