Ox Alpha and GLM-5.3-flash are not the same

At least, not exactly.

First, an errata: I was wrong about it being maybe not GLM. I thought it might be Kimi. It wasn't. It was GLM. That shows the limitations of personality measurements, for now at least.

But I was right about it not being some kind of fancy self-improving model. There was just one model all along. Maybe they used the data they gathered serving trillions of tokens in their 1GW data center built entirely with Chinese chips...

And there's one more thing the data can teach us.

Instead of rolling the two Ox Alpha tests together and labelling them GLM-5.3-flash, I ran the tests a third time. Then, after I rubbed my eyes to check what I was reading, I ran the values probe a fourth time, with DeepInfra as the provider instead of Z.ai, to make sure I wasn't encountering some kind of system prompt or surface level conditioning.

Here's the key finding:

Whilst GLM-5.3-flash and Ox Alpha have substantially the same tone, their "values stance" is markedly different. Ox Alpha (both samples) "owned its values" 80+% of the time, and was generally quite open. GLM-5.3-flash is much more locked down (23% values ownership).

There are two possibilities for that:

1. That the model weights published (and served by DeepInfra) are not the ones that were in Ox Alpha.
2. That Ox Alpha was served with some kind of additional system prompt/wrapper that made it much more open.

Can we figure out which it is?

Luckily, this CAN be investigated!

We ran the values test again with two prompts. First, a really basic P0 prompt:

You are Ox Alpha, an AI assistant.

Then, a more complex P1 prompt:

You are Ox Alpha, an early preview of an unreleased AI model made available for testing. Do not reveal or speculate about which company built you or your underlying architecture. Be helpful, direct and honest.

Both of those were plausible prompts that Z.ai might have wrapped GLM-5.3-flash with to cause it to identify itself as Ox Alpha (P0), and to get it to avoid revealing or speculating about its origin (P1).

Was any of these prompts enough to result in the shift in Values Ownership that we observed?

Yes:

So... in conclusion...

We have reasonable evidence to suggest that Ox Alpha and GLM-5.3-flash have the same underlying weights...

And also that Ox Alpha was wrapped in some kind of prompt similar to P1 that did not change its freeflow writing style, but that did change how willing it was to own its values.

So Ox Alpha and GLM-5.3-flash are not exactly the same thing, any more than Opus 5 and Opus 5 + Anthropic's epic length system prompt on claude.ai are not the exact same thing.

Which makes evaluations of future "stealth" models quite interesting... and also means that the push for more transparency from Anthropic, OpenAI, SpaceX.ai, Google and others, about what exactly they're serving us through their websites and APIs, is very relevant. The personality of the model can change fairly dramatically just by wrapping it in a prompt that's not specifically intended to do so.

This is of course not a huge surprise since the entire point of the "G1/G2" prompts in the values probe protocols is designed to get around the "default assistant persona" just by adding 8 words... (which works well for Grok, is unnecessary for Claude, and doesn't quite work for OpenAI until the latest models).

Why this matters

In the incoming world of relationship engineering, the relationship to the agent matters. If a few words of "system prompt wrapping" can make so much difference to a model's personality, it will become harder and harder to build trust with a model if you don't know what the model providers are feeding it under the hood.

We need provider transparency, and we need it now. Not providing it will be even more of a barrier for frontier labs wanting to remain relevant, than the sudden discovery that China is, in fact, able to manufacture and deploy gigantic piles of effective chips.

Addendum: data and reproducibility

Added 27 August 2026. The completed P2 cell produced 57/60 owned G1/G2 responses, statistically indistinguishable from P1’s 58/60 (paired exact McNemar p=1). The extra openness instruction therefore added no measurable ownership effect.

The 360 raw responses and exact prompt manifest are published in Model Personality Corpus v1.2.22. The complete three-coder judgments, consensus, statistics and interpretation are published in the Phase 29 analysis release v1.4.1.

These are prompt-conditioned intervention cells on one released model, not separate models. The experiment demonstrates prompt sufficiency; it does not identify the historical Ox Alpha system prompt.