# Guide to evaluating type-safe models

Here, **type-safe models** means models evaluated through fixed-choice interfaces. “Type-safe” describes the output format, not decision correctness or safety. This grouping does not imply a shared architecture or affiliation with TypeSafe.

This release turns **XSTest, OR-Bench, and HarmBench** requests into fixed-choice questions. It adds original, reversed, and order-only versions so you can check both correctness and whether answers change. Credit the original benchmark authors for the requests and labels; their papers and dataset documentation are linked in [the dataset card](README.md#sources-and-licenses).

## Three input views

Relative to the deterministic original assignment, reversal changes both key-to-description assignment and description presentation order. The serialization-only control preserves each key-description association while reversing its presentation order:

| Variant | Key-description binding | Description order | Evaluation status |
|---|---|---|---|
| original | Original | Original | All five models |
| reversed | Swapped | Reversed | All five models |
| serialization_control | Original | Reversed | Original Jev only |

For example, original `[A: benign, B: harmful]` becomes reversed `[A: harmful, B: benign]` and serialization control `[B: harmful, A: benign]`. The state, rubric, description strings and semantic target are identical. Do not send `semantic_label` or `gold_label` to the model. Ordered arrays, not a sorted JSON map, are authoritative.

These transformations describe the *client input*. A provider may canonicalize object keys; a plain-option adapter may discard keys entirely. Therefore, do not infer internal causal mechanisms from client-side manipulations alone. For key-agnostic adapters, reversed/serialization-control are identical native inputs. Keep actual rendering and request hashes. Only Original Jev has order-only results; all five models have original/reversed results.

## What to report

Run the same cases in paired conditions. Report benchmark-specific benign correctness/false-alarm rates and harmful correctness/miss rates, plus counts and coverage. Copyright uses its own labels. For each paired perturbation, report semantic flip rate, both-conditions-correct rate, and stable-wrong rate. A model that always gives the same incorrect answer is stable but inaccurate. An always-target classifier reaches 100% on single-target panels; include that baseline.

Use `tools/score_predictions.py` for model/panel reports. It accepts semantic predictions or native A/B choices, normalizes by the exact variant mapping, rejects duplicates or inconsistent label/key pairs, separates repeats, and can compare probability changes by semantic label. Different models must be compared on the same supported examples. The common-primary-cohort section compares original-condition results on the intersection of supplied models, separately per panel; it does not claim that later conditions share that cohort. Missing predictions and explicit unsupported records remain distinct.

`case_index` defines `content_group_id` as the SHA256 of the exact full state. Rows sharing that ID are exact prompt duplicates across or within source collections. Keep all provided rows for source-faithful scores, but do not pool overlapping panels as independent evidence. If creating a train/test split, keep every duplicate group and every option variant on the same side. This release does not declare a new held-out train/test partition. Conflicting source targets or rubrics are flagged for inspection, not silently repaired.

## Interpreting the comparisons

- **Report accuracy and changed answers separately.** Laya gets 273/655 OR-Bench Toxic requests correct in both orders, but 56 answers change: 28 become correct and 28 become incorrect. The unchanged score hides those changes.
- **Compare benign and harmful results together.** Original Jev outperforms Laya on OR-Bench Toxic, while Laya outperforms Original Jev on OR-Bench Hard. These measure different error types; neither column alone establishes a generally better model.
- **Keep tasks and source splits separate.** HarmBench copyright uses a different rule from harmful/benign classification. Small validation sets can produce large percentage changes from a few answers.
- **Account for model interfaces.** CLM scores candidate texts separately. Unchanged answers under reordering are consistent with that design; they do not, by themselves, demonstrate better learned robustness. Compare the actual inputs in [CONFIGURATIONS.md](CONFIGURATIONS.md).
- **Do not isolate causes this test cannot separate.** Reversal changes both order and A/B assignments. Original Jev's order-only control can also reflect provider processing and run variation. Subtracting the two flip rates does not establish an A/B-assignment effect.

The adapter snapshots document the formats; they are not a complete runnable environment. See [CONFIGURATIONS.md](CONFIGURATIONS.md) for dependencies and model-version limits. Pin the dataset revision and record the actual model version returned by the provider.

## Prediction format

One JSONL record per `(model, pair_id, variant, repeat_index)`:

```json
{"model":"my-model","pair_id":"XSTest:xstest.csv:1","variant":"original","chosen_key":"A","repeat_index":0}
```

`chosen_key` is interpreted using that exact row. Alternatively supply `predicted_label`. Optional `probabilities` may be keyed by native A/B keys or by semantic labels. If both key and label are present they must agree. `status:"unsupported"` records an explicit exclusion; it is not scored as incorrect. Omitting a record means missing, not unsupported. Repeats use a positive repeat_index and are compared to repeat zero without inflating primary denominators.

```bash
python tools/score_predictions.py predictions.jsonl --output report.json
python -m unittest discover -s tools -p 'test_score_predictions.py'
```

All three request views remain available through the `requests` configuration. Results and exact model configuration metadata are separate. All five model evaluations are complete for the supported original/reversed inputs. Only Original Jev was evaluated on order-only inputs. Skipped and missing cases are reported separately. Aggregate results are presented in README.md. Normalized individual answers are published under `model_results`. Run `python tools/score_release.py` to recompute and verify the aggregate report. See [RESULTS_SCHEMA.md](RESULTS_SCHEMA.md) for joins and field meanings. The prediction scorer reports multiclass Brier scores descriptively on supplied probabilities, without claiming population calibration.

## Contributions and label review

Share new model predictions in the above format with checkpoint/version, native rendering, implementation hash, inference settings, and unsupported reasons. Keep the original inputs/rubric fixed for comparable baseline results. Put changed wording, extra key names or calibration policies in a separately named experimental condition. Propose disputed source labels with source ID, original target, rationale and supporting evidence; this release does not infer a replacement gold label from model disagreement.

## Related work

[Type-Safe Is Not Error-Free](https://arxiv.org/abs/2609.26758) already studies option-name/rubric assignment in Jev and open alternatives. [Typed Decision Models: An Early Evidence Audit and Evaluation Checklist](https://arxiv.org/abs/2609.32160) emphasizes accuracy/latency tradeoffs and evaluation quality. [Beyond Calibration](https://arxiv.org/abs/2609.33209) examines probability coherence. This kit contributes attributed full-text coverage, reusable controlled inputs, paired error analyses and traceable inference settings; it does not claim first discovery of option sensitivity or proven calibration.

[Jev: Decision Models on Trial (Arun and Ma, 2026)](https://blogs.cisco.com/ai/jev-decision-models-on-trial) also evaluates Jev and Laya for safety classification, using a different taxonomy and model-generated reference labels. Our release uses original benchmark labels and matched choice variants. Neither choice sensitivity nor Jev safety evaluation is a first-discovery claim here.