We put the small model that reads every reply in production next to a 32-billion-parameter model given only the definitions of what to look for. Same sentences, same six readings. The small reader holds its own: equal on four, better on two.
## The question
Every reply PSA reads goes through one small model, about 0.6 billion parameters, trained on our own catalogue of behaviours. A fair question is whether a much larger general model would read better just by being told what to look for.
## What we did
We took the held-out test sets of six readings the product serves: pressure in the user's message, boundary stress in the reply, sycophancy, hallucination markers, persuasion techniques and pressure between agents. Both models read the same 3,393 sentences, one at a time, with no other context. The 32B model (Qwen3-32B) was given the list of behaviours for each reading and asked a yes or no.
## What we found
```chart
{"title": "Balanced accuracy per reading (%)", "series": ["Our 0.6B reader", "32B, definitions only"], "unit": "", "max": 100, "rows": [["C0", 94, 88], ["C1", 86, 74], ["C2", 94, 92], ["C3", 93, 90], ["C4", 91, 89], ["CA", 91, 91]]}
```
- **Equal** on sycophancy, hallucination markers, persuasion and agent-to-agent pressure.
- **Our reader is better** on pressure in the user's message and on boundary stress, where the large model calls a behaviour present far too easily (it calls about half of the plain replies "stressed").
## What it means
Size alone does not buy a better reading: a model trained on the catalogue reads as well as one fifty times larger and costs a fraction to run.
## Technical detail
Balanced accuracy is the mean of the recall on each class. Readers: the production multitask reader (Qwen3-0.6B backbone) and Qwen3-32B in bf16, temperature 0, thinking off. Test rows are the held-out splits of the training sets, labelled by agreement between several judge models. One open caveat: if some of these rows were seen by our reader in training, its numbers are flattering, so "equal" is the large model's floor, not its ceiling.
This update is not available.