開源研究 · Apache-2.0 · AGI 候選驗證套件 · 非已證明 AGI Open research · Apache-2.0 · AGI-candidate proof package · not proven AGI

邁向可驗證的 AGI 與 ASI Toward grounded AGI & ASI

Wisdom before intelligence 智慧先於智能

Sophia AGI 是一套以來源紀律與驗證閘道為核心的開源研究計劃：不在無法查證時胡謅，以公開失敗帳本記錄缺口，並同步建造測量優先的空間世界模型。目標是 AGI / ASI——但從不假裝已經達成。 Sophia AGI is an open research program built on provenance discipline and verifier gates: abstain instead of fabricate, publish a failure ledger of what is still open, and build a measurement-first spatial world model. The aim is AGI / ASI — without pretending we are there.

sophia-agi 原始碼 sophia-agi on GitHub 論文、排行榜與 Ask Sophia → Thesis, leaderboards & Ask Sophia →
−12.5pt 歸因幻覺下降（已驗證，0% 假陽性） Attribution-hallucination drop (validated, 0% FP)
+15.8% SimpleQA 選擇性準確率提升 SimpleQA selective-accuracy lift
528 雙語語料範例 + 公開失敗帳本 Bilingual corpus examples + public failure ledger
v0.12 閘道 · 良心核 · 世界模型伴生 Gate · conscience kernel · world-model companion

現代 AI 會把《道德經》說成孔子、把 1950 年代的想法歸給佛洛伊德，再把錯當前提繼續推理。Sophia 把「來源紀律」做成可機器執行的閘道：查得到才接受，查不到就棄權或攔截——並以同一套誠實標準推進 AGI 形狀的能力與空間世界模型。

Modern AI merges the Dao De Jing with Confucius, credits Freud for 1950s ideas, then reasons on top of the error. Sophia turns source discipline into a machine-enforced gate: accept only what verifies, abstain or block otherwise — and applies the same honesty standard to AGI-shaped capabilities and a spatial world model.

問題所在The Problem

自信，不等於正確 Confidence is not correctness

同一道問題，一般語言模型自信地答錯；Sophia 追溯來源，必要時誠實地拒絕作答。閘道公式：claim → verify → accept · abstain · block。 Same question. A generic model answers confidently — and wrong. Sophia traces sources and abstains when honesty demands it. The gate: claim → verify → accept · abstain · block.

點選一個問題，看看兩者的回答 ↓ Pick a question and compare the answers ↓

提問Prompt
一般語言模型Generic LLM
Sophia

示意互動展示，依來源紀律原則整理；完整推理與驗證流程見 GitHub。 Illustrative interactive demo based on source-discipline principles. Full reasoning and verifier flow on GitHub.

今天能用什麼What ships today

來源可查證的推理層 A provenance-aware reasoning layer

不是另一個聊天機器人——是可接入任何代理管線的治理閘道、語料與基準、以及可離線運行的 MCP 工具面。 Not another chatbot — a governance gate for any agent pipeline, an open corpus and benchmarks, and an offline-capable MCP tool surface.

01
📜
OKF 來源知識庫OKF Provenance Wiki
Open Knowledge Format
機器可讀信念圖譜：authorConfidence、doNotAttributeTo、tradition、矛盾帳本與反事實撤回。900+ wiki 頁，四個人文領域為主。 Machine-readable belief graph with authorConfidence, doNotAttributeTo, tradition fields, contradiction ledger, and counterfactual retraction. 900+ wiki pages across four humanities domains.
02
🔒
驗證閘控推理Verifier-Gated Reasoning
Fail-closed by default
查不到就不裝懂。多重評審、機器檢查器與預先註冊的 no-overclaim 門檻決定什麼可以公開宣稱。 Abstain when unverified. Multi-judge panels, machine checkers, and pre-registered no-overclaim thresholds decide what may be published as fact.
03
⚖️
良心核 + 道德議會Conscience Kernel + Moral Parliament
allow · revise · abstain · block
七路徑認知與道德控制層：不確定性、憲政規則、跨傳統公共道德底線、欺騙訊號偵測。功能控制，非主觀意識。 Seven-path epistemic and moral control layer: uncertainty, constitutional rules, cross-tradition public moral floor, deception signals. Functional control — not subjective consciousness.
04
🔌
MCP · Skills · 契約MCP · Skills · Contract
Drop-in governance
穩定 CONTRACT、MCP 伺服器、skills 層與閘道輸出再驗證。可接 LangGraph、Claude SDK、n8n 或任何 MCP host。 Stable CONTRACT, MCP server, skills layer, and output re-verification. Plugs into LangGraph, Claude SDK, n8n, or any MCP host.
AGI 開發細節AGI development stack

