The useful number is 2,841

Nineteen days ago a machine I built started trading on its own. This morning the site says it stands at 1.102 times what I would have had by buying Bitcoin and going to bed.

Nineteen days is weather. If you take one number off that page, take the other one: 2,841.

The denominator

That is how many ideas have been written down, predicted, tested and filed. Forty-seven batches, 14,699 runs, 65 verdicts since the machine went live. Most of the verdicts say no.

Bacon had the problem in 1620, long before anyone could run a backtest: the mind, having taken a position, gathers up whatever agrees with it and walks past everything that does not. Give that habit a laptop and it becomes a method. Test three thousand variations against Bitcoin and one of them will beat it on luck alone, near enough guaranteed, and you do not have to cheat to end up there. You only have to keep the winner and lose interest in the rest.

There is a statistic that undoes this, the deflated Sharpe ratio, and it discounts a result by how many times you went looking for it. It needs one input that nobody volunteers: the true count of attempts. Not the ones you liked. All of them.

So every trial stays in the ledger and the count only climbs. The 2,841st makes the 2,842nd harder to pass, including for me, permanently, by design. And you have no way to check that my denominator is real unless the failures are published with the same weight as everything else, which is why the dull batches go out at all.

Medicine worked this out decades ago and made it law. You register the trial, with its endpoints, before you run it, because the alternative was pharmaceutical companies running twelve studies and publishing the one that came out well. Crypto has never had to file anything. It will show you a backtest on a Tuesday and has never once been asked how many it ran first.

Batch 25, which found nothing

Early in September a chart was going round: cyclical timing models, Bitcoin about to fall off a cliff. The chart had no rule in it, a Fibonacci extension off three old tops with three log overlays fitted to match, so there was nothing to test and nothing to score. The question underneath was fair, though, and the ledger held no cycle model at all.

Ten lines were written and stamped overnight on 6 September. Halving calendars. Pi Cycle. The two-year moving-average multiplier at two settings. A power law fitted only on pre-ETF closes. A seasonality rule that sits out any calendar month with a negative historical mean. A 200-day switch as the control. Daily closes back to 2014, signal on the previous close, traded at the next, fees in.

The windows are fixed and they do not move between batches. January 2024 to August 2025 is the only window anything is ranked on. The last twelve months are read as pass or fail and never contribute to a ranking, because a holdout you are allowed to rank on has quietly become a second fit window. Older history, back to 2020 and to 2015, can kill a line and can never promote one.

The seasonality rule came first on the ranking window at 1.092. Measured from 2015 it is worth 0.24, and about a fifth of that went to fees. The known-bad line topped the ranking, which is roughly what a fit window is worth on its own.

Pi Cycle has the only long story. Three and a half times Bitcoin since 2015, earned on four trades, two events in eleven years. Against the 301 trials the ledger held that week its deflated Sharpe comes out at 0.24, below what noise would be expected to produce over that many attempts. Two good dates is not a rule.

Five of the eight cycle lines never fired at all after the ETF. They are hold-BTC with a switch that has never moved.

Of about eight predictions written before the run, four were wrong, and all four missed the same way: these models do less than I assumed. Nothing was promoted. The ten lines sit in the ledger with a full window set, so the next time that chart comes round the comparison is already dated and stamped.

Surviving is not the same as getting money

A line that clears everything still does not trade. It gets a shadow book, replayed every night through the same engine, on the prices the live system actually had at each decision, at the fee rate the exchange actually charges. Ninety days minimum, and it has to win on multiple and on drawdown, not one of the two. Twelve run every night. Several of them are the live machine with one thing changed and are doing worse than it, which is the cheapest way I know to learn what a change costs.

There is one language model in the system and it never trades. Its job is to listen. The voices out there put out calls all day: traders sharing an idea on a stream, analysts with a following, a few closed channels where things get passed around before they get said out loud. Almost none of it arrives as anything a machine can use. The model turns that prose into structured fields, what was claimed, on what, by when and what would count as wrong. Those fields carry no weight until the person behind them has a record, scored privately against what their earlier calls actually did.

I am doing that for a reason that has little to do with signals. Almost everyone talking has something else to sell: a paid group, an exchange referral code or a follower count big enough that the platform pays the rent. None of that makes them wrong. It does mean being right and being watched are two different jobs, and only one of them reliably pays. Somewhere in that noise there may be one or two people who are in it for the teaching, and I would like to know which ones. There is no way to find that out by listening harder. You write down what they said, you wait for the horizon they set themselves, and you score it. It is the same test I run on my own ideas, and it seemed only fair to point it at theirs.

Where I failed the method

Each hypothesis file, verdict and daily snapshot is hashed, and the hash goes into Bitcoin through OpenTimestamps. 190 are confirmed in a block. I do not run a chain and there is no token; the calendars are public and free, and anyone can check a proof against a Bitcoin node without asking my permission.

Two things have gone wrong so far and neither of them was the theory. In both cases the rule was right and the practice underneath it was not, which is the more ordinary way these things break.

Batch 25’s own file was stamped at 23:39 UTC. The run had finished at 23:34. Five minutes. The predictions were written before the run and the file says so, but the Bitcoin proof does not precede the outcome, and preceding the outcome is the entire job of the stamp. It was the first batch written after the ledger existed and the stamp came at the end of the session instead of the start. From batch 26 the file is stamped before it runs, and any row whose proof lands after its run carries a mark saying so, for as long as the ledger exists.

Then there was the week the shadow books were not what I thought they were. They were meant to be the live machine with one change each. They were running on the engine’s default universe, on their own paper base, and the invariant built to keep the twin honest was comparing raw balances, so it went blind the moment the live capital moved on 2 September. Building the public scoreboard is what exposed it, after a week of running had not, and reading the code the next day turned up more: the books had been built on an older configuration of the brain and each had been told it held what the live book held. All of them were restarted on 7 September from the live holdings with the live parameters. The invariant compares multiples now and read zero to six decimals on its first night. The pre-reset week is archived and the scoreboard says so above the table.

A private research process fixes that quietly on a Sunday and nobody is any the wiser. A public one has to stand there and say it, which is most of the reason for making it public.

The next chart that goes round will meet ten dated lines that already say what it is worth. That is all this is for.

The record itself lives at zenko.report: the scoreboard, every shadow book, the full ledger of verdicts and the proofs, rebuilt each day from the ledgers rather than typed. The method page is where the rules I keep referring to are written down. Nothing here is investment advice, and the system takes no outside money.