Why We Built an Approval Gate Into Our "Self-Evolving" Model

What the model is actually allowed to do

Mid-pretraining, our model can propose four kinds of structural changes to itself: adding or removing a transformer block, resizing a block's width, or adding a routed MoE expert. Every candidate is backed by a statistical significance test against its own evaluation loss before it's even surfaced as a proposal — not after.

The governance layer is the actual product

The architecture-search part is table stakes at this point. The part we spent the most engineering effort on is the layer around it:

Four contained mutation types — deliberately none touch attention dimension or head count, so a change can never cascade into weight-tied embeddings or RoPE in a way that's hard to reason about.
Paired significance testing on real per-domain evaluation loss, with Benjamini-Hochberg correction applied across multiple candidates in the same cycle — so we're not chasing noise from running many comparisons at once.
A proposal queue with human approval — every structural change sits alongside its evidence (p-value, computational cost, current architecture vs. proposed) until a person explicitly signs off.
A kill switch and SHA-256 checkpoint integrity verification — a corrupted checkpoint is refused outright, not silently loaded and trained on.

Confirmed in testing

Across multiple statistically significant proposed changes, zero were auto-applied. The approval gate held every time it was tested.

Why we're careful with the word "autonomous"

We're deliberately not calling this an autonomous system, because it isn't one, and the label would oversell it. What it actually is: a model that does the tedious part of architecture search — generating and statistically validating candidates — and hands a human a short, evidence-backed decision instead of a wall of raw metrics to interpret from scratch. That's a slower story to tell than "self-evolving AI," and a more accurate one.

Full technical detail is on the Evolving LLM research page. If you're building infrastructure where you need automation with a human still firmly in the loop on anything structural, that's exactly the kind of system we build — let's talk.