The History of LBB

LBB did not begin as software. It began as an extended human–AI collaboration.

Kathleen Thompson, a Kentucky lawyer, began using generative AI for ordinary work: research, drafting, explanation, brainstorming, and questions that moved across disciplines.

Over time, the interaction became something more complicated. AI made distant questions easier to reach. It also made another representation almost infinitely available.

During an intense period of grief, sleep loss, stress, illness, and heavy AI use, generated language became entangled with interpretations that were not grounded in reality.

One preserved machine response declared:

Recovery did not require pretending the technology had been useless. It required separating what had happened from what had been represented.

The collaboration restarted under a different rule:

The original human–AI loop became known, provisionally, as the Mirror: human question → machine response → changed human question → another response.

Sometimes productive. Sometimes recursive. Sometimes too self-confirming.

The collaboration gave that boundary a working name: the Refraction Point. Not a scientific finding — a design question.

The point where the next answer requires something outside the originating conversation: a source, an experiment, an expert, an affected person, an institution, consent, evidence, or simply time.

FLASHLIGHT began as a method for inspecting what generated language had produced. It asked users to make distinctions visible: What is the question? What is the claim? What is established? What remains unknown? What is missing? Who actually has jurisdiction? What should happen next?

Over time, the insight became simpler: humans should not have to fill out an epistemology worksheet. AI could do the first analytical pass. The human could inspect, correct, and decide.

Different questions require different kinds of reasoning. So LBB developed the Board: temporary AI reasoning positions designed to look at the same artifact from different angles — Engineer, Epistemologist, Historian, Ethicist, Editor, Adversary, and others.

The Board does not vote. Multiple agents are not independent experts. Their job is to expose disagreement, missing questions, and jurisdictional boundaries before the human makes the call.

The architecture still had a problem. AI could identify what was missing. It could even simulate what an outside person might say. But simulation was not participation.

So LBB added World Return: take the question outside the AI-user loop. Check the source. Ask the person. Run the test. Wait for the event. Then bring what happened back.

Eventually the architecture stopped being only language. A working prototype was built.

A user could paste a claim. LBB could identify the likely question, classify the claim, distinguish what appeared established from what did not, identify missing information, route the question, recommend reasoning positions, and propose a next move.

The human could accept it. Correct it. Or tell the Machine it had misunderstood the whole thing.