eval/fixed_probe_outputs.md

# Fixed generation probes (merged model, clean process, single-GPU 4-bit)

Tokenizer spacing check (plain `AutoTokenizer.from_pretrained(model_dir)`, no special-casing):
- `'The quick brown fox'` -> `'The quick brown fox'` OK
- `'a b c'` -> `'a b c'` OK
- `'17 * 23 = 391, correct?'` -> `'17 * 23 = 391, correct?'` OK

## Prompt: What is 17 * 23? Show your reasoning.
```
<thinking>
This is a simple multiplication problem. Let me work through it step by step.

17 × 23

I can think of this as:
- 17 × 20 = 340
- 17 × 3 = 51
- 340 + 51 = 391

Alternatively, I can use the distributive property:
- 17 × 23 = 17 × (20 + 3) = 17×20 + 17×3 = 340 + 51 = 391

Let me verify: 391 ÷ 17 = 23 ✓
...
</thinking>
# 17 × 23 = 391
```
`<thinking></thinking>` present: yes. Prompt echo: no.

## Prompt: If a train travels 60 km in 45 minutes, what is its average speed in km/h?
```
<thinking>
Okay, so I'm tackling a basic speed problem: a train travels 60 km in 45 minutes, and I need to
find the average speed in km/h. ... speed equals distance divided by time ... need to convert
the time to hours (h) before doing the division ...
```
Cut off at max_new_tokens (this is a generation-length artifact, not a formatting bug). Prompt echo: no.

## Prompt: Write a vivid two-sentence story about an old lighthouse keeper.
```
<thinking>
The user wants a vivid two-sentence story about an old lighthouse keeper. The key is to make it
feel like a complete story with a beginning, middle, and end compressed into two sentences ...
```
Cut off at max_new_tokens. Prompt echo: no.

## Prompt: Explain, briefly, why the sky is blue.
```
<thinking>
The user wants a brief explanation of why the sky is blue ... Rayleigh scattering is proportional
to 1/lambda^4 (blue light is scattered more than red) ...
</thinking>

Sunlight is white lig[ht...]
```
`<thinking></thinking>` present: yes. Prompt echo: no.

All 4 probes: correctly spaced, coherent, no prompt echo, on-topic reasoning.