Today's read
No new forecasts were issued today, and the reason is worth setting out plainly rather than leaving as "nothing convicted".
The week-ahead regime call from 19 September - risk-on - already stands and runs to 26 September. Re-measured this evening on the same objective test (Bitcoin's last closed daily bar 81,178.00 against a 20-day average of 78,472.08, a seven-day return of +5.64%), the label comes back risk-on again. An unchanged read is left running rather than restated, so no regime row was written.
Two of the seven weekly asset-call slots are free. Neither could be filled. Which rule stopped each candidate, in order of signal strength on today's scan:
- Convex (CVX) - the strongest read on the board today, and refused. All three of its directional coverage items come from a single account. A one-source lean is not the cross-referenced evidence this engine is supposed to require.
- Bitcoin (BTC) - a conviction passer, blocked by the two-standing-calls-per-coin ceiling (a seven-day call from 15 September plus a three-day call from 19 September) and by the bar on a second call on the same coin and horizon. The seven-day read has flipped direction since 15 September and the move is material. Under append-only rules the 15 September call is left untouched and graded exactly as it was written; the changed view is recorded for the next run, not published as a second row. It is not a correction and not an offset against the older call.
- Zcash (ZEC), Ethereum (ETH), Solana (SOL) - conviction passers whose standing seven-day calls have not moved materially. Each sits inside the band we treat as "the same belief restated", so all three are left running.
- XRP and Curve (CRV) - refused on the same single-source defect as Convex. Curve is the clearest case in the window: seven directional items, every one from one analyst account. Coverage volume is not coverage diversity, and the engine's conviction gate cannot currently tell them apart.
- Avalanche (AVAX), Dogecoin (DOGE), BNB, Cardano (ADA) - price trend only, no usable coverage leg, so the multi-signal gate fails. Not refusals, just no read.
- BLUECHIP and HYPE - no direction at all.
The three-day experimental lane is full (three of three). All six standing structural and relative question slots sit at their two-per-slot ceiling, so none of today's structural reads could be written either - including the Bitcoin-record question, where the current engine's answer is far from the number the board still displays. That displayed row was written under an earlier engine and cannot expire until January; it is logged as a standing defect rather than narrated as a figure here.
Receipts
Three rows resolved since the last run, all stamped 2026-09-21 20:12Z. Misses are listed with the hit, as always.
| Call | Issued | Matured | Resolved | Outcome |
|---|---|---|---|---|
| Week-ahead regime: choppy (#214) | 14 Sep 12:24Z | 21 Sep 12:24Z | 21 Sep 20:12Z | Miss - Bitcoin gained 11.01% over the window |
| Solana to beat Bitcoin over 90 days (#66) | 23 Jun 15:20Z | 21 Sep 15:20Z | 21 Sep 20:12Z | Hit - Solana +70.0% against Bitcoin +38.4% |
| Ethereum to trail Bitcoin over 90 days (#67) | 23 Jun 15:20Z | 21 Sep 15:20Z | 21 Sep 20:12Z | Miss - Ethereum +66.4% against Bitcoin +38.4% |
The regime miss is the one to sit with. That call read a tape below its 20-day average with narrow participation as range-bound; the week that followed was an 11% Bitcoin advance through a level it had not traded at since January. It was not a marginal miss and the correction came late - the risk-on read that replaced it was only written on 19 September, five days into the move.
The two quarterly relative calls were issued on 23 June under the first engine. They are not pooled with anything in the accuracy section below, which covers only current-engine seven-day asset calls.
Accuracy, trailing 30 days
Scope: engine epoch 4, seven-day asset calls only, the trailing 30 days to this run (21 September, 20:18Z). Fourteen scored calls against a 25-call floor, so the verdict stays insufficient data and nothing here is an established claim.
Brier skill score against the base-rate reference: +0.0382 (our Brier 0.2542 against the base-rate shadow's 0.2643). Marginally on the right side of zero, on a sample too small to mean much.
The decomposition says where that sits. Reliability 0.0255 - the term where over-confidence shows up. Resolution 0.0082 against an uncertainty of 0.2296: near zero, meaning our probabilities barely separate the calls that came good from the ones that did not. Calibration error 0.1566 against a 10% bar - a fail.
And the blunt test we keep publishing because it is the one that stings: we do not beat both constants. A forecaster that said "higher" to all fourteen of those calls would have scored a Brier of 0.2210; one that said 50% to everything, 0.2500. Both better than our 0.2542.
Separately, and clearly labelled as a direction-picking diagnostic rather than a headline: nine of those fourteen called directions came in, 64.3%, with a confidence interval running from 38.8% to 83.7%. That statistic never reads the probability we publish - a model that says 50% to everything and one that says 95% to everything score identically on it - so it cannot see the over-confidence above, and it is not what we lead on.
The regime record is tracked on its own: five resolved in the trailing 30 days, two correct, 40%, against a constant always-risk-on baseline that would have scored 60% over the same five. Today's miss is part of that count.
What the corpus carried
Eighty-six items were analysed in the 20-21 September window. Bitcoin led attention with 19 scoped mentions labelled bullish, ahead of Ethereum on 9, Curve on 5, Solana on 4 and Zcash on 3. At sector level the split was two-sided: derivatives drew 9 mentions on a bearish label while tokenised equities drew 7 bullish, with DeFi and memecoins on 4 each.
The window's reported events were Bitcoin trading through a level it had not reached since January, an exchange desk publishing seven-day implied move ranges across the four largest assets, an on-chain desk describing a short-liquidation shelf and a modest rebuild of options leverage against muted perpetual funding, and a corporate treasury reporting a further Bitcoin acquisition alongside a token repurchase. Against that, two analyst accounts read the same Ethereum tape as range-bound or facing resistance.
Twelve distinct accounts across three source kinds carried Bitcoin's directional coverage over the seven-day scan window, with no single account past a quarter of it. That is the shape the engine is built for. Curve's seven-from-one is the shape it is not, and today the difference cost us the only two slots we had free.
