# Rocket Relay Model Authenticity Monitoring

Last reviewed: April 27, 2026

## Goal

Rocket Relay gives buyers continuous operational evidence that advertised model access is not an opaque black box. The platform records model quality, active probe, and route health signals so operators can detect downgrades, unhealthy upstream accounts, routing drift, and performance divergence.

These checks are operational evidence, not a cryptographic proof of model identity.

## Recorded checks

The `model_quality_checks` table records:

- Provider.
- Model ID.
- Check type: fingerprint, eval, latency, or route.
- Status: pass, warn, or fail.
- Score.
- Latency.
- Expected fingerprint.
- Observed fingerprint.
- Sample prompt hash.
- Error message.
- Metadata.
- Checked timestamp.

## Observed route health

The API process records hourly route health checks from real production metadata:

- 24-hour request count.
- 24-hour failure count.
- 24-hour average latency.
- Active upstream account count.
- Problem upstream account count.

## Active probe checks

The API service can run scheduled active probes through the internal proxy route:

- Low-token verification requests.
- Temperature set to 0.
- No customer API key and no customer billing.
- Prompt hash stored instead of prompt text.
- Response hash, HTTP status, latency, provider/model metadata, and match result stored in `model_quality_checks`.

Operators configure probes with:

- `MODEL_QUALITY_PROBES_ENABLED`
- `MODEL_QUALITY_PROBE_INTERVAL`
- `MODEL_QUALITY_PROBE_TIMEOUT_SECONDS`
- `MODEL_QUALITY_PROBE_MAX_MODELS`
- `MODEL_QUALITY_PROBE_MODELS`

## Fingerprint and eval checks

External scheduled jobs may still run deeper controlled evaluations against provider routes and submit results to:

`POST /api/admin/model-quality/checks`

The admin console displays active probe, fingerprint/eval, production traffic, and upstream pool status. The public trust endpoint exposes only redacted summary metadata.

## Operator workflow

1. Add or sync upstream accounts.
2. Configure active model pricing rows.
3. Enable internal active probes for high-value models.
4. Review `/admin/model-quality` for pass, watch, warn, and fail status.
5. Disable unhealthy upstream accounts or routes when checks fail.