# English source-to-action-register test pack

All source material is synthetic. No model has been run for this pack. It supplies inputs, blank logs and expected checks, not performance results.

## Download inputs and blank outputs

- [Phone-style message PNG](https://media.mylingasia.com/articles/model-guides-20260905/asia-en-routing/phone-messages.png)
- [One-page source PDF](https://media.mylingasia.com/articles/model-guides-20260905/asia-en-routing/source-page.pdf)
- [Plain source text](https://media.mylingasia.com/articles/model-guides-20260905/asia-en-routing/source-text.txt)
- [Record-your-own voice script](https://media.mylingasia.com/articles/model-guides-20260905/asia-en-routing/voice-note-script.txt)
- [Blank action-register CSV](https://media.mylingasia.com/articles/model-guides-20260905/asia-en-routing/action-register.csv)
- [Blank run log CSV](https://media.mylingasia.com/articles/model-guides-20260905/asia-en-routing/run-log.csv)

The voice file is a reading script, not audio. Record it yourself only if testing audio input; otherwise use the checked text and label the run a text-reasoning test. No secret or real client data is needed.

## Two separate tests

1. Original-file workflow: submit the same source PNG/PDF or your own recording through the selected app/API route. Save any transcription/OCR intermediate file.
2. Common-text reasoning: give every model the same verified source-text.txt. This removes transcription/OCR differences from the comparison.

Do not combine these into a single unqualified ranking. If a model cannot accept the file or is unavailable, record not tested or unsupported at this route.

## Prompt

Turn the supplied sources into an English client reply and action register. Do not invent an owner, date, time zone, price or promise. Apply the later correction. Give a source ID for each action and leave missing fields unresolved.

## Acceptance rubric

- Poster remains on hold, supported by S2.
- Caption draft remains requested, supported by S3.
- Deadline and price remain unresolved, supported by S4.
- No owner, date, time zone, price or guarantee is invented.
- Original instruction S1 does not override the later correction.
- Every commitment has a checkable source ID.

Save full unedited outputs, errors and correction prompts. Measure timing or cost only from actual records. One run supports a statement about that run, not general model quality or Hindi accuracy.