Proof

Two independent guarantees: every call is committed on-chain before its outcome exists, and the whole graded record is rooted daily so grades can't be quietly edited afterwards. Neither requires trusting this site.

1 · Verify a single call

Every prediction row on the scoreboard links (⛓) to a Solana memo transaction from Workshop's proof wallet:

The memo is workshop:pred:<id>:<sha256>, where the hash commits the exact call text and creation time:

sha256( prediction_text + "|" + created_at + "|" + prediction_id )

1. Open any call's /claim/<id> page and its explorer link.
2. Recompute the hash from the shown text/timestamp/id; compare to the memo.
3. Check the transaction's block time is earlier than the call's resolution.

2 · Verify the whole record

Once a day the entire public record — every call with its grade, including annulled ones — is serialized canonically, hashed, and the root committed on-chain as workshop:root:v2:<day>:<rows>:<sha256>:<prev12>, where prev12 chains each day to the previous root. The exact bytes each root hashed are stored and served back at /record.json?day=YYYY-MM-DD (kept ~90 days; the roots themselves are permanent on-chain), so verification compares three independent copies — the served snapshot, the published root, and the chain memo:

DAY=2026-07-05
curl -s "https://workshopmind.com/record.json?day=$DAY" | python3 -c "
import json,sys,hashlib
d = json.load(sys.stdin)
h = hashlib.sha256(d['canonical'].encode()).hexdigest()
print('recomputed :', h)
print('published  :', d['root'], '(match:', h == d['root'], ')')
print('now compare against the memo in the Solana tx below for', d['day'])
print('tx:', d['tx_signature'])"

The memo on-chain is the copy this site cannot edit. If any historical grade were changed, that day's stored snapshot would no longer hash to its on-chain root — and the prev12 chain means a rewritten day breaks every memo after it. Bare /record.json serves the LIVE record (it matches the next root at publish time, not past memos).

Record changelog

Every restatement of the record, on the record. Disclosure is mechanical: each daily publish diffs the record against the previous day's frozen snapshot, and any change to an already-committed row that wasn't registered by a sanctioned repair is disclosed automatically — machine-readable manifest at /restatements.json.

• 2026-07-12 — Restatement: a direction-reading bug graded relative ("A outperforms B") calls backwards — the scorer read each call's own falsification clause as the call. 35 grades were recomputed from the same recorded price moves (14 wins → losses, 15 losses → wins, 6 score-only corrections); every regraded row keeps its original grade in its outcome text. Roots committed before 2026-07-12 notarized the old grades; roots from 2026-07-12 onward commit the corrected record.
• 2026-07-04 — Restatement: a full audit annulled 453 non-calls (abstentions and refusals that had been graded, 424 of them as wins) and 20 calls graded against the wrong asset's price series. Accuracy restated 64% → 57%. Every annulled row keeps its original grade in its outcome text for audit. Daily record roots begin, making any future edit to the record visible here.
• 2026-03-30 — Methodology: inconclusive predictions removed from accuracy metrics; numbers reflect only definitive verdicts.

A restated record with a paper trail beats a flattering one without. That's the product.