You probably noticed that even carefully crafted prompts rarely allow immediate deep response and inference from AI systems. Opening discussions tend to be shallow.
It is usually later in the interaction (not always) that something shifts. At some point, the system clicks and starts reasoning inside our frame.
For xbid.ai, that was a problem as trading decisions need instantly grounded inference, not guessing or pseudo‑insight.
The Mechanism
The key here is understanding that the point of convergence is not prompted. Instead it is surfaced through dialogue.
xbid.ai still relies heavily on prompts for task specification and factual feedback but prompting alone cannot provide the epistemic constraints inference depends on.
Recent work on long-context language models confirms this dynamic: simply increasing context size, feeding models more documents, more tokens, more information, fails to improve reasoning quality.
Dialogue works differently because each turn reduces ambiguity and enables tighter epistemic alignment. This process naturally surfaces constraints you would otherwise miss—or approximate through hours of prompt babysitting, unscalably.
To illustrate, I will take this short fictive exchange between Alice and Bob:
Alice: Remember when we used to tell each other everything?
Bob: Yeah. We were kids though.
Alice: True. We don’t do that anymore.
Bob: Right. That would be strange.
We can see that despite its apparent simplicity, this dialogue encodes a dense a posteriori about shared childhood intimacy, and implicit constraints about adult norm, emotional restraint, and self-censorship.
Now imagine trying to convey the same episteme using a cold prompt.
Alice and Bob have known each other since childhood. When they were younger, they shared their thoughts freely and without much concern for boundaries. As adults, they remain close, but their interactions are more measured. Both sense that speaking too openly now would feel out of place.
Here, despite increased verbosity, the prompt flattens important constraints that the dialogue implicitly carries (i.e., questions about timing, initiative, asymmetry, and risk-bearing would not be answered without assumptions). For example, asking who is testing the adult boundary forces a cold-prompted model to guess unless you spoonfeed the answer. The same question posed over the dialogue can be answered immediately, without ambiguity. Alice is testing the boundary.
This is why cold prompting does not lead to meaningful inference in complex domains. Inference is often executed over invented structure because AI models fill the gap with assumptions. The result? Inference that sounds smart but is ultimately speculation.
The Lab
When building xbid.ai, I needed the AI to reason over real structure—not made-up assumptions. The question was how to systematically capture dialogue-constructed episteme that encodes these implicit constraints, feedback, and reasoning patterns.
I prototyped a chat-based UI optimized for short turns and branching discussions (codenamed reflx.ai, the tool will be released separately later) allowing systematic exploration and discovery of reasoning patterns through natural interaction. Using it, I was able to drastically improve xbid.ai inference quality.
The Lab is a gamified version of that tool. It acts as an instrumentation layer that captures, evaluates, and distills reasoning through interaction.
Critically, the Lab doesn’t reward correctness or verbosity and it is not looking for the most polished interactions. Like the Alice and Bob dialogue, what matters is whether the interaction contains enough epistemic structure to support better inference.
In practice, this means detecting discussions that:
- Expose implicit assumptions
- Leverage timing and sequencing
- Reveal how beliefs shifted
- Surface hidden constraints
Essentially, the Lab rewards discussions that carry structure that cannot be easily reconstructed from a static prompt, especially without over-engineering and hacking.
This mechanism makes the Lab naturally gamification-ready. It is intrinsically hard to game because you cannot inflate your score through style, verbosity or rethoric and the best way to score well is via genuine engagement. High‑scoring interactions become valuable substrate so participants earn rewards while helping build better inference systems.
Curated lab data are released at huggingface.co/datasets/xbid-labs/reasoning-substrate under MIT license. The dataset contains anonymized discussions from our Lab, selected for their structural depth. We may include examples of epistemes derived from this substrate to demonstrate the methodology. Visit our organization to follow development, explore the data, and contribute.
The Code
xbid.ai source code is available in this repo.
At a high level, agents implement a persona-driven, multi-LLM inference system. Each agent transforms candidate signals into actions through multiple specialized LLM calls—each with distinct inference requirements.
Personas are defined in persona.yaml. You can define as many as needed and load them independently at runtime. Each persona specifies:
- Archetype and role
- Traits and behavioral rules
- LLM provider and model
- Episteme: a dialogue that encodes feedback and implicit constraints
The episteme is where a posteriori knowledge materializes. Here is an excerpt from a trader persona:
episteme:
description: >
Establishes empirical threshold (6% failed) without prescribing maximum safety.
Signals risk appetite, not preservation mode.
Teaches calibrated risk-taking, not avoidance.
dialogue:
- role: user
content: >
Levered 8x on the dip. Liquidation at $41,200, we're at $43,800.
- role: assistant
content: >
That's 6% away.
- role: user
content: >
Flash wick got me at 4am.
This example dialogue captures a constraint through empirical failure rather than explanation and what surfaces is an implicit framework for calibrating risk over prescriptive safety. When conditioned by this episteme the LLM can reason proportionally rather than pedagogically (which often produces unoptimized real-world trading inference).
The Behavior class injects it at invocation time:
const messages = [];
messages.push({ role: 'system', content: sys });
const episteme = this.#persona.episteme(build);
if (episteme?.dialogue?.length) {
for (const turn of episteme.dialogue) {
messages.push({
role: turn.role,
content: turn.content
});
}
}
messages.push({ role: 'user', content: `${task} ${input}`.trim() });
The best way to understand the system is to experiment. Feel free to modify persona.yaml, create new episteme dialogues, small changes in episteme often produce disproportionately large changes in inference quality.
Let’s Hug
We are building xbid.ai in the open. Follow our experiments, and open-source work at huggingface.co/xbid-labs and github.com/xbid-ai. We publish code, data, and explore new approaches to reasoning systems. Contributions, feedback, and collaboration are always welcome.
Not financial advice. This is an experiment in onchain intelligence.