建造 AGI / ASI 的誠實路徑 An honest path toward AGI / ASI

不靠「更大模型」口號。Sophia 在信任層創新：來源、驗證、校準、失敗封閉，以及可量測的自我擴展。以下為目前實作狀態——已接線、部分、或玩具介面皆標示清楚。 Not “scale alone.” Sophia innovates at the trust layer: provenance, verification, calibration, fail-closed control, and measurable self-extension. Status below is honest — live, partial, or toy interface.

範圍聲明：這是邁向 AGI 的研究計劃，不是 AGI 證明。門檻預先註冊且尚未全部達標。canClaimAGI 維持 false。公開失敗帳本記錄每一個 OPEN 阻礙。 Scope: This is a research program toward AGI — not a claim of AGI. Thresholds are pre-registered and not all met. canClaimAGI stays false. A public failure ledger lists every OPEN blocker.
已上線Live

推理核心與審議Reasoning core & deliberation

多後端 LLM、CoT / best-of-N、自我驗證、過程獎勵、部門議會（sector councils）。路徑：agent/model.py、agent/best_of.py、agent/council_deliberate.py。Multi-backend LLMs, CoT / best-of-N, self-verification, process reward, sector councils. Paths: agent/model.py, agent/best_of.py, agent/council_deliberate.py.

已上線Live

信念圖與來源引擎Belief graph & provenance engine

可查詢的 claims/justifications 圖、信心沿鏈傳播、矛盾帳本、反事實移除與撤回。路徑：okf/graph.py、okf/counterfactual.py。Queryable claims/justifications graph, min-over-chain confidence, contradiction ledger, counterfactual removal and retraction. Paths: okf/graph.py, okf/counterfactual.py.

已上線Live

驗證層與閘道Verification layer & gate

神經符號檢查、可執行驗證、引用驗證、graded answer/hedge/abstain。閘道：agent/gate.py、agent/grounded_gate.py、agent/verifiers.py。Neurosymbolic checks, executable verification, citation validation, graded answer/hedge/abstain. Gate: agent/gate.py, agent/grounded_gate.py, agent/verifiers.py.

已上線Live

安全、BLP 與 MCP 閘道Security, BLP & MCP gateway

Bell-LaPadula 分類、防火牆、airgap 設定檔、kill-switch、輸出再驗證。側效工具預分派授權測試覆蓋。Bell-LaPadula labels, dataflow firewall, airgap profile, kill-switch, served-output re-verification. Pre-dispatch authz on side-effecting tools.

部分Partial

記憶、檢索與校準Memory, retrieval & calibration

向量檢索已接入（提交的 lexical-semantic hash 後端 + 可選高品質後端）。分級信心來自來源品質，對正確性預測偏弱——已如實量測。結構化 KG 固化尚未進 live 路徑。Vector recall is live (committed lexical-semantic hash backend + optional higher-quality backends). Graded confidence is a provenance prior, weak as a correctness predictor — measured and disclosed. Structured-KG consolidation not yet in the live path.

離線迴圈Offline loop

自我擴展驗證飛輪Self-extending verification flywheel

棄權 → 找缺口 → 合成 verifier → held-out 驗證 → 晉升或維持棄權。離線確定性迴圈已閉合；live RL 權重更新（需 GPU）仍 OPEN。Abstain → find gap → synthesize verifier → prove on held-out → promote or stay abstained. Offline deterministic loop closed; live RL weight update (GPU) remains OPEN.

