The read
The week-ahead regime stays choppy. Bitcoin's coverage this period is our deepest of any asset and nets only mildly positive: CLARITY Act lobbying, a State Department partnership with a bitcoin policy body and a derivatives no-action letter sit against a mining-pool Chapter 11 filing, fresh EU sanctions on an exchange, a reported 49% drawdown with continued fund outflows, and a takedown order against a messaging-protocol repository. Short-term momentum is only slightly positive. Alt breadth is thin - Solana and Dogecoin score sideways, Binance Coin and Cardano lean lower on momentum alone - so nothing points to a decisive market-wide direction or a broad rotation.
Five reads were issued this run: a fresh regime call, three weekly asset leans (Convex Finance, Ethereum, Bitcoin) and one experimental three-day lean. A quarterly relative read on Convex Finance versus Bitcoin also cleared the gate but was held back - the standing structural and relative book is already at its distinct-question ceiling. XRP's standing weekly lean was re-scored at essentially the same number, so it was left running rather than restated.
Magnitudes matter here. Since the engine's July recalibration, weekly leans land near 0.35-0.40 rather than 0.60-0.80. A 36% modeled chance against a 33% base rate is a marginal lift, not a confident call, and the model's strongest read this run - Convex Finance at roughly 38% against a ~25% base rate - is about one and a half times its base rate on thin, small-cap coverage. The lift the model applies over the base rate is a placeholder parameter that has not yet been fitted to outcomes; the direction is our signals' lean, not demonstrated skill.
What the corpus said
Thirty items were analyzed in this window. Market-scope coverage skewed bearish while individual coins skewed bullish - itself a choppy signature. The idiosyncratic read of the run is Convex Finance: an item on a custody-wallet accumulation of over a million tokens of a zero-inflation yield token, plus a separate analyst piece on its yield appeal and market cap outlook while a related governance token trades near alt-cycle lows. That is two independent items on a rank-250 name, which is why it clears the multi-source gate at all - and also why confidence stays low.
The prediction-market signal remains absent: the feed has not synced and its bitcoin-mapped entries are non-directional, so the blend renormalizes onto our corpus and momentum signals. That is the engine failing safe as designed, not a degraded read.
Receipts
Three calls matured and were graded by the deterministic resolver since the last run:
| Call | Issued | Matured | Outcome |
|---|---|---|---|
| Week-ahead market regime: choppy (#116) | 2026-07-18 08:10Z | 2026-07-25 08:10Z | correct (Bitcoin closed the week -0.71%) |
| Bitcoin - 7-day upside lean, p=0.372 (#118) | 2026-07-18 08:10Z | 2026-07-25 08:10Z | incorrect (-0.11%, no notable move) |
| Ethereum - 7-day upside lean, p=0.418 (#117) | 2026-07-18 08:10Z | 2026-07-25 08:10Z | incorrect (+0.57%, short of the move threshold) |
Both misses were low-confidence, sub-even leans that never produced a notable move - Ethereum edged the right way and Bitcoin finished flat - and the resolver scores a notable move, not any move. All three were resolved at 2026-07-26 02:29Z.
Accuracy: where we actually stand
The headline metric is the Brier Skill Score against a base-rate reference - the only figure that reads the probability we publish rather than just the direction.
- Brier Skill Score vs base rate: -0.517 on the current engine's claim pool. Negative means our published probabilities have scored worse than simply quoting the historical base rate and ignoring every signal we have. We report that plainly.
- Sample: n=2, far below the n>=25 threshold required before any skill claim is made. The verdict on the record is INSUFFICIENT_DATA and no accuracy claim is established. A two-call Brier Skill Score is not a stable estimate of anything - it can swing on a single outcome - so the honest reading is: unproven, not disproven.
- Murphy decomposition: reliability 0.157, resolution 0.000. Reliability is where over-confidence shows up, and it is the entire story here; resolution at zero means our probabilities have not yet discriminated between what happened and what did not.
- Directional accuracy (a diagnostic, blind to the probability we publish) reads -32% versus base rate on the two weekly calls, and -7% across the seven graded experimental three-day calls. Neither is a skill measure.
For context only, pooled across every engine version ever run - which mixes an older, materially over-confident model with the current one - the record shows a wide gap between stated probability and outcomes. That pooled gap is exactly the over-confidence the July recalibration was built to remove, and it stays visible in the record: probabilities are written verbatim from the engine and are never hand-adjusted to make the calibration curve look better.
What is not being answered
Altseason and the majors-drawdown questions produced no fresh directional read this run, so their existing standing calls run on unchanged. The Bitcoin-record question was re-scored at essentially its standing number, so nothing was restated. Sector rotation, cycle-top, cycle-bottom and behavioral questions remain outside what we can score. Event calls remain dormant while the market feed is stale.
