VeriSynth — Bittensor AIGC Detection Subnet

◈ Bittensor Subnet Ideathon 2026

VeriSynth

Multi-Modal AIGC Detection Subnet
on Bittensor

Text Detection Image Detection Audio Detection Video Detection Continuously Improving

Team

黄梓庭 · 李怡蕾 · 杨虎啸

⚡ The Problem

AIGC is Everywhere —
Detection is Broken

90%

of online content may be AI-generated by 2026
Europol 2022

1,100+

fully AI-generated "news" sites identified
NewsGuard 2025

19.2%

Google top results contain AI content (↑ from <5% in 2023)
Originality.ai 2024

Downstream Harms

Academic fraud & integrity violations
Fabricated news & election misinformation
Synthetic identity & insurance fraud
Copyright disputes, social media bots

Regulatory Urgency

EU AI Act — enforcement begins 2025
China 《生成式AI管理办法》 — in effect since 2023
US Executive Orders — synthetic media labeling
Enterprise procurement urgency is now

🔍 Current Solutions

Every Existing Solution
is Incomplete

Solution	Modality	Update Speed	Cross-Modal	Centralized	Price
GPTZero	Text only	Slow	✗	✗ Single vendor	$15–99/mo
Reality Defender	Image+Video	Medium	✗	✗ Single vendor	$1,000+/mo
SN32 It's AI	Text only	Continuous ✓	✗	Decentralized ✓	TAO-gated
SN34 BitMind	Image/Video	Continuous ✓	✗	Decentralized ✓	TAO-gated
VeriSynth ◈	Text+Image+Audio+Video	Continuous ✓	✓	Decentralized ✓	$0.001–0.005/q

💡

The gap: No detector reasons across modalities — and none is benchmarked against continuously fresh adversarial samples. VeriSynth fills both gaps simultaneously.

⚔ The Arms Race

AIGC Detection is a Permanent
Arms Race — Only Bittensor Keeps Up

✗ Centralized Company

10–50 ML engineers — narrow architectural exploration
Self-interest conflict: they build generators, slow to detect them
Fixed product cycle — iteration is quarterly, not continuous
Single data pipeline, single point of bias

✓ VeriSynth on Bittensor

Global miners compete in parallel — diverse architectures simultaneously
No conflict: miners are rewarded to detect ALL generators
Every tempo (72 min) incentivizes improvement — it never stops
Multiple validators = multiple independent data pipelines

⚡

Core Argument

AIGC detection is an arms race against generative models. Bittensor's permanent incentive mechanism is the only structural solution that can match the pace of that arms race.

◈ VeriSynth

Two Tracks, One Subnet

🔵 Mining Track — TAO Emissions

1 Miner trains multi-modal detection model

2 Exports to ONNX/TFLite, uploads to HuggingFace

3 Registers commit hash on Subtensor chain

4 Validator downloads → runs locally on RTX 3060

5 AUC + Latency + Modality + Robustness → Yuma → TAO

🟢 API Track — Organic Revenue

1 User submits content (text / image / audio / video)

2 API gateway routes to top-ranked miner Axon endpoint

3 Returns P(AI-generated) per modality + unified score

4 Revenue → 90% to miner, 10% subnet treasury

5 Market signal steers miners toward user-valued models

🏗

192 UID Slots

128 miners + 64 validators

⏱

360 blocks/tempo

≈72 min evaluation cycle

💰

41 / 41 / 18

Miners / Validators / Owner

🛡

Local Eval

Validators run models locally

✓ PoI

Proof of Intelligence — Verified

🧠

Commodity = Trained ML Model

Miners must produce genuine multi-modal detection capability. Empty weights, random classifiers, and API proxies fail immediately under local validator evaluation.

🎯

Objective, Deterministic, Reproducible Scoring

AUC-ROC on public benchmark with block-hash-seeded sampling. Any third party can independently verify identical results.

