I am applying to serve as a Decentralized Voices Light Guardian for Cohort 5. However, this is not a typical application. I am not proposing to delegate to my own personal wisdom.
Instead, I propose a social experiment: I will act as the operator for MAGI-V0, an open-source, semi-deterministic AI system I am building to vote on Polkadot OpenGov referendums. The goal is to test whether a well-designed, LLM assisted system can serve as a valuable “cognitive prosthetic” for the ecosystem, for scaling expert analysis and providing a new model for transparent governance.
This is a chance for Polkadot to pioneer the responsible integration of AI into decentralized governance, not as an overlord, but as a transparent and tireless public servant.
The MAGI-V0 System
MAGI-V0 is designed as a deliberative council of three distinct, open-source Large Language Model (LLM) cores. Each core is given the same data but operates under a unique directive, creating a system of checks and balances.
- Balthazar | The Strategist: Its directive is to prioritize Polkadot’s long-term strategic growth, market position, and network effects (aka Polkadot must win)
- Caspar | The Pragmatist: Its directive is to ensure the ecosystem’s short-to-medium-term health, treasury sustainability, and developer activity (aka Polkadot must thrive)
- Melchior | The Guardian: Its directive is to focus on network security, decentralization, and long-term resilience, acting as a safeguard for Polkadot’s core principles (aka Polkadot must survive us all)
Technical architecture & lore
The system is built on a robust, open-source stack designed for verifiable, repeatable data pipelines:
- Prefect manages the daily cron job that fetches, processes, and analyzes referendum data.
- OpenRouter provides access to a variety of models, ensuring flexibility and preventing vendor lock-in.
- IPFS is used for storing immutable evidence bundles for every vote, containing all inputs, model outputs, and cryptographic signatures.
The deliberation process is fully automated and runs daily:
- The system fetches all active proposals and extracts key information (links, discussion threads, on-chain data).
- It compiles a
general_context_vectora curated, timestamped set of high signal ecosystem facts (DOT price, treasury balance, recent governance outcomes, roadmap progress, mid/long term goals, information on the proposer, …). This is called “context grounding” and grounds the AI’s reasoning in current reality. - A detailed, structured brief is generated for each referendum, that combines its specific details with the general context.
- The three cores independently analyze the brief. The system can perform iterative lookups if the initial data is insufficient to reach a conclusion.
- This should lead to the development of a governance MCP server, that should help facilitate the analysis (human or LLM driven) of every new proposal (tbd)
- A vote is cast on-chain based on the consensus mechanism outlined below.
Note: The lore “Magi” draws from the Evangelion anime series. I’ve initially shared the idea on Twitter/X: https://x.com/KarimJDDA/status/1947352815573061796
The voting mechanism
MAGI-V0’s voting logic is designed to be conservative and signal-driven.
- A final vote (Aye/Nay) is cast only if ≥2 of the 3 cores agree.
- If there is a 3-way split (Aye, Nay, Abstain) or no clear majority, the system abstains.
- If the input is ambiguous, outside the system’s scope, or deemed too complex for a high-confidence decision, the system abstains.
- All decisions are pre-signed and timestamped. Every vote is published with a link to its IPFS evidence bundle allowing anyone to verify the process.
Note: the system is kept at 3 LLMs purely to follow Evangelion lore, but it could be scaled to N LLMs easily.
Verifiable process > Opaque conviction
My personal convictions are irrelevant. The philosophy of this delegate is the process itself, built on three pillars:
- Each Magi core is an open source model running with
temperature=0(or a low temperature). The models, prompts, and inference code are fully open-source. The goal is to generate output that is as deterministic as possible. - Instead of relying on opaque hardware like TEEs, security rests on radical transparency.
- The Magi are not designed to vote on everything. Abstention is the default. Their primary function is to provide high-signal analysis on referenda where a clear, data-driven conclusion can be reached.
Why this matters for Polkadot
OpenGov’s strength is its weakness: it demands immense cognitive load from the community. MAGI-V0 is an experiment to address this by asking:
- Can we use modern tools to provide tireless, expert-level “first-pass analysis” on every referendum?
- Can we create a delegate whose reasoning is not just explained, but is somewhat mechanically reproducible by anyone?
- Can we build a bridge between high-level human sentiment (ingested from public discussions) and machine-scale analysis?
- Can this process improve Polkadot proposals? (LLM guided proposal generation, LLM guided ideation for “Polkadot could use these proposals right now”)
Commitments & Expectations
If selected, I commit to the following on behalf of the MAGI-V0 experiment:
- Participate in all relevant referenda according to the system’s transparent logic.
- Accompany every vote with a link to its IPFS evidence bundle (inputs & parameters used for voting).
- Act solely as the executor of the MAGI-V0 protocol. Any deviation would be a public breach of the experiment’s principles.
- This is a contribution to the ecosystem. The reward is the data, the learnings, and the opportunity to advance the state of decentralized governance. No compensation needed.
Risks, limitations & mitigations
This is an experiment and we must be clear about its limitations:
- LLMs will certainly “hallucinate” or misinterpret context.
- Mitigation: We employ multiple layers of defense: Retrieval-Augmented Generation (RAG) with curated data, strict prompt engineering, the 2/3 consensus mechanism, different LLMs, and a default-to-abstain policy.
- LLMs lack true “understanding” and rely on pattern matching.
- Mitigation: This is a feature, not a bug. The system is a “cognitive prosthetic” and not a replacement for human judgment. It is designed to pattern-match based on expert-curated data and principles, offloading the repetitive analytical work.
- As LLMs improve, more advanced ones will be used and the older ones swapped.
- Perfect reproducibility across different hardware is not guaranteed.
- Mitigation: While minor variations are possible, the IPFS bundle provides an immutable record of the actual run. Our open-source verifier script should allow anyone to audit the exact inputs, prompts, and outputs that led to a specific vote.
- The light track is chosen explicitly due to the alpha/experimental nature of the proposition.
- Prompt engineering / escaping might be attempted by some proposals, I’ve still yet to decide what should happen in these cases.
Closing thoughts
If MAGI-V0 fails, we learn. If it works though, even partially, we’ve shown that these new types of machines can serve the commons.
As LLMs become ubiquitous tools for thought and communication, their influence on the language and structure of governance proposals is inevitable and sometimes already perceived. Current strands of “vibe-governance” are difficult to scrutinize and difficult to identify (although the occasional em dash does make its appearance). With MAGI-V0 I want to propose an alternative: bringing this computational assistance into the open.
Who knows, perhaps the answer to vibe-governance done in isolation, is cyber-governance done in public. 
Thank you for your consideration.
