What this is, and what it is not. This review does not — and is not intended to — show that the BSI AI Catalyst Scanner is a reliable tool to be used on its own for investment decision-making. It is one input in a research process. Everything below is a backward look at what the screen surfaced between April and June 2026, measured honestly, including the parts that do not flatter it.

Three quarters of data now exist. That is enough to say something that was not sayable after one — and the timing matters, because the weekly scanner stops at the end of August. More on that at the bottom.

The full breakdown by week and by month is in the embedded report above. What follows is what the numbers actually mean.

Q2 2026 at a glance

Thirteen weekly scans, April WK1 through June WK4. 130 picks, of which 129 have usable price history. Every 45-day window closed on August 5, so nothing here is a partial measurement.

• Two out of three picks gained 15% or more within 45 days — 86 of 129, a 66.7% hit rate.
• More than a third gained 30% or more — 46 of 129.
• The average pick peaked at +35.7% inside its window.
• The average pick also traded 20.8% underwater at some point inside that same window.
• Held to today, the average pick is +6.5%, and just over half — 66 of 129 — are still green.

Read the first bullet and the fourth together, because that pairing is the whole story. The gains were real and widely available. So was the pain. Both happened to the same basket, often to the same stock.

The consistent part, and the part that isn't

Three completed quarters are now on the record, and one number has barely moved:

• Picks gaining ≥15% within 45 days: 73.5% in Q4 2025, 67.9% in Q1 2026, 66.7% in Q2 2026.

Roughly two out of three, quarter after quarter, across three very different tapes. Q4 2025 ran hot, Q1 2026 was a grind, Q2 2026 recovered. The screen's ability to put a name in front of you before it moves has not meaningfully changed in nine months.

Now the number that moved violently — the return if you simply bought and held every pick:

• +26.6% in Q4 2025, −8.0% in Q1 2026, +6.5% in Q2 2026.

Same screen, same method, three wildly different outcomes. The difference is not skill. It is which quarter you happened to start in.

The stable number is the one that requires you to act. The unstable number is the one that requires you to do nothing. That is not a coincidence, and it is the single most useful thing in this dataset.

A note on Q3 2026: it is deliberately absent. All 60 picks from July and early August are still inside their 45-day windows — the earliest matures in mid-September. Any Q3 figure quoted today is a partial measurement that structurally understates the peak. It will be published when it closes, not before.

The gap between the peak and the finish

The average Q2 pick peaked at +35.7% and sits at +6.5%. That 29-point giveback is not a rounding error, and it is not the scanner being wrong. Catalyst-driven biotech names spike and fade — that is the market structure they trade in.

Where the giveback happened:

• 29 of 129 picks (22%) touched +15% or better and are negative today.
• 15 of 129 (12%) touched +30% or better and are negative today.
• 50 of 129 (39%) drew down 20% or more at some point in the window.
• 26 of 129 (20%) drew down 30% or more.

Four names from the quarter make the point better than any average:

• EDSA (April WK4) peaked at +295% — nearly a quadruple — and sits at +2.5% today. It gave back essentially all of it.
• TVRD (June WK3) more than doubled at +111%, fell as far as −44%, and ended at −31%.
• REPL (May WK1) peaked at +258% and is +298% today. It never stopped. Buy-and-forget was exactly right on this one.
• JAGX (April WK3) never offered more than +11%, dropped −82% inside the window, and is −93% now.

You cannot know in advance which of those four a name will turn out to be. You can decide in advance what you will do when it moves.

The real edge is not the scanner

This is worth being blunt about. The screen's job is narrow: put a name with a dated catalyst in front of you while it is still cheap enough to matter. On that specific job the numbers have held for three quarters. That is where its usefulness ends.

Everything that decides whether those numbers become your numbers happens after the pick appears.

Set the stop before you enter. Nearly 40% of Q2 picks drew down 20% or more; a fifth drew down 30% or more. If you have not decided where you are wrong before you buy, the market will decide for you, and it will do so at the worst available price. JAGX did not require a forecast. It required an exit rule.

Take profit into the spike. Two-thirds of picks offered 15% or more, and over a third offered 30% or more. Those gains were real and they were reachable. A fifth of the basket handed back a 15%+ gain and finished red. Selling a portion into strength is not leaving money on the table — on this dataset, it is the table.

