Algolia-Central2 — Operation Log

[2026-06-18] new | Algolia-Central2 Answer-Quality Lab

Files touched: index.md, log.md, CLAUDE.md, wiki/overview.md, wiki/requirements.md, wiki/open-questions.md, wiki/dev-log.md, wiki/decisions/2026-06-18-pivot-to-2x2-four-panel-lab.md, wiki/decisions/2026-06-18-flagship-app-and-four-indices.md, wiki/decisions/2026-06-18-content-source-multi-agent.md, wiki/decisions/2026-06-18-judge-3dim-per-panel.md, wiki/decisions/2026-06-18-daily-delta-sync.md Notes: Wiki initialized and populated directly from an in-depth design discussion (interview skipped — content already gathered). Captured the pivot to a 2×2 four-panel lab, the move to FLAGSHIP_Accelerator_Program_APP with four indices, the content-source multi-agent design (data-validated against the live Visibility index), the judge model, and the new daily delta-sync requirement. Five ADRs written. Build not yet started.

[2026-06-18] record | Finalized agent roster + v2 principle

Files touched: wiki/decisions/2026-06-18-content-source-multi-agent.md, wiki/requirements.md, wiki/open-questions.md, index.md, log.md Notes: Agent names final — Maverick (orchestrator) · Technical · Marketer · Academy · Support. Locked the governing principle: decisions are data-driven, not legacy-driven; rc2/rc3 = legacy POC, Algolia-Central2 = v2. Dropped rc2's SE/SA persona carry-over.

[2026-06-18] record | Lab UX + judge panel design + dedup profile

Files touched: wiki/decisions/2026-06-18-ux-and-judge-design.md, wiki/dev-log.md, wiki/open-questions.md, index.md, log.md Notes: Full lab UX designed (2×2 grid scoreboard; score-click slide-in pinnable judge drawer; lean judge panel = 3 always-on + 2 expanders, grounding cards on violation). Judging tiers: 1 judge live (indicative) / 3 judges batch (authoritative, supermajority gate). Grounding = gate at ×1 (recommended, pending confirm). Source index dedup-profiled (en + url-dedup ≈8,100). Full mockups in repo docs/refactor/ux-design-v1.md. Pending: Scorecard view design + grounding-×1 confirm, then write plan → UltraCode.

[2026-06-18 evening] record | THE PLAN written + hardened; app/key/naming corrected

Files touched: repo docs/superpowers/plans/2026-06-18-answer-quality-lab-2x2-refactor.md (new), repo SESSION.md, repo CLAUDE.md, memory project-build-home-central-app.md Notes: Build plan WRITTEN — 7 phases, per-phase model tags, self-contained, real code + acceptance criteria. CORRECTIONS that override these ADRs + the brief: (1) build home = CENTRAL app 0EXRPAXB56 (ALGOLIA_* keys), NOT a separate FLAGSHIP app — NeuralSearch + Agent Studio verified live by probe; ALGOLIA_ADMIN_API_KEY in place + write-verified (lacks only addApiKey). (2) Index naming = AC2_WWW_<ARCH>_<RETRIEVAL> (P1 SINGLE_KEYWORD / P2 MULTI_KEYWORD / P3 SINGLE_NEURAL / P4 MULTI_NEURAL); agents ac2-*. Replaces ALGOLIA_WWW_PROD_V2_*. (3) Grounding gate ×1 CONFIRMED. Then 5 review gaps closed (Arijit): Maverick = CODED coordinator (lab/server) + native source-scoped specialist Agent Studio agents (8) · register BOTH OpenAI+Gemini providers on 0EXRPAXB56 (per-app IDs) · unified /api/answer SSE endpoint (single answer path, keys server-side) · multi-turn = GENERATED follow-up judged via followUpQuality · FAIRNESS INVARIANT = identical model+provider+grounding across all 4 panels + coordinator. NEXT = fresh UltraCode build session (Phase 0→7), /mode acceptEdits. NOTE: the wiki ADRs/brief still say "FLAGSHIP" + old index names — the repo PLAN is authoritative.

[2026-06-28] record | Backlog A (expandedQuery drop) + B (native memory) + 2 ADRs

Files touched: index.md, log.md, wiki/dev-log.md, wiki/open-questions.md, wiki/decisions/2026-06-28-drop-expandedquery-from-retrieval.md, wiki/decisions/2026-06-28-stateless-replay-vs-redis.md Notes: A shipped (raw turn to agent) — main dc03bc7; B negative (keep manual replay, no Redis) + ADR-002 — main 03e9184. Lessons in repo docs/sop/lessons-log.md. Only P2b (human-rank calibration) remains.

[2026-06-28 PM] record | content-source routing spike verdict + data-hardening + honed-prompt phase

Files touched: wiki/decisions/2026-06-28-content-source-routing-spike-verdict.md (new), wiki/dev-log.md, wiki/open-questions.md, index.md Notes: Spike KILLED content-source SCOPING (fair baseline +0.25 noise; first run's +0.76 was baseline-blindness). Data-hardened n=1000/source: depth only in Support + "Other"; corpus is metadata. Honed data-realistic prompts drafted + warm-baton design. NEXT: push Support honed prompt + adherence smoke, then full warm-baton multi-turn A/B.