About the Methodology · Infra Atlas

The sourcing rules

Every factual claim in the Atlas must satisfy four rules before it is published. No exceptions.

Primary source only

Data comes from the vendor's authoritative documentation, pricing page, or a public API — not a blog post, a press release, or a third-party article. When a source is secondary (e.g. a community dataset built from vendor APIs), it is named explicitly on the instrument card and in the raw data.json field "source".

Source is named and linked

Every instrument lists its sources. Matrix cells carry a per-cell source URL: click any cell to open the detail drawer and see the exact page the claim is sourced to. Decision pages have a numbered source list; comparison rows carry inline src comments in the generated HTML.

GA only

A region, feature, or SKU is included only when listed as generally available in the vendor's authoritative documentation — not when announced, in preview, or "coming soon". The exact definition of GA per provider is documented in docs/data-policy.md.

Stale data is a bug

Wrong data is not a style issue — it is a defect. If you spot something off, open an issue with the upstream source that contradicts the site. The Atlas will be corrected and the verified date re-stamped. Open an issue →

Data tiers

Not all data refreshes the same way. The Atlas uses three tiers, visible on every instrument card and the data API page.

Tier	How it works	Instruments	Cadence
Tier 1 — Live	A `refresh.sh` script pulls data from a public API or dataset at 06:00 UTC daily. No credentials. The generated timestamp in `data.json` is the actual pull time.	EC2 Observatory, Azure VM Atlas, OCI Compute Observatory, OVH Instance Catalogue	Daily
Tier 2 — Curated	Hand-maintained data that is re-serialized to `data.json` by `refresh.sh`, but the underlying data only changes when a human edits the source HTML and re-stamps the verified date. The script proves the JSON faithfully mirrors the HTML; a CI region-drift guard pins the set to a vendor-verified reference.	Region Map, GCP Compute Index	On change (human-triggered)
Tier 3 — Dated snapshot	Fully hand-curated reference content. Each page carries a Verified · YYYY-MM-DD stamp. CI freshness guards alert when a snapshot is older than its threshold (45 days for the fast-moving Generative AI Atlas; 180 days for the compliance and cross-cloud matrices; 365 days for Decisions).	All APIM instruments, Kubernetes, IAM, Networking, Compliance, Confidential Computing, Observability, Sovereignty, AI Atlas, Egress, Equivalent-SKU, Decisions	On re-verification

The hero "Updated daily · 06:00 UTC" refers specifically to the Tier 1 instruments. The manifesto and colophon on the homepage state this correctly. Do not generalise it to the rest of the Atlas.

What "Reviewed · date" means

On a Tier 3 instrument, the Verified · YYYY-MM-DD stamp is an assertion that, on that date, a maintainer:

Checked each cell

Every claim in the matrix or table was re-verified against its cited primary source. Cells where the source had changed or was ambiguous were updated, marked ◐ (partial), or annotated with a caveat note.

Updated stale claims

Anything found to be incorrect was corrected before the stamp was bumped. The stamp is not bumped on a schedule — it is bumped only after verification has actually happened.

Linked the sources

Every cell in a matrix carries a src field (visible in the detail drawer). A verified page has no unsourced claim. If a source cannot be found for a claim, the claim is removed or downgraded to ◐ with a note.

Automated freshness guards (CI)

The following checks run on every push and fail loudly if data drifts. They are the mechanical layer of the sourcing commitment.

✓

Region drift guard — check_region_drift.py pins the Region Map's provider sets to a dated, vendor-verified region-reference.json. Any region added or removed without updating the reference fails CI.
✓

Data-JSON sync guard — verify-data.yml re-extracts data.json from the curated index.html for Tier 2 instruments and diffs it against the committed file. Proves the JSON faithfully mirrors the HTML source.
✓

Landing-page stat guard — check_landing_stats.py derives instrument counts from data.json and fails if a homepage chip shows a stale number. sync_landing_stats.py self-heals the chips after each daily refresh.
✓

Instrument count guard — check_instrument_count.py counts the live instrument cards in the HTML and fails if any static count literal (hero stat, section header, filter chip) diverges. Prevents the "21 vs 22 instruments" self-report drift.
✓

OG card guard — check_og_card.py fails if the social share card's instrument count diverges from the live cards. build_og.py self-heals it.
✓

Freshness guard — verify-freshness.yml fails when a Tier 3 instrument's Verified stamp is older than its threshold, forcing a re-verification pass. Thresholds: AI Atlas 45 days; most matrices 180 days; Decisions 365 days; European Sovereignty 183 days.

Honesty about what we don't know

The Atlas deliberately uses three levels in its comparison matrices:

✓ Yes / Supported

The claim is verified against a primary source. The source is linked in the detail drawer.

◐ Partial / Caveated

The feature exists but with significant limitations, or the evidence is mixed. The drawer note explains what the asterisk means. This level is used intentionally — it is not a placeholder for "we didn't check".

✗ No / Not available

The claim is verified as absent or unsupported. Not a default — requires explicit evidence of absence.

Per-region family membership in the EC2, Azure and OCI compute instruments is sourced from public datasets and is best-effort, not independently credential-verified. The exact wording is in docs/data-policy.md §3. GCP compute has no credential-free price source; the instrument carries no pricing column rather than invent one.

Who runs it

One engineer, working in public. The Atlas is open-source on GitHub. All data sources are public, credential-free, and documented. Corrections are welcome — the fastest path is an issue with the contradicting upstream source.

There is no company, no investor, no ad network, no vendor relationship. The Atlas makes no money from the data it shows. The funding model is reader-supported donations; the independence model is "the methodology is public and auditable".