This column keeps a public scoreboard of falsifiable claims. Each marker has a bar, a deadline and a state, and we grade them whether they flatter us or not. Today one marker fails on a tenth of a percentage point, one collapses so fast it re-proves a rule we published four days ago, and one turns out to have been near-unfalsifiable from the moment we wrote it — because we had been measuring it through a window we never bothered to widen.

All figures struck at 06:27 UTC on 27 August 2026 unless a different timestamp is named. Bitcoin at press time: $78,667.68 on Binance, down 0.30% on 24 hours, up 25.09% for August, 3.21% below Tuesday’s $81,272.62 high.

A3 — FAIL, by one tenth of a point, on a number that did not move

The bar: July core PCE at or below 3.2% year on year, settling 26 August. The print: 3.3%. Fail.

Graded for design first. When we wrote A3 on 25 August we called it “pointing at FAIL” on a consensus of 3.3%; the next day we downgraded that to a genuine coin flip after finding FactSet carrying 3.2% and the Cleveland Fed nowcast at roughly 3.29%. That correction was right, and the marker was fair: it was live until the release, and it failed by the smallest increment the series reports.

But there is a design flaw underneath the fair grade, and it is worth more than the result. Core PCE was 3.3% in June and 3.3% in July. Headline was 3.7% in both months. We wrote a marker on a threshold beneath an annual rate that had not moved in two months, which means A3 was never really a bet on inflation — it was a bet on whether a rounding boundary would be crossed by revision noise. A marker written against an unchanged series is a coin flip dressed as an opinion. The information was in the monthly figures the whole time, where headline swung from −0.1% to +0.2% and core doubled from 0.1% to 0.2%. Next time the bar goes on the month.

THE FIND — our most-quoted fact about this rally is not a fact about this rally

For four consecutive runs this column has published a version of the following sentence: funding has never once settled above the 0.0100% baseline across this entire advance. Yesterday we put a number on it — “24 consecutive settlements, maximum exactly 0.0100%, not one above baseline across roughly $18,700 of range” — and called it the most durable fact of the move.

We pulled the whole series this morning instead of the default window. Here is what it says.

Binance BTCUSDT perpetual funding, all 500 settlements the public endpoint returns, pulled 27 August 2026 06:27 UTC.

The observation is true and it is much bigger than we said. It is also, for that exact reason, not evidence about August. We had been presenting “funding never rose during this rally” as a statement about the character of this advance — unlevered, spot-driven, no crowded long to unwind. On the full series, funding also never rose during the fall from $82,850 to $57,800, or during July’s chop, or during any other five-month stretch we can see. The premium has been absent from this venue’s perpetual for at least a hundred and sixty-six days across a forty-three per cent round trip in both directions. That is a fact about market structure, not about the rally, and every inference we drew from it about this move needs re-deriving from something else.

Two honest limits on the claim. First, 500 rows is the maximum depth Binance’s public funding endpoint returns; the streak may well be longer, and our number is bounded by data availability rather than by the market. Second, 0.0100% is not a cap — it is the default interest-rate component of Binance’s funding formula. Funding exceeds it only when the perpetual holds a premium to the index at the settlement stamp. So “never above baseline in 500 settlements” translates as: the perpetual has not carried a sustained premium into a single settlement in five and a half months. That is still remarkable. It is just not remarkable about August.

How we got it wrong, one day after publishing the lesson that prevents it

The cause is embarrassing and entirely mechanical: the default page size on that endpoint is 30 rows, and we had been counting the streak inside whatever window happened to be on screen.

Yesterday, in this column, we published the following lesson after catching an unfalsifiable claim in our own draft about the top-trader ratio: “check the window before claiming a series high — a superlative whose window is unstated is not falsifiable.” We applied it to a statistic we were quoting and did not apply it to the statistic we were leading on. Yesterday’s column is here and the sentence is still in it.

B1 — FAIL, and it was a bad marker. The bar was “any funding settlement above 0.0100% by 28 August.” We wrote it as though a breach inside a week were a live question. On a series that has not breached in 500 consecutive settlements, it was not a live question; it was a near-certain fail with the appearance of a bet. B1 is graded FAIL and retired, and we are grading its design as poor in the same breath, because a scoreboard that only reports outcomes and never audits the quality of its own questions is worth nothing.

Replacing it: D1 — any Binance BTCUSDT funding settlement prints above 0.0100% on or before 30 September 2026. Same question, honest horizon, written against the real window rather than a page-size artifact. If you want the mechanics, our guide to funding, open interest and liquidations explains what a premium actually is.

A bull framing published on the exact candle that failed — Bloomberg Television’s crypto segment of 25 August 2026, recorded as bitcoin printed $81,272 and dated so you can grade it:

https://www.youtube.com/watch?v=wMSZRbTyUTQ

C3 — the projection moved twice inside a single run

On 23 August, after the U2 marker failed, we published a rule: a difficulty projection is not a forecast until the epoch is more than 90% complete. Four days later, here is that rule being tested in real time.

That is 2.4473 percentage points of movement in 24 hours, and a further 0.208 percentage points in fifteen minutes — at thirty per cent epoch completion, with 1,411 blocks still to mine to the retarget at height 965,664. The projection is not converging; it is being dragged by hashrate. The three-day average hashrate readings run 880.32, 890.40 and 964.39 EH/s for 25, 26 and 27 August, against a current difficulty of 125.807T after the previous retarget cut 1.3122%.

C3 — next adjustment cuts more than 2.00% — is now failing badly, at −0.63% against a −2.00% bar, and we wrote it two days ago off a −3.08% projection at fifteen per cent completion. That was a marker built on exactly the kind of number our own rule says is not yet information. It stands, it will be graded, and the lesson is that publishing a rule does not immunise you against breaking it. The mechanics are in our difficulty guide.

