🧅 Onion · Pure-symbolic AGI

Self-contained · Transparent · Auditable · 0% pre-trained LLM

Phase 4 COMPLETE 99.33% honesty (measured) KG 16,067,852 nodes Phase 5.5 Wikidata in flight

📊 Live Bench Scoreboard 🛡️ vs Claude · 7-Dim Moat ⭐ GitHub

📐 The Moat · Why Onion is Different

Per CLAUDE.md design rule "เหนือ Claude 10+ มิติ" — these are structural advantages, not vendor-marketing. All Onion numbers measured from internal benches; refresh on every bench run.

KG Nodes

16.07M

Lossless · grows forever · no prune

Honesty

99.33%

vs ~15-30% Claude halluc. (est.)

Trace

100%

Source-node + evidence + confidence

Cost / query

Owned hardware · no API tokens

Privacy

100% local

NO_EXTERNAL_API audited

Determinism

structural

Same input → same output (CI gate Phase 6.7)

Frames / Lex

512 / 19.7K

FrameNet-style · TH + EN coverage

Lifelong learning

continuous

+89,189 nodes today

📊 Public Dashboards

Two live dashboards — refreshed by scripts/refresh_bench_scoreboard.py from coverage JSON snapshots.

📊 Bench Scoreboard

Phase 4/5 production benches · KG state · rd 327 object-reasoning · regression gates 4/4 · composer EN/TH · live API health card.

Open dashboard →

🛡️ Onion vs Claude · 7-Dim Moat

Defensibility scoreboard with honest framing · measured Onion metrics · estimated Claude metrics · weaknesses surfaced without spin.

Open dashboard →

✍️ Composer Demo

Frame-semantics-driven NLG composer · TH + EN · 12 example frames offline · pure-symbolic · zero-LLM.

Try the demo →

🔗 Resources

⭐ GitHub · ocorpt/onion-data Open architecture · CLAUDE.md design rules · pipeline modules · 0% pre-trained LLM commitment. 💬 Telegram · @ONION_SystemBot BrainV3 self-contained pipeline · ask Onion directly · no external API in path. ✍️ Web Composer Demo Frame → surface NLG · interactive · offline · TH + EN. 📊 Bench Scoreboard Live Phase 4 / Phase 5 metrics · KG state · regression gates · API health. 🛡️ 7-Dim Defensibility Moat Honesty · trace · cost · privacy · learning · determinism · no-forgetting. 📖 API Docs REST endpoints for the public macllm-api surface.

🏗️ Architecture in One Line

Core reasoning = pure symbolic (rule · KG spreading · Frame Semantics · CBR · Conceptual Blending · Markov n-gram + KN · graph algorithms). Neural supplement allowed only if (a) architecture from neuroscience (Hebbian / STDP / SDR / HTM / SNN), (b) weights from-scratch on our data, (c) supplement role only, (d) degradable — system still answers when neural is off. No Transformer / BERT / LoRA / distillation from closed APIs. Storage lossless · no node prune · no data decay.