The read
The standing week-ahead regime call is choppy (issued 14 September, expires 21 September) and it is not re-issued today: nothing in this window changed the view, and a call is never edited once written - it runs to its own expiry and is graded as-is.
Underneath it, two seven-day slots refill and the experimental three-day lane opens for the first time since 28 August. Both new reads lean lower, and both are honest about where the lean comes from.
Bitcoin - leans lower over the next seven days, roughly 58% against a base rate near 54%. The coverage leg is well sourced: 38 items, 29 of them directional, from nine distinct accounts across three source kinds, with no single account contributing more than a quarter. It is also genuinely split - an exchange desk flags elevated seven-day implied volatility and heavily call-biased September options interest, an on-chain desk reports profit-taking by long-held cohorts, one analyst frames the week around rate-hike odds, another cautions on a near-term correction ahead of the policy meeting, while one analyst, across two posts, reads the same range as resolving higher. Centred on Bitcoin's own coverage history across 2,238 prior items, that leg nets only barely positive - close to inert. The seven-day price trend is clearly negative and carries the blend even though coverage is weighted two-to-one over trend. So this is a trend-led call, not a coverage-led one, and we would rather say that than dress it up.
Pump.fun - leans lower over the next seven days, roughly 58% against a base rate near 54%. Here the model disagrees with our own corpus. Four items from two independent analyst accounts, three of them directional, lean constructive: one frames fee-funded buyback-and-burn revenue models as the leading alt-coin theme of the cycle, the other reads the recent pullback as a long-standing support area. Centred on the token's own history that leg nets positive. It is outweighed by a firmly negative seven-day trend. The coverage behind it is thin - four items, one source kind - and it points the other way, so treat this as the trend speaking, with our corpus on record against it.
Bitcoin, three days - roughly 56% against a base rate near 51%. Same evidence, shorter clock. The three-day lane is a pre-registered experiment: visible and graded, but excluded from the headline accuracy claim, which counts seven-day calls only.
Receipts
One call matured and resolved since the last run.
| Call | Issued | Matured | Resolved | Outcome |
|---|---|---|---|---|
| Ethereum (ETH) - leans higher over the next 7 days (stated 51.7%) | 8 Sep 01:52Z | 15 Sep 01:52Z | 15 Sep 13:37Z | Correct - close 2,516.84, move +0.69% |
A hit, and a small one: under the current rule a call is graded on the sign of the move alone, so +0.69% counts exactly as a larger move would. The four rows reported in the previous edition (one correct, three incorrect) are unchanged and not re-counted here. Every row in this list was issued by the current engine epoch, so there is no mixed-grading caveat to attach.
Accuracy - the number that reads our probabilities
Trailing 30 days, engine epoch 4, seven-day asset calls only - the pool the public claim is scoped to:
- Brier Skill Score versus the base-rate reference: +0.0144 (our Brier 0.2533 against the base-rate shadow's 0.2570) on n = 21, against the 25 the claim requires. Verdict: insufficient data, and a skill score this close to zero would establish nothing even at n = 25. Our probabilities are, so far, about as good as quoting the historical frequency and ignoring every signal we have.
- Murphy decomposition: reliability 0.0220, resolution 0.0174, uncertainty 0.2449. Reliability is low, which is good - the over-confidence that dominated earlier epochs has come down. Resolution is also low, which is the problem: our probabilities barely discriminate between the calls that land and the calls that do not.
- Calibration error 12.45%, against a 10% bar - still failing.
- A flat forecaster still beats us. Always saying higher scores 0.2361 and always saying 50% scores 0.2500, both better than our 0.2533. That is the honest headline, and it has not moved much in weeks.
- Directional diagnostics, for completeness and not as a skill claim: hit rate 57.14% (95% CI 36.55% to 75.53%), direction-picking skill +0.0671. Neither reads the probability we publish, so neither can see over-confidence; they are not the headline.
- Of the five go/no-go bars, one passes: regime, 60% correct across five resolved reads against 40% for an always-risk-on baseline. Brier, calibration error, edge and sample size all fail.
Per-kind over the same 30 days: asset calls n = 25, 56% correct, average stated probability 59.4%; regime calls n = 5, 60% correct.
What we refused, and why
The slate scored 12 assets. The top two names by signal strength were both refused on evidence quality, not on the model:
- Uniswap ranked first by a wide margin, and every one of its six directional coverage items comes from a single analyst account. Third consecutive run refused. Its underlying subject - a fee-switch vote and a large token burn - is also not fresh news; the source describes an ongoing buyback engine, not a new catalyst. One additional independent outlet is all it needs.
- Convex carries the same shape: five items, one account.
- BNB likewise - three items, one account, one source kind.
Five more names (Ethereum, Solana, Curve, XRP, Avalanche) came back with no directional read at all this run, and two (Dogecoin, Cardano) cleared the strength gate on price trend alone with no corpus to corroborate them, which is not the multi-source read we publish.
The longer-horizon board
Nothing new there, and the reasons differ by slot - worth naming individually rather than lumping them together:
- New Bitcoin record within six months: the scorer did emit a read with a directional edge this run, but the question already carries its maximum of two standing rows. Blocked by the slot ceiling, not by the size of the change.
- Broad altseason within six weeks: emitted on a historical frequency with a flat signal blend, and it sits about one point from the standing read - inside the band that means "leave it running".
- A 20%-plus drawdown in a major: no read at all this run; the scorer's basket fell below its history floor. The standing call is untouched.
- Alt-versus-Bitcoin over 90 days: Solana re-scored with no directional edge, and each of the three covered names already carries two standing rows.
- Event calls: three standing, which is the cap.
The prediction-market leg contributed to none of today's rows - the standing state for short-horizon price calls, and the blend renormalises onto our corpus and price trend as designed.
Window
This edition reads 13 to 15 September: 65 analysed items against 126 in the preceding three-day window. Bitcoin leads coverage with 14 mentions, Ethereum 8 (down 4), Uniswap 4 (up 3), Solana 3 and XRP 3. Market-scope coverage of crypto as a whole is labelled bearish, down 8 mentions; the defi and derivatives sectors lead the sector table, with exchange coverage the biggest riser on the back of two venue wind-down stories.
