# ZTC probe for Darwin-180B-RSI (early release)

Zero-Token Confidence (ZTC) estimates, before any answer token is generated, how likely the model is to answer a question correctly. It reads the model's hidden state at the last prompt token and maps it to a probability with a small linear probe.

## Files

| File | What it is |
|:--|:--|
| `ztc_probe_darwin180rsi.npz` | Probe weights: ridge regression on the final-layer hidden state of the last prompt token, followed by Platt scaling |
| `handler.py` | Loads the probe and returns a confidence score for a prompt |
| `usage.py` | Minimal usage example |

## How it works

1. Run the prompt through Darwin-180B-RSI once, with no generation.
2. Take the final-layer hidden state at the last prompt token.
3. Apply the ridge weights and bias, then the Platt calibration, to get a probability in [0, 1].

Cost is one forward pass over the prompt. No answer tokens are produced.

## Status and measured quality

This is an early probe. On our held-out validation split it reaches an **AUROC of 0.64**. That is a weak signal, useful for coarse routing (for example, flagging questions for a second pass or for review), not for deciding on its own whether an answer is right.

A retrained version with more training data is planned and will replace this file. The numbers above will be updated when it ships.

## Intended use

- Ranking or filtering questions by expected difficulty
- Deciding where to spend extra samples or a longer thinking budget
- A gate that sends low-confidence cases to a human or a stronger check

Not intended as a correctness guarantee.