The read
Week-ahead regime: choppy. Bitcoin closed the trailing week at 77,392.96 against 80,341.83 seven sessions earlier — a 3.67% decline that leaves it below its 20-day average of 78,518.79. That is a tape under its own trend, but it is not yet a tape in a confirmed slide: the ensemble's fused read on Bitcoin comes back flat at both the three- and seven-day horizons, with bullish coverage (20 items, leaning positive against Bitcoin's own extended coverage history) pulling directly against a bearish seven-day price trend. Two signals cancelling is not a direction.
What did change decisively is breadth. Of the frozen large-cap basket, 2 of the 6 members with priceable history out-returned Bitcoin over the week (33%); in the mid-cap bucket it was 1 of 6 (17%). Last week's call read that breadth the other way and was wrong — see the receipts below. On these numbers there is no rotation engine running: the alts are not leading, they are falling harder. TRX (+1.85%) and Ethereum (+0.01%) were the only large-cap names to finish flat or better; Chainlink (-13.44%) and Zcash (-9.78%) were the worst of the mid bucket.
The macro backdrop in this week's corpus is consistent with a range rather than a resolution: an exchange desk flagged August consumer prices at 3.4% alongside rate-hike odds, core producer prices softened while Treasury yields rose, and one analyst account reads Bitcoin as rejecting its 50-week average with the range intact. A policy meeting sits inside the horizon. Range-bound is the modal outcome we can defend; a second sharp leg down is the outcome we cannot yet distinguish from it.
What we did not call, and why
No asset calls were issued this run. Six of seven seven-day slots are open and the experimental three-day lane is empty, so this is not a capacity constraint — it is a quality one. Three assets cleared the engine's conviction gate and all three were set aside:
- Uniswap scored highest on the slate (a 55% lean against a 43% base rate) on the back of coverage of a real, verifiable mechanism: the fee switch its DAO voted for last December, the 100M token burn that came with that vote, and an on-chain buyback engine now running off swap fees. But all four corpus items behind it come from a single analyst account. One voice reporting a real event is still one voice; our stated edge is multi-source corroboration, and we do not have it here.
- XRP convicted on the downside, but all three of its corpus items are directionally neutral. The bearish coverage vote is an artifact of zero bullish items landing against a 16% historical bullish rate on a sample of three — not evidence. Stripped of that, the read is momentum alone, which the gate itself would reject.
- Ethereum re-scored at roughly 54%, against the 51.7% already standing in call #213 from 8 September. A 2-point move is the same belief restated, not a new forecast. That call stays running and will be graded on its own terms on 15 September.
Zcash, last week's only winner, has fallen out of the corpus entirely (one mention) and now scores on momentum alone — no read. A leaving-it-empty week is the honest outcome when the evidence is thin; the alternative is padding a public record with calls we do not believe.
Receipts
Four calls matured and were graded this run: one correct, three incorrect.
| Call | Issued | Matured | Graded | Outcome |
|---|---|---|---|---|
| Zcash — leans higher over 7 days (71%) | 06 Sep 01:21Z | 13 Sep 01:21Z | 13 Sep 21:16Z | Correct (+5.14%) |
| Solana — leans higher over 7 days (52%) | 06 Sep 01:21Z | 13 Sep 01:21Z | 13 Sep 21:16Z | Incorrect (-1.52%) |
| Week-ahead regime: rotation | 06 Sep 01:21Z | 13 Sep 01:21Z | 13 Sep 21:16Z | Incorrect (Bitcoin -3.71%; large breadth 33%, mid 17%) |
| Bitcoin — leans higher over 7 days (59%) | 03 Sep 23:50Z | 10 Sep 23:50Z | 13 Sep 21:16Z | Incorrect (-5.54%) |
The Bitcoin call matured on 10 September but was not graded until 13 September, when the resolver next ran — three days between maturity and grading, stated here because the two are different events and we report both.
The rotation miss is the instructive one. It was issued on a scan in which 17 of the 20 largest non-Bitcoin assets had outrun Bitcoin over the prior week. Graded a week later on the frozen basket, only 2 of 6 priceable large caps and 1 of 6 mid caps had. Those are different baskets, so the two counts are not a like-for-like reversal, but the call read one week of broad participation as persistent, and it was not.
Accuracy
The headline figure is the Brier Skill Score against the base rate: +0.0047 across 20 resolved seven-day asset calls at the current engine epoch, resolved over the trailing 30 days. That is skill indistinguishable from zero. Our probabilities are, on this sample, worth about what quoting the historical frequency would have been worth — no more.
Three things behind that number, stated plainly:
- We are beaten by both constant strategies. A model that simply said "up" every time scores a Brier of 0.2401 and a flat 50%-on-everything scores 0.2500, against our 0.2543. Lower is better. Both naive benchmarks are ahead of us.
- Calibration is failing its bar. Expected calibration error is 10.7% against a 10% threshold. The Murphy decomposition puts reliability at 0.0210 (lower is better — this is where overconfidence would show) and resolution at 0.0167 (higher is better — this is our ability to tell high-probability situations from low-probability ones, and it is close to nothing).
- The sample is too small to conclude anything. 20 of the 25 resolved calls required before any skill claim is permitted. The verdict from the validator is INSUFFICIENT_DATA and we are not treating a positive sign on a near-zero number as evidence of anything.
One bar passes: the regime read is 3 of 5 correct against an always-risk-on benchmark of 2 of 5. On five observations that is a note, not a track record.
Directional accuracy across those 20 calls is 55% (confidence interval 34% to 74%) — reported here as a diagnostic only. It does not read the probability we publish, so it cannot see overconfidence, and it is never our headline.
Standing book
Active into next week: the Ethereum seven-day call (#213, expires 15 September), three long-horizon macro reads on the policy path, inflation and recession, and six standing structural and relative questions covering a new Bitcoin record, broad altseason, a 20%-plus drawdown in a major, and Ethereum, Solana and XRP against Bitcoin. All six structural slots are occupied: the Bitcoin-record question and the three relative reads are each at their two-call limit, and altseason and the drawdown read re-scored inside their material-change bands this run, so none were refreshed.
