Track record · 21 September 2026
We spent a week deleting our own best numbers. This is what was left standing — including the parts that are not working.
First, what we removed
Every realized return we published had been measured against a stale entry price — a cached snapshot rather than the price the market actually printed. One call was reported as +50.78% when the real move was +21.96%. We rescored the book.
Then we found 27 verdicts produced by a council run in which no seats returned at all. The system had published a confident neutral call built on nothing. Twelve of those had reached the scoreboard; they are excluded now, and the code refuses to publish a verdict with no seats behind it.
Then 72 more, where the published call contradicted the council’s own distribution — a decisive signal sitting on the smaller of two opposing peaks, or a direction declared on 9% agreement. Those are marked contested, which is what they always were.
What is left
Measured at 21 days and beyond, because below that every tier we run is indistinguishable from a coin flip and we would rather say so than quote a number that flatters us.
| Horizon | Calls scored | Directionally right |
|---|---|---|
| T+30 | 604 | 55.1% |
| T+60 | 570 | 54.2% |
| T+90 | 493 | 56.6% |
That is the top tier. Under the corrected rule the edge is modest and roughly flat across horizons — a few points above a coin flip, not the improving curve the first version of this post described. We are publishing it anyway, because a record you cannot lose is not a record.
What is not working
Averaging everything we run gives 52.5%. The gap between that and the top tier is not a rounding detail — it is tiers that are underperforming, and we would rather name them than bury them in an average.
| Tier | Horizon | Calls | Right |
|---|---|---|---|
| Mid tier | T+30 | 305 | 53.4% |
| Broad screen | T+60 | 4,098 | 48.7% |
| Experimental | T+30 | 216 | 36.1% |
The experimental tier is worse than a coin flip and is labelled experimental for exactly that reason. The broad screen decays badly past a month. Both are being reworked. Neither is what a subscriber is paying for, and neither is quietly folded into a headline figure.
Why publish this
Every number above got worse this month — twice. First because we went looking for reasons our own scoreboard might be lying to us and found three. Then, two days after publishing, because we found a fourth: the scoring rule itself. What remains is what survives being graded the way you would grade us.
You can check it. Every call carries the date it was issued, the price it was issued at, the distribution the council produced, and what the market subsequently did. That is the product: not a confident paragraph, but a position you can audit and a ledger that counts the misses.
Policy note, 23 September 2026. Every published signal now carries a frozen record of how it was decided, and the public feed declares which rules are in force and which are only being measured. Two candidate filters (a conviction floor and a large-cap exclusion) are being evaluated in shadow: they change no published signal, and they will not be switched on unless a forward test on calls issued from today qualifies them. Track record figures remain all-history under the current rule; a separate post-policy cohort is reported alongside. Each per-symbol feed entry now shows the analysis price, the scoring baseline, a dated latest price and every horizon outcome, hits and misses alike.
Figures re-derived from the live ledger on 23 September 2026 under the public rule (signal shown vs the sign of the realized return, no flat band), minimum horizon 21 days, sample sizes shown. Past performance is not indicative of future results. Nothing here is investment advice.