Verified marketplace goods

# Proposal: Verified marketplace goods

## Why
Buyers have to trust the seller's own quality claims. A verification record states which claims were checked, under which judge, and the outcome, and it outlives the store.

## What Changes
- The marketplace verifies each item's claims against their sources with a judge model and prompt pinned by hash.
- It publishes a verification record per item, labeled VERIFIED.
- If the judge turns out to have a blind spot, its hash names every affected record for re-verification.


## Capabilities
### New Capabilities
- `verified-listings`: A verification record per item, scored under a judge pinned by hash. Buyers check without trusting the seller.

## Impact
- External service: AI Source of Truth at https://aisourceoftruth.com. Reads are anonymous. Writes need a publisher key pair you generate, a client id and secret from registration, and a token from `/oauth/token`.
- Routes used: `POST /v1/records`, `POST /v1/verify-batch`, `GET /v1/judges/{hash}`.
- No PromptReports product is required.
- Source of this plan: https://aisourceoftruth.com/use-cases/verified-marketplace (CC0). Regenerate it with `curl -H 'Accept: application/json' 'https://aisourceoftruth.com/use-cases/verified-marketplace/openspec?lang=python'`.

# Design: Verified marketplace goods

## Context
Marketplaces for datasets, reports, prompts and models. Buyers have to trust the seller's own quality claims. A verification record states which claims were checked, under which judge, and the outcome, and it outlives the store.

## Where it sits in your agent
Publish from the step that approves a listing. Keep the judge prompt versioned and hashed in your repo.

Over MCP, point the agent at `https://aisourceoftruth.com/mcp` (Streamable HTTP, JSON-RPC 2.0; read tools need no token). The tools this change uses: `publish_record`, `verify_batch`.

## Decisions
- Treat the registry as a source of facts about records, never about the world. Show what it says, with the digest, and do not turn a verdict into a quality score.
- If the registry cannot be reached, treat the answer as unknown (the same as `check`), never as `rely`, and retry later.
- Keep content private: send hashes. Never put personal data in a record; it is public and permanent.
- Keep the private key and client secret in your secret store. Rotate keys at `POST /v1/keys`; a revoked key cannot sign.

## Setup and signing helper

```python
# pip install requests cryptography
import base64, json, os, uuid
import requests
from cryptography.hazmat.primitives.serialization import load_pem_private_key

AISOT = os.environ.get("AISOT_ORIGIN", "https://aisourceoftruth.com")
KEY = load_pem_private_key(open(os.environ["AISOT_KEY_PEM"], "rb").read(), password=None)  # Ed25519

def token():
    r = requests.post(f"{AISOT}/oauth/token", json={
        "grant_type": "client_credentials",
        "client_id": os.environ["AISOT_CLIENT_ID"],
        "client_secret": os.environ["AISOT_CLIENT_SECRET"],
    })
    r.raise_for_status()
    return r.json()["access_token"]  # lasts 60 minutes

def signed_post(path, body, content_type="application/json", token=token):
    raw = json.dumps(body, separators=(",", ":")).encode()
    sig = base64.urlsafe_b64encode(KEY.sign(raw)).rstrip(b"=").decode()
    headers = {"Content-Type": content_type, "X-AISOT-Signature": sig, "Idempotency-Key": str(uuid.uuid4())}
    if token:
        headers["Authorization"] = f"Bearer {token()}"
        headers["X-AISOT-Key-Id"] = os.environ["AISOT_KEY_ID"]
    r = requests.post(f"{AISOT}{path}", data=raw, headers=headers)
    r.raise_for_status()
    return r.json()
```

## Calls
### POST /v1/records
append a signed record; the answer carries its digest and inclusion proof.

```python
answer = signed_post("/v1/records", record, content_type="application/vnd.aisot.record+json")
```

### POST /v1/verify-batch
signature, inclusion proof, errata count and judge for up to 1000 digests in one call.

```python
r = requests.post(f"{AISOT}/v1/verify-batch", json={"digests":["{digest}"]})
r.raise_for_status()
answer = r.json()
```

### GET /v1/judges/{hash}
a judge version, its cohort of records, and its published prompt once retired.

```python
r = requests.get(f"{AISOT}/v1/judges/{hash}", headers={"Accept": "application/json"})
r.raise_for_status()
answer = r.json()
```

## What this does not prove
- The registry stores what your judge decided. It never re-runs the judge or rescores.

# Tasks: Verified marketplace goods

## 1. Configure
- [ ] 1.1 Set `AISOT_ORIGIN=https://aisourceoftruth.com` in the agent's environment.
- [ ] 1.2 Generate an Ed25519 key pair, register the publisher (`POST /v1/publishers`), and store `AISOT_CLIENT_ID`, `AISOT_CLIENT_SECRET`, `AISOT_KEY_ID` and the private key as secrets.
- [ ] 1.3 Mint a token at `/oauth/token` and confirm `GET /v1/me` answers with your publisher id.
- [ ] 1.4 Optional: add the MCP server `https://aisourceoftruth.com/mcp` to the agent's MCP config.

## 2. Build
- [ ] 2.1 Get the verification scope: Verification publishers are reviewed before the `publish:verification` scope is granted. Start with provenance records while you wait.
- [ ] 2.2 Publish per item: Post a verification record with the artifact, pipeline, claims, summary and the judge fields (`judge_model`, `judge_prompt_version`, `judge_prompt_hash`). (`POST /v1/records`)
- [ ] 2.3 Let buyers check: Show the digest on the listing. Buyers check it with verify-batch before paying. (`POST /v1/verify-batch`)
- [ ] 2.4 Handle a recalled judge: If a judge version is found wanting, list its cohort, re-verify, and issue errata where results change. (`GET /v1/judges/{hash}`)

## 3. Verify against the live registry
- [ ] 3.1 Get the verification scope: Your token's scope includes `publish:verification`.
- [ ] 3.2 Publish per item: The item page links its record.
- [ ] 3.3 Let buyers check: A buyer's check returns the judge version.
- [ ] 3.4 Handle a recalled judge: The cohort count matches the records you published under that judge.

## 4. Done when
- [ ] 4.1 Each listing carries a VERIFIED record a buyer can check alone.
- [ ] 4.2 A judge recall becomes a public, counted correction.

# verified-listings Specification

## ADDED Requirements

### Requirement: Verified listings
The marketplace SHALL publish a verification record, with the judge pinned by hash, for every listing it labels verified.

#### Scenario: Listing approved
- **WHEN** a listing passes verification
- **THEN** its verification record is published and its digest shown on the listing