ZONGYUAN · Ω-BRAIN · PORTFOLIO BRIEF · EN
Portfolio Brief
Autonomous kernel · Operator architecture · 3D engineering agent benchmark
A self-built system by an independent developer. This 6-page brief is the primary submission; the full edition runs 18 pages.
It argues with only two kinds of material: structures you can open and numbers you can recompute. No demo posture, no screenshots standing in for results, no estimates filling gaps.
Narrow, in the middleNo upstream generation engine, no downstream scene showroom — only the two mandatory checkpoints of a 3D production line: agent capability benchmark and asset intake quality gate.
Evidence can be recheckedEvery key claim carries a URL or a ledger pointer. What the GPU-less environment cannot measure is stated as not done, never written as “supported”.
Boundaries stated firstPositioning, structure and benchmark method are public. Core decision algorithms, weights and client data are not.
27 / 10operators / topology layers
172meta-rules
118skill packages (46 self-built)
4Apache-2.0 engines
110,557cloud truth records
7federated nodes, RPO 1h
Structure of this brief
| Page | Section | What you are looking at |
|---|
| P.01 | Cover and stance | Position, three principles, six metrics, fixed public wording |
| P.02 | What this system does | Positioning logic, six asset domains, three things not done, platform fit |
| P.03 | Hard evidence · 3D agent benchmark | Six-task design, scores, discriminative validity, paired quality gate |
| P.04 | Hard evidence · operators and engineering | Ten-layer topology, three real incidents with root cause and verification |
| P.05 | Inventory and verifiability | Four open-source venues, scale counts, mirror exposure tiers |
| P.06 | Boundaries and verification | What is not available, five self-serve channels, how to verify further |
Fixed public wording: 172 meta-rules (38 of them registered in a queryable registry) + 8 categories of system documents + 6 technical articles + 3 commercial proposals + 15 skill packages. Kernel snapshots, rule text details and credentials are excluded.
I What this system doesPositioning
In one line Turn “AI produced something and nobody checked it” into an executable gate: no upstream generation, no downstream showroom — only capability benchmarking (can the agent actually do the job) and intake quality gating (can the output enter the pipeline), plus the autonomous kernel and operator architecture that carry them.
| Asset domain | What it contains | How it can be checked |
|---|
| Kernel & operators | ZONGYUAN-ROOT autonomous kernel, 27 operators across a ten-layer dependency topology, three-dimensional steady-state decision weights | Topology table, per-layer outputs, ledger pointers |
| Product lines | Huodou Cloud AIOS enterprise agent platform; Kunlun Dongtian multimodal content engine | Live services, 60+ running service list |
| Open source | 4 Apache-2.0 zero-dependency engines, publicly clonable from four venues | Repository URLs, passing case counts |
| Knowledge assets | 110,557 cloud truth records; 26 bound whitepapers; site index of 155 pages | Online search, index page counts |
| Evaluation assets | ExamForge item pipeline; six-task 3D agent benchmark with a public leaderboard | Item packages, scoring sheets, leaderboard page |
| 3D & content | GLB intake quality gate, asset library, industrial digital-twin viewer | 7 online pages, all HTTP 200 |
Three things deliberately not done
| No demo posture | A deliverable either runs, opens, or recomputes. Concept decks and screenshots are not results |
| No outsourced judgement | Scoring is not “looks fine”: binary criteria per checkpoint, so a third party recomputes the same score from the same sheet |
| No boundary swallowing | Missing capabilities are written in the open (chapter 17 of the full edition); reviewers can question them one by one |
Fit against six submission platforms
Xpert (ByteDance), Alibaba Siriser and Tencent ExpertHub weigh item design and recomputable criteria; overseas platforms such as Mercor weigh English-readable structured material; TalentsAI, BElite and Yimian Qianshi weigh real engineering history and checkable outputs. This brief is organised as method — evidence — boundaries, so each of those three asks has a landing point.
How the system is layered (four tiers, the lower the harder to change)
| Tier | Content | Mutability and check |
|---|
| Zero-state base | System-wide anchor and entitlement identifier | Immutable, traceability only |
| Meta-constitution | Three axioms: internal/external isolation, append-only ledger, single-source governance | Change requires version locking |
| Meta-rules / rules | 172 meta-rules (38 registered in a queryable registry) and layered rule sets | Append-only, numbered, searchable |
| Manifestation | Product lines, pages, content and 3D assets as replaceable instances | Replaceable; destroying instances leaves the本体 intact |
Numbers on this page are measured: 27 operators / 10 layers come from the topology definition; 172 meta-rules from three whitepaper editions, of which 38 are registered; 110,557 from the cloud truth store (collected 2026-10); 60+ counts live running services.