Decide the exit when you decide the entry. For a catalyst trade the exit is structural, not emotional: before the binary print, on the print, or on a defined stretch of the run into it. Pick one and write it down while you are still calm.

Do not buy and forget. This is the behaviour the data punishes most consistently. The hit rate is stable at two-thirds; the hold-forever return swings from +27% to −8% depending on nothing but your start date. Passive exposure to this list converts a repeatable edge into a coin flip.

There is one more pattern worth knowing, because it sharpens all four rules. Sorted by size, the micro caps — $10–50M — produced the highest average peak of any tier in the quarter at +48.9%, and the worst outcome for anyone who held them, at −21.7%. The large caps above $1B produced smaller spikes, +31.4%, and kept nearly all of it, +24.7%. The smaller the name, the more the entire result depended on you selling it.

The scanner finds the setup. Position sizing, the stop, the profit-take and the exit are what turn a 66% hit rate into a result — and all four are yours, not the model's.

Data quality, stated plainly

The scanner has produced data errors in every quarter of its existence, and pretending otherwise would make this whole exercise worthless.

The recurring ones: cash and runway figures scraped wrong in both directions, occasionally by enough to invert a thesis. Catalysts attached to the wrong ticker after an acquisition or spinoff. A currency bug that inflated a foreign issuer's cash balance roughly a thousandfold. Trial phases mislabelled. In July an automated fact-check layer was added to catch exactly these; in August it still waved through a company whose lead trial the FDA had publicly stripped of pivotal status, because the press release led with the word "constructive."

Two honest observations. First, these errors are caught in manual review before publication — every featured name is checked against filings and primary sources, and corrections are printed in the issue itself rather than quietly fixed. Second, and more importantly: the errors have at times affected which names ranked where. A stock has been ranked #1 on a balance-sheet figure that was simply wrong. That is not a footnote. It is the clearest possible argument for treating the output as a starting list rather than a conclusion.

The performance above is what the screen produced including those errors, and it has been consistent anyway. That suggests the edge lives in the catalyst-timing structure rather than in any individual data field being right.

One exclusion this quarter: CING (May WK1) has no usable current price and is left out of the statistics, leaving 129 of 130 picks measured. Two Q2 picks are reverse-split adjusted.

How to actually use this data

1. Treat the list as a research queue, not a buy list. Verify the catalyst, the date and the balance sheet yourself before risking anything.
2. Size for a 20–30% drawdown, because 39% of picks delivered one.
3. Write the stop and the profit-take down before entering. Not after.
4. Tie the exit to the catalyst, not to a price you are hoping for.
5. Expect roughly two in three to work and one in three not to — and make sure the two pay for the one. That is a function of sizing and exits, not of the screen.

What happens next

The weekly AI Catalyst Scanner ends after August 2026. A small number of issues remain, and then the weekly cadence stops. The reason is unglamorous: it costs real money in API and data-feed fees every week, and that cost has been climbing.

Two things continue. Deep-dive pieces and one-off research will keep coming — those were always the part where the analysis mattered more than the screen. And there will be a final Q3 performance review once those windows close in mid-September, so the record gets finished properly rather than trailing off mid-measurement.

Beyond that, the intention is to come back with a better system. Three quarters of honest measurement have made it fairly clear where this one falls short: the weak link is the data layer, not the ranking logic. Cash figures, trial phases and catalyst attribution need to be verified against filings rather than scraped and hoped for, and the automated fact-check bolted on in July proved that a language model reading a press release is no substitute for that. A better version fixes the inputs first.

The main finding in this review points the same direction. Surfacing a name with a dated catalyst before it moves has worked at a stable rate for three quarters. What has never been systematised is the half that actually determines the outcome — the stop, the profit-take, the exit. If there is a next version, that is where the work belongs.

Thank you to everyone who has read along. It has been a genuinely useful thing to build in public, errors included.

Disclaimer: This report is for informational and educational purposes only. It does not constitute investment, financial, or medical advice. Past performance does not guarantee future results. Biotech catalyst trading involves substantial risk of loss, including total loss of capital. Conduct your own due diligence before making any investment decision.