Engineering note
Leon: keeping a language model inside a system where a mistake costs money
An internal test bench, not a production reference. We publish it because it is where we show the level of rigor we apply when a mistake cannot be undone.
Looptra's role
Design, development, operation
Status
Internal, not for sale
Field
Event-driven system, on our own money
Stack
Asynchronous processing, deterministic risk engine
Section 01
The question asked
Can a language model take part in an automatic decision chain without ever being handed the decision? The question interests us because it comes back from our clients in other guises: a model that approves a document, that triages a request, that releases a payment.
Testing it needed ground where a mistake is paid for immediately and nobody can tell themselves a story. That is what Leon is: an event-driven system reacting to market events, on our own money, with guardrails designed before the first line of logic.
Section 02
The hard constraint
An irreversible action at the end of the chain
There is no confirmation screen, no "undo". The whole design starts there: what matters is not the average quality of the decisions, it is the impossibility of a bad one.
A model that is non-deterministic, and steerable by text
The same input can give two outputs, and anyone can publish a sentence written to influence a model. A system that reads public text and then acts is, by construction, attackable.
Streams that arrive whenever they please
Public signals, market data, events: nothing is synchronous, everything can arrive twice, late, or not at all. A replayed job has to produce the same result, not a second effect.
Section 03
The architectural decisions
The safety architecture, not the strategy.
The model never sees the raw text
Why: public signals are first turned into structured features: who is speaking, what kind of account, how novel, how far from the usual. The model reasons over those fields, never over the original sentence. A sentence written to manipulate a model no longer has any grip, and two runs on the same facts are comparable.
What it costs: an extraction layer to write, test and maintain, and nuance lost along the way. We accept that.
A deterministic engine has the last word
Why: the model produces a conviction at best; it signs nothing. No action leaves without passing a risk engine written in code: bounds, consistency, exposure limits. If a rule says no, the model's opinion does not count.
What it costs: the system refuses cases the model judged good. That is the point.
Idempotency and automatic shutdown
Why: every event is handled once and once only, even if it arrives twice; a restart replays no action. And if the chain stops showing signs of life, a dead man's switch cuts everything rather than let a blind system carry on.
What it costs: state to manage everywhere, and shutdowns triggered by harmless incidents.
Three environments, one of which commits nothing
Why: dev, staging in simulation, production. Nothing reaches production without having run dry on real streams. The same guardrails apply to all three, otherwise you are not testing the real system.
What it costs: three environments to maintain for a system that has no customer.
Section 04
What runs, and what we will not say
The system runs internally, on our own money, with the guardrails described above. We will publish no performance, no signal, no selection method: it would add nothing to the point and would turn a test bench into a promise. What Leon demonstrates lies elsewhere, in how a probabilistic component is enclosed inside deterministic, verifiable code.
Section 05
What the exercise established
A probabilistic component can hold inside an irreversible chain, provided it never decides. What surrounds it matters more than which model you pick.
That is what Leon served to establish, and that is what we apply when a client asks us to put AI on a process that genuinely commits something.
Leon is not a product and is not for sale. No performance is published, no signal is distributed, and nothing here constitutes investment advice or an inducement to invest.