Not Every Hallucination Needs More Electricity

Sorry for the longer post. I am a bit frustrated, but I think the frustration points to a real structural issue.

The new wall is not always in front of the idea. Sometimes it is in front of the sender.

I am working on DESi , the Dynamic Epistemic Sequencer. The basic question behind it is simple:

Does this answer actually follow from the given input?

If not, maybe the system should not simply keep generating. Maybe it should stop, route, verify, reduce, reject, or ask for clarification. In a world where billions of AI requests are processed every day, this is not only a quality issue. It is also an efficiency issue.

Not every question needs a giant model. Not every answer deserves a long context. Not every claim should be carried forward. Not every hallucination needs more electricity.

DESi is an attempt to connect reasoning cost to epistemic justification: use stronger computation where it is needed, but avoid wasting inference on unsupported output, wrong methods, missing evidence, or tasks that should have been routed to deterministic tools in the first place.

What frustrates me is that work like this is often not filtered primarily by content.

Reddit asks for karma. arXiv asks for endorsement. LinkedIn may decide I am not a real person. Hacker News can flag faster than anyone can seriously read. Emails disappear because the sender does not have an academic domain.

This is not a conspiracy. It is routine.

Modern platforms mostly do not ask first:

• Is the idea clear?
• Is there code?
• Are the tests reproducible?
• Are the claims smaller than the evidence?

They ask:

• Does the sender fit the expected trust pattern?

That may be necessary for spam defense. But it also means that independent technical work can become invisible before it is evaluated.

And this is the strange loop: a system like DESi could, in principle, help replace some of these crude social filters with more content-aware checks. Instead of asking only whether the sender has karma, institutional affiliation, or the right reputation pattern, one could ask whether the claim is supported, whether the evidence is present, whether the method fits the task, and whether the output can be audited.

But DESi itself remains hard to make visible because the existing filters do not evaluate that kind of structure first. They evaluate the sender.

The irony is sharp: we are building AI systems that should become better at separating evidence from assertion, while the discourse around them is still gated by reputation signals, institutional addresses, account age, karma, and social normality.

DESi is not a world-saving claim. It is a technical attempt to make LLM workflows more procedurally disciplined: route tasks to the right method, keep claims and evidence separate, detect unsupported output earlier, preserve audit trails, and reduce waste where possible.

Maybe DESi is wrong. Then it should be criticized technically. I am not asking for status, conference applause, or better coffee. I am asking for the idea to be judged by its claims, code, tests, and limitations.

If you think this direction is useful, I would appreciate it if you shared the project or pointed it to someone working on efficient, evidence-aware AI systems.

DESi outline:

hstre.github.io

DESi — Deterministic governance for LLM pipelines

Replay-stable. Read-only. Auditable. DESi watches LLM pipelines from the outside — and reports its own failures too.