[feature] Best-of-N / verifier-in-the-loop selection over the gate

[yuyu04]

What problem does this solve

cladding owns something most harnesses lack: a strong, deterministic, execution-based verifier (the 15-stage gate + spec-conformance). But it's used as a single-pass PASS/FAIL on one attempt. The 2025–26 consensus is that the verifier — not the generator — is now the bottleneck and the moat, and that solution coverage scales with the number of candidate attempts as long as you have a good verifier to select among them. cladding has exactly that verifier and isn't exploiting it.

Proposed shape

A best-of-N mode for the drive loop / clad run:

• Generate K candidate implementations of a feature (varying seed/temperature/persona framing).
• Run the gate on each (already isolated per-feature by modules).
• Select the green candidate — or, when several pass, rank by a spec-conformance rubric (oracle coverage, fewest warn-level findings).
• Keep the winner, discard the rest; record the selection in the audit log.

This makes cladding's gate a selector, not just a judge, and turns its verification rigor into higher first-pass conformance.

Versioning scope (GOVERNANCE.md §2)

• Minor — new drive-loop mode
• possibly Major if it touches the clad run / drive-loop contract — deferring to maintainer scoping, which is why this is an issue first.

In-scope check (GOVERNANCE.md §4.1 / §4.2)

• Not regressing Iron Law conformance — it raises selected-candidate quality
• Not bypassing the anti-self-cert guard — selection is by the gate + an independent rubric, not self-judgment by the author persona
• Not forking the Ironclad spec
• Not cosmetic-only — ships with a test harness exercising K-candidate selection on a fixture feature

Alternatives considered

• Single-attempt + reflect loop (retry on failure) — complementary, not a substitute. Reflect fixes a candidate; best-of-N explores several and selects. They compose.
• LLM-judge selection — rejected as primary: cladding's deterministic gate is a better, cheaper, non-self-certifying selector. An LLM rubric only tie-breaks among gate-green candidates.

Willing to implement?

• Yes
• Open to either — would want maintainer agreement on scope/contract before coding (flagship idea, larger surface).

---

The strategic headline of a competitive-gap analysis: cladding's verifier is its differentiator; best-of-N is how it compounds.