Supplementary Data 1: CI-DeceptiveTest metadata and selected evaluation results

Manuscript: CI-DeceptiveTest: An Execution-Paired Benchmark for Test Assurance
in AI-Assisted Development
Corresponding author: Qining Li (castelu@qq.com)
Prepared: 21 October 2026

Contents
1. benchmark_metadata/: the existing manifest-first metadata package, preserved
   without modifying its files or checksums. It contains 600 pair records,
   1200 evaluation metadata records and answer keys, code hashes, execution-log
   manifests, Static Rule predictions/results, schemas, and verification/scoring
   scripts. The upstream package's candidate status and review notes are retained.
2. results/mimo_v2_5_pro/ and results/mimo_v2_5/: 1200 saved predictions per
   model, mapped to public view IDs, plus the original numerical score reports.
   Failed parses retain the original source_control_static fallback and are
   identified by parse_ok=false (6 and 2 cases, respectively). Raw model text
   and reasoning are excluded. The parse flags are checked separately from
   scoring because the normalized labels already contain the original fallback.
3. results/mutation/: the frozen 42-pair mutation overlay and inventory reports.
   Fifteen supplemental complete pairs are reported separately. The 57-pair
   inventory does not replace the frozen 42-pair main evidence set.
4. tools/reproduce_scores.py: offline verification of the saved Static Rule
   and MiMo results against the included answer key; Python standard library only.
5. provenance.json and MANIFEST.sha256: source provenance and file checksums.

Run from this directory
  python3 benchmark_metadata/tools/verify_public_candidate.py
  python3 tools/reproduce_scores.py

These checks require no API key, model call, network connection, or new
experiment. They reproduce scores from stored predictions, not model inference.

Access boundary
The supplement does not contain original/transformed upstream test code, raw
execution logs, raw LLM requests, or raw model outputs. Clearance for
redistribution of third-party source material and raw logs is incomplete.
Requests for the underlying data and additional experiment artifacts should be
sent to Qining Li at castelu@qq.com, describing the research/review purpose and
the required subset. Access is subject to applicable third-party permissions
and non-redistribution conditions. The legacy candidate package is not a
standalone archived public raw-data release, and no repository DOI is asserted.

The supplementary materials allow score reproduction and integrity checks.
Full detector evaluation and execution reproduction require the raw inputs and
the corresponding execution environment through the stated access path.