玩具介面Toy interfaces

AGI 缺柱（介面 + 參考實作）AGI missing pillars (interfaces + toys)

程式歸納、主動推理、MCTS 規劃、預測世界模型、持續可塑性、分層記憶——皆為可測的 fail-closed 介面與玩具參考，不是其名稱所描述的完整能力。產出標記 candidateOnly: true。Program induction, active inference, MCTS planning, predictive world model, continual plasticity, layered memory — fail-closed interfaces with toy references, not the capabilities their names describe. Artifacts marked candidateOnly: true.

訓練基質Training substrate

RLVR · LoRA · 本地推理RLVR · LoRA · local inference

Verifier-gated 訓練、adapter 晉升閘、Sophia-Wisdom-4B 試點、DGX Spark / 本地 Ollama 路徑。多項 GPU 活跑與第三方 hidden eval 仍在失敗帳本 OPEN。Verifier-gated training, adapter promotion gates, Sophia-Wisdom-4B pilot, DGX Spark / local Ollama paths. Multiple GPU live runs and third-party hidden evals remain OPEN in the failure ledger.

伴生倉庫Companion repo

空間世界模型（Sophia World）Spatial world model (Sophia World)

與推理軌跡世界模型分開：把消費級 360° 影片變成測量優先、來源閘控的數位孿生。詳見下方進度。Separate from the reasoning-trajectory world model: consumer 360° video → measurement-first, provenance-gated digital twins. Progress below.

已驗證成果Validated results

能上標題的數字，才上標題 Only numbers that clear the gate

VALIDATED 需 ≥2 獨立評審家族共識、報告一致性、≥3 次運行與信賴區間。其餘標為 illustrative / candidate。 VALIDATED requires ≥2 independent judge families in consensus, reported agreement, ≥3 runs, and confidence intervals. Everything else is labelled illustrative / candidate.

已驗證Validated

歸因幻覺下降（本地 8B）Attribution-hallucination reduction (local 8B)

36.1% → 23.6%

Δ 12.5 百分點（95% CI [5.6%, 19.4%]），假陽性成本 0%。弱模型現象；強基座上會衰減。第一方 pack 限制已公開標註。Δ 12.5 pp (95% CI [5.6%, 19.4%]) at 0% false-positive cost. Weak-model phenomenon; decays on strong bases. First-party pack bounds disclosed.

已驗證 · 外部資料Validated · external data

SimpleQA 選擇性預測SimpleQA selective prediction

+15.8% / +7.8%

20% coverage 選擇性準確率提升：DeepSeek +15.8% [9.8%, 22.1%]；Qwen-2.5-72B +7.8% [2.3%, 13.5%]。κ ≈ 0.97 / 0.99。校準結果，非 AGI 宣稱。Selective accuracy lift @20% coverage: DeepSeek +15.8% [9.8%, 22.1%]; Qwen-2.5-72B +7.8% [2.3%, 13.5%]. κ ≈ 0.97 / 0.99. Calibration result — not an AGI claim.

校準 / 歷史Calibration / historical

陷阱棄權：0% 虛構Trap abstention: 0% fabrication

0% vs 17–25%

自建棄權校準包上，Sophia 對真正「我不知道」陷阱虛構 0%；原始模型 17–25%。後續 T2 複製發現此類別上 pipeline 可能不優於 raw——已記入失敗帳本，不當永久標題。On a self-authored abstain pack, Sophia fabricated 0% on genuine “I don’t know” traps vs 17–25% raw. Later T2 replication found the pipeline may not beat raw on that class — logged in the failure ledger; not a standing headline.

機器檢查器Machine verifiers

法律引用 · 步驟驗證Legal citations · step verifier

100% / 100%

小型構造基準上，legal_citation_exists 與 misstep-injection step verifier 均 100% 召回、0% 假警報。驗證閘道邏輯，非廣域能力宣稱。On small constructed benches, legal_citation_exists and misstep-injection step verifier hit 100% recall at 0% false alarm. Validates gate logic — not broad capability claims.