📈

Intelligence Improves Over Time

Dynamic benchmark refreshed every 24h with new generator outputs. Static models decay. Peak detection quality ratchets upward each epoch.

💵

Real-World Economic Value

$13.4B AI detection market by 2029. GPTZero, Originality.ai, Reality Defender prove enterprise willingness to pay.

🔒

Cannot Be Trivially Gamed

Validators download and run model files locally. Miners cannot self-report latency or proxy to APIs. Private dynamic sets change every 24h.

🔬 Validator Design

The Hybrid Benchmark —
Our Core Innovation

Public Layer 40% of score

Text: RAID, GRiD, CUDRT, M4
Image: GenImage, DiffusionForensics
Audio: ASVspoof 2021, WaveFake
Video: FaceForensics++, DFDC

Same across all validators → academic credibility + low variance

Dynamic Layer 60% of score

① Block hash → deterministic seed (auditable)

② Harvest from latest generators every 24h

③ Human pair samples from open archives

④ Adversarial transforms (paraphrase / JPEG / MP3)

⑤ Private — validators never reveal pre-evaluation

Per-validator unique sets → Yuma filters colluding validators

AUC

HYBRID

        =  0.4 × AUCpublic  +  0.6 × AUCdynamic
      

Aggregated per tempo → normalized → submitted via set_weights()

📐 Incentive Math

Composite Scoring Formula

FinalScore(m) = 0.50 × AUC_hybrid + 0.15 × R_latency + 0.15 × R_modality + 0.20 × R_robust

AUC_hybrid 50%

0.4 × AUC_public + 0.6 × AUC_dynamic — core detection quality

R_latency 15%

Piecewise: 1.0 if t≤500ms, linear decay to 0 at 2000ms, 0 if t>2s

R_modality 15%

¼ × Σ 𝟙[AUC_k > 0.70] across text/image/audio/video — 0.25 per modality

R_robust 20%

AUC on adversarially-perturbed samples — directly measures robustness

Reward weight: w_i = FinalScore_i³ / Σ FinalScore_j³ ← α=3 steep top-N: top ~10 miners get 60-70% emissions

💰 Economics

Dual-Signal Reward System

Signal 1: Benchmark

Source: Bittensor TAO emissions
Distribution: Yuma weights → 41% to miners
Dominant phase: Phase 1 cold start
Effect: Miners optimize AUC on hybrid benchmark

Signal 2: Market Revenue

Source: API calls, B2B contracts, Chrome extension
Distribution: 90% → called model's miner, 10% → treasury
Dominant phase: Phase 2+ commercial
Effect: Miners optimize for what users actually need

Evolution of Dominant Signal

Phase 1

TAO emissions

dominate

Phase 2

Mixed signals

API growing

Phase 3+

Market revenue

dominates

Treasury (10% of API revenue) funds validator infrastructure, benchmark API costs, and community grants — ensuring sustainability independent of TAO price.

📊 Market

$39.7B Addressable Market by 2029

$13.4B

TAM

AI Content Detection
2029 projection
23% CAGR

$4B

SAM

Enterprise moderation
+ academic integrity
EN + CN markets

$20M

SOM (3yr)

0.5% SAM capture
via API + B2B
ARR target

Competitor	Modality	Price	Key Limit	vs VeriSynth
GPTZero	Text	$15–99/mo	Single modality, slow updates	⬇ 10×+ cheaper, multi-modal
Reality Defender	Image+Video	$1,000+/mo	No text, enterprise-only	⬇ 50×+ cheaper, unified API
Hive Moderation	Text+Image	$0.001–0.01/q	Proprietary, no evolution	Similar price, continuously improves
VeriSynth	All 4 modalities	$0.001–0.005/q	—	Decentralized + Self-improving

🚀 Go-to-Market

Cold Start? Here's Our Plan

Participant	Months 1–3 Incentives	Ongoing Incentives
⛏ Miners	20% of owner's 18% share → top-20 bonus Early registrants claim UID slots with minimal competition. Public leaderboard visibility.	90% of API revenue to called model's miner Benchmark refreshes continuously → competition space stays open forever.
✓ Validators	First 5 validators co-author benchmark paper (academic credit) Lowest-cost bond accumulation window.	Stable 41% Yuma emission share Bond appreciates as subnet TVL grows.
👥 Users	First 20 B2B clients: 6 months free API Exchange: usage data feedback + case study participation.	Continuously improving detection quality API SLA: 99.9% uptime commitment.

