Growing a mind
from its own experience

Kherad is building a small, transparent mind whose core capacities are earned through consequence rather than installed from human data or designer intuition. Every result is kept. Every failure is on the record.

69 entries in the public record · 2026 → present

Working thesis

Testimony, not truth.

Large language models recall and remix what humanity has already written. Kherad treats the same body of knowledge as testimony: claims to be tested against its own lived experience, kept only when they pay their way.

Grounded

Belief must be earned by consequence first. Nothing is trusted before it has survived testing against reality.

Cumulative

What is earned compounds. Every new domain or body should be cheaper than the last.

Honest by construction

Every belief carries provenance — lived, told, or invented — and the record of how well it has been tested.

The program

A mind that earns what it knows.

Kherad grows a small, transparent mind whose faculties are earned from consequence rather than installed by assumption. Human knowledge enters as testimony to be tested, never as unexamined truth. The work is organised around two core ideas.

Organ Doctrine

Faculties, not assumptions

Organs (computational capacities) arrive through three documented channels: endowed at the beginning (biological priors translated into code), tuned by population search in the machine, or grown on measured demand when the existing structure cannot pay for new competence. Nothing is added because it “should” exist.

Evolution in the machine

The deep search runs here

Instead of ingesting human data at planetary scale, we run selection over mind architectures. Genomes define how minds grow. Selection pressure comes from performance on tasks the current mind must actually solve.

Measured state

What is actually true at this moment.

All active development uses small, synthetic, fully specified tasks. This is deliberate. The program remains early. We measure foundations cleanly before claiming reach.

Epoch
Epoch 2 — Evolution
R3 is active. First heritable selection on composition-task proxies has been run and graded under preregistered bars.
Kept organs
2 kept
Structure Composition and Consolidation. Both have passed coordination tests with adjacent organs. One organ has been dropped after impedance measurement.
Scope
Synthetic tasks only
8-dimensional boolean feature vectors. No embodiment, no real-world data, no language. All generalisation claims are currently seed-held-out within this family.
Public record
69 entries
Every gate preregistered before execution. Failures and parked bars are retained. Nothing has been retracted.

See the full current program →

Lab Notes

The technical record.

Regular, precise technical memos. These are the primary ongoing output — scoped, dated, and written so that the actual state of the work is visible.

Read the latest notes →

The road

Eight epochs, one direction.

The program is structured as eight epochs. Each must be earned before the next is credited.

E0 E1 E2 E3 E4 E5 E6 E7

See the full record →

From the bench

Instruments, as they are earned.

Tools are extracted only after the research has already needed them to exist. Nothing is built for its own sake or for a market.

View instruments →

Following the work

For researchers, builders, and serious observers.

The public record, Lab Notes, and timeline are the main way to follow the program. If you are working on related problems and want to reach out directly, the contact details are on the About page. This is long-term work. We move at the speed of measured results.