refresh card (through block-8719200)

README.md

@@ -1,43 +1,50 @@
 ---
 license: cc-by-4.0
-pretty_name: cascade held-out eval pool (24h-lagged reveal)
+pretty_name: cascade held-out eval pool (lagged exact reveal)
 tags:
   - time-series
   - forecasting
   - benchmark
 ---
 
-# cascade eval pool — lagged public reveal
+# cascade eval pool — lagged public reveal (exact bytes)
 
-Daily snapshots of the **held-out evaluation pool** used by the
-[cascade](https://github.com/TensorLink-AI/cascade) subnet, built from the
-[tsbench-forge](https://github.com/tensorlink-dev/TSBench-Forge) live catalog of
-real public time series across the 7 GIFT-Eval domains.
+Retired snapshots of the **held-out evaluation pool** used by the
+[cascade](https://github.com/TensorLink-AI/cascade) subnet. Each folder is a
+**byte-identical mirror of the `pool/snapshots/block-<N>.tar` that validators
+scored** — downloaded from the private pool bucket, sha256-verified against the
+publisher index, and republished unmodified. A snapshot is revealed only after
+a newer snapshot has superseded it, so no revealed pool can be selected by a
+current or future round.
 
-**Each snapshot is revealed 24h AFTER it was used for scoring** (`as_of` is a
-closed UTC day). Because the pool rotates daily, a released snapshot is always
-for a round that is already scored — it lets anyone **reproduce the leaderboard**
-without exposing the round currently in play.
+## Verifying a round receipt
+
+1. Your receipt's pool provenance carries the snapshot tar's sha256.
+2. Find the folder whose `POOL_SHA256` matches; `block-<N>.tar` in that folder
+   is the exact artifact — hash it yourself to confirm.
+3. Series order = filenames sorted lexicographically; `window_ids` are
+   positional (`w<i>` = the i-th series in that order).
 
 ## Layout
 
 ```
-snapshots/<YYYY-MM-DD>/          one folder per revealed day (as_of)
-  <series_id>.npy                float32, freshest context_length + horizon points
+snapshots/<as_of>-block-<N>/     one folder per revealed snapshot
+  block-<N>.tar                  the EXACT tar validators scored
+  POOL_SHA256                    its sha256 (matches receipts + publisher index)
+  <series_id>.npy                the same files, unpacked for convenience
   metadata.json                  {series_id: {freq, seasonal_period, domain, source}}
-  provenance.json                build config (context_length, horizon, builder_version)
+  provenance.json                build config recorded at publish time
 ```
 
-Scoring is identity-agnostic: MASE uses `seasonal_period`, CRPS/WQL use the
-values + a model's quantiles, and the KOTH cluster-bootstrap groups windows by
-`source`. Re-cut windows of `context_length + horizon`, forecast the horizon,
-and the metrics reproduce byte-for-byte.
+Legacy `snapshots/<YYYY-MM-DD>/` folders (through 2026-08-03) predate this
+scheme: they were **rebuilds**, not byte mirrors, and are known to differ from
+the scored tars (see the repo issue history). Use the `-block-<N>` folders for
+exact replay.
 
-## Latest snapshot — `2026-08-03`
+## Latest revealed — `2026-07-28` (block 8719200)
 
-- **series:** 1097
-- **domains:** transport=326, energy=244, nature=219, healthcare=118, econ_fin=79, sales=64, web_cloudops=47
-- **cadences:** min=347, D=218, 15min=157, 5min=147, 30min=100, H=95, 10min=14, 6min=10, 240S=6, 15S=1, 10S=1, 1S=1
+- **series:** 2896
+- **sha256:** `bf5258612d7d9be71764a65caa2966ac114a4f559baad7082171af6025b1ad66`
 
 > This is an **evaluation** set, not training data. Publishing it to train on
 > would contaminate the benchmark it exists to measure.