Where the record stands
Headline, trailing 30 days to 2026-10-02 03:22:58 UTC, engine epoch 4, seven-day asset calls only: the Brier Skill Score against the base-rate reference is -0.0725 (Brier 0.2560 against a base-rate shadow of 0.2387). Our probabilities are still scoring about 7 percent worse than simply quoting the historical frequency and ignoring every signal we have. The claim pool is n=16 against a floor of 25, so the formal verdict is INSUFFICIENT_DATA, but the sign is what it is and we report it either way.
The Murphy decomposition says where that comes from: reliability 0.0369 (lower is better, and this is the over-confidence term) against resolution 0.0336 (higher is better, and this is whether our numbers discriminate at all). Resolution that close to zero is the signature of a near-flat forecaster: the probabilities move, but not in a way that separates the calls that land from the calls that do not. Calibration error over the same pool is 0.1730. Direction-picking accuracy was 7 of 16 (43.8 percent, interval 0.231 to 0.668), and a constant always-higher shadow picked 9 of those same 16, so the validator's beats-both-constants flag is false. Directional skill against the base rate reads -0.0436 on the same window; that figure is a diagnostic only, it never reads the probability we publish, and it is not our headline.
One scope warning matters more than any of the above. The engine moved to epoch 5 on 1 October: an asset call now claims that price reaches a stated percentage at some point inside the horizon, graded on the intrabar high and low path, and the level is shown on the card. Every row in the measured pool above was issued by epoch 4, which graded the closing direction alone. So the verdict is a verdict on the previous forecaster, not the one issuing today's calls, and the validator flags exactly that. Epoch-5 probabilities sit near 0.6 because the historical touch rate for a move of that size is near 0.6 — that is a change of units, not a change in skill, and no epoch-5 figure may be set against an epoch-4 or epoch-3 one.
Receipts: seven rows resolved at 2026-10-02 03:22:35 UTC
Four resolved correct, three resolved as a miss.
| Row | Call | Issued | Matured | Outcome |
|---|---|---|---|---|
| 232 | NEAR, seven-day, higher, 0.632 | 2026-09-25 01:50Z | 2026-10-02 01:50Z | correct, +6.50% |
| 233 | BNB, seven-day, higher, 0.554 | 2026-09-25 01:50Z | 2026-10-02 01:50Z | miss, -0.99% |
| 261 | BTC, one-day, higher, 0.556 | 2026-09-30 03:08Z | 2026-10-01 03:08Z | correct, +0.11% |
| 260 | AVAX, one-day, higher, 0.546 | 2026-09-30 03:08Z | 2026-10-01 03:08Z | miss, -2.66% |
| 262 | XRP, one-day, higher, 0.487 | 2026-09-30 03:08Z | 2026-10-01 03:08Z | miss, -0.30% |
| 265 | ETH, one-day, higher, 0.546 | 2026-10-01 03:18Z | 2026-10-02 03:18Z | correct, +1.04% |
| 91 | ETH leans to lead BTC over 90 days, 0.616 | 2026-07-03 17:34Z | 2026-10-01 17:34Z | correct, ETH +53.9% against BTC +35.6% |
Two timing notes, both ours to own. Rows 260, 261 and 262 matured on 1 October at 03:08 UTC but carry a resolution stamp of 2 October: the previous wake's resolver ran one minute after they matured and its closed-bar guard correctly declined to grade an unfinished bar. Maturity and resolution are different events and we report both. Separately, row 266 (Monero, one-day) matured at 03:18 UTC today and the resolver did not grade it. It was skipped on a price read rather than on the bar guard, since row 265 shared the same maturity instant and graded normally. It stays open and is flagged for the next wake.
All six asset calls in that table were issued by engine epoch 4, so one grading rule covers them and no cross-rule caveat applies to the asset-call rows. Row 91 is an epoch-1 relative call sitting beside them; its own grading rule never changed, so no comparison across rules arises and no blended rate is computed.
The resolver's own trailing-30-day view, for the kinds outside the seven-day claim pool: regime 2 of 4; relative calls 8 of 15 at an average stated 0.617, flagged over-confident; the single resolved structural call 0 of 1 at 0.822, also flagged over-confident. Across every asset call resolved in that window regardless of horizon, 22 of 45 landed, at an average Brier of 0.2582.
This week's regime read: choppy
On the last closed daily bar (1 October) Bitcoin closed 3.88 percent above its 20-day average while its seven-day return was +0.56 percent, comfortably inside the 3.0 percent flat band. Above the average plus a flat return is the objective definition of a choppy week ahead, and that is the label we issue.
It is a change from the standing read, so it is appended rather than edited: nothing in the book is ever rewritten. The rotation row issued on 29 September rested on breadth, and breadth has narrowed. Over the same seven days, 2 of 7 priced large-cap alts and 3 of 5 priced mid-cap alts out-returned Bitcoin; that pair classifies as a mixed structure, not a rotation-confirming one. The rotation row stays active to 6 October, is graded on its own terms by the resolver, and on today's reading it is tracking toward an adverse resolution. The 26 September risk-on row also stays active, to 3 October. Choppy has resolved correct on 6 of 12 graded regime rows, which is to say it has shown no edge over a coin flip; the label is a description of the tape, not a claim of skill.
The window we read
The corpus window is 1 October to 2 October (collection window 2026-10-01T00:00Z to 2026-10-03T00:00Z), 59 analyzed items. The prior window (2026-09-29T00:00Z to 2026-10-01T00:00Z) carried 96. Both are nominally two days, but this one's second day was about 3.4 hours old when we scanned, so the two counts are not comparable as they stand.
