Kairos Signal / revision ledger / sample case — generated 2026-09-15 01:45 UTC from live state

Arweave replaced its own history by −14.0%. Did your model notice?

On 2026-09-13 our detector recorded Arweave's network blocks count for 2026-08-14 as 1,980,116. Days later, the same upstream API answered 1,702,204 for that same already-observed date — a -14.9% backward revision, published by the source itself, with no announcement, no changelog, and — at every other data vendor — no trace. They overwrite the number and move on. Downstream models keep consuming whichever value they happened to cache, and nobody tells them the ground truth moved under them.

The four Arweave restatements we hold

Upstream swung between ~1.98M and a stale ~1.70M height four times over three weeks. Each row is one detected transition; the first-detection stamp never moves, and revision_id is a deterministic hash of the revision identity, so a subscriber can de-duplicate replays safely.

Resolution rule: LATEST OBSERVATION WINS for current-state reads; the superseded value is retained, never overwritten — the retained before-value is the only reason the restatement is detectable at all.

Verify it yourself

The current upstream value (a live height, always advancing):

curl -s https://arweave.net/info | python3 -c "import json,sys; print(json.load(sys.stdin)['height'])"

The value upstream published for 2026-08-14 after the rewrite is visible in our own daily archive:

curl -s "http://100.65.160.3:8123/?query=SELECT+value+FROM+depin_daily+WHERE+network='AR'+AND+metric='BLOCKS'+AND+snapshot_date='2026-08-14'" -d ''

Every ledger row carries a verify_url — the exact upstream endpoint each series came from — so a buyer never has to take the diff on faith. Query the full ledger:

curl -s "https://kairossignal.com/v1/revisions?symbol=AR"

Who else is on the ledger

27 restatements over the archive we hold (26 backward revisions of cumulative metrics + 1 sub-noise case), across GRT, STORJ, SOGNI, AUKI, R1 and EIGEN — including an all-time fees counter that moved backwards by 23,480 and three token-transfer counters that did the same. Detection is mechanical; the full queryable feed is /v1/revisions, and you can subscribe to restatements per symbol via POST /v1/alerts/subscribe (revision_restate, webhook).

The product, in one paragraph

If a fund's model consumed the pre-restatement Arweave number, it was wrong for days and had no way to know: the vendor it bought from had already overwritten. A risk team needs the reverse question answered too — did any vendor history move under me, and when? That record does not exist anywhere else, because recording the before-value has negative value to a vendor whose pitch is a clean, stable history. We publish the before, the after, both stamps, and the URL to check the current upstream value yourself.