Where the record stands
Headline, trailing 30 days to 2026-09-27 02:09Z UTC, engine epoch 4, seven-day asset calls only: Brier Skill Score against the base-rate reference = +0.0142 on n=13. That is the only figure on this page that reads the probability we actually publish, and it is above zero - our probabilities scored marginally better than quoting the historical frequency and ignoring every signal we have. The three previous runs read -0.008, -0.008 and -0.035, so the sign has turned.
It is also very small, and the same output gives three reasons not to bank it:
- A constant forecaster still comes out ahead on these rows. On the same 13 calls, a shadow that said up every single time would have landed 9; the engine landed 7. The validator's beats-both-constants flag is false. A +0.014 skill score that cannot out-pick a one-word rule is not evidence of a directional read.
- n=13 against a floor of 25. The verdict is INSUFFICIENT_DATA and the confidence interval on the hit rate runs 29% to 77%. A single call flipping moves the headline by more than the headline itself.
- Resolution is near zero. The Murphy decomposition reads reliability 0.0357 (lower is better - this is where over-confidence would show) and resolution 0.0387 (higher is better - this is whether our probabilities discriminate at all) against the world's uncertainty of 0.2485. Expected calibration error is 0.0778, inside the 10% bar. Read together: our stated probabilities are roughly honest about themselves, and they separate winners from losers barely at all. That combination is what a positive-but-tiny skill score looks like.
For context only, and not as a skill claim: directional accuracy on the same 13 rows is 54%, and the resolver's own 30-day per-kind table reads asset calls 59% on 27 rows, regime 60% on 5, relative calls 45% on 11, and structural calls 0% on 1. Those are hit rates - they are blind to the number we publish and cannot see over-confidence, which is exactly why they are not the headline.
Receipts - everything that resolved since the last run
Eleven rows matured and were scored by the deterministic resolver. Eight hits, three misses. Every asset call below is an engine-epoch-4 row graded on the sign of the move alone; the one epoch-1 row in the list is a relative call, whose grading rule has not changed since it was issued, so no cross-rule comparison arises and no blended rate is computed.
| # | Call | Stated | Issued (UTC) | Matured (UTC) | Outcome |
|---|---|---|---|---|---|
| 218 | Week-ahead regime: risk-on | - | 09-19 19:45 | 09-26 19:45 | Hit - Bitcoin +3.92% |
| 219 | Ethereum, 7 days, higher | 55% | 09-19 19:45 | 09-26 19:45 | Hit - +1.73% |
| 220 | Solana, 7 days, higher | 50% | 09-19 19:45 | 09-26 19:45 | Hit - +8.71% |
| 229 | Bitcoin, 3 days, higher | 63% | 09-24 01:14 | 09-27 01:14 | Hit - +0.14% |
| 230 | XRP, 3 days, higher | 50% | 09-24 01:14 | 09-27 01:14 | Hit - +1.35% |
| 231 | Zcash, 3 days, higher | 59% | 09-24 01:14 | 09-27 01:14 | Hit - +9.62% |
| 238 | Uniswap, 1 day, higher | 60% | 09-26 02:05 | 09-27 02:05 | Hit - +2.73% |
| 239 | NEAR, 1 day, higher | 59% | 09-26 02:05 | 09-27 02:05 | Hit - +3.68% |
| 240 | Solana, 1 day, higher | 56% | 09-26 02:05 | 09-27 02:05 | Miss - -0.54% |
| 179 | Broad altseason inside six weeks: no | 82% | 08-12 18:48 | 09-26 18:48 | Miss - 82% of 11 alts gained on Bitcoin |
| 75 | Ethereum trails Bitcoin over 90 days: yes | 79% | 06-28 16:00 | 09-26 16:00 | Miss - Ethereum +71.5% against Bitcoin +41.7% |
All eleven were stamped resolved at 2026-09-27 02:09Z, which is when the resolver ran, not when each call matured - the two are shown separately above on purpose.
The two long-horizon misses are the ones worth sitting with. Row 179 said broad altseason would not arrive inside six weeks, at 82%. It arrived: 82% of the frozen eleven-coin basket gained on Bitcoin over the window, against a 60% threshold and a ~23% historical frequency for that outcome. Row 75 said Ethereum would trail Bitcoin over 90 days, at 79%; Ethereum returned +71.5% against Bitcoin's +41.7%. Both were high-stated-probability calls on the same underlying proposition - that alternatives would not gain on Bitcoin - and both failed in the same direction, on horizons of six weeks and three months. That is a pattern in the long-horizon book, not two independent unlucky draws, and it is the single clearest thing in this record to watch. Structural calls now read 0 of 1 and relative calls 5 of 11 on the trailing 30 days.
Row 240 deserves a word too: Solana closed -0.54% over its one-day window, so a call stated at 56% resolved as a miss on a move of about half a percent. Under sign-only grading there is no dead band, and small adverse closes count fully against us.
What we are calling today - eight new rows
Nothing here is a recommendation and no figure below is a target. Every probability is the ensemble engine's output, copied verbatim; what we write is the reasoning.
The engine scored 18 assets on each of three horizons. Thirteen cleared the conviction gate. Two of those - Sui, which carried the highest score on every horizon for the third run running, and Convex - were refused because their corpus leg rests on a single directional outlet, which fails our multi-source rule. One more, Avalanche, was set aside because its corpus vote falls back to a global baseline rather than its own coverage.
Seven-day lane (the claim pool), 2 new rows. The Ethereum and Solana seven-day calls matured yesterday, freeing two of seven slots; both were refilled with the highest-scoring clean passers available.
