Response-stage analysis materials

The enclosed paired seed summaries and seed-subject error counts accompany
the separate response-stage results in the article. They retain their own
score/context versions and analytical normal-reference intervals. The primary
per-source results elsewhere in this archive retain their bootstrap methods.

Summary analysis (Python standard library; no model weights):
    python code/reproduce_materials.py --command response-summary

Text-quality evaluation (requires the local configured RoBERTa-large encoder):
    python response_summary/code/quality_metrics.py --input texts.jsonl --encoder /path/to/roberta-large --output-dir quality_output --device cuda:0

Each text input row contains id, source, prediction and ten references. Saved
decoded texts and references are in the main records/ directory. The evaluator
outputs sentence scores, corpus SARI and metric configuration separately.
RUN_AND_METRIC_CONFIG.json gives the encoder revision and metric settings;
code/requirements_probe.txt lists the evaluation dependencies. No encoder
weights are bundled. Providing the evaluator does not imply a new run occurred
during packaging. VERSION_COMPARISON.csv distinguishes the two result versions.

E3 diagnostics retain the supplied summary and calibration-state level; context
IDs identify that version's population slots. Vendored EASSE retains its
attached license and source commit. The authors' MIT grant does not supersede
the upstream terms.
