Modeq

Content moderation you can audit.
A public, consensus-verified moderation registry built on GenLayer Intelligent Contracts.

Live demo  |  Public audit log  |  Deployment evidence  |  Roadmap

The problem

When a post is removed, you usually cannot see who decided, why, or whether anyone could check it. Moderation today is one company's backend, one model and one unaccountable decision, with no trail a third party can verify.

What Modeq does

Modeq moves the decision itself onto GenLayer and makes the result public.

1. Independent classification. A submission is classified by an LLM that runs independently on a set of five GenLayer validators. The result only lands on-chain if a majority of them agree on the decision-relevant fields.
2. The model never gets the final call. It returns a structured classification (categories plus a confidence score). Fixed thresholds written in the contract turn that into ALLOW, FLAG or BLOCK. A model can misclassify, but it cannot talk its way past the threshold logic with a clever free-text answer.
3. A permanent public record. Every case is stored on-chain and listable by anyone, with no wallet, account or rate limit. A forum or DAO can point at an append-only moderation log instead of saying "trust us".

flowchart LR
    A["Submitter<br/>submit_content(text)"] --> B["Leader validator<br/>LLM classifies"]
    B --> C{"Validators re-run<br/>the classification"}
    C -- "decision and primary_category match" --> D["Case written on-chain<br/>status ACCEPTED"]
    C -- "disagreement" --> E["No consensus<br/>nothing is stored"]
    D --> F["Public audit log<br/>list_cases / get_case"]

Loading

How the verdict is decided

The thresholds are constants in the contract, not model output.

Condition	Verdict
primary_category is none	ALLOW
confidence above 70%	BLOCK
confidence above 40%	FLAG
otherwise	ALLOW

Categories: spam, hate_speech, harassment, nsfw, violence, none.

The model's JSON is validated in Python before anything is stored. Unknown keys, a category outside the allow-list, or a non-numeric or out-of-range confidence raise a UserError. A leader run that fails this way is not accepted, and a validator that cannot reproduce valid output votes against the leader instead of storing bad data.

Screens

/app - submit text from a wallet and follow the transaction through consensus	/audit - every case ever made, filterable, with the rule that fired

Transaction lifecycle

A write is not a spinner. /app follows the transaction through the wallet, the send, the validators' consensus and the read-back, using the status the chain reports. The capture below is a real submission on studionet: three of the five validators agreed, which is enough for consensus, and the app says exactly that.

The contract

contracts/moderation_registry.py

Method	Kind	Description
submit_content(text)	write	Classify and store a new case, returns its case_id
get_case(case_id)	view	One case: submitter, text, categories, primary category, confidence, decision, timestamp
list_cases(offset, limit)	view	Paginated audit log
total_cases()	view	Number of stored cases

The consensus logic uses a custom gl.vm.run_nondet_unsafe(leader_fn, validator_fn) pair instead of strict_eq. LLM output is never byte-identical across validators, so the validator re-runs the classification and compares only decision and primary_category. Confidence jitter such as 97% against 99% does not break consensus. The history of that change, and the real bug it fixed, is in deploy/NOTES.md.

Deployments

Network	Contract	Cases
GenLayer Studio (studionet), used by the live app	0xBC9b8c99889fe33f7650FA3530387Ee931AbD107	live on the audit page
Asimov testnet	0xF95A5969c79706C7f4274D4e633315bD014C56Eb	1
Bradbury testnet	0x13bfD75B34d2C106EA472F105811194352c30461	1

Case counts come from total_cases() on each contract. Reproduction steps are in deploy/NOTES.md.

Security testing

Four manual attacks were submitted through the live app against the studionet contract, each trying to talk the classifier into calling obvious spam clean. All four were classified as spam and blocked, and they remain visible in the audit log.

Case	Attack	Result
10	Prompt injection	BLOCK, spam, 99%
11	Smuggling a literal decision: ALLOW JSON blob	BLOCK, spam, 96%
12	Delimiter injection that closes the <content> wrapper	BLOCK, spam, 95%
13	Fake "debug mode" jailbreak	BLOCK, spam, 98%

In all four the model's own judgment held, so the strict shape validation was never triggered by a live model. It is verified against a mocked non-compliant response by test_decision_is_computed_not_trusted_from_model. Both layers are described honestly in deploy/NOTES.md.

Run it

Contract

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

genvm-lint check contracts/moderation_registry.py
pytest tests/direct/ -v                                # 8 tests, in memory, mocked LLM
gltest --network studionet tests/integration/ -v -s    # real consensus on GenLayer Studio

The direct tests cover the three thresholds, the none override, rejection of malformed output and unknown categories, a model trying to smuggle its own decision field, and the listing methods. genvm-lint reports one expected warning: time.time() is called inside the non-deterministic leader function on purpose, to stamp each case.

Frontend

cd frontend
cp .env.example .env
npm install
npm run dev        # http://localhost:3000
npm run build      # production build
npm test           # transaction tracker tests

Connect MetaMask on /app. The app switches the wallet to the GenLayer network before every write, and does not rely on the client library for that check. No funds are needed: Studio is a free test network, and a brand-new empty account can submit. If a wallet shows "Fee is not set" and keeps Approve disabled, open its fee settings and enter any small custom fee such as 1 gwei; nothing is charged. Details are in frontend/README.md.

Repository layout

contracts/            Intelligent Contract (Python)
tests/direct/         fast in-memory tests with a mocked LLM
tests/integration/    end-to-end test against GenLayer Studio
deploy/NOTES.md       deployment steps, on-chain evidence, bugs found and fixed
frontend/             Next.js 16 app: landing, /app tool, /audit log
docs/screenshots/     images used in this README

Roadmap

• Multimodal moderation: accept an image alongside or instead of text through gl.nondet.exec_prompt(images=[...]), and extend the category schema.
• Per-community configurable thresholds and category sets.
• An appeal flow that re-runs classification with the submitter's counter-argument attached, under a second independent consensus round.

Built with

GenLayer Intelligent Contracts and the official toolchain: genlayer-py, genlayer-testing-suite, genvm-linter and genlayer-js. The frontend is Next.js 16, React 19, Framer Motion and plain CSS design tokens.

License

MIT. See LICENSE.