CONNECTING

Training Provenance
Registry

Public, immutable, cryptographically anchored observations of AI training data disclosure. Every day the chain grows. Every omission is recorded.

CROVIA is a public evidence register, not a demo dashboard. This page exposes live operational artifacts and timestamped absence records.

TEMPORAL CHAIN
Merkle blocks anchored
Models monitored
Organizations
Days continuous
TPA proofs
Integrity: checking
Shadow mode: checking
Live artifact — Absence Clock
open standalone ↗
⚠ Methodology Correction — Feb 24–25, 2026

We identified and corrected two bugs in our detection engine that affected compliance scores and NEC# observations published prior to Feb 24, 2026.

Bug 1 — Compliance Mapper v1
Keyword detection used incorrect case-sensitivity logic. Some models received inflated scores (~72%) when correct score was much lower, and vice versa. All 6,018 reports recalculated with Mapper v2.0.
Bug 2 — Oracle Section Detection
NEC# detection used narrow heading patterns (## evaluation only) missing # Evaluation, ## Benchmarks etc. Models with documented safety, evaluation, and usage sections were incorrectly marked absent for NEC#5, #6, #7, #9, #11, #12, #17, #20.

We sincerely apologize to all model authors whose documentation was accurately written but incorrectly flagged as missing by our system. This was our error, not theirs. The TPA oracle (cryptographic chain) uses structured API fields and was not affected by Bug 1. Bug 2 has been corrected and TPA observations will update in the next nightly run.

Correction applied: Feb 24–25, 2026  ·  Oracle version: v1.1.0  ·  Compliance Mapper: v2.0  ·  Changelog on GitHub ↗
Total Observations
Today (UTC)
Unique Targets
CRITICAL severity
Avg compliance score
Shadow pressure index
Shadow events observed
Query latency
Intelligence Summary
Avg Risk Score
Models Gated
High+ Risk
Avg Exposure
Orgs Engaged
Issues Classified
Compliance Ranking
View full ranking →
Top models by score
Top organizations (min 5 models)
How Crovia works
Four steps. One continuous record.
01 — OBSERVE
Document audit
Every day, Crovia reads the public model card of each monitored model and checks 20 necessary disclosure elements — training data, evaluation, intended use, risks, and more — against EU AI Act, GDPR, and 9 other jurisdictions.
02 — CERTIFY
Temporal proof
Each observation is sealed into a Temporal Proof of Absence (TPA) — a cryptographic record with Ed25519 signature, Pedersen commitment, and Merkle chain anchor. What was missing, and when, is now immutable.
03 — MEASURE
Public signal scan
Model Sonar v2 scans public metadata (model cards, dataset tags, config.json, base‑model chains). It does not run inference or access private data — it records declared vs unresolved signals.
04 — PROVE
Lineage graph
The lineage graph links organizations → models → datasets using declared sources and base‑model links, exposing shared provenance gaps and inherited omissions. GDNA research is under validation; only curated samples are published.
Explore the Registry