Research
Kherad is an independent research program growing a small, transparent mind that earns its capacities from its own experience. It is not a scaled language model and makes no claim to generality yet. It is a long-term research instrument.
Core ideas
The Organ Doctrine
Faculties (organs) enter through three channels:
- Channel A — Endowed: translated biological priors installed at the start, each subject to retention gates.
- Channel B — Tuned: parameters and topology adjusted by evolution in the machine.
- Channel C — Grown: new organs composed or added when the existing structure demonstrates demand.
Organs are capacities, not answers. No facts are installed. Only performance under test keeps an organ.
Evolution as pretraining
Instead of ingesting human data at scale, we run population search over mind architectures. Genomes define growth rules. Selection pressure comes from performance the mind must actually solve.
Current measured state (sanitized)
All active development uses small, synthetic, 8-dimensional boolean feature vector tasks. This scope is deliberate.
- Epoch 2 (Evolution) is active. First heritable selection campaigns on composition-task proxies have been run and graded.
- Four organs have passed independent preregistered trials and are kept. One more is carried by credit.
- Working memory was measured and dropped due to substrate impedance.
- No embodiment, no language, no real-world data yet. All generalisation claims are currently seed-held-out within the registered task family.
Research areas
How the mind discovers and compounds reusable structure from primitives. Current main line for Epoch 1–2.
Running population search in the machine as the long prior. Genomes, harnesses, heritability, and coordination with grown organs.
Preregistration, parking, reproducibility tiers, coordination sweeps, Weismann audits, and provenance tracking. This is the main thing being hardened now.
Methodology
Every experiment is preregistered with a written prediction, pass/fail bar, and kill criterion before any code runs. Results are kept regardless of outcome. Parks are documented when a frozen bar is discovered to be unmeasurable.
Coordination sweeps, reproducibility tiers, and Weismann barrier audits are standing practice.
What is public and what is not
The structure of the program, the methodology, the high-level measured state, and every graded outcome in the public record are public. Genomes, exact hyperparameters, full primitive sets, and running implementations are withheld.
See the Lab Notes for technical detail and Timeline for the full public record.