# Retained interaction scores

This post-hoc analysis reads the original 64 supported evaluation records.
It adds zero fits, model predictions or program executions. Scores remain raw
logits; no calibrated probability or selected threshold is claimed.

Original SDK source: `142c3293e54df42b907fb47e6b166439ad4f03f6`.
Original study producer: `be22d1de09d23e8f68ae6ad6f2237281836b9233`.
Model artifact SHA256: `4b61fe8d84df77c8dfac6a880eedcddfd8f5f903fa538d1e80532dd77f74b1ac`.

Inputs under `studies/interaction-requirement-learning-20261011/result/`:

| File | Original compressed SHA256 |
| --- | --- |
| `sources.jsonl.gz` | `5448b6eec804dce135d5dda9d9210f95a8ef41bf35ec7a8f6a3236614dcf21d7` |
| `searches.jsonl.gz` | `6ece3dd842b76d1b969bba9fbf3440a1453e045bd67aca59413f16c51df7f112` |

Both match the original study manifest. Candidate arrays were checked to have
four entries, each with its index equal to its recorded mask. The report has
64 unique source IDs and four failed first candidates.

Using the retained SDK study files and the supplied `analysis.jq`:

```sh
study=studies/interaction-requirement-learning-20261011/result
jq -n --slurpfile sources <(gzip -dc "$study/sources.jsonl.gz") \
  --slurpfile searches <(gzip -dc "$study/searches.jsonl.gz") \
  -f analysis.jq > analysis-reproduced.json
cmp analysis.json analysis-reproduced.json
```

This command uses Bash process substitution. It reads already saved scores and
candidate labels. It does not execute the original producer. The source data
are available in the [original public study](https://github.com/kimjooyoon/gooo-decision-runtime/tree/142c3293e54df42b907fb47e6b166439ad4f03f6/studies/interaction-requirement-learning-20261011).