ArcAsha (Akasha-OS)

An AI Operating System for Modular Reasoning and Runtime Intelligence

ArcAsha is not a model. It is an operating system that runs on top of neural models — it configures, controls, measures, and explains AI reasoning at the OS level.

• We do not modify the model.
• We place an OS layer outside the model to manage intelligence: routing, memory, reasoning, scheduling, and self-improvement.

Core research question: Can we compose, control, and measure intelligence at the OS level — and prove it reproducibly — rather than scaling the model alone?

---

🎯 Why ArcAsha

Models (GPT / MoE) perform all reasoning inside the neural network — a black box.

ArcAsha moves reasoning outside the model:

Task → Compiler → AILSM IR → Kernel → Executive → Hypothesis → Search → Experts → Memory

• AILSM / AILSA: AI-specific IR & ISA (the "machine code" of reasoning)
• AVM: AI Virtual Memory (only the needed context is loaded, like demand paging)
• Executive / Meta Executive: who commands the whole reasoning process
• Intelligence Attachments: advanced intelligence loaded only when needed (like optional kernel modules)

---

🏗️ Architecture (3 Layers)

Layer 3  Intelligence Attachments
         Reflection / Debate / Planning / Search / Creativity / Simulation / Coding
Layer 2  Executive Runtime
         Executive / Meta Executive / Expert Evolution / Intelligence Scheduler
Layer 1  Fast Runtime
         Kernel / AVM / Expert Runtime / ODAR / Device Tree   ← realtime, always fast

• Fast vs Deliberation: Fast keeps realtime control (robot: 30.3 fps), Deliberation loads only when needed (research / long reasoning).

---

✨ Key Features

---

🚀 Quickstart

# Install (after publishing) or run from repo
npm install arcasha

# Full benchmark suite (Simulation) + Decision Explanation + Real Device plan + reports/
arcasha benchmark

# "Why did the AI choose this?" — replay the decision process step by step
arcasha replay

# OS Policy Learning — decisions become training data for the Meta Executive
arcasha policy

Or run from the repo (workspace root — convenience scripts delegate to the core package):

npm run setup           # npm install --prefix akasha-master
npm run selftest        # AILSM 89 deterministic tests
npm run benchmark       # full benchmark + reports/ (json/csv/md)
npm run arcasha -- benchmark

Or work directly in the core package:

cd akasha-master
npm install
npm run ailsm:selftest    # 89 deterministic tests
npx tsx examples/quickstart.ts   # 5-minute tour

---

🧩 Two Projects

This repository hosts two independent projects (boundaries clarified, git split pending):

akasha-link/PROTOCOL.md は両プロジェクト間の共有ワイヤ契約です（48B ヘッダ + f32[] ペイロード）。 注意: 現行の client-web ワーカーは従来の 20 バイト層ヘッダ（txId + layerId）を使用しており、 PROTOCOL.md 形式への移行は別途追跡されます。

---

📁 Repository Layout

akasha-master/        Project B: ArcAsha-Core / MetaOS (TypeScript / AILSA / AILSM / Kernel / AVM / Executive / Attachments / Caravan)
akasha-link/          Project A: Akasha-Link (distributed inference / tensor transport)
  ├── client-web/     Web client (WebGPU inference)
  ├── kernel-native/  Native kernel prototype (Rust: GPU compute / QUIC / TCP / memory pool)
  └── PROTOCOL.md     Akasha Wire Protocol (48B header + f32[] payload — the shared contract)
examples/             Attachment examples (code / math)
.github/              Issue templates + CI workflow
AI_*.md               Specifications (ArcAsha-Core, see below)

---

📚 Documentation

---

📊 Phase 4 — Real-API Validation

Phase 4 validates each component with real API calls (deepseek-v4-flash, measured, no fabricated numbers), comparing configurations on the same tasks and model. Full data: akasha-master/reports/ablation/.

