# Demo receipts: a sample claim packet, checked before the agent acts

Sample case, synthetic files. Not a real claim.

## The run the post cites: production, 2026-10-02

Both files were sent to Scam AI's **public production API**, `POST
https://api.scam.ai/v1/detections` (multipart `file`), from an internal Scam AI
account: the same detection the MCP server's `detect_media` tool calls. The
responses are saved byte for byte.

| File | Response | Verdict | Model | Credits |
|---|---|---|---|---|
| `samples/storefront.png` | `storefront-photo.production.json` (id `240b03bc-ffe2-4e9b-8fa2-94d7039b73ae`, 2026-10-02T02:22:33Z) | `LIKELY_AI`, "This image shows strong signs of being AI-generated or edited." | Eva V1.6 | 1 |
| `samples/video-statement.mp4` | `video-statement.production.json` (id `12680a6c-03d7-4fed-bf99-4cafb3557731`, 2026-10-02T02:22:45Z) | `LIKELY_AI`, "This video shows strong signs of being a deepfake or manipulated."; 8 frames scored, one a second, every one above `threshold_used` 0.5 | Eva V1.6 | 8 |

The SHA-256 of the bytes sent is the same as in the table below
(`1c1294f6…5fda` and `07270a3f…718c`).

`../evidence-brief.webp` is the post's result figure (4:5, so it reads on a
phone and in a LinkedIn or X feed): the two verdicts above, laid out as the
agent's brief. Its bars are the eight
per-frame values of `video-statement.production.json`, and its dashed line is
that response's `threshold_used`. The next steps and the recommendation are
the brief's routing under the starter Skill, not output of the detector. Its
source is `loops/drafts/blog/partners/aident/evidence-brief.html` in the
repository.

## The earlier run: development, 2026-10-01

Kept for the record. The same two files on the development router gave the
same verdicts.

What the post "Deepfake-detection MCP server for AI agents, with Aident" ran,
on what, and what came back. Every file here is the exact one the post cites.

**What ran.** Scam AI's development detection router, which reports `Eva V1.5`
(the public API runs Eva V1.6, which can score the same file
differently), one `POST /detect?detail=compact` per file. This is the detection that
the public API and the MCP server's `detect_media` tool put their own field
names on. Each JSON file is the raw router response, byte for byte; on this
release both verdicts came from Scam AI's own models (`ai_image`, and
`faceswap` for the video, listed in each response's `work`).

**0.98 is the ceiling, not a measurement.** The image detector caps every score
at 0.98, so a 0.98 here is the highest score the service reports, not a
confidence of 98%.

| File | In the packet as | Provenance | SHA-256 of the bytes sent | Response |
|---|---|---|---|---|
| `samples/storefront.png` | Photo of the insured shop front | An AI-generated image from the example set of Scam AI's image detector, filed there as a Gemini 2 output (`ai-models/ai-image-detector/assets/gemini2.png`); no visible watermark | `1c1294f674a822a87c34070f4fb7cb3ace8e42697fccc9a7dbe72b7ae5285fda` | `storefront-photo.json`, 2026-10-01T00:18:55Z: `fake`, 0.98, `use_case: general` |
| `samples/video-statement.mp4` | Short clip from the claimant | An 8-second, 720×720 clip, a still rather than a filmed video, made by slowly zooming into one of Scam AI's own AI-generated faces (`/playground-faces/ai-black-man.jpg`, SHA-256 `9f9f8dff149b682c891c98cd139a46b539d498a7ed15a25a06b084ca654fc681`) with `ffmpeg … zoompan …`; no audio, no real person | `07270a3fb3013ab8ad8e28caba7d31c940fd6a6f05c0a03ce766c952a337718c` | `video-statement.json`, 2026-10-01T00:19:14Z: `fake`, 0.98, `use_case: face`; 8 frames scored, one a second, each 0.98 |

A third file in the post's sample packet, a repair estimate PDF, was not sent:
`detect_media` takes images, video and audio, and document forgery detection is
a separate Scam AI plan. The post routes it to that plan.

`../claim-packet-samples.webp` is the figure in the post: the storefront photo
and the frame at 4 seconds of the video, side by side, resized, with nothing
added. It was not itself sent to the detector.

**How the API and `detect_media` spell this answer.** The router's `label`
`fake` is published as `verdict: "LIKELY_AI"`, `uncertain` as `ALERT` and
`real` as `LIKELY_REAL`; `probability` is published as `score`; the summary is
the platform's fixed sentence for that verdict and media type ("This image
shows strong signs of being AI-generated or edited." and "This video shows
strong signs of being a deepfake or manipulated."). The video's frame values
are published as the per-frame `frames` series.

**What the run does not show.** Both checked files are AI-made, so the run
shows the escalation branch, not how often a genuine claim photo is flagged.
Eva V1.6 gave both files the same verdict (the production run above).