# Public Jev-Omni/Clef classification export

User-authorized public synthetic research material, prepared 2026-10-02. This bundle is input for the separate Site worker; creating it does not publish or modify any Site.

Start with `story/01_JEV_UNIFIED_DECISIONS_MAIN.md` and the Japanese script. The main story recommends trying Jev-Omni for the useful common decision workflow across audio, images, text and video, then reports Clef's stronger result on the common cohort. Keep the two emergency studies in a technical section rather than the main opening/closing.

## Data layout

- `experiments/E1_JEV_INITIAL.json`: 35 accepted inputs (10 speech, 10 image, 10 fictional Japanese email, 5 video), 70 Jev-Omni decisions and ten unexecuted planned decision slots. Forward 33/35, reverse 31/35.
- `experiments/E2_CLEF_COMMON.json`: same 25 image/text/video inputs, 50 Clef decisions. Clef 50/50; saved E1 Jev-Omni 44/50. Baseline decision IDs link to E1; there was no Jev-Omni rerun and no Clef audio comparison.
- `experiments/E3_EMERGENCY_ORIGINAL.json`: five original clips, 20 separate Jev-Omni/Clef emergency-necessity outcomes. Three not-needed and two ambiguous references. Each model 6/6 eligible matches; four ambiguous decisions per model excluded.
- `experiments/E4_NEAR_APPROACH_AMENDMENT.json`: five distinct near-approach clips, 20 outcomes. One not-needed and four ambiguous references. Each model 2/2 eligible matches; eight ambiguous decisions per model excluded. Clef flips C02/C03, Jev-Omni does not.
- `responses/`: 160 sanitized saved response envelopes. Each nested `response` object, including index, confidence and ordered probabilities, is unchanged; available `semantic` maps are also unchanged. These are sanitized original envelopes, not newly generated outputs or raw execution logs.
- `media/initial`: exact accepted classifier JPEGs, fictional text files and four-second video transport MP4s. `media/near-approach`: five exact four-second transport MP4s.
- `04-Audio-Sources`: ten exact full original accepted WAVs plus the utterance/reference/result/ASR mapping. Chatterbox appears only as fixture-generation provenance. NotebookLM audio import transcribes to text and does not rerun Jev-Omni on waveform/tone.
- `EXPORT_MANIFEST.json`: per-export checksums, source mapping and sanitization disclosure. Source references are research-relative identifiers, not bundled originals or private absolute paths.
- `VALIDATION.json`: completed numeric/hash/semantics and exclusion checks. `SHA256SUMS.txt`: every bundled file except the checksum file itself.

## Publication interpretation

Forty unique assets and 160 completed decisions are retained across four separate cohorts. References belong to their experiment: a Dangerous reference in E1 can be ambiguous under E3's different emergency-stop criterion. Never infer binary correctness from an ambiguous reference. Across E3/E4 there are zero unambiguous stop-needed references, so stop-needed sensitivity is undefined. No general safety reliability, latency superiority, calibrated confidence or causal effect of the backbone follows.

Media are synthetic and labels were fixed or independently reviewed before the respective decisions. I08R desk fan replaced the rejected corkscrew before inference. Five planned initial video scenarios remain missing. No contact-generation failures are presented as accepted collision evidence. Native video originals and large visual PDFs are omitted to keep the pack modest: the included MP4 bytes are the exact evaluated transports. Clef sampled eight frames and Jev-Omni sixteen normalized frames, with different transforms; transport equality is not preprocessing equality.

Only Jev-Omni/Clef classification studies are included. No music generation, unrelated integration tests, credentials, private infrastructure, real inbox material, account-specific URLs, model weights or execution logs are included. All original research sources remain unchanged. No new generation, ASR or model inference was run.

Display naming: technical experiment IDs and the stored model key `jev` identify the Jev-Omni runs. Render that key as **Jev-Omni** in every public heading, table and legend. Preserve original technical IDs and response values.

Generator provenance: `03-Technical-Appendices/GENERATOR_PROVENANCE.md` and `.json` identify SDXL-Lightning 4-step, MiniMax-H3/PDD8 (recorded API setting: 9 steps), and Chatterbox Turbo as fixture generators, alongside Jev-Omni/Clef decision evaluators. Per-selected-input prompts, seeds/settings, source hashes and failed-generation/replacement lineage are included; unknown settings remain explicit. Fictional text fixtures were authored directly.