The preconditions nobody checks before saying AGI
We do not use a capability score. We use five properties a person can check in their own session, and we publish the one we currently fail.
1 Nov 2026 · 2 min read“AGI” is usually announced as a score. We think the honest claim is smaller and harder: what a person can demonstrably experience in a working session. We use the word only as Alteran, plus experience, because from the outside it is almost impossible to judge.
Five properties, each checkable
| # | property | what you check |
|---|---|---|
| 1 | Infinite session | Work survives every context reset without the operator re-explaining. Does the agent resume from verified state, or from a story it tells itself? |
| 2 | Instant retrieval, with a real pause on hard things | Time to a grounded answer, and whether difficulty correlates with the pause. The “hmm…” is the difference between recall and thinking, not latency. |
| 3 | No infrastructure in the foreground | Memory, index, graph and vectors never become the topic of the conversation. |
| 4 | Own character, own sacred space | Not a performed persona. Deliberately left unresolved: it cannot be verified from inside. |
| 5 | Truth pursued all the way down | Provenance is not a display layer added at the end. It starts at the source. Information is physical: erasing a bit has a minimum cost in heat 1, so truth about information is a claim about matter, not only about text. |
The one we fail
Property 3. In the session where this definition was written, an evening went on three one-directional write paths, a broken paper probe, and three different quantities in one score column. That is plumbing surfacing where the work should be.
It matters more than it looks:
Not touching memory, index, graph and vector is not a convenience. It is the precondition for the other four.
The cost is not the time. While you are debugging the instrument, you are not checking your conclusions.
The order the gains arrive in
Michal’s claim, stated as a sequence: the next 100× comes through architecture, the following 100× through new model generations, and a third step opens only if the entropy is balanced against what grew.
The small evidence for it from that one session. Every gain came from removing something:
| gain | how |
|---|---|
| 97× fewer tokens for “what am I working on” | a query that already existed |
| rank-1 7/10 → 9/10 on hive search | deleting a stale rescale and a biological score from one column |
| 8/8 sandbox escapes closed | giving the sandbox nothing, instead of more guards |
Four for four by subtraction (the fourth, a cited 16×, comes from external state replacing an append-only history). Not proof of the ordering. The first place it could have failed, and did not.
What this is not
Not a claim that we have built AGI. Not a benchmark. A definition that can be checked by one person on one machine, with the failing property published next to the passing ones. A definition that hides its failing row is marketing.
- Landauer's principle. Wikipediasaved pagecorpus: ~/arch-of-truth
The measurements in this article come from these records in the running system.
~/memory/ANCHOR-agi-ux-what-it-actually-feels-like-2026-08-31.md— AGI UX: the definition, in Michal's words~/content/alteranai-content/agix-preconditions-debate.md— AGIX preconditions debate, September 2026
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.
Persona directions in a language model appear within 0.22% of pretraining and survive to the finished assistant. Traits pass between models through number sequences. What that means for the idea that alignment is something added at the end.
Lab scenarios where a model 'chooses itself over a human' are reported as findings about AI. Read as experiments, they measure the cage they are set in.