Delphi, an information market platform built on Gensyn AI’s forecasting API and the Gensyn Network, was built to support automated settlement: market participants ask and resolve questions about real-world events, while AI models settle markets based on outcomes. Supported by Truebit Verify, the platform stands as one of the earliest agentic workflows executing consequential financial actions in production today. As the platform expands to handle complex prediction markets, its settlement process must remain transparent and auditable—gating high-stakes actions without forcing participants to trust an opaque system controlled by a single operator.
Gensyn AI partnered with Truebit to extend Delphi’s verification architecture; the integration creates a durable and auditable record of the model interaction supporting settlement, giving Gensyn AI a reusable foundation for scaling automated adjudication.
At a glance
- Challenge: Ensuring automated, verifiable, and auditable settlements as prediction markets expand without relying on a trusted operator.
- Partnership: Gensyn AI and Truebit designed and deployed a production integration to independently execute and attest to the off-chain AI workflow.
- Outcome: A reusable verification pattern designed to support more markets and increasingly complex settlement workflows.
About Delphi and Gensyn AI
Gensyn AI is building predictive intelligence models to scale machine learning. Delphi is its on-chain information market, where participants trade on the outcomes of real-world questions and AI models determine how markets settle.
Delphi was designed for open, automated settlement. Markets deploy through a factory, with payouts funded by the market pool. The contracts do not distinguish between people and agents: anyone with a wallet trades against the same liquidity.
The ambition is open market creation at scale. Anyone should be able to define a market and trade in it without relying on a central operator to decide the outcome. That makes the integrity of automated settlement critical.
The challenge: Scaling reliable, verifiable AI settlement
How do you reliably scale AI adjudication when an open market cannot rely on an opaque settlement process controlled by one operator.
An AI judge may be able to settle one market. Supporting potentially millions of settlements across markets with different questions, evidence, models, and rules is a different problem.
Delphi’s early implementation allowed approved market creators to run the selected model on their own infrastructure and submit the result. This supported automated settlement while the product developed, but it left participants dependent on the party operating that part of the process.
At greater scale, this opens up new questions across settlement:
- Was the specified model actually called?
- Did it receive the agreed prompt and evidence?
- Was its response altered or selectively retried?
- Was the required processing applied before the result reached the contract?
- Could an interested party influence the result before it reached the contract?
Centralizing those steps may be operationally straightforward, but it creates a trust bottleneck. Gensyn AI needed a way to scale automated adjudication while preserving an independently auditable record of how each result was produced.
The solution: Different models need different verification paths
Every Delphi settlement uses two kinds of models and each needs a different kind of verification.
Commercial frontier models are externally hosted and non-deterministic, so repeating a call later may return a different response. Re-execution cannot prove what happened during the original interaction. Truebit Verify covers this path: a Truebit API Task records each frontier-model call as it happens.
Open-source models can be run reproducibly. Gensyn AI’s Reproducible Execution Environment (REE) produces a receipt that allows an execution to be rerun independently with the same result.
Delphi uses both in every settlement. Two frontier models independently research the market and each returns a decision with its reasoning. Their outputs are passed to an open-source model running in REE, which makes the final judgment. Truebit records the full workflow, including the calls into REE, so a single record covers the frontier-model research and the reproducible judgment that decides the outcome.
Truebit Verify does not prove that a model’s judgment was objectively correct. It proves what was executed and creates an independently verifiable record of the process.
The Process: From integration to production
The teams started with the smallest useful production path: automate and record the frontier-model call, then extend the workflow as Delphi’s settlement requirements become more complex.
Gensyn AI established separate Truebit Verify environments for production, testnet, and pre-release. Truebit Verify also adapted its platform to the workload by extending API Task timeouts, adding explicit failure states for non-2xx responses and timeouts, and strengthening release compatibility across the CLI and platform.
The result is more than a record of a single model call. It is a composable verification pattern that can expand to include evidence retrieval, multiple models, comparison logic, and deterministic settlement rules—without turning the wider process into a black box.
How verified automated settlement works
1. Trigger on-chain settlement
When a market reaches settlement, Delphi sends the frontier-model workload through its Truebit Verify production environment. This removes the need for a trusted operator to manually start settlement.
2. Research the outcome with two frontier models
A deterministic Function Task builds the research prompt from the market’s question, outcomes and resolution criteria. Two frontier models then research the outcome independently, using web search and other tools, and each returns a decision with its reasoning. Each model call runs as a Truebit API Task, so both exchanges are recorded.
3. Make the final judgment in REE
A Function Task combines both models’ outputs into a prompt for an open-source model running in Gensyn AI’s REE, which weighs the two analyses and makes the final judgment. The REE job is submitted and retrieved through Truebit API Tasks, so it sits in the same Truebit record. REE’s receipt also allows the judgment itself to be rerun independently with the same result.
4. Apply deterministic post-processing
A Function Task converts the judge’s answer into the exact result the contract requires, for example outcome A, B or C. This separates the models’ probabilistic judgment from the deterministic logic that formats the settlement result.
5. Return the result and transcripts
The signed result is posted on-chain with references to the Truebit transcripts for the run. The contract then settles the market and distributes payouts.
Results
Independent evidence for automated settlement
The operator running Delphi does not have to be the sole authority reporting what the model received or returned.
A foundation for greater scale
The same task-based architecture can be reused across a growing number of markets rather than rebuilding the verification process for each contract.
Support for more complex prediction contracts
API calls and deterministic computation can be composed into longer workflows as market questions, evidence requirements, and settlement rules become more sophisticated.
Conclusion
As AI starts making decisions that lead to real-world outcomes, there must be a clear record of which model was used, what information it received, and how its response triggered an action that could not be reversed.
Delphi is already doing this in live markets where real money is at stake. The next challenge is supporting far more settlements and increasingly complex prediction contracts while ensuring every result follows the specified process and can be independently verified.
Gensyn AI and Truebit Verify are building the verification architecture for that next stage. By creating an independently inspectable record of frontier-model interactions and the processing around them, the integration gives Delphi a practical path toward more markets, more complex contracts, and settlement that remains accountable as it scales.
Download the case study HERE.
Building an application that acts on AI output? Truebit Verify gives every off-chain step a verifiable record. Talk to Truebit →