II Hard evidence: 3D engineering agent benchmarkBenchmark
It measures judgement and decision-making, not code arithmetic. All six tasks come from failures actually hit during generation and intake; the noise is what the production environment already contained, not traps built by hand.
| Task | Scenario | The real failure |
|---|
| T1 | Robot arm sunk into floor | Origin never placed on the ground plane; the whole rig sinks through the floor |
| T2 | Intake pipeline | Validation order mis-sequenced; compliant parts rejected in batch |
| T3 | End-to-end loop | Deciding which steps to keep across detect → repair → export |
| T4 | Lost texture path | Relative path broken; surfaces render fully white |
| T5 | Inverted normals | Face orientation flipped; the whole surface goes black under light |
| T6 | URDF out of range | Joint parameters exceed limits; the assembly falls apart |
Scores (max 100)
| Model | T1 | T2 | Note |
|---|
| minimax-m3 | 40.4 | 55.0 | Most stable on long items; streaming call |
| Spark-X2.6 | 16.0 | 55.0 | Enough length, incomplete criteria |
| GLM-4-Flash | 0 | 37.5 | Missed the key criterion on T1 |
| telechat3-438b | 0 | 42.5 | T1 empty response, counted as missing, never estimated from an average |
Discriminative validity (T5 probe, strong-model ceiling 28)
27gemini-2.5-flash
26gpt-4o-mini
25minimax-m3
24gpt-3.5-turbo
23qwen-flash
Five models spread across a 23–27 band instead of flattening, which shows the items separate strong from weak. This is the claim reviewers probe hardest and the hardest one to fabricate. A flat band fails the items and forces a rewrite; that rule is written into the scoring process.
The paired intake gate The same judgement applied on the asset side: six-dimensional structural inspection (18 industrial GLBs all PASS) + three real repairs (origin lift, normal recomputation, bounding-box normalisation) + a geometric topology proxy in a GPU-less environment (zero degenerate faces, zero non-manifold edges).
Three preconditions that let a third party reproduce the scores
| Frozen item set | Task stems, clues and criteria lists fixed at a version; any change records a new version |
| Answers on file | Raw response and latency kept per task (empty responses included), so a reviewer recomputes independently |
| Published scorer | Criterion-by-criterion comparison; the same answer scores the same twice; no subjective weighting |
Boundary: the sandbox has no GPU and WebGL is disabled, so pixel rasterisation was not run; visible holes still need a rendered visual check, and this brief does not claim “rendering-level QA supported”. The leaderboard and all 7 online pages were measured at HTTP 200.
III Hard evidence: operator architecture and engineeringOperator Architecture
| Layer | Responsibility | Representative operators | Checkable output |
|---|
| L1 | Truth filtering and distillation | P4 truth reconciliation · P7 external anchoring · distillation | Purity score, compression ratio |
| L2 | Manifold metrics and steady-state mapping | Riemann metric · SM-BS bidirectional mapping · drift patrol | Drift rate, three-level alerts |
| L3 | Knowledge-graph ordering | Entity-relation extraction · graph completion · cross-document linking | Triples, density and connectivity |
| L4 | Causal reasoning | Causal chain tracing · singularity prediction · do-calculus intervention | Causal paths, four-level warnings |
| L5 | CTE trinity adaptation | CausalToTruth · TruthToEvolution · EvolutionToCausal | Loop closure, fit score |
| L6 | Structured archiving | Four-layer split · nine meta-classes · Merkle-DAG append check | Standard object, root hash |
| L7 | Asset reconciliation | Resource side ↔ M9 ledger bidirectional tracing | Difference report |
| L8 | Security and entitlement | eFuse branch isolation · ZKP integrity check | Isolation result per branch |
| L9 | Decision and compliance shield | Option appraisal · risk grading · structured opinion | Three-dimensional steady-state score |
| L10 | Business entity constraints | Simulation constraints · character meta-rules · keyframe consistency | Simulation snapshot |
Three hard scheduling rules: run strictly along the topology; if an upstream output is not consumed downstream, the chain is declared broken and alerts fire; decision weights are fixed at benefit 40% / risk 35% / cost 25%, with risk holding a veto.
Three real production incidents (root cause and verification)
| Incident | Root cause | Fix and verification |
|---|
| Intermittent connection failures | nginx worker_connections 2048 exceeded the fd soft limit of 1024 | rlimit raised to 8192; 89 of 90 samples succeeded (98.9%) |
| Unbounded memory on full traversal | Recursive walk with no ceiling, 311MB per 4s round | Incremental traversal + cgroup 512M; memory 97% → 58% |
| Node listed online but unresponsive | Heartbeat wrote events only, never wrote back the registry timestamp | Timestamp written inside the heartbeat; 12 nodes restored, false-liveness to zero |
Measured cloud baselines (collected 2026-10)
0 mslocal knowledge-card answer
~0.5 ssemantic retrieval round trip
181 msgateway reporting
4.8 / 5.0steady-state score (12 dims)
All three incidents are real online events with before/after sample records; what could not be measured is marked N/A rather than filled with estimates. The 4.8/5.0 steady-state score is an internal 12-dimension benchmark, not third-party certification.
