# Datasheet — Crovia Continuity Observatory

Following the *Datasheets for Datasets* template (Gebru et al., 2018, [arXiv:1803.09010](https://arxiv.org/abs/1803.09010)).

## Motivation

**For what purpose was the dataset created?**
To create a public, cryptographically signed observation log of how AI vendors disclose — or fail to disclose — material information about their models over time. The dataset documents *silence*: the gaps in vendor disclosures that conventional transparency indices (which rely on vendor self-reports) cannot capture.

**Who created the dataset?**
Crovia Trust, an independent continuity observatory project. Maintained by the founding engineer with no commercial sponsorship. No vendor input.

**What support was needed?**
None institutional. Built and operated on a single Hetzner CPX42 instance with Internet Archive Wayback Machine as a public secondary source. Cost order of magnitude: tens of EUR/month.

## Composition

**What does the dataset represent?**
Each row is an *observation* of a target (a specific AI model, dataset, organization, or repository) at a specific timestamp. Observations are typed:

- `presence`  — public artifact reachable, content fingerprinted
- `absence`   — public artifact reachable, but a class of disclosure (e.g., training data, licenses, eval) is empirically missing or unchanged for ≥ N days (`AX.ABS`)
- `change`    — content changed compared to the prior observation
- `signature` — administrative event (Merkle anchor, OTS stamp, ledger seal)

**How many instances?**
At deposit time: 111,054 observations · 3,500+ unique targets · spanning approx. 2025-Q4 to present.

**What schema?**
Per row (JSONL):

```json
{
  "observation_id": "<uuid4>",
  "target_id": "org/model",
  "tipo_target": "model" | "dataset" | "organization" | "repository",
  "type": "presence" | "absence" | "change" | "signature",
  "observed_at": "<ISO-8601 UTC>",
  "source": "wayback" | "hf_card_http" | "arxiv" | "robots_txt" | "github_api" | "...",
  "evidence": { "...": "..." },
  "claim": { "...": "..." },
  "signature": { "alg": "Ed25519", "sig": "<base64>", "pub": "<base64>" },
  "merkle_leaf_hash": "<sha256>",
  "ledger_offset": <int>
}
```

**Sensitive content?**
None. All targets are public artifacts (HF model pages, GitHub repos, arXiv papers). No personal data of unidentified individuals. Vendor org names and product names are present (these are public).

**Recommended uses?**
- Studying the temporal behavior of AI disclosures over time
- Calibrating compliance frameworks (EU AI Act Art. 53/55)
- Cross-validating other transparency indices (Stanford FMTI, MIT AI Index)
- Anchoring legal/insurance claims that require third-party-witnessed silence
- Research on “proof-of-silence” cryptographic primitives

**Discouraged uses?**
- Inferring intent. The dataset records what is publicly observable. It does not ascribe motive.
- Naming and shaming individual employees. Crovia osserva, non giudica.

## Collection process

**How was data acquired?**
Three primary observation channels, each scheduled by independent timers:

1. **Wayback channel** — query Internet Archive TimeMap for the canonical URL of every target. Each capture becomes one observation. This is the strongest channel because the source is a third party.
2. **Direct channel** — polite, respectful HTTP fetch of the canonical page (HF model card, arXiv abstract, robots.txt). Only used when wayback coverage is sparse.
3. **Co-attestation channel** — Crovia submits each target to wayback Save Page Now, ensuring the Internet Archive captures it independently. This produces a *future* secondary witness.

**What target population was sampled?**
The unified target list at deposit time contains ~4,300 entries seeded from: the EU AI Office GPAI systemic-risk list, the Stanford CRFM Foundation Model Index, the Hugging Face top-1000 by downloads, the GitHub topic:llm/topic:ai stars>500 set, and frontier-focus models with active legal exposure.

**Sampling bias?**
The list is intentionally biased toward (a) high-traffic, (b) policy-relevant, (c) regulatory-disclosed models. It does not attempt to represent the long tail of community fine-tunes. This is a *load-bearing* sample, not a *representative* one.

**Time period?**
Observations begin 2025-11-24 (genesis of the axiom ledger). New observations are appended continuously.

## Preprocessing / cleaning

- Test-only and demo targets are filtered before silence indexing.
- Duplicate observations of the same `(target, capture_url)` are deduplicated by hash.
- The full append-only ledger is preserved unmodified; cleaning happens only at the index layer.

## Uses

**Has the dataset been used yet?**
- Live website at <https://croviatrust.com> for public exploration.
- Per-model dossier pages at <https://croviatrust.com/m/&lt;org&gt;/&lt;model&gt;/>.
- Embeddable widgets and SVG badges via <https://croviatrust.com/embed.js>.
- Weekly leaderboard publication on GitHub: <https://github.com/croviatrust/crovia-evidence-lab>.

## Distribution

**License**
Data: CC-BY-4.0. Attribution: cite the deposit DOI and link to <https://croviatrust.com>.
Code (collectors, generators): Apache-2.0 / MIT (see individual repos).

**How will it be distributed?**
- Public website (continuous)
- Weekly Zenodo deposit (versioned DOIs)
- GitHub repository (snapshots)

## Maintenance

**Who maintains?**
Crovia Trust founder. Single-maintainer at present. Help is welcome.

**How will errors be communicated?**
Issues at <https://github.com/croviatrust/crovia-evidence-lab/issues>. The append-only ledger means errors are not *deleted*; they are *amended* with a follow-up observation that supersedes them.

---

*Crovia osserva, non giudica.*