Provenance you can hand to a regulator.
"Trust us" is not an audit trail. Every weight, dataset version and inference Xvian produces is hashed and anchored, so what ran in March can still be proven in three years' time.
01 / What gets anchored
Four artefact types
- WModel weightsHash of the exact binary that served the prediction.
- DDataset versionsContent hash of the training and calibration corpus.
- IInference batchesSigned digest, batched hourly to keep cost sane.
- OClinician overridesDisagreements recorded with the same weight as agreements.
◈ledger.feedlive
✓ 0x7a1f… model:thorax-v4.2 weights anchored
✓ 0x3c92… dataset:retina-2026Q2 hash
✓ 0xb14e… inference batch 4,182 signed
✓ 0x9dd0… clinician override logged · SG-04
✓ 0x2f77… audit pack exported · TGA
chainxvian-attest L2
anchor cadencehourly
verificationpublic
02 / Registry
updated 22:45 AEST
Model registry — public slice
| Model | Domain | AUC | p95 | Anchor | Status |
|---|---|---|---|---|---|
| thorax-v4.2 | Chest CT | 0.992 | 41ms | 0x7a1f… | CLEARED |
| retina-v3.8 | Fundus | 0.981 | 28ms | 0x3c92… | CLEARED |
| histo-v2.1 | Pathology | 0.964 | 112ms | 0xb14e… | REVIEW |
| cardio-sig-v1.6 | ECG stream | 0.958 | 9ms | 0x9dd0… | CLEARED |
| onco-fusion-v0.9 | Multimodal | 0.941 | 210ms | 0x2f77… | TRIAL |
03 / Verify
Check it yourself
Anchors are public. Given a model hash from a report, anyone can confirm the weights that produced it were registered before the prediction was made — without Xvian in the loop.
◈verify.sh
$ xvian verify --anchor 0x7a1f…
✓ anchor found · block #8,412,907
✓ weight hash matches registry
✓ timestamp precedes inference
VERIFIED
✓ anchor found · block #8,412,907
✓ weight hash matches registry
✓ timestamp precedes inference
VERIFIED
04 / Access
Audit-grade by default
If your compliance team wants to pull the thread, we'd rather they did it early.
Request documentation →