Track record

The consensus mirror's backtest shows +2.2 pts/quarter net vs the S&P 500 over 35 complete quarters (2017Q3 to 2026Q1), directional, NOT statistically significant (t≈1.3; 95% CI includes zero). Most of that average is one quarter: 2026Q1 posted a +47.5 pt edge, roughly half of it two AI-hardware names (SNDK +258%, AMD +186%) marked at what proved to be a local peak; excluding it the edge is +0.9 pts/quarter.

The consensus mirror's backtest shows +2.2 pts/quarter net vs the S&P 500 over 35 complete quarters (2017Q3 to 2026Q1), directional, NOT statistically significant (t≈1.3; 95% CI includes zero). Most of that average is one quarter: 2026Q1 posted a +47.5 pt edge, roughly half of it two AI-hardware names (SNDK +258%, AMD +186%) marked at what proved to be a local peak; excluding it the edge is +0.9 pts/quarter.

Risk-adjusted performance

The edge above is a raw return number. Professional allocators judge a strategy on a different question: how much risk did you take to earn it? A book that returns a lot by swinging violently is worse than one that returns a little, smoothly. So here is the whole risk panel over 34 quarters, net of cost, next to the S&P 500, with nothing hidden.

Timestamped calls, graded

These are the calls that were hashed and timestamped on Bitcoin BEFORE the filing landed, then graded against what the filer actually did. A call is scored only once at least 16 days have passed, because a call stamped yesterday cannot have been confirmed yet and counting it as a miss would be arithmetic rather than evidence. Record so far, 2026-07-07 to 2026-09-21.

Read this honestly, because it is the number that counts against us. On JP 4 of 22 landed, which is above the 3.5% base rate but too few calls to claim anything. 66,319 daily receipt rows collapse to 85 distinct bets, and the distinct count is the one reported here.

Holdout backtest, which is not the record above

A different measurement, and a weaker kind of evidence. The model is trained on the older 70% of each market’s filing history and scored on the 30% it never saw, and the rows below are the calls it made there at a predicted buy probability of 60% or higher, each checked against the filing that actually followed. Hits and misses both shown. These calls were NOT stamped in advance: the filings already existed when the model ran, so this shows the model can rank, and it cannot show what the section above shows. Where the two disagree, the stamped record is the one that counts.

Live calls made from today forward are logged with their filing-due date and confirmed as each filing lands: 10 confirmed so far and 83 awaiting their next filing, counted as distinct bets rather than as the 152 daily receipt rows they were stamped across. Receipts are the model's own out-of-sample calls (P(add)>=0.60), dated and scored against the actual next filing. Hits and misses both shown. JP/UK confirm in days, so a call is checkable almost immediately. Research, not advice.

The ledger, call by call

Every stamped call with its market, filer, issuer, attested stamp, window, maturity date, outcome and the rule that graded it. The table loads the latest grading from the ledger route and falls back to the copy built into this page. One count is published: graded rows and hits among them. A stranger can pick any row and check it against the artifact named on the methodology page.

Cryptographically timestamped

Anyone can backdate a self-reported track record. We don’t. Each day’s confident calls are frozen into one file, hashed, and timestamped on the Bitcoin blockchain via OpenTimestamps before the filings land. On 2026-09-22 that was 1559 per-investor calls plus the US Q2026-09-30 board.

Verify: download commit_2026-09-22.json + commit_2026-09-22.json.ots from the public repo and run ots verify commit_2026-09-22.json.ots. The timestamp proves the calls existed on 2026-09-22 and were not edited after the filings. Self-reported receipts can be backdated; this one cannot.

Corrections

What we got wrong and fixed. We publish our mistakes because a research service you can trust is one that shows its corrections, not one that never admits them.