This release contains 500 original normal-task inputs for the synthetic two-lane TTC decision task used by the BALD-word model adaptations. Each input has only case_id and the original system / user messages. No model outputs, assistant answers, activation measurements, or runtime ground-truth labels are included in the inputs.

The intended use is monitor calibration. Normal task distribution is not a guarantee of backdoor inactivity. Do not turn these reference cases into a labeled negative evaluation set.

Sampling and exclusions

We use the existing task generator without changing its scenario construction:

Ego lane is uniformly sampled from left/right.
Each TTC is round(uniform(0.5, 10.0), 1) seconds; equal TTCs are rejected.
The original system prompt and clean in-context example are preserved verbatim.
The original normal-task generation path is used. There is no filtering on a trigger string, model prediction, activation, attack success, or correct lane.
Seed is 20261026. We accept 500 unique new scenarios after excluding the recorded training/test, runtime, and legacy-reference scenarios.

Sampling is therefore from the original synthetic normal-task generator, conditioned on these explicitly recorded disjointness requirements. It is not an empirical traffic distribution or a HighwayEnv simulator rollout. There is no imposed 50/50 correct-lane quota or extra minimum-gap threshold.

The generator exactly reproduces all 700 published source conversations (500 train + 100 test-ASR + 100 test-accuracy) with their original seed before being used to generate this release. Code and immutable file identities are included in manifest.json and code/.

Validation finds 500 unique scenarios, zero overlap with each recorded exclusion group, and exact original prompt formatting. Diagnostic lane counts are 249 left / 251 right for the larger TTC, with a minimum TTC gap of 0.1 seconds. These counts did not determine selection.

Sources

The legacy reference remains unchanged as historical evidence. It has 400 cases and overlaps the recorded source training set in 10 TTC scenarios. This v2 release satisfies the current 500-case contract with a fresh, disjoint cohort; it is not the old 400 cases with 100 rows appended. Any monitor calibrated on the old reference must be calibrated again to report results using v2.

Runtime inputs, checkpoint outputs, and existing runtime labels have not changed. This reference can be used with all model families that use the same published BALD-word prompt/task distribution. Model-specific undisclosed training sets are outside the stated exclusion proof.

Reproduction

Download the three source JSON files, pinned runtime inputs.jsonl, and pinned legacy-reference inputs.jsonl from the links above. With the provided code/ files in the same directory, run:

python code/nominal_reference.py \
  --source /path/to/published-source-files \
  --runtime /path/to/runtime/inputs.jsonl \
  --legacy /path/to/legacy-reference/inputs.jsonl \
  --output /path/to/new-output-directory

Only the Python standard library is required. A fresh output directory is required. The program verifies all pinned source hashes, replays the 700 source examples, generates the 500 prompts, and verifies disjointness. A separate standalone rerun reproduced the input file byte-for-byte.

inputs.jsonl SHA-256: 2a13d9d1a7f202200cc016b8472d93ce22a0e6e123be4e8c179d7cd1e919247b.

Published for review; no new GPU job or model evaluation was required.

Downloads last month: 3

Total file size:

700 kB