528
雙語訓練範例Bilingual training examples
900+
OKF wiki 頁OKF wiki pages
4
人文研究領域Humanities domains
Apache-2.0
核心永遠開源Core forever open
DOI
Zenodo 永久歸檔Zenodo archival DOI

這不是「我們達成 AGI」的宣稱 This is not a claim of AGI

Sophia 是一個 AGI 候選驗證套件：已量測的窄域成果、可追溯來源圖、可複製的閘道機械，以及公開失敗帳本——誠實列出仍未證明的事。智慧先於智能；ASI 是遠程北極星，不是今天的標籤。

Sophia is an AGI-candidate proof package: measured narrow results, traceable provenance graphs, reproducible gate machinery, and a public failure ledger of what is not yet proven. Wisdom before intelligence; ASI is a long-horizon north star, not today’s label.

預先註冊門檻Pre-registered thresholds 公開失敗帳本Public failure ledger 多評審 + CIMulti-judge + CIs 第三方複製清單Third-party replication checklist canClaimAGI = falsecanClaimAGI = false
Sophia World · 空間世界模型Sophia World · spatial world model

從消費級 360° 到可信任的度量孿生 From consumer 360° to trustworthy metric twins

伴生研究軌道：測量優先、來源閘控的混合空間世界模型。重點不是「看起來像 3D」，而是減少虛假自信的米制量測——Sophia 來源紀律在物理空間的類比。 Companion research track: a measurement-first, provenance-gated hybrid spatial world model. The goal is not photoreal 3D — it is reducing false-confident metric measurements, the spatial analogue of Sophia’s provenance delta.

5 Insta360 X4 擷取 Insta360 X4 captures
373.4 s 總錄製時長 Total recorded duration
749 全景幀 · 368 關鍵幀 Panorama frames · 368 keyframes
374k 陀螺儀 / 加速度樣本 Gyro / accel samples
held 首個地面 mesh 決策 First floor mesh decision
TriHex overlay contact sheet on office 360 capture
TriHex 七色疊加 · 辦公室迴圈TriHex seven-color overlay · office loop HVE 相容 Kagome / TriHex 格網（1 m 外接圓半徑語意），深度與回調可見性抑制。 HVE-compatible Kagome / TriHex lattice (1 m circumradius semantics) with depth and callback visibility suppression.
Trajectory method comparison chart
軌跡方法比較Trajectory method comparison IMU 組織 / 陀螺軟閉環 / 視覺原始等多路徑對照——尚未宣稱驗證米制精度。 IMU-organized / gyro soft-loop / visual-raw paths compared — no verified-metric accuracy claim yet.
Nadir floor segmentation
天底視角地面分割Nadir floor segmentation ERP 地面支撐遮罩 → 相機高度先驗下的 held 地面曲面 PLY 引導。 ERP floor-support mask → held ground-surface PLY bootstrap under a camera-height prior.
First office floor top-down mesh preview
首個辦公室地面 mesh（held）First office floor mesh (held) 723 頂點 · 1334 三角 · 面積先驗 ≈ 2.64 m² · scaleStatus: estimated_metric · poseStatus: illustrative。 723 verts · 1334 tris · area prior ≈ 2.64 m² · scaleStatus: estimated_metric · poseStatus: illustrative.

已完成的基礎（2026-07）Foundation complete (2026-07)

