# Customer importer demonstration

**Designed evidence replay of a recorded run · not live screen footage.**

This 45-second silent video presents actual before-and-after output from a synthetic customer-import fixture. It is a deliberately broken demo application, not the Embercrank interface. The displayed contact names and `.test` email addresses are fictional.

The original importer accepted six rows, damaged a quoted company name, and failed to reject duplicates and invalid contacts. The repaired importer, given identical input, produces three contacts, skips one duplicate and identifies two invalid rows. The company name “North, South” remains intact.

## Recorded result — 8 September 2026

- Selected model: Anthropic Claude Sonnet 5, low reasoning effort, balanced tier.
- Successful run: one attempt, no automatic escalation; 87.365 seconds of provider execution, condensed in this replay.
- Acceptance checks: zero of six passed before; six of six passed after.
- Only the importer implementation was editable. Four protected files remained unchanged.

The six checks cover quoted commas and escaped quotes, duplicate emails, missing names and invalid emails, line endings and quoted newlines, malformed records, and the complete sample outcome. Passing these checks establishes that bounded fixture outcome, not production CSV completeness or general reliability.

This successful run followed earlier development failures and corrections to language extraction and the runner's permission instructions. It is not a claim that the original development process succeeded on its first try. No billed-dollar savings, subscription-allowance savings or native Claude desktop integration was measured.

The video is intended for private review. It does not establish public installer or paid-service availability.

## Transcript

0–9 seconds: “The import finished. The list is wrong.” The original six-row output highlights the broken company name, duplicate and invalid records.

9–17 seconds: “A useful job. A visible model choice.” A shortened task summary is followed by the actual model, editable file, configured deadline and attempt count.

17–25 seconds: “The checks cover the messy cases.” The six acceptance categories and before/after result appear.

25–36 seconds: “Three contacts ready. Two rows to correct.” The actual repaired output shows the intact company name, duplicate count and invalid-row messages.

36–45 seconds: “See what ran. Check what changed.” The original duration, protected-file result and limitations are shown.