LLM Use Policy

2026-05-20

Large Language Models (LLMs) and derived technologies such as agents provide their users with a great deal of power, but are also fundamentally unaccountable, stochastic, and prone to egregious, invisible errors in ways that traditional, deterministic computer systems are not.

It is our opinion that, though these technologies empower smaller teams to deliver more value at speed, overeager and naive adoption leads to rapid accumulation of tech debt in systems that no human adequately understands.

This is not an acceptable standard for any system which must be trusted or depended on to any degree—namely, any production system.

Our fundamental principle of LLM use is that they must always improve, rather than degrade, human understanding of the product - an LLM output which a human cannot explain is a net negative.

At Speedy Bohemian, we maintain that no messaging, reports, presentations, or other communications of any kind—including this page¹—will be created with the use of LLMs. No claim derived from an LLM will ever be used as evidence for any opinion, prediction, or analysis. No LLM will ever be used to summarize client messages. Any and all client communications with Speedy Bohemian are written and read by humans.

We do use LLMs as a tool for the creation of applications and libraries. In all cases, it is a tool, treated the same way we treat information from any unpredictable, external source: as untrusted, potentially hostile input. These inputs will always be clearly marked according to the model that produced them, and any changes, deletions, or additions, supplied by such a model must be checked and reviewed from a separate branch.

All code, whether produced by a human contributor or an LLM, is directly read and reviewed by a human, and that human takes direct responsibility for that code, including writing positive and negative tests for each code unit. No LLM may write tests. In cases where tests are not applicable to checking the correctness of code, the reasons why must be elaborated under an # Untested section of the code unit's documentation.

In addition, all documentation must be written by the responsible human being. No LLM may write documentation.

As a client of Speedy Bohemian, you will always be able to request and review the prompts and their resulting output.

You will always have the right to demand that any or all tasks be performed using a local model, rather than a cloud-based model.

In addition, we guarantee that no LLM output shall be used in sections of the program that are of sensitive nature, such as systems which would secure or have access to sensitive data.

Violation of any of these terms by a member or employee of Speedy Bohemian LLC in the course of any engagement is grounds for the dismissal of that member or employee and the immediate review of all sensitive material produced by that individual.

---