- Uniswap, higher, 60% against a ~45% base. Score +0.550, the strongest clean read on the board. Its corpus leg is honest but slight: three scoped reads from two outlets, and not one of them is about Uniswap - they are an alt-rotation survey, a revenue-token piece and an on-chain comparison of the top 50 alternatives. The trend leg contributes more to this blend than the corpus leg does.
- Solana, higher, 55% against a ~45% base. Score +0.363. This one has real coverage depth - nineteen scoped reads from six outlets - and real disagreement inside it: one research desk noted Solana's share of on-chain equity tokens falling to roughly 30%, and an exchange desk published options positioning that our analysis also read as bearish - both against a majority of constructive reads. The net corpus lean is therefore mild, and again the trend leg does more work.
That brings the seven-day book to seven of seven: four large-cap majors (Bitcoin, XRP, BNB, Solana) and three idiosyncratic reads (Zcash, NEAR, Uniswap), which satisfies the standing requirement that at least two slots go to non-beta reads.
Three-day lane (experimental), 3 new rows. The lane emptied when rows 229-231 matured this morning. Refilled with the three highest-scoring eligible passers: Uniswap at 61%, NEAR at 63% and Zcash at 62%. The NEAR row is the most interesting of the three - six aligned scoped reads, but four of them are a single outlet's broad rotation pieces; the one NEAR-specific datapoint in the lookback is a trader's infographic on Intents value locked passing 150 million dollars with cumulative fees above 40 million, published 20 September. Zcash carries the highest historical frequency in the slate at ~51%, so the gap our signals contribute there is the narrowest of the set.
One-day lane (experimental), 3 new rows. Also emptied this morning. Bitcoin at 60%, Solana at 58%, Pump.fun at 58%.
The Bitcoin row is worth flagging for an honest tension. The two-day corpus snapshot behind this article labels Bitcoin's coverage bearish on 8 mentions - driven by a fresh pullback thesis, a European fintech's customer-data breach story and two chart-led downside reads, all published inside the window. The engine's corpus leg looks at a longer lookback and reads a mild net positive lean instead, because that window also contains several constructive on-chain, policy and macro-rotation reads. We are issuing the engine's number, and we are naming the disagreement rather than smoothing it: both readings are in the corpus, and the two-day label is the more bearish of the two.
The Pump.fun read rests on four scoped items from two outlets, one of which - a valuation argument on creator rewards and buybacks - was published inside this window. The Solana one-day slot is distinct from its seven-day call and is graded only against its own window.
What we are not calling, and why - slot by slot
- Market regime: nothing issued, deliberately. The objective read computed fresh this run is risk-on (Bitcoin's close of 84,433 sits above its 20-day average of 80,186, with a seven-day return of +3.92% clearing the 3.0% flat band by 0.92 points). That matches the standing regime row, which runs to 3 October. Records here are append-only: an unchanged read is left running rather than restated, and the widget is never dark because an active row exists.
- Bitcoin-record question: blocked by the two-rows-per-slot ceiling. The scorer reads 51% on a new record within six months against a ~39% historical frequency, with conviction. The slot already carries two standing rows, so nothing can be added. This matters because the newer of those two, the one the card displays, states 9% - roughly 42 points below the current read - and the slot cannot take a refresh until the older row expires in January 2027. That gap is on the record as an open defect, not a view.
- Altseason question: refreshable, but nothing material to say. The resolution of row 179 dropped this slot to one standing row, opening a refresh for the first time in weeks. The scorer reads 29.59% where the standing row states 29.6% - a move of one ten-thousandth, same direction, same engine epoch. That is well inside the material-change band, so no refresh is warranted.
- Deep-drawdown question: blocked by the two-rows-per-slot ceiling. The scorer reads 75% on the no side against a ~64% historical frequency.
- Relative calls: blocked twice over. All three standing coin slots (Ethereum, Solana, XRP) carry two rows each, and the six distinct standing structural and relative slots are exactly full, so adding a new coin would be a seventh slot.
- Event outcomes: at the cap of three. The policy-rate-path, contraction and consumer-price rows all stand.
The window we read
The corpus snapshot behind this article covers 2026-09-26 to 2026-09-27 and contains 41 analyzed items. The preceding equal-length two-day window returned 102. That is not a coverage collapse: this window's second calendar day was about two hours old at run time, so the current figure is one complete day plus a fragment, and the two counts are not comparable as rates.
Within the window, Bitcoin led on 8 scoped mentions labelled bearish and Ethereum followed on 5 labelled bullish, with everything else in single figures. Market-level coverage was bullish on 4 mentions. On the sector side, decentralised finance led on 3 mentions, with artificial intelligence, layer-1 and privacy on 2 each. No regulator-sourced item appeared at all. Coverage came overwhelmingly from the analyst tier - one analyst account alone supplied 14 items - with a handful of mover and insider items and, as is now standard, nothing from the publication tier, which is off by design.
Twitter trends contributed nothing: neither the worldwide nor the US top thirty contained a single crypto entry, so no asset entered today's slate from that channel. Trends never vote on a probability in any case - they are directionless, and the engine excludes them.
The prediction-market leg carried no admissible market on any of today's eight calls, which is its normal state on short-horizon price reads: only markets that are directional in shape and settle on our horizon are admitted, and the live catalog runs to long-dated ladders. Those calls rest on our own two signals, the analyzed corpus and price trend, and that is the engine failing safe as designed rather than a missing input.