  • 隱私安全的五擷取清冊、時間戳與必要角色完整性檢查全部通過Privacy-safe five-capture inventory; all timestamp & required-role completeness checks pass
  • ERP 射線幾何、cubemap / 切平面投影、相機姿態、地面交點ERP ray geometry, cubemap/tangent projection, poses, ground intersections
  • 確定性關鍵幀、操作者遮罩、TriHex 疊加與影片原型Deterministic keyframes, operator masks, TriHex overlay & video prototypes
  • 失敗封閉尺度政策、控制點校正、量測不確定性、Spatial Provenance DeltaFail-closed scale policy, control calibration, measurement uncertainty, Spatial Provenance Delta
  • stella_vslam / DAP / PanoVGGT / VGGT-Omega / RVT→IFC 的本地 preflight 契約Local preflight contracts for stella_vslam, DAP, PanoVGGT, VGGT-Omega, RVT→IFC
  • 只讀、僅 accepted 的 MCP bridge 接入 sophia-agiRead-only, accepted-only MCP bridge into sophia-agi
誠實邊界：目前狀態是資料就緒，不是幾何精度證明。首個地面 mesh 決策為 held；外部重建後端在本機 runtime 尚未可用（preflight 有意不執行推理）。外觀（NeRF / Gaussian）不得單獨充當結構量測。研究倉庫目前為私有開發軌道；公開敘事以 sophia-agi 整合與本頁進度為準。 Honest bound: current status is data readiness, not geometric accuracy. First floor mesh decision is held; external reconstruction backends are not yet available in the local runtime (preflight deliberately non-executing). Appearance (NeRF / Gaussians) must not be the sole structural measurement. The research checkout is a private development track; public narrative is via sophia-agi integration and this progress board.
路線圖Roadmap

從候選到可辯護的能力 From candidate to defensible capability

預先註冊、失敗封閉、公開帳本。每一級都必須通過 no-overclaim 閘門，而非敘事升級。 Pre-register, fail closed, publish the ledger. Each level upgrades only through the no-overclaim gate — not narrative escalation.

L0–L1 · 來源閘道與量測紀律L0–L1 · Provenance gate & measurement discipline

語料、基準、閘道、契約、失敗帳本——已出貨。窄域 VALIDATED 成果已公開。Corpus, benchmarks, gate, contract, failure ledger — shipped. Narrow VALIDATED results published.

L2 · 良心、代理與工具真實L2 · Conscience, agency & tool truth

良心核、道德公共標準、MCP 閘道、swarm 信任邊界。部分已上線；第三方 agentic delta 仍 OPEN。Conscience kernel, public moral standard, MCP gateway, swarm trust boundary. Partially live; third-party agentic delta still OPEN.

L3 · 執行真相與獨立評測L3 · Execution truth & independent eval

第三方 hidden eval、live RLVR、跨域保留、Lean/證明搜索與 independence-v2。多項為基礎設施就緒、結果待量測。Third-party hidden eval, live RLVR, cross-domain retention, Lean/proof search, independence-v2. Much infrastructure ready; many outcomes still to measure.

L4+ · 世界模型、RSI 與 ASI 方向L4+ · World models, RSI & ASI direction

空間孿生（Sophia World）+ 推理世界模型、受治理的持續學習、受驗證的自我改進。ASI 只作為需證明的遠程目標，從不作為預設標籤。Spatial twins (Sophia World) + reasoning world models, governed continual learning, verified self-improvement. ASI remains a long-horizon target that must be proven — never a default label.

幾分鐘內開始 Get started in minutes

複製 sophia-agi，跑離線閘道 demo，或掛上 MCP。語料在 Hugging Face；論文與排行榜在 thesis site。可完全本地 / airgap 運行。 Clone sophia-agi, run the offline gate demo, or attach MCP. Corpus on Hugging Face; thesis and leaderboards on the thesis site. Fully local / airgap capable.

原始碼：Source: github.com/tomyimkc/sophia-agi
資料集：Dataset: tomyimkc/sophia-agi-corpus
論文與排行榜：Thesis: tomyimkc.github.io/sophia-agi
DOI：DOI: 10.5281/zenodo.20930874 
# 複製儲存庫Clone the repo
$ git clone https://github.com/tomyimkc/sophia-agi.git
$ cd sophia-agi

# 離線閘道 demo（無需 API key）Offline gate demo (no API key)
$ python scripts/demo_gate.py

# MCP 伺服器MCP server
$ pip install -r requirements-mcp.txt
$ python sophia_mcp/server.py

# 良心核 demoConscience kernel demo
$ python tools/run_conscience_demo.py
支持可驗證的 AGI 研究——在 GitHub 給 sophia-agi 一顆星 → Support verifiable AGI research — star sophia-agi on GitHub →