ADR: Judge scores all four panels (3-dimension composite)

Date: 2026-06-18 Status: Accepted

Context

With the pivot to a 2×2 over our own system, the judge is the mechanism that decides which architecture wins. The judge + autocorrect loop already exists (provider-agnostic, zero-flicker, temp-0, claim-recurrence gate, 3-dimension composite).

Decision

Keep the judge and the autocorrect loop. The judge rates each of the four panels and gives a score/confidence per panel using the existing 3-dimension composite: grounding (hard-floor + scored) · confidence · breadth/depth, as the mean of 3 judges. Output is a per-query live 4-lane comparison plus a batch aggregate 2×2 scorecard over the locked question set (batch = authoritative, full sources; the live panel is thin-source-indicative).

Rationale

• The dimensional breakdown is what makes "neural beats keyword / multi beats single" defensible rather than asserted.
• Reuses proven, reproducible infrastructure; the analysis rail already renders it.
• A single confidence number per lane was considered but rejected — it hides why a panel wins.

Alternatives Considered

Consequences

• UI analysis rail extends from 3 lanes to 4.
• The question set must include items where neural/multi-agent can differentiate, or the scorecard honestly shows "no difference."
• Live judge stays fast/streamed (flash, 1 round); batch judge stays authoritative (pro, full sources).