---
name: aris-subq-technoedge-2026-05-16
description: 2026-05-16 STEP 1156-followup-10. chat-Claude TechnoEdge weekly proposal (ARIS Auto-Research-In-Sleep + SubQ 12M context LLM) Rei honest filter. ARIS = research-radar 新 entry (Pattern 5
metadata: 
  node_type: memory
  type: feedback
  originSessionId: 31de7a8f-b68c-479d-b055-229c7945fc2c
---

# ARIS + SubQ TechnoEdge proposal — honest filter

## chat-Claude proposal (2026-05-16 藤本さん経由)

TechnoEdge AI weekly 記事から 2 件:
- **SubQ** (Subquadratic Miami startup): 12M token context, SSA (Subquadratic Sparse Attention), 2026-05-05 release, $29M / $500M valuation, waitlist only
- **ARIS** (Auto-Research-In-Sleep): Claude Code × Codex MCP (GPT-5.5) cross-model research framework, Markdown skills, v0.4.6 2026-05-14 release

## fact-check 結果 (WebFetch + WebSearch verified)

| Claim | Status |
|---|---|
| `github.com/wanshuiyin/Auto-claude-code-research-in-sleep` 実在 | ✅ 9.4k stars, MIT |
| ARIS-Code v0.4.6 2026-05-14 release | ✅ accurate |
| ARIS cross-model (Claude executor + GPT reviewer) | ✅ accurate |
| ARIS skills system (Markdown-only) | ✅ accurate, **67 skills** (chat-Claude 「62」 は older changelog) |
| SubQ 実在 (Subquadratic, $29M / $500M val) | ✅ VentureBeat / SiliconANGLE / 公式 subq.ai 確認 |
| SubQ 12M context / SSA / 2026-05-05 release | ✅ accurate |
| SubQ FlashAttention-2 比 1M 52x speedup | ✅ 公式 claim (independent verify 未) |
| SubQ "researchers demand independent proof" 懐疑 | ✅ VentureBeat headline accurate |

## chat-Claude Pattern 2 stale 数値 (累積 7 件検出)

| chat-Claude | 実態 verified |
|---|---|
| SubQ cost「**1/5**」 | ❌ 実 **1/20** (VentureBeat: "~1/20th of Claude Opus") |
| ARIS「**62 skills**」 | ❌ 実 **67 skills** (62 は older changelog) |
| ARIS reviewer「**GPT-5**」 | ❌ 実 **GPT-5.5** (WebFetch 確認) |
| SEED_KERNEL「**1,500** 理論」 | ❌ 実 **1,601** (累積 stale, 既観測) |
| Lean4「**31,000** 定理」 | ❌ 実 **2,255** (28,745 delta, Mathlib cumulative 誤帰属の可能性) |
| 「Zenodo 論文 **61本**」 | ❌ 実 **153本** (92 delta) |
| 「**Paper 62-65** + Paper **66** 候補」 | ❌ 実 Paper 153 publish 済 (88 delta) |

→ `feedback_chat_claude_hallucination_warning.md` Pattern 2 累積 18 → **25 候補** (前 batch 含む).

## Pattern 5 累積 15 例目 — ARIS philosophy 既実装

chat-Claude framing: 「Claude駆動 × 外部 LLM 批判レビュアー」 + 「世界の最前線と並走している証拠」

**実態 (verified)**:
| ARIS feature | Rei 既実装 |
|---|---|
| Cross-model review loop | `src/aios/auto-prover/private/ensemble-voter.ts` (REI-PROVE 5-prover: Goedel + DeepSeek-Prover-V2 + Vampire + LeanHammer + BFS, 92% benchmark, STEP 1064 production-ready) |
| Multi-model coordination | `src/axiom-os/multi-agent-orchestrator.ts` (chat-Claude も認知済 ✅) |
| Self-evolution `/meta-optimize` | `src/axiom-os/infinite-evolution-engine.ts` (chat-Claude 正確に identify ✅) |
| Local LLM (Ollama / LM Studio) | DeepSeek-Prover-V2 + Goedel-Prover-V2:8b + bfs-prover GGUF 既 install (STEP 1057) |
| Research wiki | SEED_KERNEL 1,601 + META-DB 4,306 |
| External AI dialogue | `src/aios/external/external-ai-dialogue-engine.ts` |
| MCP integration | 30+ MCP tools (mcp__rei-project__moe_route 等) |

= 「並走」 でなく **「同じ場所」**. ARIS が独立に同 architecture 到達 = validation, Rei 採用 path は full fork でなく **skill design inspiration** (Markdown-only skill UX layer).

## Pattern 5 累積 16 例目 — SubQ は v1.9 で既 watch 化

chat-Claude proposal: 「SubQ は種だけ植えて経過観察」

**実態**: `data/research-radar/collatz-watch.json` line 349 (v1.9, 2026-05-14, STEP 1135) で **既 watch 化済**:
```json
{
  "fullName": "subq.ai",
  "category": "post-transformer-attention",
  "priority": "medium",
  "notes": "Subquadratic SubQ 1M-Preview (2026-05-05 release, $29M / $500M valuation)..."
}
```

= chat-Claude session memory なし default で「種だけ」 提案 = Pattern 5 systemic 16 例目.

## 採用判断 (option A 採用 2026-05-16)

- ✅ **option (A)**: research-radar watch (collatz-watch.json) に ARIS 追加 + memory seed
- 工数: ~15 min (JSON edit + 本 file)
- SubQ は既 entry の状態確認のみ (no edit)
- ARIS fork sandbox は **defer** (Rei 既 REI-PROVE ensemble と重複 risk 大)

### collatz-watch.json v1.10 update

- Schema バージョン v1.9 → v1.10
- ARIS 新 entry (`wanshuiyin/Auto-claude-code-research-in-sleep`):
  - category: `multi-agent-tooling`
  - priority: medium
  - notes: 9.4k stars / MIT / v0.4.6 / 67 skills / Pattern 5 重複多数明示 / inspiration 価値 mention
  - factCheckedAt: 2026-05-16 STEP 1156-followup-10 (WebFetch 9.4k stars / GPT-5.5 訂正記録)
- description 更新で v1.10 changelog 追加
- repos count 45 → 46

## 起動 trigger (ARIS adoption を defer から繰り上げる条件)

- Rei REI-PROVE ensemble が specific use case で ARIS skill design に劣ると判明 OR
- ARIS skill design pattern が Rei MCP + ReiLauncher 強化 inspiration として有用と判断 OR
- 藤本さん明示 commit

それまでは watch 観察のみ.

## 関連 reference

- `data/research-radar/collatz-watch.json` v1.10 (本 STEP で update)
- `src/aios/auto-prover/private/ensemble-voter.ts` (REI-PROVE 既稼働)
- `src/axiom-os/infinite-evolution-engine.ts` (chat-Claude も identify)
- `src/axiom-os/multi-agent-orchestrator.ts` (chat-Claude も identify)
- `feedback_chat_claude_hallucination_warning.md` (Pattern 2 累積 25 候補)
- `feedback_vscode_agents_window_2026-05-16.md` (前日 STEP 1156-followup-8 同 pattern)