The window in one line
Macro crowded out crypto-native news this window. The US July employment report showed a loss of 23,000 positions against an expected gain of 80,000, the central bank left its policy rate unchanged, and equities rallied while Bitcoin sat flat. Ninety-eight items cleared analysis between 2026-08-07 and 2026-08-08, and 38 of them scored importance 4 or above.
Attention keeps narrowing. Bitcoin drew 15 mentions on a bullish label, down 24 from the prior window; the broad crypto-market scope drew 13 on a bearish label, down 30. Equity-market coverage fell 11 to five items on a neutral read. Of the ten tracked sectors, only layer-2 gained (+1); security fell 13 to three items as the hardware-wallet firmware story aged out, and decentralised finance fell 10 to two. Regulators produced six items, including a Treasury sanctions action against two digital-asset exchanges tied to Iran's Revolutionary Guard.
What resolved since the last run
Two calls matured and were scored by the deterministic resolver. Both came back correct, and the second one is only half a compliment, for reasons worth stating.
| Call | Issued | Matured | Resolved | Outcome |
|---|---|---|---|---|
| #165 Dogecoin, leans higher over 3 days (stated 49.0%) | 2026-08-05 04:39Z | 2026-08-08 04:39Z | 2026-08-08 16:25Z | Correct - close at expiry 0.0702, +0.40% |
| #65 A broad altseason within six weeks remains the less-likely outcome (stated 74.2%) | 2026-06-23 15:20Z | 2026-08-07 15:20Z | 2026-08-08 16:25Z | Correct - 42% of 12 alts beat Bitcoin against a 60% bar |
Three further calls resolved just after the previous run's cut-off: a three-day Zcash call correct, a three-day BNB call incorrect, and an earlier altseason call incorrect.
That last one deserves the spotlight rather than the footnote. Call #57 said an altseason inside six weeks was near a coin-flip and was graded incorrect on 2026-08-07; call #65 said the opposite and was graded correct a day later. Both were our own rows, issued a day apart in June, and both ran to their own expiry under the append-only rule - nothing was withdrawn once the view changed. One of them had to lose. Read the pair as evidence of how the ledger works, not as evidence of a read.
Where the accuracy record actually stands
The public accuracy claim is scored on seven-day asset calls issued by the current forecaster. That pool currently holds zero resolved calls, because the engine version changed on 2026-08-05 and the clock restarted with it. So the Brier Skill Score against the base-rate reference - the only headline figure that reads the probability we actually publish - is not computable this run, and the claim remains unestablished against its 25-call minimum.
What we can say is what the earlier forecasters measured. Engine epoch 1 scored a Brier Skill Score of -0.4565 across 27 resolved calls; epoch 2 also scored negative. Both mean the same thing, plainly: our published probabilities scored worse than simply quoting the historical base rate and ignoring every signal we have. Nothing since has overturned that, and a fresh epoch counter should not be read as having fixed it.
The thirty-day per-kind tallies do not rescue that picture: asset calls 14 of 51, market regime 3 of 5, structural calls 1 of 3 at an average stated probability of 0.66, an overconfident segment. These are direction-picking counts pooled across every horizon and engine version in the window. They cannot see whether a stated 66% belonged at 66%, which is the one defect this engine has actually demonstrated, so they appear here as context and never as skill.
What we issued today
One call. The seven-day book is full at seven, the events book is full at three, the six structural and relative reads we carry are all current, and the standing regime read runs until tomorrow, so the only opening was a single three-day experimental slot.
NEAR Protocol, leans lower over the next three days. Coverage is thin and points mostly one way: a multi-asset fund rebalancing trimmed NEAR out of one of its portfolios, a market digest ran bearish, and only a constructive treasury-management proposal from a co-founder pulls the other way. Measured against the wider corpus baseline, that mix nets to almost nothing, so the lean is carried by the three-day price trend alone rather than by two signals agreeing. The engine puts it near 62% against a base rate around 55%, a lift of under seven points, on the experimental horizon that is excluded from our published accuracy record.
Two reads ranked above it and were passed over deliberately. Both traced their entire corpus signal to a single analyst's chart commentary on two related tokens. One voice restated across four posts is not the cross-source agreement this engine is built to fuse, and issuing it would import that analyst's conviction as though it were ours.
What we are watching, and where it bites
The employment miss and the unchanged policy rate pull against each other on the standing year-end policy-rate read, which sits near 89%. A labour market shedding positions is exactly the condition under which that call gets tested; the read is not being adjusted here, and it will be graded as written at year-end.
Elsewhere, centralised-exchange futures volume printed a 2023 low while equities rallied, a curious pairing, with risk appetite returning to stocks but not to the crypto derivatives bid. The tape is thin rather than directional, which is what the standing regime read already says.
