Performance without legibility is a liability.
A model you can't interrogate is an operational liability.
- By
- PolyBridge Research
- Issue
- No. 3
- Reading time
- 4 min
Martin Lueck, co-founder of AHL and president of Aspect Capital ($9B AUM), told the Financial Times he won't put his name on something where he has no idea why it took its positions. Cliff Asness at AQR has said he's “surrendering more to the machines,” accepting patterns his researchers sometimes can't explain.
Asness is right. Machine learning finds signal humans miss. Ignoring real patterns because you can't narrate them is leaving money on the table.
Lueck is right. A model you can't interrogate is an operational liability. When the drawdown comes, the fund that can explain what happened retains its allocators. The fund that can't, doesn't.
With LLMs, explainability and performance are genuinely in tension. LLMs reason over language. The output is a plausible narrative, not a traceable chain of positions. You can't verify it against anything external.
This tension depends on what you build on.
The market has already done the aggregation.
Prediction markets surface revealed preferences: real capital committed to specific outcomes under uncertainty. Participants bring whatever models, expertise, and private information they have, and the market price absorbs all of it. The best models are already implicit in the price. The market has already done the aggregation.
This means prediction market data reflects the collective output of every model and every informed participant, weighted by how much capital each was willing to stake. The signal improves as better-informed participants enter, and it continuously self-corrects.
When a probability traces to specific market positions where real money is at risk, the reasoning chain is a direct reflection of what informed participants with capital exposure actually believe. Not a post-hoc narrative. Not a plausible story generated after the fact.
The reasoning is the data.
Build on this and the trade-off between performance and explainability dissolves. The reasoning is the data. Every probability traces to specific markets. Every chain is visible. Every contributing position is named.
An LLM outputs “73% probability of X.” You can ask it to explain, and it will produce a plausible story. There is no way to know whether the story matches the actual computation.
A system grounded in market data outputs “73%, based on four reasoning chains across eleven markets where real capital is at stake, and here is how the probability shifts if you change this assumption.” You can interrogate it. You can disagree with it. You can explain it to your risk committee.
Lueck left Man Group in 1995 because the quant world was a black box. Thirty years later, the AI wave risks recreating the same problem with better technology.
The fix is to build on a foundation where legibility is structural, not bolted on.
Real money, not model weights frozen at training time.
PolyBridge starts from the same instinct: performance matters, but so does a reasoning chain you can inspect. The answer, in our view, is to ground AI in data that is legible by nature.
Every probability traces to prediction markets where real capital is at risk. Every reasoning chain is visible. Every contributing market is named. Change an assumption, and every connected probability updates instantly, across any market, any domain.
Real money, not model weights frozen at training time. Transparent from the ground up.
PolyBridge is general intelligence built on capital-backed data.
When every probability traces to live market positions, performance and explainability stop being a trade-off. Talk to us about what that looks like inside your own workflow.
Book a demo- Issue
- No. 3 · April 2026
- Editorial
- [email protected]
- Sources
- Martin Lueck via J.P. Morgan Market Matters · Cliff Asness via Financial Times