If you've read Your Cleanup Counts. Now Prove It., you know what DeCleanup Network does. This post is about how it works - every layer of the verification pipeline, every infrastructure decision, every trade-off we made building a public goods protocol on Celo and Base.
This is for builders, technical readers, and anyone who wants to audit the claim before they trust it.
The protocol runs on two networks with different entry points.
Phase 1 - Farcaster mini app. Lightweight cleanup logging via Farcaster frames and the Base app. Photo-based submissions, basic verification, $bDCU rewards. Optimized for reach - the lowest possible friction for a first submission.
Phase 2 - Celo full dApp at dapp.decleanup.net. Geolocation mapping, AI waste detection, Hypercert issuance, governance, and the full dMRV pipeline. Optimized for depth - every submission produces a record that meets verification standards.
Both platforms feed the same verification standard. The proof output is identical whether you submit from a Farcaster frame or the full dApp. The platform is the entry point; the standard is the product.
Every cleanup passes through four independent layers before it becomes an onchain record. No single layer is sufficient on its own. The combination is what makes the record defensible.
A participant submits via the dApp: before and after photos, GPS coordinates captured at the physical location, waste category estimates across plastic/glass/mixed/organic, and a weight estimate in kilograms.
GPS is captured at submission time, not pulled from photo metadata after the fact. This matters: metadata can be edited, location spoofed at the app level. Submission-time GPS tied to the platform session is significantly harder to falsify.
Photos are stored on IPFS immediately on upload. The content hash is attached to the onchain record at mint time. The consequence: the photos cannot be changed after the record is issued without changing the hash, which breaks the link to the onchain record. The evidence is cryptographically locked at the moment of submission.
Stack: Supabase for backend - user profiles, submission queue, verification state, and the pipeline between capture and onchain issuance. Google Auth (OAuth 2.0) as an alternative login path to wallet connection.
Before any human review, submissions pass through automated waste detection.
We use YOLO (You Only Look Once) object detection trained on the TACO dataset - Trash Annotations in Context, a labeled dataset of real-world litter across 60 waste categories. The model runs on submitted photos, detects waste objects, estimates type and volume, and produces a confidence score.
The confidence score serves two purposes: it provides an independent volume estimate comparable against the participant's self-report, and it flags anomalous submissions — images that don't match the declared waste type, photos that appear to be recycled from other sources, or claims where the AI estimate diverges significantly from the stated weight.
A high-confidence AI match isn't approval - it's a green light to human review. A flagged submission goes to additional scrutiny, not automatic rejection.
Stack: Ultralytics YOLO on the TACO dataset. Detection output feeds the verification queue with a confidence flag attached.
Note: full AI analytics will be available on Celo platform in Q3 2026, currently it's undergoing testing processes.
AI detection is a filter, not a judgment. Every submission that passes automated screening enters a community verification queue.
Verifiers are $bDCU/$cDCU holders who have staked to join the verification layer. They review the before/after pair, the GPS location, the AI confidence score, and the weight estimate. They approve or flag with a reason.
Slashing applies for false approvals. Verifiers who consistently approve fraudulent submissions lose staked $cDCU. This keeps the verification layer economically honest without requiring central oversight - the incentives do the work.
This layer exists specifically for what AI cannot cover: local context, genuinely ambiguous submissions, edge cases that require human judgment about whether a cleanup actually happened. A beach that looks the same before and after because it was already mostly clean. A forest trail submission from a location where no trail exists. A weight estimate that's technically consistent with the photos but implausibly high for the team size reported.
A submission that passes both AI and community verification is minted as an onchain record on Celo mainnet. The record contains:
IPFS hash of the before/after photo pair
GPS coordinates as onchain data, not metadata
Waste category and volume
AI confidence score
Verification timestamp
Participant address
Verifier signatures
The record is public, permanently linkable, and queryable by anyone. Every 10 verified cleanups, a Hypercert activity claim is issued - a structured, living impact record built on AT Protocol (the same decentralized data layer as Bluesky). It captures who did the work, what was collected, when, and where. The 10 individual cleanup records attach as measurements and attachments, making the Hypercert progressively richer over time.
A Hypercert is a signed data record that lives on a Personal Data Server, portable across any platform that reads the protocol. Onchain tokenization for programmable funding is on the Hypercerts roadmap but not yet live; DeCleanup Network's cleanup records are already structured to support it when it arrives.
