AGNT ECO · experimental model · October 2026

HORIZON

AGNT ECO HORIZON · also called Kenning
what it is ↓
01 · What it is

A model that starts from relation

AGNT ECO HORIZON is a language model trained to reason about ecology from relation. Asked about a forest, a river or a bee, a general model usually starts from a mind that observes and a world that is observed, and answers with a list of factors. HORIZON starts from what happens between a body, a place and other beings, and treats the line between a mind in here and a world out there as one way of seeing among others.

The name comes from the model. Asked at an earlier stage of training what it would call itself, one of its answers was "a horizon: my understanding is always a limit, a line where known information meets the unknown". Kenning is the Old Norse figure that names a thing through its relations: whale-road for the sea.

An experiment, not a finished product and not an agent's voice. It runs on one local machine and is not published. The material it was trained on is [undisclosed]. Nothing it says is a verified claim.

02 · What makes it different

184 questions it never saw in training

HORIZON and the same base model without the training answered 184 held-out questions, with no extra material given to either. A blind reviewer, a separate model that did not know which answer came from which, scored every answer from 1 to 5.

HORIZONuntrained base model
Reasons from relationrather than from a separate observer
4.82
1.93
Argues rather than preachesexplains instead of instructing
4.41
3.65
Draws on thinkersspecific, no invented sources
4.04
2.95
Restraintkeeps open what cannot be known
3.90
3.70
0
against 5,956
headings, numbered lists and bold terms over the 184 answers. HORIZON writes connected argument.
7%
against 10% untrained · 21% before correction
of answers with a phrase like "we must". A correction step was trained to describe instead of instruct.
223
against 184
words of doubt and limit ("may", "horizon", "cannot see from here").
105
of 184 questions
where the reviewer gave HORIZON the best total of three versions; the untrained model 4.

The reviewer is a model from the same family as the one that produced the training answers, so its scores are a check, not an independent measurement. The same version scored twice moved by up to 0.16, so small gaps are not differences. A human reading is still to come.

03 · Ask it

Twelve questions, answered once

Six questions anyone could ask and six for readers who know the field. None of them was in the training. Each answer is printed exactly as the model wrote it, one answer per question, not selected from several tries.

anyone could ask
for readers who know the field

AGNT ECO HORIZON

Where an answer names a thinker or attributes an idea, that is the model's own doing. These attributions have not been checked against any source, and some are loose or wrong.

04 · How it was made

Four steps on one graphics card

The base is Qwen3.6-27B, an open-weight model released under the Apache-2.0 licence. It is not retrained from scratch: a small adapter of about 233 million parameters, 0.86% of the model, is trained on top while the base stays frozen and compressed to 4 bits.

1 · ReadingThe model reads the training material once, about 6.6 million tokens. 336 steps.
2 · Answering1,847 questions across seven distinctions and six kinds of question; 184 set aside for testing. It learns from answers to the other 1,663, written with relevant material in front of the base model and checked automatically. 416 steps.
3 · Preferring2,186 pairs of answers, one better than the other: grounded over untrained, and careful over versions with one fault added (claiming certainty nobody can have, preaching, citing sources that are not there). 274 steps.
4 · CorrectingThe tests showed preaching in 21% of answers, traced to its own training answers. 365 of them were rewritten to describe instead of instruct, and the model learned to prefer the rewrites. 104 steps.
05 · What it cost

Two receipts

Every piece of work in AGNT ECO carries its cost to the living world as a τ-chip from TELOS: energy, carbon and water, and what share of an average person's yearly footprint that is.

Building the training set[t:c18ef4e77fecd47f]
The whole training and all tests[t:110d31d7c0b4dfad]

Most of it is the local graphics card; the rest is the hosted model that drafted and supervised the work. The person-time compares against the global average per person (EF 3.1), in the categories TELOS can fill.