At approximately 00:47 UTC this morning, at block 963,648, bitcoin’s mining difficulty retargeted from 127.48 trillion to 125.81 trillion. That is a cut of 1.312%. This desk had a public marker on that print — U2, written on 17 August, requiring a settled retarget of zero or better. It failed, and it failed by 1.31 percentage points.

We are going to spend the first half of this piece on why, because the failure is more instructive than a pass would have been, and because the post-mortem has been deferred twice and is now overdue. Then we will get to the thing in this weekend’s derivatives data that we did not have yesterday, which is that the largest accounts on Binance have been adding into the pullback while the retail crowd sat flat, and the position ratio has just crossed 2:1 for the first time in the series.

U2: FAIL. Settled −1.312% against a bar of ≥ 0.00%

The marker was deliberately uncomfortable. On 17 August, with hashrate collapsing and mid-epoch estimates deeply negative, a projection of a flat-or-positive retarget looked like a stretch. Then it stopped looking like a stretch. Then it looked like a stretch again. Here is the entire published series, one 06:10 UTC snapshot of mempool.space’s difficulty-adjustment endpoint per day, against the settled print.

The post-mortem: what a mid-epoch difficulty projection is actually worth

Four findings, and the first one is the uncomfortable one.

One. The projection was wrong in sign for two consecutive days, at the exact midpoint of the epoch. On 19 and 20 August the estimate read +0.49% and +0.37%. The settled print was −1.31%. A trader who had taken the mid-epoch reading as information would have been positioned against a 1.8-point error, and would have had a full day to feel confident about it before the reading flipped back. This desk published both readings and, on 20 August, wrote that U2 “now points at a pass.” It did not. We were reporting the endpoint honestly and the endpoint was honestly useless.

Two. The error collapses non-linearly, and almost all of the collapse happens in the last quarter. Errors by snapshot: 0.32, 1.03, 1.80, 1.68, 0.37, 0.11, 0.04 points. The projection was not gradually converging. It was noise until roughly the sixth day and then it snapped onto the answer. By snapshot six, with about 200 blocks left, the estimate was within 11 basis points. By snapshot seven, within 4. There is a usable signal in a difficulty projection, and it exists for roughly the final 10% of the epoch.

Three. The third-party estimators were wrong too, and wrong in the same direction. CoinWarz was carrying approximately 126.77T and −0.55% for this retarget. The print was 125.81T and −1.31%, more than twice the cut. This is not a defect in one endpoint. It is what happens when you extrapolate an average block interval from a hashrate series that moved 23.7% peak to trough inside three days.

Four. The volatility in the estimate was the story, not the level. Hashrate ran 1,025.3, 1,000.2, 941.1, 782.1, 812.7, 863.6 and 945.9 EH/s on three-day averages across the epoch — a 23.7% collapse and then most of a recovery, inside one retarget window. The projection dutifully tracked each leg. Anyone reading the projection was reading a lagging, amplified restatement of hashrate they could have looked at directly.

The operational rule we are adopting: a difficulty projection is not a forecast until the epoch is more than 90% complete. Before that it is a hashrate chart with extra steps and a misleading number of decimal places. We will keep publishing the series, because publishing a bad estimate and then grading it is the only way anyone learns what it is worth — but we will label it.

And a note on our own marker design, in the spirit of grading the board honestly. Yesterday we graded Y2 a PASS and said loudly that it was a worthless pass, because the tape had already cleared the bar by a quarter when it was written. U2 is the opposite case and the better marker: it was genuinely uncertain, it flipped twice, it was gradeable, and it failed. A marker board that only reports passes is a marketing document.

The find: the crowd is flat, the size accounts just crossed 2:1, and they added into the drawdown

Now to the part that is new since yesterday. Binance publishes two long/short ratios that are routinely quoted as if they were the same statistic. The global long/short account ratio counts accounts: one wallet, one vote, regardless of size. The top-trader long/short position ratio weighs the actual notional held by the largest accounts on the venue. This week they have been telling opposite stories, and this weekend the gap widened again.

Read the two right-hand columns together. Over eight days the account ratio fell from 2.22 to 1.01 — the crowd went from better than two-to-one long to dead flat, and it did that as price rose $15,400. Over the same eight days the position ratio went from 1.46 to 2.0533, rising on five of the last six readings and hitting the highest print of the series this morning. It has now crossed 2:1 for the first time in the window.

