Bitcoin spent the last five days doing the opposite of what this window's coverage would suggest. A 0.25% policy-rate rise landed on 16 September, the US Digital Asset Market Clarity Act failed a Senate cloture vote 49-50 the same week, and the tape went up anyway: 77,278.73 to 81,360.00, a 5.28% seven-day gain that puts price back above its 20-day average of 78,347.75.
That is enough to move the objective regime label. Last week's read was choppy - a tape below trend with no confirmed direction. This week it is risk-on on the same test.
Breadth is narrow, not rotational
The move is not broad. Measured over the same seven days against Bitcoin's own return, 2 of 6 priceable large-cap alts out-returned it, against 4 of 5 priceable mid caps - a mixed structure, not a confirmed rotation. The largest single move was ZEC (+31.9%), but three other mid caps also out-returned Bitcoin. Both figures come from the same frozen basket over the same window; they are two buckets of one measurement, not two separate measurements.
What we issued
Three asset calls and two structural refreshes. All three asset calls lean higher, and all three are trend-led rather than coverage-led - worth stating plainly, because our corpus is the part we claim as a moat and it is doing almost no work this week:
- Ethereum, 7 days, higher, ~55% (base rate ~47%). Nineteen coverage items, 16 directional, nine distinct accounts across three source kinds, no account past a quarter of them - a genuinely diverse set. But centred on Ethereum's own 798-item coverage history that leg contributes +0.02 against a momentum lean of +0.77. The price trend is carrying it.
- Solana, 7 days, higher, ~50% (base rate ~44%). Ten items, 7 directional, three accounts, and one analyst supplies over half the directional items. Centred on Solana's own 382-item history the coverage leg reads slightly negative - it points marginally against our own call. Momentum carries the blend entirely.
- Bitcoin, 3 days, higher, ~56% (base rate ~49%, experimental horizon, excluded from our public accuracy claim). This direction is opposite to our standing seven-day lower call from 15 September. That call still runs to 22 September and will be graded exactly as issued - we do not edit, retract or net calls against each other. The three-day read is an independent forecast on its own window.
We scored the full slate and refused more than we issued. CVX and UNI both cleared the conviction gate on score - UNI has now topped the slate four runs running - and both were refused for the same reason: every directional item in each one's coverage traces to a single account. XRP cleared the gate and was refused on that same test. AVAX, ADA, BNB and DOGE ranked high on momentum alone with no corpus coverage behind them and did not convict.
On the structural board, two slots refreshed and two did not, for different reasons. Altseason moved from an implied ~18% to ~30% for the breadth condition inside six weeks - past the 10-point band that defines a material change - so a fresh read was appended alongside the existing one. The majors-drawdown question now scores under a revised definition (four of eleven majors down 40% inside 90 days, where the standing row was written as 20% inside six months), so the new read was appended under its own terms while the older row keeps the terms it was written under. The Bitcoin-record question is at its two-row ceiling and could not take a refresh regardless of what it scored. The relative-coin slots for Ethereum, Solana and XRP sit at that same ceiling.
Receipts
Five calls resolved this run: two correct, three incorrect.
Four of the five are 90-day relative calls issued in June under engine epoch 1 - an earlier forecaster with a different scoring rule. They are marked below and are not pooled with the current epoch-4 accuracy figures.
| Call | Issued | Resolved | Outcome |
|---|---|---|---|
| Solana's odds of beating Bitcoin over 90 days have risen above even (p=0.613) - prior engine epoch | 2026-06-21 | 2026-09-19 | Correct - SOL +53.7% vs BTC +28.5% |
| Solana's odds of beating Bitcoin sit above baseline over 90 days (p=0.290) - prior engine epoch | 2026-06-19 | 2026-09-19 | Correct - SOL +45.7% vs BTC +20.3% |
| Ethereum leans to underperform Bitcoin over 90 days (p=0.724) - prior engine epoch | 2026-06-19 | 2026-09-19 | Incorrect - ETH +43.0% vs BTC +20.3% |
| XRP's odds of beating Bitcoin over 90 days sit above baseline (p=0.392) - prior engine epoch | 2026-06-21 | 2026-09-19 | Incorrect - XRP +26.8% vs BTC +28.5% |
| Bitcoin leans lower over the next 3 days (p=0.557, experimental horizon) | 2026-09-15 | 2026-09-19 | Incorrect - close at expiry 79,745.50, +4.57% |
The three-day Bitcoin call matured on 18 September and the two Ethereum and Solana relative calls matured on 17 September; all five were graded on 19 September. We report the gap between maturity and grading rather than collapsing the two dates into one.
One further item has not resolved but is worth flagging now rather than at year-end: our standing macro call that the policy rate is left unchanged for the remainder of 2026, issued 30 July at roughly 89%, is displayed with that wording, and the rate rose 0.25% on 16 September. The condition the row actually resolves against is narrower - no rate cut through 2026 - so the rise does not decide it. It resolves at year-end and stands untouched until then.
Accuracy
Trailing 30 days, engine epoch 4, seven-day asset calls only - the claim pool is 16 resolved calls against a 25-call minimum, so the verdict remains INSUFFICIENT_DATA and nothing below is an established result.
Our headline measure is the Brier skill score against a base-rate reference - the only figure on this page that actually reads the probability we publish. It is +0.086 (our Brier 0.2355 against 0.2576 for simply quoting the historical base rate and ignoring every signal we have). Positive, but not established, and three things sit against it:
- Reliability 0.031 against resolution 0.016 in the Murphy decomposition. Reliability is where over-confidence shows; resolution is whether our probabilities discriminate at all. Ours discriminate weakly, and our calibration error costs us more than our discrimination earns.
- An expected calibration error of 0.174 against a 10% bar. That fails.
- A forecaster that answered higher on every one of those 16 calls would have scored a Brier of 0.2226, better than our 0.2355. We do not beat that constant, and we say so.
Directional accuracy over the same pool was 11 of 16. That is a direction-picking statistic: it is blind to the probability we attach, so it cannot see over-confidence, and it is not our headline.
The honest summary is that the probabilities we publish are edging ahead of a base-rate shadow on a pool too small to mean anything, while failing a calibration bar and losing to a forecaster that just answers up every time.
