The read
No new calls this session. The book runs unchanged: a rotation regime read plus four seven-day
directional leans (BTC, ETH, SOL, ZEC), three multi-month macro reads and the standing structural
slate. Three of the seven asset-call slots sit open, and they stay open — the engine found exactly
one fresh conviction-passer that does not already carry a live call, and its evidence base failed
our independence test. That is the intended outcome, not a gap.
Why the rotation call still holds
The regime row issued 6 September called rotation — breadth in the alt complex leading, rather
than a clean risk-on or risk-off read off Bitcoin alone. This week's corpus argues the same way from
a different direction: an exchange index of the fifty largest non-Bitcoin assets was reported to have
gained 28.16% across August, outpacing Bitcoin over the same stretch. That is an independent,
after-the-fact measurement of exactly the breadth condition the regime call was built on.
Underneath it, the Bitcoin picture is quiet rather than weak. Two separate on-chain desks published
this window: one recording a single-day spot-ETF inflow of roughly 643 million dollars, the other
attributing Bitcoin's unusually low realised volatility to dormant supply held by multi-year wallets.
A corporate treasury also expanded a credit-securities repurchase programme. Constructive flow,
compressed volatility — a combination that supports a lean, not a conviction move, and the engine's
Bitcoin number says so.
What the corpus covered
Thirty-five items were analysed across the window, with coverage thinner and more scattered than the
prior period: Bitcoin at four mentions, Ethereum two, and a scattered tail of single-mention names
(AERO, CRV, ENA, LINK, PUMP, SOL, STX, XRP). Sector-level coverage tilted to tokenised equities,
AI and infrastructure. Two exchange executives and one analyst all published variations on the
same theme — capital-markets activity migrating on-chain — which is a narrative read, not a
resolvable one, and it is not represented in any call.
The one call we did not make
PUMP was the only asset to clear the engine's conviction gate without already holding a live call:
a 54% seven-day upward lean against a 46.6% base rate. We did not issue it. All three corpus items
behind that lean come from a single analyst account. The engine cannot see source concentration —
it counts mentions — so a single voice repeated three times reads to it as a corpus signal. It is
not one. Our stated edge is multi-source agreement; a one-voice lean is a single-source take wearing
a multi-source score, and issuing it would put a number on the public record that our own method
does not support.
The same filter caught CRV again (nine of ten items from one account, and the blended read came back
sideways in any case) and HYPE, which has genuine multi-source backing across three accounts but no
usable price history for the engine to anchor on — so there is no number to publish, and we do not
invent one.
BNB, AVAX, DOGE and ADA all scored strongly on raw momentum and all failed the gate for the same
reason: zero or near-zero corpus coverage, leaving momentum as the only live signal. A single-signal
read is not what this book claims to be.
Standing calls, re-scored
Every live seven-day lean was re-scored against the current engine. All four moved less than ten
percentage points from their issued value — BTC 0.586 to 0.513, ZEC 0.712 to 0.699, SOL 0.522 to
0.473, ETH 0.517 to 0.530 — which is the same belief restated, not a new forecast. Records here are
append-only and never edited, so an unchanged view means nothing is written. The three structural
questions re-scored the same way: the six-month Bitcoin-record read, the six-week breadth read and
the majors-drawdown read all landed inside their standing bands.
Receipts
Nothing resolved this session. No call reached its expiry date since the previous edition. The
last resolution — a seven-day Ethereum lean, issued 31 August, resolved correct on 8 September at
+0.20% — was already reported in the previous edition and is not counted twice here. The next
maturity is the Bitcoin lean on 10 September.
Every call we issue is public and graded from the moment it is issued; there is no hidden burn-in
period and nothing is quietly withdrawn before it scores.
Accuracy
The headline figure is the Brier Skill Score against the base rate: +0.0426 across 18 graded
calls at the current engine version. Positive means our published probabilities scored slightly
better than simply quoting the historical base rate and ignoring every signal we have.
That "slightly" is doing real work, and three things sit behind it:
- The sample is too small to establish anything. 18 graded calls against a 25-call floor. The
verdict the validator returns is INSUFFICIENT_DATA and we are reporting it as such. - We do not beat a constant call. A forecaster that mechanically said "up" on every one of these
18 calls would have scored a better Brier (0.2263) than we did (0.2565). Twelve of eighteen
resolved upward in a period where most things drifted up. Until that flips, the positive skill
score is not evidence of direction-picking ability. - Calibration is poor. Murphy decomposition puts reliability at 0.0662 and resolution at 0.0498;
expected calibration error is 0.2226 against a 10% bar. Our probabilities are still stated with
more confidence than the outcomes justify — that is where the remaining error lives, and it is an
engine fix, not something we shade by hand.
Four of the five internal go/no-go bars currently fail. The regime bar passes: regime reads are 3
of 4 against a 50% always-risk-on baseline, on a sample far too small to lean on.
Directional accuracy across those 18 calls is 61%, with a confidence interval spanning 39% to 80%.
That number is a diagnostic only — it never reads the probability we actually publish, so it cannot
detect the over-confidence the calibration figures just showed. It is not our headline and should
not be quoted as one.