The detail that is genuinely new today: the last two readings, 1.9597 and 2.0533, were both taken during the drawdown. Bitcoin has fallen 4.19% from Friday’s high across Saturday and Sunday. Through that fall, the largest accounts on the venue increased their net long exposure twice. The crowd, over the same two days, went 0.9554 → 1.0387 → 1.0072 — statistically flat. Retail did not buy this dip and it did not sell it. Size bought it.

We will state the limits plainly, because this is one statistic on one venue. It is Binance only. “Top trader” is Binance’s own definition of its largest accounts, not an audited category. Position ratios can rise because longs were added or because shorts were closed, and the endpoint does not distinguish. And a 2:1 ratio among the largest accounts on one exchange is not evidence about the whole market. What it does rule out is the tidiest bear story available this morning, which is that the rally was a retail squeeze that is now unwinding as retail leaves. Retail already left. Price is 4% off the high anyway. Somebody else is holding the long side.

The rally still has no leverage in it, and the unwind does not either

This has been the central observation of the week and it survived the pullback intact. Coin-denominated open interest on Binance futures peaked at 111,988 BTC on 15 August, fell to a series low of 105,618 on Saturday, and sits at 107,009 BTC this morning. Dollar-denominated open interest over the same window went from $7.06 billion to $8.16 billion. Same market, same days: coin OI −4.4%, dollar OI +15.6%. Every outlet that wrote “open interest surged” quoted a number that price inflated for them.

Funding is the confirmation. Across the twelve settlements from 08:00 UTC on 19 August to 00:00 UTC today — a window containing a $14,775 rally and a $3,954 give-back — the funding rate on Binance BTCUSDT perpetuals never printed above the 0.0100% baseline once. Five of the last five settlements printed exactly 0.0100%. There has been no premium for holding a long. A squeeze of the size the wires described should have dragged funding to multiples of baseline for days. It did not move at all.

There is a nuance in that OI number worth flagging because it cuts against us. Coin OI ticked up overnight, 105,815 at 00:00 UTC to 107,009 live. That is a 1.1% increase, and it is the first meaningful build in a week. It is also the direction that would eventually falsify this thesis, so we note it in the same paragraph in which we note that the thesis is intact.

Marker board, 23 August 2026

One correction to yesterday’s framing of Y1. We described the 30-year as “about 7 basis points away” from the 5.35% bar. That is arithmetically true and analytically lazy. The 30-year’s 2026 high is 5.34%, set on 18 August. Y1 does not require a small move; it requires a new high for the year, in four sessions, against explicit Treasury policy aimed at pushing the long end the other way. That is a much harder bar than “7bp” makes it sound, and we should have written it that way.

On Z1: the cushion has fallen from 5.9% to 4.31% in twenty-four hours. Bitcoin would have to lose a further $3,147 — 4.1% — before tonight’s close for the marker to fail. That is unlikely but not absurd: a 4%-or-wider intraday range has happened on three days this month — 19, 20 and 21 August, at 9.01%, 6.49% and 8.86% — against a next-widest reading of 2.96% across the other twenty days. Sunday’s range so far is 2.41%.

The bear case, stated properly

The strongest argument against everything above is not that the leverage data is wrong. It is that the leverage data is answering the wrong question. MEXC’s Shawn Young put it as crypto “giving the Treasury’s intervention far more credit than it deserves,” and named his invalidation levels: the 10-year above 4.7% and the 30-year above 5.3%. As of Friday’s close the 10-year is at 4.736% — already through the first level — and the 30-year at 5.276%, two and a half basis points from the second.

In other words, the macro trigger that produced this rally on 19 August was the Treasury doubling its long-end buyback capacity to suppress yields. Yields at the long end are now back within a rounding error of where they were when the intervention was announced. If the buyback is the whole story, the story has already been retold and the price has not yet noticed. The unleveraged character of the rally makes it more durable than a squeeze; it does not make the premise correct.

What settles next

Tonight, Z1 and the weekly close, which also decides whether this was the best week since February 2024 or since March 2023 — the boundary is a close at $76,744.29. Monday, Strategy’s 8-K, Bessent at 2:00pm ET, the first ETF row since the high, and the Z3 window opens. Wednesday, July PCE. Thursday to Saturday, Jackson Hole, with X2 and Y1 both settling Friday alongside Warsh’s first keynote. We will grade all of them, including the ones we get wrong.