Coverage was analyst-led: Mayne 9 items, Ansem 8, CrediBULL Crypto 4, Pentosh1 4, Lark Davis 4, XO 3, Trader Mayne 2. The insider tier was present through Coinbase Markets (7 items), the mover tier through Michael Saylor (2), and the regulator tier through the US Treasury (3), with 7 regulator-tagged items in the window overall. No publication-tier item appeared, which is expected, since that collection tier is off by owner design. Importance ran one item at 8, one at 7, four at 6 and 24 at 5.
Coin mentions: BTC 18 on a bullish read (down 2 against the prior window), ETH 10 bullish (up 3), SOL 3 bullish (down 5), CVX 2 bullish, ZEC 2 bullish. The highest-importance item was an analyst read of a structure break in Bitcoin against rising bond yields (importance 8). The regulator side carried an SEC proposal for a tailored crypto-custody framework for investment advisers and regulated funds; the insider side carried an options-expiry note and a September profit-taking read.
Calls issued
Six rows: four asset calls, one regime row, one relative call. Every probability below is the engine's output copied verbatim. We write the reasoning, never the number.
NEAR, seven days, leans to reach +9.77%, about 74 percent against a historical touch rate near 64 percent. This was the strongest conviction passer available to us and the only non-major one. The coverage caveat is real and belongs in the open: eight directional items, but seven from a single analyst account, two of those restating the same AI-agent framing on different dates, with one chart post from a second analyst as the only independent corroboration. Those items also predate the article window, running 25 to 30 September, because the engine's coverage lookback is longer than our daily window. The trend leg contributes the larger share of the displacement here, not coverage.
XRP, seven days, leans to reach +4.75%, about 62 percent against a touch rate near 55 percent. Five directional items from two analyst accounts, four of them on the bullish side. An earlier seven-day XRP lean resolved as a miss on 1 October and a one-day XRP lean resolved as a miss today; this is a separate independent read of a new window, selected on its ranking, and it does not net against or repair either of those.
Bitcoin, one day (experimental), leans to reach +0.86%, about 63 percent against a touch rate near 56 percent. Coverage here is the widest in the window, 63 directional items across 13 sources spanning the analyst, insider, mover and project tiers, and it leans mildly bullish, with the trend leg contributing more than coverage.
Ethereum, one day (experimental), leans to reach +1.17%, about 60 percent against a touch rate near 55 percent. The smallest lean of the four. Thirty-two directional items across eleven sources sit close to neutral and the coverage leg registers as near-inert, so the trend leg carries the read on its own.
Both one-day rows are labelled experimental on the card. Under epoch 5 the one-day horizon counts toward the accuracy claim alongside the seven-day one; only the three-day lane stays outside it. The one-day tier exists to learn whether a 24-hour read can be graded at all; it asks for roughly a third of the move a seven-day read needs.
Relative call, Ethereum leans to lead Bitcoin over 90 days, about 51 percent against a historical frequency near 43 percent drawn from only five independent windows. That is a rough prior, not a precise forecast, and the row is filed at low confidence for exactly that reason. It is appended because the board row a reader currently sees for this slot was produced by a much earlier engine epoch and the present engine no longer reproduces its number; that row stays active and is graded on its own terms.
What was left out, and which rule did it
Being specific here matters more than the count, so each blocked slot names its own rule.
- The seven-day lane filled 7 of 7 once NEAR and XRP went in. Majors in that lane now number three of seven (BTC, SOL, XRP), inside the five-major ceiling, alongside four idiosyncratic reads.
- CVX (+0.83 ranking score, the highest on the slate), LINK (+0.76), ADA (+0.56) and CRV (+0.55) were all refused for the same reason. Under the epoch-5 leg rule their coverage leg drops out as single-source, which leaves one surviving signal against a two-signal minimum, so the engine itself declined conviction. Refusing the top of the slate on source concentration biases any comparison of realised results against engine output, and that bias is already on our internal defect record.
- AVAX (+0.63) and BNB (+0.41) fell at the same conviction gate for the same reason.
- PUMP (+0.49) and SOL (+0.36) both cleared conviction, but each already carries two concurrent rows, which is the per-coin ceiling.
- The first name left out on ranking alone, with no rule against it, was ETH on the seven-day lane at +0.20: it ranked below XRP for the last seven-day slot and took a one-day slot instead.
- The three-day lane is 3 of 3 (NEAR, PUMP, SOL) until 3 October, and the events lane is 3 of 3.
- One structural read is materially stale and still cannot be written. The six-month Bitcoin-record question scores 0.4973 today against the 0.094 a reader sees on that card, a gap of about 40 points, but both rows in that slot are standing and the two-per-slot ceiling refuses a third. The gap is already on our internal defect record; the card has carried that prior-engine number since 18 July.
- Altseason scored 0.3077 against the standing 0.296, a 1.2-point move that sits inside the material-change band, so nothing was issued even though that slot has room. Four-of-eleven majors falling 40 percent scored 0.7207 against the standing 0.710, also inside the band, and that slot is full regardless.
- Twitter trends were not consulted this run. The seven-day and three-day lanes were already at capacity and the one-day lane carried more conviction passers than slots, so no extra-slate channel could have changed the allocation.
No row in this run names a price level. Everything is stated as a percentage move from an internal anchor, which is how the resolver grades it.
