Requirements — Answer-Quality Lab (2×2)

Panels (UI: 4 lanes, each independently judged)

Matrix orientation: rows = retrieval, cols = architecture; numbering row-major. - P1 — Single · keyword · ALGOLIA_WWW_PROD_V2_SINGLE_KEYWORD - P2 — Multi · keyword · ALGOLIA_WWW_PROD_V2_MULTI_KEYWORD - P3 — Single · neural · ALGOLIA_WWW_PROD_V2_SINGLE_NEURAL - P4 — Multi · neural · ALGOLIA_WWW_PROD_V2_MULTI_NEURAL (bottom-right, the full-power cell)

Removed entirely: the algolia.com website panel and the Ask AI panel (and their wiring).

Indices

• App: FLAGSHIP_Accelerator_Program_APP (NeuralSearch ON).
• Seed: fresh copy of Visibility ALGOLIA_WWW_PROD_V2 (1QDAWL72TQ, read-only), English only, latest snapshot available at build time, deduped by url.
• Keyword vs neural = retrieval config (mode) on identical records.
• Per-index configuration is independent and data-proven per setup; common baseline first.
• Case-3 tuned config (synonyms + keyword levers) is the starting config for SINGLE_KEYWORD only; keyword-only levers do NOT go onto the neural indices. See 2026-06-18-flagship-app-and-four-indices.

Single-agent flow (Panels 1–2)

• Reuse the current Case-3 Agent-Studio agent (hardened grounding instructions + the false-refusal/query-reformulation fix), repointed at the SINGLE indices.
• Same agent config across Panels 1 and 2 — only the index/retrieval differs.

Multi-agent flow (Panels 3–4) — content-source specialists

• Maverick = orchestrator/interface: answers first to best ability, asks follow-ups ("peel the onion"), extracts entities, routes; keeps a broad/unscoped index view; can fan out to multiple specialists in parallel and synthesize.
• Specialists = shared MULTI index scoped by source facet filter. Locked roster (names final):
• Technical: source ∈ {Documentation, Developers, Customer Stories} — docs, APIs, integration guides, best practices, customer case studies.
• Marketer: source ∈ {Website, Blog, Resources, Other} — website/marketing, news, blogs, pricing, security.
• Academy: source = Academy.
• Support: source = Support.
• Specialists query the MULTI keyword index (Panel 3) or MULTI neural index (Panel 4). Same partition both panels.
• See 2026-06-18-content-source-multi-agent.

Judge

• Judge + Autocorrect loop carried over (provider-agnostic, zero-flicker, temp-0, claim-recurrence gate).
• Scores each panel: grounding (hard-floor + scored) · confidence · breadth/depth; mean of 3 judges.
• Output: per-query live 4-lane comparison + batch aggregate 2×2 scorecard over the locked question set (batch = authoritative, full sources). See 2026-06-18-judge-3dim-per-panel.

Daily delta-sync (NEW)

• Incremental, data-only sync from Visibility ALGOLIA_WWW_PROD_V2 → all four FLAGSHIP indices.
• Never modifies index config (each index keeps its own settings/mode); upserts record deltas only, English-only, deduped by url.
• See 2026-06-18-daily-delta-sync.

Acceptance (to refine)

• All four panels answer the locked question set, fully grounded (bait-query clean).
• Judge produces a reproducible 2×2 scorecard.
• Daily sync runs unattended and preserves per-index config.
• Question set must contain items where multi-agent/neural can actually differentiate (else "no difference" is the honest result).