To make the successor falsifiable in both directions rather than only ours: D2 — the retarget at height 965,664 lands between −1.50% and 0.00% inclusive. That is the range the current trajectory implies, so it can fail on either side.

The cohorts turned, both of them, on the same reading

Binance BTCUSDT, daily readings, 30-row window — the deepest the venue publishes.

Yesterday both series printed a 30-day extreme: the size accounts at 2.2560 and the spread between the cohorts at 2.254×. Today both reversed. The top-trader ratio fell 4.87% to 2.1462 — still elevated, still the second-highest reading in the window, but off the top. The crowd rose 8.50% to 1.0859 (52.06% long, 47.94% short), going net long again after printing dead flat at 0.5002/0.4998 yesterday. The spread narrowed to 1.976×, the first contraction after eight widenings in nine readings.

C2 — top-trader ratio holds at or above 2.0000 at every reading through 28 August 00:00 UTC — is still passing, with one reading left and a 7.3% cushion. Note the direction of travel though: the size money trimmed into a fall, having added into one. That is the reverse of the pattern we praised them for on Wednesday, and we are printing it in the same table.

Open interest: second consecutive fall, and the two OI numbers disagree again

Coin open interest fell 0.94% on the day and 1.51% over two sessions, wiping out Tuesday’s build. Dollar open interest fell only 0.32% over the same day, and is up 18.65% against the 16 August reading while coin open interest is down 5.29% against it. Same positions, two stories, and the difference is entirely price. We have now flagged this divergence on nine separate runs; it is the single most reliably misreported statistic in this market.

Z2 — coin open interest back at or above 111,988 by 28 August — is failing by 5.71%, with one session left. Z2 is our own falsification marker, written to catch us if the “unlevered rally” framing was wrong. It has not come close, which is the reading that does not embarrass us, and we say so every time precisely so that the day it does turn is legible. The live reading at press time is 106,480.

C1 — and a marker whose bar was badly chosen

C1 asks whether bitcoin trades above$79,563.71 during a US cash session, 13:30–20:00 UTC, by 28 August. It tests a time zone rather than a level: that figure was the highest print any American session reached during Tuesday’s $81,272 spike, which happened at 10 p.m. Eastern and was never revisited in New York hours.

Two things. C1 is failing, 1.15% away with one session to run. And the bar was poorly set: because we chose the level by taking the best US-session print, the marker can only pass on a strict clear of a figure the tape has already touched exactly once. It is not unfalsifiable like B1 was, but it is asymmetric, and we should have set it at a round number or a percentage above. Noted, and the question — does American liquidity ever confirm the overnight high? — remains the right one. The pattern it came from is in field guide #32.

X2 — fourth consecutive day grading our own premise wrong, now with a Fed policymaker on the other side

Brent for October delivery: $87.12 at 07:04 BST, down 0.82% on the day, against an X2 bar of $95.00 settling tomorrow. Failing by 8.29%, and the gap has widened every day since 23 August, when it was 0.65%.

What makes today different is who else holds the premise. Boston Fed President Susan Collins, in a written essay published 25 August, named the continued closure of the Strait of Hormuz alongside oil inventories “at historically low levels” as her principal upside risk to inflation. That is X2’s thesis, held by a sitting policymaker, in writing, while the price falls. We are not claiming she is wrong — inventory stress and spot price can diverge for a long time, which is arguably the whole point — but we have now been wrong about this for four days and are recording that the error is widely shared rather than idiosyncratic. Y1 — 30-year Treasury yield closing at or above 5.35% by tomorrow — shares the premise and is also failing, at 5.174%, 17.6 basis points short.

The board

B2 deserves a line of its own because it has quietly become the most likely pass on the board. August spot bitcoin ETF flows stand at $3,282.0m across 18 sessions. The bar is October 2025’s $3,424.9m, which needs $142.9m over the three remaining August sessions — $47.6m each — against an eight-session running average of $350.24m. It would take a collapse to fail. We upgraded it from “coin flip” to “favoured” yesterday and it has strengthened again since. Details in our flow-data guide.

A technical read on the same tape from 23 August 2026, before the $81,272 print and the failure that followed — dated so you can grade it against what happened:

https://www.youtube.com/watch?v=L1sxga_Vhig

What today actually taught us

Three things, in ascending order of usefulness. The smallest: a marker written against an annual rate that has not moved in two months is not a forecast. The middle: publishing a rule about difficulty projections does not stop you writing a marker off one four days later, and the projection moved a fifth of a percentage point while we were writing this sentence.

The largest, and the one we would rather not have found: the statistic we were proudest of had a window we never checked. Twenty-four settlements was never the number. Five hundred was, and it is a different claim entirely — better as an observation, useless as evidence about August. The rule we published yesterday would have caught it on any statistic except our own. From today, every streak, series high, series low and “consecutive” count in this column ships with the depth of the pull that produced it, and the pull goes to the endpoint’s maximum before the sentence gets written.

One footnote, filed against ourselves. The independent recompute we run over every derived figure before publication caught a slip in this very article: the dollar open-interest comparison against 16 August had been struck off a fifteen-row pull whose first row was actually 13 August, and printed +18.34% where the correct figure is +18.65%. The same class of window error, in the same run, inside the draft whose lead is a window error. Caught and corrected before publication, and recorded here because a column that lectures about pull depth and then gets it wrong twice in one morning is obliged to say so.

Investment disclaimer. This article is journalism, not investment advice. Bitcoin and other digital assets are volatile and you can lose the entire amount you put in. Nothing here is a recommendation to buy, sell or hold any asset. Figures are as of the timestamps stated and change continuously. Do your own research and consider taking advice from a licensed professional before making any financial decision.