What survived the corpus sweep: two measurements out of 4,595
We ran every public Pine strategy we could find against six markets on daily bars, under four guards published before the run. Two script-market measurements cleared all four — and neither has fills we can confirm.
The Alpha Desk publishes twenty scripts we wrote ourselves, and the obvious objection to that is that we chose them. So there is a second exercise, and it is a different one: scrape every public Pine script we can find and measure the whole set, with the rejection counts published beside the results. The desk scores a script on a composite across eighteen market-timeframe cells. The sweep below runs each market separately on daily bars against four guards. Different questions, different rulebooks — they are not two views of one number.
The snapshot holds 12,424 scripts. An AST pass sorts them by what they actually do — 3,797 declare strategy() or place a strategy.* order, 4,058 emit discrete signals through alertcondition, alert(), or a plot marker, 2,787 only plot a series, 1,470 are visual or utility tools, and 312 are libraries. Only the first group can be backtested without wrapping it in trading rules we invented on its behalf, so it is the group this post is about. The other tracks are measured differently, and those results live on /corpus.
Each of the 3,797 ran against six markets — BTCUSD, ETHUSD, SPX500, QQQ, GOLD, EURUSD — on daily bars. That is 22,782 script-market measurements. The unit matters and everything below uses it: a script can be ineligible on one market and perfectly measurable on another, so counting scripts would force a judgment the data does not support.
Eligibility removes most of the pool. 5,478 of the 22,782 measurements produce a result the guards can even read; the buckets the other 17,304 fall into are published in full on /corpus. Near-identical uploads are then collapsed by returns correlation at 0.99 or above, which brings the pool to 4,595. Duplicates are the norm in a public corpus, and leaving them in would inflate every number that follows.
Four guards apply to what is left, each fixed before the run. Positive Sharpe in the last 30% of history, the window the author did not fit to: 3,030 of the 4,595 survive. A breakeven cost of at least 20 basis points, roughly twice a realistic round trip: 1,990 survive. Beating buy-and-hold over the identical window: 626 survive. Deflated Sharpe of 0.95 or better, which asks whether the result is simply the best of a large number of tries: 4 survive.
All four at once: two, both on ETHUSD. Neither is tagged as having verified fills — their fill sequence depends on an intrabar path daily bars cannot confirm — so even those two are a label, not a finding. Per-market ranked rows are published fifty deep; 63 scripts appear in at least two of those lists and 4 in four or more, which is the closest thing this sweep produces to a cross-market result.
None of that says public Pine is worthless. It says the combination — surviving a deflated Sharpe test, a cost floor, an out-of-sample split and a buy-and-hold benchmark simultaneously — is rare enough that the correct prior for any individual script is that it does not clear them. The desk catalog reads the same way from the other direction: twenty audited scripts, zero Verified, sixteen Conditional, four No alpha, alpha scores between 23 and 49, and seventeen of the twenty trailing buy-and-hold.