Second opinion

Beyond the automatic checks, the Second opinion asks an AI to critique your model for quality — vague names, missing pieces, over-wiring, a workflow that never completes. It's exactly that: a second opinion, always yours to accept, act on, or reject. (The automatic checks catch structure; this reads for meaning.)

There are two ways in:

• Per layer — the Second opinion button in a stage's action bar (next to Generate) reviews just that layer; its findings appear in the stage's own issues panel. This is the quick, in-context check.
• Whole model — the Second opinion card (and the Review whole model button) on Home open the full dashboard below, which reviews every layer top-down plus the cross-cutting pass.

You can also chain it to generation: Settings → AI engine → After generating → "Run the Second opinion automatically" makes each Generate run the review on that layer as soon as it finishes (findings land in the stage's issues panel). It's scoped to the layer you generated — generating resets the layers below it to placeholders, so there's nothing valid to review below yet.

Who reviews. By default the Second opinion uses the same model that generated each layer (at higher effort). Under Settings → AI engine → Reviewer you can point it at a different engine — the LLM-as-judge pattern: e.g. have Anthropic review layers you generated on OpenRouter. A fresh, independent (often stronger) model tends to catch more than the one that wrote the model.

The loop

The Second-opinion dashboard is a launcher and status board, not a place to edit findings. It opens with a one-line summary of what to do next (where to start, or "all reviewed — looks clean"), a thin progress gauge showing how many layers are reviewed and how many are flagged, and one headline action:

1. Review all layers → runs every reviewer top-to-bottom, read-only — a second opinion on the whole model that never changes it. Each layer's row then shows a status and finding count.
2. Click a flagged layer → the dashboard jumps you to that layer's own stage, where its findings are listed in context (next to the map/entities they're about). There you close each one:
   • Fix — apply the AI's suggestion;
   • Ignore — accept it for good (see Ignoring concerns);
   • or click the finding to jump to the element and edit it yourself.

Findings live on the stages, not in the dashboard — so there's one place for each thing, in context. The one exception is the cross-cutting section at the top, whose findings span the whole model and belong to no single stage; those stay in the dashboard.

Everything else is a power tool, tucked behind the Advanced toggle: per-layer model/effort labels and Auto-fix all — a review-and-regenerate loop that drives every flagged layer to clean but changes your model (see Progress & loops). For a single-layer check without the dashboard, each stage also has its own Second opinion button.

It's advisory, and it doesn't converge to zero

The reviewer is calibrated to always find something — a few subjective suggestions is a fine place to stop. Findings are judgment calls, not a checklist that hits zero. The panel says so, and nudges you when you've refined a layer enough.

Only generated layers are reviewable

Until you run Generate on a layer, what you see there is a placeholder — a rough draft derived automatically from the layer above, not a model Kiln has actually worked out. Reviewing a placeholder would just have the AI critique the draft, so the panel gates it: a not-yet-generated layer is shown dimmed, marked "Placeholder — not generated" with "Generate first", and can't be reviewed. Generate it on its own stage and it becomes reviewable. (Review all layers likewise skips placeholders; if every layer is still a placeholder, it tells you there's nothing to review yet.)

Cross-cutting root issues

Above the per-layer list sits a Cross-cutting section — the whole-model pass. It looks across every layer at once for a broken chain: a capability with no entity, a role that owns nothing, an entity no command ever touches. These are root causes — fixing one at its source often clears several per-layer symptoms below, so it's the first thing worth checking. Each finding links straight to the element on the map.

Where to start — work top-down

Your model is a stack: each layer builds on the one above it, so a finding on a lower layer only makes sense once the layers above it are sound. The dashboard's roll-up reflects that dependency order instead of a flat list:

• Layer rows are listed top-down, each showing its status and finding count.
• The highest layer with a real problem (a concern, not just an optional suggestion) is marked Start here — open that one first.
• Layers below an unresolved one are dimmed with "Resolve X first". That's not just tidiness: regenerating an upstream layer (or letting Auto-fix all rebuild it) resets the layers beneath it, so fixing a lower layer first can be undone the moment you fix the one above.
• A layer that was reset by an upstream change is flagged "changed upstream — re-review" rather than a bare "not reviewed" — so you can tell at a glance which layers a quick re-check will confirm. Nothing is re-scanned automatically (that would spend tokens); re-running the review is always yours to trigger.

Open the highest flagged layer, fix its findings on its stage, re-review, and the ones below reappear — usually a shorter list, because fixing a root cause upstream often removes its knock-on findings.

Fix vs. Auto-fix all

There are two "make it better" actions, and the difference matters:

• Fix (per finding, on the stage) applies one precise edit and nothing else — it converges. This is the everyday path: open a flagged layer, and Fix or Ignore each finding.
• Auto-fix all (whole-model, behind the dashboard's Advanced toggle) reviews and regenerates every flagged layer in a loop. It's powerful for driving a sparse model to clean, but because it rebuilds layers it changes your model and can churn — undoing earlier fixes as it rebuilds. Use it early; prefer per-finding Fix to reach a stable, closed state.

The next pages cover both, and how to tell progress from churn.

Cost

Each review, Fix, or Auto-fix that calls the model costs a small amount (shown in the session spend). Running a single-layer review shows a quick cost confirmation first; ticking "don't ask again" (or toggling Settings → AI engine → Reviewer → Confirm cost off) turns it off for good. The whole-model Review all layers confirms up front regardless. This feature needs a real model, so it's available when you run your own instance with an API key — not on the keyless public demo.