Research period August 2026 · James Kennedy, founder

Research notes

We looked at 57 races and 41 got no read

Fifty seven races went through the read on a single race day. Sixteen qualified and were scored. Forty one were published as No Read. Exactly one race reached full convergence.

The numbers, without decoration

57 assessed. 16 scored. 41 refused. 1 converged.

I want to leave those sitting there for a moment before explaining any of them, because the explanation is usually where a racing company starts smuggling comfort back in.

Read them cold and they look like a bad day. A full card went past and the clear majority of it produced no output whatsoever. If you were paying by the race you would feel short-changed. If you ran a tipping line you would not survive publishing that ratio twice.

It is the most honest day of work we have put on record.

This was a development day, and I want to be plain about it

Nobody paid for those 57 races. No customer received a selection. Not one of the 16 that scored was issued to anyone as a tip, and the single race that converged was not sent anywhere either.

I am saying that first rather than last because the number is the interesting part and I do not want it quietly doing work it has not earned. A refusal rate from a development run is evidence about a method. It is not a track record and it is not a performance claim.

What the day actually was: the whole stack run end to end against a real card, under conditions fixed in advance, with nobody nudging it. That is the only way to find out what a method does when it is not being helped. You are not testing whether it can produce something impressive. You are testing what it does on an ordinary Saturday when large parts of the card are simply not legible.

Refusing is a decision, not an absence of one

There is a lazy version of No Read that is really just a shrug. We are not sure, so we will say nothing, so nobody can hold us to anything. That version is worthless and it deserves the suspicion it gets.

The version we run is a set of conditions that either hold or do not. Is the field complete and is every runner correctly identified. Do we actually hold run style history for these horses, or are we looking at blank cells and quietly treating blanks as zeroes. Do the sources describe the same race, or have we ended up with two decent accounts of two different contests. Is the shape of the thing legible at all, or is this one of those races that could be run three ways with a different horse suited each time.

Every one of those is answerable before the stalls open. None is a mood.

And the timing is the whole point. The gate conditions were written down before the card, not after it. The score arithmetic was applied afterwards, while the score itself was still being validated, so this is a test of what the gates do on a real day rather than a prospective record. Anybody can decide in hindsight that a race was unreadable once it has embarrassed them. Deciding in advance is harder and it is the only version that means anything.

What forcing a view would have cost

Imagine we had required an output on all 57. It would have been trivial. Every one of those races has runners with numbers attached, and numbers can always be ranked. You will never fail to produce a top-rated horse. The machinery does not stop working when the evidence runs out, which is exactly the problem.

So we would have published 57 reads. Sixteen would have been what they appear to be. The other 41 would have been the same typeface, the same layout, the same confident little ranking, generated out of gaps.

The damage is not really the 41 bad reads. It is what they do to the 16 good ones. Once manufactured output and genuine output look identical on the page, nobody can tell them apart, including us. You have not added 41 opinions, you have subtracted the meaning of the other sixteen.

An instrument that fires on every input is not measuring the input. It is reporting that it ran.

What the sixteen had in common

Not quality. This is the part people expect me to get wrong, so let me get it right early. The sixteen were not the best races on the card, or the most likely winners, or the ones where we fancied something.

They were the readable ones.

Conceptually, a race qualified when the basic facts about it held together. The population was complete and identified, so we were looking at the race that was actually going to be run rather than a stale declaration. The runners had enough history for their preferred way of racing to be a fact rather than a guess. The independent descriptions of the contest pointed at the same underlying shape instead of quietly disagreeing. And the race had a shape you could state in one sentence without hedging it three times.

That is a much lower bar than good. It is closer to the bar a scientist sets before agreeing that an experiment is worth running. Legible first. Interesting afterwards, if at all.

The one that went the whole way

Sixteen races cleared the bar to be scored. One reached full convergence, which is the strictest state the read has.

Convergence means the separate lines of evidence, which are perfectly capable of disagreeing and frequently do, ended up describing the same thing about the same race. That is rare by construction. It should be rare. A method where everything converges has not built independent lines of evidence, it has built one line and copied it out four times in different fonts.

And it is worth being precise about what that state does not claim. It is not a probability. It is not confidence. It is certainly not a price opinion. A converged race where the horse gets beaten has not failed, because nothing in the convergence ever said the horse would win. It said our evidence agreed. Those are different sentences and the gap between them is where most racing products live.

Why 41 refusals is not the embarrassment it looks like

Because of what we are trying to add to.

Across 60,671 races with one clear favourite, the favourite won 35% of the time and did not win 65%. Set against what those prices were implying, that is close to what the market said would happen. The market is not a soft target. It is thousands of capable people, with money at stake, being publicly wrong for a living and correcting quickly.

If that is true, then most races on most days genuinely contain nothing we can usefully add. Which means a high refusal rate is not the method failing. It is the method agreeing with something we already know about racing and declining to pretend otherwise. A stack that produced strong reads on 50 races out of 57 would be telling me about its own design, not about the card.

The harder version is the one worth building

I know what silence costs commercially. Volume sells. Restraint does not. Every instinct in this industry pushes towards covering more races, filling more of the day, giving people a reason to open the app at two o'clock and again at half past.

Nobody has ever renewed a subscription because of the races they were not told about.

But the reframe is straightforward once you say it out loud. You are not paying for how many reads arrive. You are paying for the difference between the races that get one and the races that do not. If every race gets a read, that difference is zero, and you are buying nothing at a price that looks reasonable because there is so much of it.

41 out of 57 is what the difference looks like when it is real. I would rather publish that and explain it than publish 57 confident-looking rankings and hope nobody asks which ones came out of gaps.

Related research

Elite Pass publishes research, not guaranteed outcomes. Findings are measured against the market and remain subject to replication and sample size. The evidence room and the public ledger carry the full record.