Component ablation (50 tasks × 3 runs)

• AVM ON vs OFF significance via McNemar test (discordant b=2 / c=0, two-sided p=0.50 — no significant difference, and no regression)
• Per-task detail: akasha-master/reports/ablation/ablation.md (authoritative). akasha-master/reports/ablation/ablation-quick.md is a pre-separation quick measurement (12 tasks)

Long-context AVM (12,668 chars / 396 pages)

• 96.5% token reduction / 94.7% cost reduction at 100% accuracy (page supply 39/396 = 9.8%)
• Boundary-crossing search misses fixed by page overlap (slide window); search precision improved by IDF weighting (akasha-master/reports/ablation-long/)

Executive bottleneck (50 tasks)

• Measurement exposed a double model-call bug under forceDelegate (12% of tasks made a second empty/duplicate call) → fixed
• After fix: every task calls the model exactly once; Executive latency delta +348ms → +37ms (+348ms = pre-fix delta measured in PR #37; +37ms matches the ablation table: ③+Executive 1334ms − ①Baseline 1297ms); TS-side overhead ≈ 0.2ms (akasha-master/reports/ablation-exec/)

🤖 SWE-bench Real-Problem Validation (coding agent)

ArcAsha's software-engineering agent (akasha-master/src/arcasha/swe/) solved real SWE-bench Lite instances with real API calls (deepseek-v4-flash, temperature=0). Details: akasha-master/README.md.

• 3 sympy instances selected from SWE-bench Lite (test split): 24213 (dimension equivalence), 23117 (Array([])), 24152 (TensorProduct.expand)
• Result: 3/3 resolved (100%) — 26/29/11 model calls, 31/44/13 tools, 93s/206s/71s per instance; agent edits source only and verifies with pytest; test files are write-protected (gold test_patch is applied at eval time)
• Honest caveats: stochastic LLM output (one instance needed a retry); 3 selected tasks is not a statistical estimate; token usage was not captured in this run (usage aggregation now added to the harness — future runs record it in akasha-master/reports/swebench/swebench-results.json)

Plain DeepSeek vs arcasha (1-instance controlled comparison, 2026-09)

To quantify what the agent/tool layer adds, we compared a plain deepseek-v4-flash call (issue + file excerpt, one-shot unified diff) against the arcasha agent on the same instance sympy__sympy-24213. Real API measurements; details: akasha-master/reports/swebench/compare-deepseek-vs-arcasha.md.

• Plain DeepSeek: 0/3 resolved — it always identified the correct fix (identical to gold), but hand-written unified diffs had wrong hunk line counts / truncated context → git apply failed every trial (43,184 tokens / $0.027 off-peak / 267 s)
• arcasha: 1/1 resolved — edits files directly via edit_file, so the diff is generated by git (no hand-written hunk math) → always applicable (741,409 tokens / $0.170 off-peak / 127 s)
• Takeaway: solving SWE-bench needs not just "knowing the fix" but a tool to apply it to real files; plain-model one-shot diff generation is possible but unreliable

---

🧪 Status

• v1.0 released — AI OS first generation (Phases 0-4: ISA/IR/Kernel/AVM → Realtime devices → Reasoning → Executive/Meta → Attachments → Validation)
• v1.1 — Decision Replay, Real Device benchmark plan (Mac / iPhone 15 Pro / iPad M4)
• Phase 4 real-API validation (2026-09) — component ablation (Baseline/AVM/Executive/Full, 50 tasks × 3) + long-context AVM (96.5% token reduction at 100%) + Executive bottleneck (double-call bug fixed: +348ms → +37ms)
• SWE-bench real-problem validation (2026-09) — SWE-bench coding agent solved 3 selected sympy instances from SWE-bench Lite: 3/3 resolved (100%) (deepseek-v4-flash; see akasha-master/README.md)
• selftest [1]-[89] all pass / golden 30 / AILSA selftest / build + dist verified

---

🔬 Research Positioning

ArcAsha is not "a bigger model". It is:

An experimental platform to compose, control, and measure AI intelligence at the OS level — reproducibly.

The most novel point: the OS can explain why Reflection / Planning / Debate were used (Decision Explanation), replay the whole decision process (Decision Replay), and learn from its own decisions (OS Policy Learning) — a training axis orthogonal to Transformer pretraining.

License

MIT — see LICENSE.