The chain of custody. Published numbers trace back to immutable evidence: a content-addressed quote snapshot, an append-only observation ledger, and the raw source payloads behind verified rows. Raw payloads are archived privately and available on request; this page renders the rest of the chain — identities, checksums, and evidence references — from committed artifacts at build time.
The capture pipeline collects every source independently, pushes the gzipped raw payload bundle to the immutable private Neon archive, verifies the stored bytes, and only then ingests observations. A failed archive write leaves the database untouched, and partial source failures become explicit coverage gaps instead of discarded or back-filled rows.
Immutability
Snapshots never mutate
Each published quote snapshot is a complete table addressed by the SHA-256 of its own contents; refreshes create a new snapshot instead of editing one, and database triggers reject inserts, updates, or deletes that would alter published data. Exact price history is append-only, and daily aggregates are immutable source-set versions.
What “verified” means
Pinned, hashed evidence
An exact observation is verified only when its raw payload is already durable in the private archive, its evidence reference pins immutable bytes, and the payload’s SHA-256 is recorded on the row. New references name a content-addressed Neon object; historical Git references resolve through the archive’s preserved commit mappings. A daily aggregate is verified only when every observation behind it is; its reference is a Neon source-set digest — the SHA-256 of those observations’ immutable IDs — with no Git object of its own. Raw payloads are archived privately and available on request: email aicomputetracker@gmail.com with a reference from the table below.
Archive writes are checked by content hash, database triggers reject updates and deletes, and the maintenance writer can only read and insert archive records. A database owner can change those protections. The references below are private identifiers, not public download links. This page is built from committed artifacts and does not query the private archive at build time.
Published quote snapshot
The current GPU quote table is one complete, immutable Neon snapshot exported to the repository as src/data/cloud-prices.snapshot.json. Its identity below is derived from that committed export.
Identity and row populations of the published quote snapshot
Recomputed at build time from the committed export using the same content addressing the database verifies on every round trip. When the committed export is the promoted snapshot — which the release gate requires before it is committed — this equals the published Neon snapshot ID. The Neon-side content checksum itself is enforced upstream by checksum verification on read and by triggers that reject changes to published data; this page cannot query Neon.
Checksum of the committed export (not the Neon content checksum)
Plain SHA-256 of the committed file bytes, as shasum -a 256 src/data/cloud-prices.snapshot.json prints it at the deployed commit. The repository is private; the export file is available on request.
Observation ledger
The exact price-history ledger is append-only. /price-history.json serves 2891 of its points as of , of which 2357 are verified against retained raw payloads. It withholds every point of the sources the site withdrew, and of the class aggregates they could have fed. Each withdrawal’s record counted them once, in the ledger export it was frozen on: 896 (864 verified) of the export of when Vast.ai was withdrawn on 2026-10-20 (correction record) and 105 (3 verified) of the export of when CoreWeave, Crusoe, Salad, RunPod, Voltage Park, GetDeploying and AIMultiple were withdrawn on 2026-10-20 (correction record). A later export can hold more of their points, and those are withheld too. Quality tiers below are the ledger’s own qualityStatus field. The table lists the 30 most recent of 118 capture days; /price-history.csv holds every observation with its quality status.
Observation counts by date and quality tier, 30 most recent capture days
Date
Verified
Provisional
Secondary
Estimated
Total
Evidence refs
52
—
—
—
52
25
52
—
—
—
52
25
52
—
—
—
52
25
52
—
—
—
52
25
37
60
—
—
97
10
52
—
—
—
52
25
37
—
—
—
37
21
37
—
—
—
37
21
37
—
—
—
37
21
37
—
—
—
37
21
41
—
—
—
41
22
26
49
—
—
75
7
41
—
—
—
41
22
41
—
—
—
41
22
41
—
—
—
41
22
42
—
—
—
42
23
42
—
—
—
42
23
42
—
—
—
42
23
30
53
—
—
83
8
45
—
—
—
45
23
41
—
—
—
41
22
45
—
—
—
45
23
49
—
—
—
49
24
49
—
—
—
49
24
49
—
—
—
49
24
34
57
—
—
91
9
37
60
—
—
97
10
52
—
—
—
52
25
39
—
—
—
39
20
16
—
—
—
16
10
Raw-evidence references
The 25 newest of the 1373 distinct references the ledger carries, with the SHA-256 of the payload each pins (the rawEvidenceSha256 recorded on each citing observation). Git-pinned references name an exact historical commit plus a bundle path and source fragment, preserved in Neon. Content-addressed Neon references pin the compressed bundle’s SHA-256, path and source fragment. Neon observation-set digests identify the exact source set behind an immutable daily aggregate version inside the database. Click a reference or digest to select it for copying, and email aicomputetracker@gmail.com to request the payload behind it.
How a requested payload is checked: gunzip the pinned bundle, find the evidence entry whose source matches the reference fragment, and hash its retained texts (single body hashed directly; paginated responses length-framed) — the digest must equal the rawEvidenceSha256 shown above and recorded on the observation in /price-history.json. Bundle paths follow quotes/YYYY/MM/price-observations-<scope>-<timestamp>.json.gz. An AWS Price List entry, from collector live-quotes-v8 on, retains a projection in place of the offer file its source names (hundreds of megabytes): that day’s region-index read and check of the file, with the file’s tracked rows as published. Its one retained text is what is hashed, so the digest is never the file’s; the entry’s originalSha256 is the file’s own SHA-256, which is not kept. Every digest is also listed by day, with a Merkle root per day, in the evidence digest manifest.
Forward raw ladders
The forward-curve archive holds 69 sessions from through . 50 live sessions have their raw Kalshi ladders retained gzipped in the private Neon archive (path convention kalshi/YYYY/MM/kalshi-<capturedAt>.json.gz) and are derived under the current estimator, 13 of them re-derived from those ladders when the estimator changed. 7 live sessions have no retained ladders, so they keep the methodology they were published under and are never compared with the current one. The 12 legacy sessions were imported from an earlier archive and predate raw retention. Current versions: kalshi-ladder-median-50-v2 + unsmoothed-v1. The table lists the 30 most recent sessions; every archived session has its permalink on the snapshot archive.
Archived forward sessions and their expected raw-ladder paths, 30 most recent
Expected paths are composed from each session’s capture stamp using the retention workflow’s naming convention; the workflow files a ladder under the year/month of its own push time, so a capture at a month boundary can land in the neighboring folder. This build-time page cannot list the private branch to confirm, which is why these paths are labeled expected rather than verified.