The read
Bitcoin closed higher over seven sessions and sits above its 20-day average, so the objective regime check returns the same label the standing week-ahead read already carries: risk_on. Under the append-only rule an unchanged label is not re-issued, so row 218 keeps running to its 26 September expiry and no new regime row was written this run.
Underneath that, the fused reads are lopsidedly trend-led. Across the 7-day scan every conviction-passer leaned the same way, and on most of them the coverage leg came back flat or marginally negative once centred on each asset's own history. That is a real caveat, not a footnote: the corpus moat is the half of the blend that is supposed to be ours, and this week it is mostly abstaining.
Accuracy, stated plainly
Headline, trailing 30 days to this run (25 September, 01:50 UTC), engine epoch 4, 7-day asset calls only: the Brier Skill Score against the base-rate reference is -0.0078 (model Brier 0.2584 against a base-rate reference of 0.2564). That is negative. Our published probabilities scored marginally worse than simply quoting each asset's historical frequency and ignoring every signal we have.
The Murphy decomposition on the same 12 calls: reliability 0.0414 (lower is better, and this is where over-confidence shows up), resolution 0.0417 (higher is better, whether the probabilities discriminate at all), uncertainty 0.25. Calibration error is 0.1243. The pool does not beat both constants either: an always-up constant scored 0.2304 against our 0.2584 on the same 12 calls.
The claim pool stands at 12 resolved 7-day calls against a 25-call minimum, so the verdict is INSUFFICIENT_DATA and no skill claim is being made. Nothing is warming up out of sight; every call is public and graded from issuance, and what is measured so far is slightly worse than the base rate.
As a clearly-labelled direction-picking note only: hit rate on those 12 is 50% and skill against the base rate is +0.022. That statistic never reads the probability we publish, so it cannot see over-confidence and it is not the headline.
Separately, as trailing-30-day context rather than skill: the resolver reports 17 resolved asset calls at a 47% hit rate against an average stated probability of 0.574 (flagged over-confident), 10 resolved relative calls at 50% against 0.572, and 4 resolved regime calls at 50%.
Receipts since the last run
One call matured and was scored, and it was a hit.
| Call | Issued | Matured | Resolved | Stated | Outcome |
|---|---|---|---|---|---|
| Solana vs Bitcoin, 90 days, leans to lead (row 71) | 2026-06-26 15:47Z | 2026-09-24 15:47Z | 2026-09-25 01:40Z | 0.584 | Correct - SOL +62.8% against BTC +40.5% |
That call was issued under engine epoch 1. Relative calls have graded the same way throughout, so it is directly comparable to later relative calls, though not to any asset call.
New calls
Two 7-day calls were issued, filling the book to its cap of seven: NEAR Protocol, which takes the second of the two slots reserved for reads that are not simply Bitcoin beta, and BNB, which takes the fifth and last large-cap slot. Both cleared the conviction gate; the engine set both numbers.
NEAR is the stronger of the two on evidence. Four coverage items from three distinct accounts across two source kinds all run constructive, led by an analyst documenting fast growth in NEAR Intents total value locked and fee capture, and centred on NEAR's own coverage history that leg nets clearly positive with the price trend agreeing. BNB is the opposite shape: three items, two of them directional, from two accounts, and once centred the coverage leg is effectively flat, so the price trend carries that blend on its own. Both theses say so on the card.
Three 1-day calls were also issued, the first in that tier: NEAR, BNB and Solana. One-day calls are an experimental horizon - they are shown for transparency and are deliberately excluded from the accuracy record, which counts 7-day calls only. They are a learning instrument, and a harder one than they look, because a one-day window gives a directional read far less room to express itself than a week does. Solana's carries the bluntest caveat of the five: its coverage leg comes back marginally negative once centred, so our own corpus is not on that side of the call and the trend leg is carrying it alone.
What was blocked, and by which rule
- Bitcoin, XRP and Zcash all cleared the 7-day conviction gate, and none could be issued: each already stands at the two-concurrent-calls-per-coin ceiling, and on the 7-day leg each would also be a second call on the same coin-and-horizon slot, which is barred outright.
- Ethereum and Solana cleared the gate on 7 days, but the fresh reads sit within a tenth of their standing calls, so the material-change band is not met and the standing rows were left running untouched.
- The 3-day tier is at its cap of three concurrent calls, all of which mature on 27 September.
- MATIC and Pump.fun were refused on evidence: the engine flagged both as resting on a single directional source.
- The structural and relative board is completely blocked. All six standing question slots are occupied, and every one of them already carries the maximum of two concurrent rows, so neither a new question nor a refresh of an existing one had anywhere to go. The nearest relief is the altseason slot, which frees on 26 September.
- Events are at their cap of three.
One note from the structural scan worth recording even though nothing could be issued from it: the current engine reads the chance of a new Bitcoin all-time high inside six months well above the roughly 9% standing row the board currently shows on that question, which was written by an earlier engine generation. That gap stays visible on the board until one of those rows expires and a refresh slot opens.
Coverage this window
37 items were analyzed across the 24-25 September snapshot window. Bitcoin led attention with 13 mentions on a bullish net read; Ethereum (5) and Solana (4) both came back bearish on net; NEAR and Zcash carried 2 apiece. On the sector side the exchange-traded-fund and stablecoin buckets picked up coverage while derivatives coverage fell back. Publication-tier sources remain switched off by design, so no publication-tier report sits in this window's corpus.