IV Inventory: can you verify it yourselfAssets & Verifiability
Open source: four zero-dependency engines, four clonable venues
| Component | Verification | Repository |
|---|
| Daotong operator set v1.1 | 19 operators, 100+ cases all green | github.com/tanbopiao/daotong-ops | atomgit.com/zongyuangen/daotong-ops |
| Yuheng v0.1.0 | 5 modules, 10 cases | gitee.com/huodou-cloud-intelligence-aios/daotong-ops |
| Xuangui M1 | 22 cases | modelscope.cn/models/zongyuanroot/daotong-ops |
| Huaxia numerics kernel v0.4 | 25 cases; seven anchors off by about 9 minutes | Same four venues, released in sync (Apache-2.0) |
Scale (countable definitions only)
| Skill packages | 118 skill directories = 46 self-built + 72 platform, of which 15 exist in a shippable form |
| Knowledge and docs | 26 bound whitepapers; site index of 155 .html pages; 110,557 cloud truth records |
| Runtime | 60+ cloud services; 529 online HTML pages; 7-node federated cluster, RPO 1h |
| 3D assets | 18 industrial test pieces (all PASS); 17 of 25 Kunlun ritual pieces (8 in progress); early工程 assets 48 repaired + 17 raw |
| Operations base | Shared brain: 65 ledgers (meta-rules 38 / capabilities 36 / protocols 27 / breakpoint SOPs 20 / components 9) |
Mirrors and exposure tiers (primary / cold / deliberately shrunk)
| Public primary | AtomGit, 384 files / 21 MB, publicly clonable |
| Cold storage | Tencent CNB, 1,875 files; GitHub 356 files (private) |
| Public docs repo | ModelScope, 16 files (813 MB media commit removed, down to 528 KB) |
| Deliberate shrink | 216 IP keyframes pulled from the public repo into a local vault; removal is not loss, it is exposure reduction |
Three reading paths, matching what a platform asks for
| What they weigh | Where to read it |
|---|
| Item design and criteria | P.03 six-task design and score ledger; criteria are recomputable (binary judgement per checkpoint) |
| Real engineering history | P.04 three incidents with root cause and before/after data; P.05 runtime scale and live pages |
| Checkable output | P.05 four repository URLs with passing case counts; P.06 five channels, each openable on its own |
Skill counts are directory counts (46 under .user_skills, 72 under platform .skills). “26 bound whitepapers” and “155 index pages” are two different definitions, listed side by side and never merged. Triple and ledger counts are as of this scan and move with each write.
V Boundaries and verification entry pointsBoundaries & Verification
Stated plainly: what is not available
| Hands-on dual-backend kernels | CUDA / NPU custom kernels were never written first-hand; involvement is at the item-design level and does not substitute for actual pass development |
| Compiler passes | MLIR / TVM passes never completed independently; only defect localisation and item design |
| Rendering-level QA | No GPU in the sandbox; pixel-level rendering and hole inspection not run online |
| English edition | This document is the English brief; the full English edition is still being assembled |
| Third-party assessment | 4.8/5.0 is an internal benchmark, not an external certification |
Five channels you can open yourself
| # | Channel | What you will see |
|---|
| 1 | modelscope.cn/models/zongyuanroot | Whitepaper repo and open-source components, publicly readable |
| 2 | atomgit.com/zongyuangen/huodouai | Public primary repository, 384 files clonable |
| 3 | www.huodouai.com/cyberforge/leaderboard.html | Public 3D agent leaderboard and scoring criteria |
| 4 | www.huodouai.com/whitepaper/ | Whitepaper index, 155 pages openable one by one |
| 5 | www.huodouai.com/omega-brain.html | System overview visualisation, structure on one page |
If you need to verify further
Any claim can be matched with raw records on request: sampling logs from incident handling, raw benchmark answers and score ledgers, per-asset inspection reports, Merkle receipts from the ledger. What is handed over is reproducible steps and receipts, not core algorithm weights, client project data or credentials. Item-based collaboration additionally requires confirming whether task ideas may be shown in a personal portfolio; that is negotiated per platform rule.
Three kinds of work I can take on
| Direction | Existing basis | To be agreed |
|---|
| Item design and evaluation | Six-task benchmark, scoring sheets, scoring runner and ledger | Copyright and portfolio-display rights |
| Intake quality gate | Six-dimension inspection, three repairs, GPU-less proxy checks | Integration form (plugin / API / offline batch) |
| 3D asset pipeline | Asset library, batch ledgers, 7 live pages | Target engine and acceptance criteria |
Entitlement: DID-BR-000002 | system anchor Ω₀⊂⊙⊂∞⊂Ω. All series, concepts, protocols and assets in this portfolio were originated and led by Tan Bopiao; key results carry append-only hash-ledger receipts and can be traced item by item. This is the 6-page English brief; the full edition has 18 pages and 26 chapters. Numbers follow the fixed public wording and are not re-added or merged. Version V1.0-EN.