You assume this is a GPT call behind a form. You would be right to — most of them are.
The skepticism
“Isn’t every AI fintech just a thin GPT wrapper?”
Mostly, yes — and that skepticism is correct engineering hygiene. A wrapper takes your inputs, prompts a frozen general model, and renders the reply as a chart. Nothing proprietary touches your distribution.
So interrogate it. The question is not “is there an LLM in the loop” — it is “what object does the tail estimation.” That single distinction is the whole of our research thesis, and it is the one most fintech demos cannot survive.
The cost
What does trusting a token-predictor with tail mass cost you?
A generic LLM predicts the next plausible token. Plausible means high-probability, so it pulls toward the consensus middle of what it has seen — the regime where you do not need help. It has no native representation of a return distribution’s tail mass; it narrates a forecast, it does not allocate probability to a –40% session.
The cost is quantifiable. A Gaussian baseline — the implicit prior a “reasonable-sounding” model regresses toward — recovers an S&P 500 kurtosis of 2.90 against a real 16.70, underestimating density everywhere beyond |r| > 5% (Fukunishi, Qiu & Takahashi, Generating Synthetic Stock Return Distributions with Diffusion Models, 2026). At the 1% level its Expected Shortfall understates the real loss by 2.73 points. A consensus-seeking estimator does not just miss the crash — it certifies the crash as improbable, in the exact session your concentration breaks. The full derivation of why kurtosis collapses under a next-token prior sits in the quantitative method.
The right object
What do you actually want a model to do here?
Allocate mass where the loss lives. You want the estimator that says the tail is fat when the tail is fat, and prints how often it is wrong — not the one that produces fluent prose and quietly assumes the middle.
How it works
How it works
The engine does not ask a language model to imagine a crash. It learns the full return distribution and samples from it.
The Sandbox Engine — diffusion, not next-token
A diffusion model reverses a noise process: from pure noise it denoises back to the data distribution, learning the shape of returns directly rather than predicting the next likely value. Mass goes to the heavy tails by construction, not by narration. The same engine drives the tail-risk diagnostic you can run on your own book.
model · tail-aware diffusion · data · ~20y S&P/FTSE/TOPIX · coverage · 50,000 paths · stated failure rate · 4.2% false-comfort
In the source benchmark the diffusion model recovers S&P kurtosis of 13.38 vs the real 16.70 and matches 1% Expected Shortfall to within 0.00, where the token-consensus baseline collapses to 2.90 (Fukunishi et al., 2026). We publish a 4.2% false-comfort rate — runs that report a tail smaller than realized — because a model that hides its error rate is the wrapper.
The role of language
“Then why is there an LLM here at all?”
Because language is the interface, not the estimator. A model may parse your three inputs and write the document in plain English — orchestration. The tail math runs in the diffusion engine, the equations export, and you can edit the inputs and re-run inside the portfolio dashboard. A wrapper hands you prose; this hands you the distribution to audit. If you still expect a hidden GPT call, read the practical objection we field most.
3 fields · 48-hour document · no call, no sequence.
Frequently asked questions
Can an LLM model tail risk?
No. A next-token LLM optimizes for the most probable continuation, which pulls every estimate toward the consensus middle — the regime that needs no help. It has no native representation of crash mass, so a Gaussian-style prior recovers S&P 500 kurtosis of 2.90 against a real 16.70 (Fukunishi et al., 2026). Tail estimation requires an object that allocates probability to the heavy tail directly.
Is Entail Capital just a thin GPT wrapper?
No, and the distinction is auditable rather than asserted. A wrapper prompts a frozen general model and renders the reply; nothing proprietary touches your distribution. Here the tail math runs in a separate diffusion engine, the equations export, and you can edit inputs and re-run — a wrapper cannot hand you a distribution to audit.
What does the LLM do versus the diffusion model?
The LLM is the interface, not the estimator. It parses your three inputs and writes the document in plain English — pure orchestration. The diffusion model does the quantitative work: it learns the full return distribution from ~20 years of S&P, FTSE, and TOPIX data and samples 50,000 paths to price the tail.
How can I verify the model is not just narrating a forecast?
We publish a 4.2% false-comfort rate — the share of runs that report a tail smaller than realized — because a model that hides its error rate is the wrapper. The benchmark also shows the diffusion engine matching 1% Expected Shortfall to within 0.00, where the token-consensus baseline understates the real loss by 2.73 points. Both figures, and the input equations, are exportable for independent review.
Why use AI for tail risk at all rather than a closed-form model?
Because real return distributions are non-Gaussian and a single closed-form prior misprices the tail by construction. A diffusion model learns the empirical shape directly instead of assuming it, placing mass on heavy tails where the loss lives. AI here is a tool for honest distribution fitting, not a substitute for stating its own failure rate.