Research
Kherad is an experiment in building a small, transparent mind that learns from its own experience — grounded first, taught second. It is not a scaled-up language model, not a claim of AGI, and not a product. It is a research instrument, and the value of an instrument is in what it honestly measures.
The thesis
Every large language model today is built the same way: ingest a vast corpus of human-generated text, and produce a system that recalls and remixes it fluently. It is an extraordinary engineering achievement, and it has a structural ceiling — it can only ever reproduce patterns that already exist in its training data. It does not test what it says against the world; it cannot invent what nobody has written.
Kherad asks a different question: what if a mind grounded everything it believes in its own consequences first, treated everything else — including the entirety of human knowledge — as testimony to be doubted and tested rather than truth to be memorized, and kept every faculty it grew only because it was earned by real demand, never installed because a designer assumed it was needed?
Methodology
Every experiment is preregistered: a written prediction, a pass/fail bar, and a kill criterion are frozen before the run, not after. Every result is logged in the permanent record — including the failures, which are never quietly dropped. As of this writing, the record holds 69 entries, spanning three eras of substrate (a hand-built grid world, a real filesystem habitat, and a replayed real-history "tape world"), with a mix of passes, failures, and mixed verdicts — see the full timeline.
One especially important entry: Entry 65, graded a full failure. A hoarding faculty was added to the creature and made every measured outcome worse. Rather than patch around it, the finding was traced to its root — the world itself could not ask a storage question, because it had only one essential variable. That failure triggered a full constitutional reset: the "Second Founding," which now governs every entry from 67 onward, with a moratorium on building new worlds until the mind itself has caught up.
What's being tested next
The current roadmap runs across eight epochs — from the mind composing its own internal structure, to evolution running in-machine as a stand-in for the deep-time search that produced biological brains, to a mind that can read human knowledge and decide for itself whether to trust it, to invention that goes beyond anything it was given, to transfer across bodies, to a shared language between multiple minds, and finally to communication with us.
An LLM tells you what humanity already wrote. Kherad is built to eventually tell you what nobody has told it yet — and to show its work.
What is public, and what isn't
The full narrative, every graded result, and the methodology are public — see the Field Notes and the Timeline. Implementation details — genomes, hyperparameters, world configurations, and source code — are held privately while this remains a bootstrapped, pre-funding effort. A curated technical paper is in preparation; in the meantime the full internal record is timestamped for priority and available on request for serious research inquiries.