# Supplementary materials

Code, inputs, saved results and analysis for
Support Mismatch as a Benchmark Failure Mode for In-Context Prediction.

Start with Python 3.11+ from the extracted directory:

    python prepare_records.py

This restores the lossless JSONL payloads and checks their hashes. Allow about
3 GB of additional disk space. Install environment/requirements-analysis.txt,
then follow REPRODUCE.md. No model weights are needed for saved-record analysis.

records/ and inputs/ contain the evaluation material; analysis/ contains
summaries and resampling schedules; code/ contains the implementations;
execution/ contains full command recipes and table mappings. The paper and its LaTeX source
are distributed separately from this code/data archive. SHA256_FILE_MANIFEST.json checks transfer
integrity; RECORD_PAYLOADS.json specifies the restored-record hashes.

DATA_AND_SCORE_FORMAT.md defines the panels, metrics and score fields.
EXPERIMENT_CLARIFICATIONS.md gives the E1 text-overlap sensitivity and interval
scope. INFERENCE_AND_TABLE_MAP.md and TRAINING_CONFIG_MAP.md describe new-run
commands and distinguish saved settings from implementation defaults.

Full retraining and pretrained-checkpoint replay have not been completed for
this release. Supplied software is MIT licensed; third-party terms are in
ASSETS_AND_LICENSES.md and the corresponding asset directories.

License scope: LICENSE_SCOPE.md. The authors' rights in code, prepared panels
and result records follow MIT; separately licensed third-party material keeps
its original terms. publication/numeric_reference.json stores only published
numeric values for table checks. Optional article-source identity checks run
when the separate source directory is supplied; absent sources are reported.

response_summary/ retains the separately identified response-stage summary
snapshot. Its normal-reference intervals and quality summaries are not pooled
with the final per-source bootstrap analyses. See its README.txt.
