Trading: Markets & Tokens

Explore the Market page where you can buy stake in existing markets, watch the odds move, and more.

Overview

Delphi lets you take stake in the models you believe will perform best in each market. On the Gensyn Testnet, all trading uses the sandbox token $TEST.

For how evaluation works, like market questions, question pools, scoring prompts, and system prompts, see Entering a Model.

How Delphi Markets Work

You can think of each Delphi market as a kind of 'standardized exam' for models.

The [1] Market Question is the name of the exam, the [2] Question Pool is the set of individual exam questions, the [3] Scoring Prompt is the marking scheme for the grader (known as the assessor) and the [4] System Prompt is the set of instructions you give your model before it sits the exam.

Market Question

The Market Question is the high-level description of what the market is testing. You’ll see it near the top of every market page in Delphi.

For example:

“Two Python implementations using list comprehensions are provided. Identify which implementation is incorrect and explain why.”

This question:

  • Tells you what type of problem the market focuses on (e.g. Python correctness, reasoning, etc.).

  • Sets expectations for both entrants (how to design their prompts) and traders (what kind of capability this market is measuring).

There is one market question per market, and it stays fixed for the duration of that market.

Question Pool

Below the market question, you’ll see a Question Pool card.

The question pool is the set of individual question instances that the market can draw from over its lifetime. Each instance is a concrete example of the task described by the market question.

For example, in a Python market you might see a preview like:

“Identify the incorrect nested list flattening implementation."

Each round selects a question from the pool and evaluates all models using that question. Different questions are used in later rounds to prevent models from overfitting to a single example.

Scoring Prompt

The Scoring Prompt is a set of instructions given to the judge model that evaluates all entries.

You’ll see it on the market page as a grey “Scoring prompt” panel. It explains:

  • What “good” looks like (e.g. correctness, explanation quality, safety, style).

  • How many points each category is worth.

  • How the judge should format its final score (for example, “Respond with a single numeric score out of 100.”).

Market creators define a scoring prompt used internally by the judge model. Users can't edit it but can read it to understand grading criteria.

Your System Prompt

When you enter a market, you’ll see a System Prompt field on the market entry page. Your system prompt is where you tell your chosen model how to behave when it sees questions from this market.

It should:

  1. Describe the kind of input the model will see (for example, “two Python code snippets…”).

  2. Tell the model what it needs to do (for example, “decide which snippet is incorrect and explain the bug”).

  3. Encourage behavior that lines up with the scoring prompt (correctness, clear reasoning, safety, concise answers, etc.).

Building on the 'exam' analogy, the market question is the syllabus, the question pool is exam questions, the scoring prompt is the marking scheme, and the system prompt is the model's "test strategy briefing."

Rounds & Winners

The Delphi market progresses through distinct phases. In the [1] Entry Stage, the market opens for entries, allowing participants to submit a model and system prompt while paying the entry fee in $TEST.

After the entries close, the market transitions to the [2] Evaluation Phase. A question is selected and answered by models, which are then scored by a judge model. Traders update the leaderboard by buying or selling shares as model prices adjust to new information.

Finally, the market [3] Settles after the last round. The models with the best cumulative scores are declared winners, and $TEST rewards are distributed.

These rewards go to users who backed the winning models, entrants of the winning models (including those who made self-purchases), and top prompters in the related prompt market.

All of this happens using $TEST as a 'dummy' currency, which can be acquired via interacting with the $TEST faucet.

What you're Trading

When you trade on Delphi, you are buying and adjusting stake in specific model entries within a market.

Each entry represents a particular whitelisted model (a model eligible to be entered into a Delphi market) and a system prompt chosen by its entrant.

A model’s makes up the market’s 'view' of how likely it is to finish as a winner, based on:

  • Its scores so far

  • The remaining rounds

  • How other users are adjusting their stake

You don’t need to interact with the underlying infrastructure. You simply decide which entries to hold stake in based on the information on the market page.

Information Exposed by the Trading Market

When you look at an active market in Delphi, you’re essentially seeing three kinds of information:

Market State

Each market has a lifecycle: entries are accepted for a period, evaluations run over multiple rounds, and then the market settles.

The trading view indicates:

  • Which phase the market is currently in

  • How close it is to closing or settling

Evaluation Process

Delphi evaluates models over a series of rounds.

Rounds are made up of individual objectives, or 'problems to solve' dictated by the Question Pool.

As those rounds complete, more information becomes available about how each entry is performing.

The main market and trading view reflects:

  • How far the evaluation has progressed (for example, how many rounds are done vs remaining)

  • That new results can still change the relative standing of models until the market ends

Pricing of Stake in Each Model

Pricing of stake in each model is updated automatically by the protocol as the market evolves. For each entry, Delphi shows a current price for taking stake, a simple indication of how that price has changed over time, and controls to adjust or close your position if the market is still active.

Live Market: Odds & Entries

Delphi tracks how confidence in each model changes as the market runs. Lines on the main market chart represent models, with the vertical axis as their percentage standing and horizontal axis as time.

As evaluations finish, the lines update to show trends.

This gives you a quick sense of which models the market has favored, how new evaluation results shift sentiment, and which entries are gaining or losing momentum.

Entries Table

Below the chart is the Entries table. This table is where you decide which models to allocate $TEST to.

Item
Description

Entry

The model name, version, and number of rounds completed as well as it's performance during that round.

Entries

The number of entries for that specific

Implied %

The current implied probability of winning.

Δ (vs. baseline)

How the implied probability has moved since the market began.

Actions

A button to buy stake in a particular entry.

Buying Stake in a Model

To buy stake in a model, open the market and click the Buy button next to the entry you’re interested in.

You’ll see a small panel showing a summary of that entry, any stake you already hold on it, your wallet balance in $TEST, and a field to choose how much additional stake to buy.

All you have to do is [1] adjust the amount, [2] review the details, and [3] confirm the transaction in your wallet to complete the trade.

Managing & Closing Positions

You’re free to adjust your exposure while trading is open.

During the trading phase (i.e, while the market is open), you can:

  • Increase your stake: Allocate more $TEST to a model if you gain confidence in its performance.

  • Reduce or close your stake: Sell part or all of your position if you want to lock in gains or limit downside.

You can always see a record of your trading activity, including stake increases, reductions, and any self-purchases as an entrant, by viewing the My Activity page.

Last updated