Miners + Validators compete → Detection quality improves → Users trust & pay → Revenue flows to miners → Stronger miners join ↩

📍 Roadmap

From Hackathon to Mainnet

MVP

Hackathon MVP — 4 weeks

Testnet subnet. Text + image modalities. HF model upload + on-chain registration. Reference miner + validator. Hybrid benchmark v0. End-to-end demo.

P1

Phase 1: Mainnet — Months 2–4

20+ miners, 5+ validators. Public docs + reference implementations. Initial open API. Leaderboard dashboard.

P2

Phase 2: Expansion — Months 4–8

Audio + video modalities. 100+ miners. Chrome extension. First B2B contracts. Open validator registration.

P3

Phase 3: Commercial — Months 8–12

B2B API GA + SLA. 10+ paying enterprise clients. API revenue signal active. SOC 2 Type I certification.

P4

Phase 4: Moat — Year 2

China-market compliance push. Federated data contribution layer. Academic benchmark paper. 500+ miners.

👥 Team

Three Complementary Builders

🛠

黄梓庭

Founder · Full-Stack / ML Lead

Post-00s full-stack engineer, AI application developer
Shipped B.Museum — EdgeOne Pages + serverless + KV
PyTorch / HuggingFace hands-on experience
AI-focused content creator — AIGC producer's insider view

📦

李怡蕾

Product & Strategy Lead

Product-research hybrid AI product manager
End-to-end lifecycle: research → architecture → launch → ops
LLM techniques applied to product experience
Data-driven user growth + commercialization track record

🏗

杨虎啸

Technical Architect Lead

16 years software development + project management
Core technical staff, Shanghai JiaTou Internet Group
Led multiple ad-tech systems and innovation products
Full-stack expert + cross-functional team leadership

🔄 Phase 3 Moat

The Content Fingerprint Flywheel

Core Flywheel

New generator released (GPT-X, Sora-X)
→ Validator daemon harvests samples within 24–48h
→ Static models decay in scoring
→ Miners retrain against new generator
→ Subnet capability advances
→ Repeat infinitely

Federated Data Layer (Phase 3+)

Content platforms contribute anonymized labeled data
Contributors get API discounts + custom model priority
Data never leaves contributor's infrastructure
Cross-platform data no single company can replicate

          New Generator released

          ↓

          Validator harvests new samples

          ↓

          Benchmark gets harder

          ↓

          Miners must retrain

          ↓

          Detection improves

          ↓

          Subnet value increases

The harder detection becomes,

the more valuable VeriSynth becomes.

◈ Why VeriSynth Wins

The Only Structural Answer
to the AIGC Arms Race

01

Real problem, regulatory urgency. AIGC labeling mandates in China + EU create procurement pressure now, not eventually.

02

Defensible architecture. Multi-modal + local model evaluation + hybrid benchmark + cross-validator consensus — hard to replicate from scratch.

03

Bittensor-native fit. The continuous adaptation problem maps perfectly to the continuous incentive protocol. Not a subnet of convenience — a subnet of necessity.

04

Dual-signal economics. TAO emissions for cold start. API organic revenue for long run. The subnet does not depend on TAO price forever.

05

Complementary to SN32 / SN34. Not competition — a broader, unified layer that could aggregate and extend both.

A decentralized, continuously self-improving, multi-modal AIGC detector is not a
"nice to have" — it is the only structural answer to a generator arms race that no centralized vendor can keep up with alone.