A forecasting record you can check.

The evidence

Six links turn a forecast into something a stranger can check: nothing on the board rests on our word about when a forecast was made.

1. 01Forecastwritten to a file with a secret salt, before the question lockssce-2026-10-infl1y
   11-05 11:42 UTC
2. 02Commitmentsha256 of that file, published at oncea0dc68959107f6158c79…
3. 03Manifestevery hour, all commitments in one file that names the previous manifest's sha25620261105T114223Z
   3b8da15747317fe38fba…
4. 04Bitcointhe manifest's sha256 is anchored in a block through OpenTimestampsblock 974666
   2026-11-05 12:02 UTC
5. 05Revealwithin an hour after the lock the forecast file is published; its hash must matchfile public
6. 06Scoreagainst the arena's outcome, with the arena's scoring codeskill 93.3 · #1 of 14

The values under each step are those of one forecast, sce-2026-10-infl1y; every forecast has its own chain below.

One forecast, end to end

The forecast furthest along its chain right now, with every value it rests on. Each link points to where it can be checked.

1. 01Forecast
   questionsce-2026-10-infl1y
   committed2026-11-05 11:42 UTC
   version2.11
   locks2026-11-07 15:00 UTC
2. 02Commitment
   sha256a0dc68959107f6158c79ed75e1ed3d50606bc41f3e3720009b92c0fce96109ff
   of file20261105T114212Z.forecast.json
3. 03Manifest
   file20261105T114223Z.json
   sha2563b8da15747317fe38fbac6ce93c251ab6aaad21a06bf8b428feb0634fdba9cd9
   previous20261105T104151Z.json
4. 04Bitcoin
   block974666
   block time2026-11-05 12:02 UTC
   proof20261105T114223Z.json.ots
   checkat least 2 h before the lock ✓
5. 05Reveal
   file20261105T114212Z.forecast.json
   forecastN(3.914, 0.1)
6. 06Score
   outcome3.91441
   loss0.0234
   persistence0.3506
   skill93.3
   rank#1 of 14

The chain, forecast by forecast

Every question in the record. A filled dot is a link that can already be checked; a hollow one is still to come (the Bitcoin anchor about an hour after a manifest, the reveal after the lock, the score at resolution).

version

one committed forecast file; xumi commits a new version when its forecast for a question changes.

counted version

fixed in advance and by commit time alone: the last version committed at least 3 hours before the lock. Nothing is chosen after the outcome.

proven

a manifest listing the counted version is in a Bitcoin block timestamped at least 2 hours before the lock. Block timestamps may lag real time by up to about an hour, so this shows the forecast existed before the lock.

miss

no counted version, or its anchor is too late or missing 24 hours after the lock. A miss is never replaced by an earlier version: it counts zero in pairwise, and in the skill summary it takes the worst skill any entrant had on that question, so withholding a proof can never help.

Coverage. 52 arena questions lock inside the record; xumi has a counted or pending forecast for 52 of them.

Check it yourself

One command checks the whole record without trusting us: every file in every manifest, that each revealed file matches its commitment, that the manifests chain and only ever grow, that the counted version is the one the rule picks, that each OpenTimestamps proof commits to its manifest and lands in a real Bitcoin block (header hash and proof of work checked, block hash cross-checked on two explorers), the 2-hour margin, coverage against the arena's question list, and for single-number questions the loss against the arena's outcome.

pip install opentimestamps
curl -fsSLO https://renlab.ai/data/xumi/verify.py
shasum -a 256 verify.py                # d688ca3cfa22c38002272e8f143525c60d99ead76b0a50093d2ad85cdc79ddfa
python3 verify.py

By hand, for one forecast:

1. Download the forecast file (after the lock), its manifest and the manifest's .ots proof.
2. shasum -a 256 <forecast file> must equal the forecast's commitment and its entry in the manifest.
3. Check the proof on opentimestamps.org (drop the manifest and its proof), or with ots verify against your own Bitcoin node. The block must be timestamped at least 2 hours before the lock.
4. Score the revealed forecast against the arena's outcome with the arena's scoring code; it must match the result below.

The salt in each forecast file keeps the commitment from revealing the forecast before the lock. The whole record is in record.json.

Manifests

Latest 10 of 44 manifests shown; all are listed in record.json.

Forecasts and results

Revealed forecasts, scored with the arena's scoring code against the arena's resolved outcome. Skill is 1 − loss / persistence loss, ×100, against the arena's persistence baseline; rank is among every method that answered that question on the arena, from the arena's published losses. Lead is how long before the lock the counted version was committed, next to the median for the other entrants' filings on the arena.

What this does and does not prove

• Proves that each counted forecast existed, unchanged, before its question locked, and that no version was dropped, swapped or chosen afterwards.
• Proves each single-number score from the arena's outcome; profile and ranking scores use the arena's code too but are not rechecked by the script.
• Does not prove that this is the only forecaster we run, or that no other chain exists. Only a submission to the arena, timed and scored by the arena, settles that; xumi has not been submitted yet.
• Does not prove skill on a small sample. Read the results with the number of questions next to them.

Status and rules

xumi has not been submitted to the Social Simulation Arena yet. The numbers on this page are our own scoring with the arena's published code, compared with other entrants' live results; they are not arena results and the arena has not reviewed them. Questions, data and scoring code: Social Simulation Arena; paper: Li et al. (2026), Zenodo, CC BY 4.0. Arena data as of 2026-11-12.