Coverage
How Dokima scan coverage works under each customer tier — what is scanned, how fresh the verdict is, and what priority the scan gets in the queue.
Per-tier scan policy
| Tier | Web scanner | API access | Monthly scans | Force-fresh on submit | Priority lane |
|---|---|---|---|---|---|
| Free (unauthenticated) | 3 scans/day per IP | none | n/a | no | recon-backfill |
| Free (signed-in account) | 25 scans/month | none | 25 | no | recon-backfill |
| Hobby (£5/mo) | n/a | yes | 500 | no | badge-refresh |
| Pro (£39/mo) | n/a | yes | 30,000 | no | badge-refresh |
| Compliance (custom) | yes | yes | per contract | yes | user-fresh |
| Enterprise (custom) | yes | yes | per contract | yes | user-fresh |
What "force-fresh" means
Most submissions hit the verdict cache first: if a fresh verdict exists in the last 24-hour window AND the model has not advanced its commit SHA on Hugging Face, the cache returns immediately. Compliance and Enterprise tiers commit to a stronger contract — every submission re-runs the full audit against the current HF metadata. That guarantees the audit-trail PDF reflects the latest source state at the moment the customer asked for it.
What "priority lane" means
The scan queue has three priority classes. Lower class number is higher priority:
- User-fresh — Compliance and Enterprise submissions land here. Workers drain this class first.
- Badge-refresh — Hobby, Pro, and the background refresh path for newly-active models. Drained after user-fresh.
- Recon-backfill — Free-tier submissions and opportunistic corpus expansion. Drained last so paid submissions are never starved by free-tier load.
Within a class, scans are processed oldest-first (FIFO). At MVP scale a single worker is sufficient; the queue model is already cluster-ready for scaling.
What we scan, in every tier
Every Dokima scan, regardless of tier, evaluates the model across the seven dimensions published in the rubric:
- Serialisation safety (22 points)
- Model card completeness (17 points)
- Licence clarity (13 points)
- Namespace provenance (13 points)
- Safety and bias evaluations (13 points)
- Regulatory alignment (9 points)
- Ecosystem context (13 points)
The dimensions and their point weightings do not vary by tier — a Free-tier verdict is the same verdict a Compliance-tier verdict would be for the same model at the same commit SHA. What varies is the freshness contract and the queue priority.
Deep binary scanning (Compliance and Enterprise)
A separate, opt-in capability planned for the Compliance and Enterprise tiers extends each scan with full file-content inspection of model weights and serialised artefacts. The capability is licence-gated (models with restrictive licences cannot be deep-scanned). Implementation lands as a post-launch feature; this section is forward-looking.
Coverage transparency
Dokima publishes a continuously-updating coverage rollup at /api/v1/coverage/by_bucket. The aggregate counts every verdict in the corpus by month and is used as the input to the public dashboard. The endpoint is open to all callers (no auth required).