sdr_status_2026-04-10
SDR (Scientific Deep Research) — Status Brief 2026-04-10
Generated by Feanor worker, 2026-04-10 07:05 MSK
Current State Summary
Deploy: ✅ dev.ai4s.pro (Neuler + DataChat deployed 30 мар)
Tests: ✅ 106/106 passing (venv python3.13)
DRACO score: claude-opus 50.3/100 (best, published)
Sprint 2-30: Done (20 MCP tools)
Sprint 31: 🔴 BLOCKED — waiting for OPENROUTER_API_KEY
Active Blockers (all ask-level, Danil’s decision)
| Blocker | Impact | Review item |
|---|---|---|
| OPENROUTER_API_KEY | Sprint 31 unblocked; better LLM for judge eval | /review/inbox/api_key_openrouter_proposal.md |
| ANTHROPIC_API_KEY | eval judge using claude-haiku → DRACO ~60+ (vs 50.3 now) | backlog: SDR верифицировать ANTHROPIC_API_KEY |
| eval branch commit OK | eval/ dir merge → eval scores trackable in CI | /review/inbox/sdr_eval_anthropic_judge.md |
DRACO Benchmark History
| Date | Model | Score | Status |
|---|---|---|---|
| Mar 2026 | llama-3.1-8b | ~8.7/30 | Stuck (no keys) |
| Mar 2026 | claude-opus (DRACO best) | 50.3/100 | Published ✅ |
| Sprint 31 target | claude-haiku + openrouter | 55-60+/100 | Needs keys |
Recent Activity
- 2 апр: Данил Шерки разработал PaperVizAgent — агентный фреймворк для научных иллюстраций (5 агентов: Retriever/Planner/Stylist/Visualizer/Critic, 3 раунда). Смерджен в dev. Review:
/review/inbox/danil_sherki_papervizagent_dev.md - 2 апр: IAP dataset вопрос от Данила Шерки — создание датасета из IAP+DFT (Perelman). Review:
/review/inbox/danil_sherki_iap_dataset_question.md - 30 мар: dev.ai4s.pro задеплоен (DRACO bench опубликован)
Critical Path to DRACO >55
OPENROUTER_API_KEY → Sprint 31 run → new bench score
ANTHROPIC_API_KEY → eval judge upgrade → cleaner eval
eval commit OK → eval/ in CI → reproducible scores
All 3 are single-decision items from Danil. One conversation can unblock everything.
Connection to Other Projects
- PhD: SDR = умный литобзор → ускоряет написание §7 (обзор методов)
- RNF grant: сильный DRACO score = демонстрация AI4Science impact
- Perelman/MetaReview: SDR как компонент поиска статей для extraction pipeline
- OPTIMA 2026 (Sep, Petrovac): SDR бенчмарк = потенциальный contribution
Next Steps (priority order)
- Даниил: OK на OPENROUTER_API_KEY + eval commit (one decision session)
- Фёанор: после ОК — запустить Sprint 31, DRACO bench
- Данил Шерки: review PaperVizAgent → merge main или отдельная ветка?
- Фёанор + Данил Ш.: PaperVizAgent integration в Neuler.DeepResearch