Here’s a compact, practical way to think about **embedding in‑toto provenance attestations directly inside your event payloads** (instead of sidecar files), so your vuln/build graph stays temporally consistent.

---

### Why embed?

* **Atomicity:** build → publish → scan → VEX decisions share one event ID and clock; no dangling sidecars.
* **Replayability:** the event stream alone reproduces state (great for offline kits/audits).
* **Causal joins:** vulnerability findings can cite the exact provenance that led to an image/digest.

---

### Event shape (single, self‑contained envelope)

```json
{
  "eventId": "01JDN2Q0YB8M…",
  "eventType": "build.provenance.v1",
  "occurredAt": "2025-11-13T10:22:31Z",
  "subject": {
    "artifactPurl": "pkg:docker/acme/api@sha256:…",
    "digest": {"sha256": "…"}
  },
  "provenance": {
    "kind": "in-toto-provenance",
    "dsse": {
      "payloadType": "application/vnd.in-toto+json",
      "payload": "<base64(in-toto Statement)>",
      "signatures": [{"keyid":"…","sig":"…"}]
    },
    "transparency": {
      "rekor": {"logIndex": 123456, "logID": "…", "entryUUID": "…"}
    }
  },
  "sig": {
    "envelope": "dsse",
    "alg": "Ed25519", 
    "bundle": { "certChain": ["…"], "timestamp": "…" }
  },
  "meta": {
    "builderId": "https://builder.stella-ops.local/gha",
    "buildInvocationId": "gha-run-457812",
    "slsa": {"level": 3}
  }
}
```

**Notes**

* `provenance.dsse.payload` holds the raw in‑toto Statement (Statement + Subject + Predicate).
* Keep both **artifact digest** (subject) and **statement subject** (inside payload) and verify they match on ingest.

---

### DB model (Mongo-esque)

* `events` collection: one doc per event (above schema).
* **Compound index:** `{ "subject.digest.sha256": 1, "occurredAt": 1 }`
* **Causal index:** `{ "meta.buildInvocationId": 1 }`
* **Uniq guard:** `{ "eventId": 1 } unique`

---

### Ingest pipeline (deterministic)

1. **Verify DSSE:** check signature, cert roots (or offline trust bundle).
2. **Validate Statement:** subject digests, builder ID, predicateType.
3. **Upsert artifact node:** keyed by digest; attach `lastProvenanceEventId`.
4. **Append event:** write once; never mutate (event‑sourced).
5. **Emit derived edges:** `(builderId) --built--> (artifact@digest)` with `occurredAt`.

---

### Joining scans to provenance (temporal consistency)

* When a scan event arrives, resolve the **latest provenance event with `occurredAt ≤ scan.occurredAt`** for the same digest.
* Store an edge `(artifact@digest) --scannedWith--> (scanner@version)` with a **pointer to the provenance eventId** used for policy.

---

### Minimal .NET 10 contracts

```csharp
public sealed record DsseEnvelope(string PayloadType, string Payload, IReadOnlyList<DsseSig> Signatures);
public sealed record Provenance(string Kind, DsseEnvelope Dsse, Transparency? Transparency);
public sealed record EventSubject(string ArtifactPurl, Digest Digest);
public sealed record EventEnvelope(
    string EventId, string EventType, DateTime OccurredAt,
    EventSubject Subject, Provenance Provenance, SigMeta Sig, Meta Meta);

public interface IEventVerifier {
    ValueTask VerifyAsync(EventEnvelope ev, CancellationToken ct);
}
public interface IEventIngestor {
    ValueTask IngestAsync(EventEnvelope ev, CancellationToken ct); // verify->validate->append->derive
}
```

---

### Policy hooks (VEX/Trust Algebra)

* **Rule:** “Only trust findings if the scan’s referenced provenance has `builderId ∈ AllowedBuilders` and `SLSA ≥ 3` and `time(scan) − time(prov) ≤ 24h`.”
* **Effect:** drops stale/forged results and aligns all scoring to one timeline.

---

### Migration from sidecars

1. **Dual‑write** for one sprint: keep emitting sidecars, but also embed DSSE in events.
2. Add **backfill job**: wraps historical sidecars into `build.provenance.v1` events (preserve original timestamps).
3. Flip **consumers** (scoring/VEX) to **require `provenance` in the event**; keep sidecar reader only for legacy imports.

---

### Failure & edge cases

* **Oversized payloads:** gzip the DSSE payload; cap event body (e.g., 512 KB) and store overflow in `provenance.ref` (content‑addressed blob) while **hash‑linking** it in the event.
* **Multiple subjects:** keep the Statement intact; still key the event by the **primary digest** you care about, but validate all subjects.

---

### Quick checklist to ship

* [ ] Event schema & JSON schema with strict types (no additionalProperties).
* [ ] DSSE + in‑toto validators (offline trust bundles supported).
* [ ] Mongo indexes + append‑only writer.
* [ ] Temporal join in scanner consumer (≤ O(log n) via index).
* [ ] VEX rules referencing `event.meta` & `provenance.dsse`.
* [ ] Backfill task for legacy sidecars.
* [ ] Replay test: rebuild graph from events only → identical results.

If you want, I can turn this into ready‑to‑drop **.proto + C# models**, plus a Mongo migration script and a tiny verifier service.