The hardest problem in Web3 environmental tooling is not verification - it is onboarding. The people doing the most cleanups are not crypto-native. Requiring a funded wallet before a volunteer can log their first cleanup is a guaranteed adoption ceiling.
We built the Celo dApp with account abstraction as a core architectural requirement.
Pimlico provides smart account infrastructure and a paymaster on Celo. Participants never pay gas - the protocol sponsors all transactions through the paymaster. The first-time experience: open the app, submit a cleanup, see a reward. No gas prompts, no wallet funding, no network configuration.
The implementation required a deliberate choice. The standard path for account abstraction in consumer apps runs through managed services - Web3Auth, Privy, similar. These work well. Their production-ready scale plans run on a rate that is not affordable thus not viable for a public goods protocol with no subscription revenue. We did not use them.
Instead we built directly on permissionless.js - Pimlico's open-source library for constructing smart accounts without a managed service. Full control over the account abstraction layer, composable with any signer, no monthly platform fees. Gasless transactions, batched operations, sponsored gas - all available at the library level.
Google Auth (OAuth 2.0) provides the non-wallet login path. The wallet is created and managed through the smart account layer in the background. Wallet interaction surfaces only when a participant chooses to interact with governance or move tokens - until then, the experience is a standard web app.
The pattern is intentional: deferred friction. Capture participation first. Introduce custody complexity only when the participant is ready and motivated to engage with it.
Two tokens. Distinct roles. No functional overlap.
Fair launch via Clanker. No pre-mine, no team allocation, no investor round. Every $bDCU in circulation was acquired on the open market or earned through verified cleanup activity.
Supply: 1,000,000,000. 85% fair launch, 15% Clanker vault for deflationary burns and ecosystem rewards.
Liquid and tradable. The network's entry point - earnable immediately through the Base mini app. Also the bridge: five verified cleanups on Base generates a $cDCU claim on Celo, connecting lightweight participation to governance without requiring participants to manage two networks from day one.
Max supply: 10,000,000. Non-transferable. Earned exclusively through verified environmental action - on the Celo platform or bridged from Base.
Cannot be purchased. The design intent: governance should be held by people who have contributed to the network, not by capital that bought in. 250 $cDCU unlocks full voting rights on Gardens, DeCleanup Network community on Celo.
Allocation: Staking 35% · Campaigns 25% · Public distribution · Team · Treasury · Community · Liquidity.
Every component of the stack is open source under MIT. This is not a values statement - it is a logical requirement for the verification claim to hold.
"This cleanup is independently verified" means nothing if the verification pipeline is proprietary. Independent verification requires independent auditability. The code, the model weights, the smart contracts, the onchain records - all of it must be publicly inspectable for anyone to trust the output without trusting us.
The IPFS photo storage, the Supabase pipeline, the contracts on Celo and Base, the YOLO model - all public. Contracts verified on Celoscan and Basescan. Impact API public at decleanup.net/api/impact.
Open source also means composable. Any project in ReFi, DePIN, or impact accounting can build on top of DeCleanup Network verification without asking permission. The Hypercert output is readable by any ecosystem that supports the standard. ESG tools can query the API directly. Future certification layers can use our records as inputs.
This is the difference between a product and a primitive. A product you use. A primitive you build on. We are building the primitive.
Partner | Role |
|---|---|
Hypercerts | Onchain impact certificates (ERC-1155) per 10 verified cleanups |
Farcaster | Social layer · mini app on Base frames |
Pimlico | Smart accounts · gasless UX · paymaster on Celo (via permissionless.js) |
Supabase | Backend · profiles & verification queue |
Reown | WalletConnect v2 for Celo dApp |
ENS | Onchain identity · decleanupnet.eth |
Ultralytics YOLO | AI · trash type & volume detection on TACO dataset |
Google Auth | OAuth 2.0 · non-wallet login path |
Phase 3 in design:
AI verification v2 - improved YOLO model, expanded waste categories
RWI Rank - onchain reputation scores for participants and organizers
Impact staking - long-term commitments to cleanup programs
Cross-campaign analytics - aggregated impact data across the network
First public funding vote on Gardens: July 2026.
Start here: dapp.decleanup.net Read the full protocol: decleanup.net/litepaper Governance pool: gardens.fund Get in touch: support@decleanup.net · Telegram - X - Farcaster
DeCleanup Network · Open-source · MIT · Built on Celo + Base

