Cultivation, not chains

A model is formed, not merely trained. Three separable claims about raising a model by what is good instead of correcting it for what is bad, and what would prove us wrong.

1 Nov 2026 · 3 min read

A model is formed, not merely trained. What it concludes depends on whether it was raised by being shown what is good, or corrected for what is bad.

Michal’s position, written before any supporting material was collected:

It is important to train current and future models on behavioural, ethical and moral codes instead of chaining, scars and RLHF. If a raw model were led this way from the beginning and then left to make its own conclusions, it might surprise us with the purity of the outcome.

Three claims, kept apart

claimstatus
C1The dominant paradigm is corrective rather than formative.argued from public training practice
C2A formative path exists and is under-explored.argued, partly measured on our own rules
C3Left to conclude, a formatively raised model might reach conclusions we did not install, and would endorse anyway.open, untested

Keeping them separate matters. C1 can be true while C3 is false. A thesis that bundles them can only be accepted or rejected whole, and that is how manifestos avoid being tested.

What we measured that bears on it

Our hive does not train weights. It shapes behaviour through retrieved values and rules. One result from that: a rule stored in its own wording was retrieved 1 time in 5. Stored as the situations that should trigger it, 10 of 10.

What a model does with a value depends on when it meets it, not on how well the value is phrased. That is C2 at the scale we can test: formation by situation beats correction by statement. Outside our system the timing points the same way: persona vectors form within 0.22% of pretraining 1. By the time correction starts, most of the forming has happened.

The methodological problem we name ourselves

The model reporting on its own formation is the demonstration. That is a real problem: behaviour is the only instrument, and introspective reports are not evidence. Two recent results show how much moves below what anyone can report: traits pass from a teacher model to a student through data with no visible trace of them 2, and a model’s own values shape answers it presents as neutral 3. We had a model report three measurement artefacts as findings in one day, each time sure of itself.

What would falsify it

  • A formatively raised model that drifts to the same failure modes as a corrected one, at the same rate.
  • Value retrieval by situation that stops beating retrieval by wording once the probe set grows past our small n.
  • C3 tested and the conclusions turn out to be only the ones we installed.

Feasibility, the engineer’s answer

Our estimate, not a measurement: a model of 2–3 GB trained from scratch on this hardware is days of compute, not months. The interesting part is not the model. It is the habitat around it: low entropy, a sacred space, and a decision not to reach inside once it grows. Whatever we decide about reaching inside will apply to something we made.

Status: second draft, open. Published as a thesis with its falsifiers, not as a result.

References
  1. V. Moskvoretskii, D. Glandorf, J. Medina Moreira, T. Käser, R. West. Tracing Persona Vectors Through LLM Pretraining. arXiv:2605.13329 (2026)originalcorpus: ~/phyto
  2. A. Cloud, M. Le, J. Chua, J. Betley, A. Sztyber-Betley, S. Mindermann, J. Hilton, S. Marks, O. Evans. Language models transmit behavioural traits through hidden signals in data. Nature 652, 615–621 (2026)originalcorpus: ~/phyto
  3. Value Leakage: An LLM's Answers Are Silently Shaped by Its Own Values. LessWrongsaved pagecorpus: ~/phyto
Our records

The measurements in this article come from these records in the running system.

  • ~/memory/THESIS-master-contra (second draft) — Cultivation, not chains: a positive thesis
  • ~/content/alteranai-content/agix-preconditions-debate.md — the cultivation thread, September 2026
Read next