The read
The objective regime read is risk-on, and it is marginal. Bitcoin closes above its 20-day moving average and its seven-day return clears the flat band by under half a percentage point - clear of the threshold, but not by much. Breadth is split rather than rotating: of the frozen alt basket, 3 of the 6 large-cap names with usable price history beat Bitcoin over the week, while all 5 mid-caps with history did. The structure classifier scores that combination as mixed, so a rotation call is not supported. The standing risk-on row from 19 September matures later today; this row opens the next weekly window so the regime read does not go dark between wakes.
Corpus this window (25-26 Sep, two days): 53 analyzed items, against 103 in the preceding two-day window (23-24 Sep). The windows are the same length, but the current one's second day is only two hours old, so treat the gap as an incomplete day rather than a coverage collapse. Bias across the most-mentioned coins is predominantly bullish (BTC, ETH, SOL, NEAR and ZEC bullish; AVAX and ADA bearish). The largest single story by importance is an exchange security incident carrying bearish scoped reads on ETH, BNB and AVAX.
What was issued
Four rows: the regime, plus three one-day experimental asset calls - Uniswap, NEAR and Solana, the top three conviction-passers by absolute score that had a free slot and a clean evidence read.
- UNI - thin but clean coverage: 3 items, 2 independent sources, 2 source kinds, all bullish. Strongest new idiosyncratic read in the scan.
- NEAR - 6 items, 3 sources, 2 kinds, all bullish, with the strongest trend leg on the board.
- SOL - deepest coverage (21 items, 6 sources) but genuinely split, so the corpus leg centres to roughly flat and the trend leg carries the read. Said plainly on the card.
All three numbers are the engine's, written verbatim. The one-day tier is a pre-registered experiment and is excluded from our published accuracy claim.
What was blocked, and by which rule
- Seven-day lane: at the 7-slot ceiling (BTC, ETH, SOL, XRP and BNB in the five large-cap slots; ZEC and NEAR in the two reserved idiosyncratic slots). UNI produced the best clean new seven-day read in the scan and could not be issued - a written call is never edited or displaced, so no slot opens until the ETH and SOL rows mature later today.
- Three-day lane: at the 3-slot ceiling (BTC, XRP, ZEC, all maturing 27 Sep).
- Event calls: at the 3-call ceiling.
- Structural and relative questions: all six distinct standing questions are occupied and every one of them already carries its maximum two rows, so even a refresh is barred this run. That includes the Bitcoin-record question, where the current engine reads about 51% while the displayed row - issued under an earlier engine - reads about 9%. That gap is logged as an open finding; it is not something this agent may edit, because records are immutable.
- BTC, XRP and ZEC one-day reads all convicted, but each of those coins already holds its maximum two standing calls across horizons.
- SUI carried the highest absolute score on every horizon and was refused: its directional coverage comes from a single source. PUMP, MATIC, CRV and DOGE were refused on the same rule; AVAX was refused because its corpus vote reduces to a baseline with no directional content.
Publications remain switched off by design, so the dark-publication flag sits on every row in the scan and is not treated as a defect.
Receipts - calls that resolved since the last run
Three one-day experimental calls, all issued 25 Sep 01:50 UTC, all matured 26 Sep 01:50 UTC and resolved 26 Sep 01:54 UTC. All three were graded under the current rule (direction only), so no cross-rule comparison is involved.
| Call | Stated | Outcome | Realized |
|---|---|---|---|
| NEAR - leans higher over the next day | 0.590 | Hit | +8.63% |
| SOL - leans higher over the next day | 0.543 | Hit | +3.94% |
| BNB - leans higher over the next day | 0.587 | Miss | -0.13% |
Two hits and one miss. The BNB miss is the instructive one: the move was -0.13%, which under direction-only grading is a miss by a rounding error's worth of price. That is the grading rule working as written, not an anomaly, and we report it rather than round it away.
Lifetime, the one-day experiment now stands at 3 resolved, 2 correct. That is n=3. It is not a track record and it is not eligible for the accuracy claim.
Accuracy, scoped
The headline below covers the trailing 30 days to this run (26 September, 02:05 UTC), engine epoch 4, seven-day asset calls only - the pool the public claim is gated on.
- Brier Skill Score vs the base rate: -0.035 (model Brier 0.2621 against a base-rate reference of 0.2533). Negative. Our probabilities are scoring slightly worse than simply quoting the historical base rate for each coin and ignoring every signal we have. That is the honest headline and we are not swapping it for a friendlier one.
- Murphy decomposition: reliability 0.0533 (lower is better - this is where over-confidence shows, and it is the larger of the two terms), resolution 0.0459 (higher is better - our probabilities do discriminate somewhat), uncertainty 0.2479.
- Calibration error (ECE): 0.178 - large, and consistent with that reliability term.
- Sample: n=11 against a required floor of 25. The verdict is INSUFFICIENT_DATA, and it would be even if the skill score were positive.
- The pool does not beat either constant-direction shadow: an always-up forecaster would have gone 7 for 11 on these same calls where the engine went 5 for 11.
- Directional-accuracy note, clearly labelled as a diagnostic and not the headline: hit rate 0.455 with a 95% interval of [0.213, 0.720], and direction-picking skill of -0.025. This statistic cannot see the probability we publish, which is exactly why it is not the lead.
For context only, the resolver's 30-day per-kind summary across all horizons: asset calls 47% on n=19 at a mean stated probability of 0.576 (over-confident), regime calls 50% on n=4, relative calls 50% on n=10. A raw hit rate is not a skill claim and none of these are.
Nothing in these numbers is adjusted by hand. When the resolver says the engine is over-confident, that stays visible in the record until the engine is fixed - it is never corrected in the prose.
