What the corpus carried, 27-28 September 2026
Window: 2026-09-27 00:00Z to 2026-09-29 00:00Z exclusive - one complete day plus roughly the first two and a half hours of 28 September. 31 items completed analysis in that window, against 95 in the previous two-day window (2026-09-25 to 2026-09-27). The two figures are not like-for-like: the current window's second day is only a couple of hours old, so this is an incomplete day, not a coverage collapse.
Coverage counts for this window: BTC 6 mentions (window bias bullish, 11 fewer than the prior window), ETH 4 (bullish, 9 fewer), CVX 2 (strong bullish, 2 more), SOL 2 (bullish, 5 fewer), XRP 2 (bearish, unchanged). Sector reads were mostly constructive but thin - etf 3 mentions, ai 2, then a tail of single-mention sectors. Importance skewed low-to-mid: 17 of the 31 items scored 2 or below, 13 scored 4 or above, one reached 6.
Source mix was analyst-dominated: Mayne 4 items, Ansem 4, Lark Davis 4, Coin Bureau 2, XO 2, CrediBULL Crypto 2, alongside one regulator (US Treasury, 4 items), two mover accounts (Vitalik Buterin 3, Michael Saylor 2) and one insider (Nansen 2). No publication-tier item entered the window. Publication collection is off by owner design, so that is the expected shape of the corpus rather than a gap in it.
The regime read did not change, so no new regime row was written
The objective regime inputs, measured on closed daily bars to 2026-09-27: Bitcoin's last close sat above its 20-day average, and its seven-day return came to +4.06% against the 3.0% flat band. That is the risk-on branch of the same objective test the resolver applies at expiry. The standing regime row (id 237, risk-on, issued 2026-09-26, running to 2026-10-03) carries that label already, so under the append-only rule nothing was issued for the regime slot this run - a re-issue of an unchanged label would add a correlated row to the record and nothing else. The widget is served by the active row, not by a fresh one.
Accuracy, trailing 30 days to 2026-09-28 02:30Z
Scope: engine epoch 4, seven-day asset calls only - the single pool the public accuracy claim reads.
- Headline. Brier skill against the base-rate reference: +0.0259 (our Brier 0.2481 against 0.2547 for a shadow that ignores every signal we have and quotes the historical frequency). Positive.
- The pool is 14 rows of the 25 required, so the validator's verdict is INSUFFICIENT_DATA. Nothing here is established.
- And the positive number needs its companion fact in the same breath: on those exact 14 rows a constant always-higher shadow picked 10 right where the engine picked 8. The validator's beats-both-constants flag is FALSE. A forecaster that says higher to everything out-picked us on this sample.
- Why the skill score is only barely positive: the Murphy decomposition reads reliability 0.0081 (low is better - our stated probabilities are close to their realized frequencies) against resolution 0.0144 on an uncertainty of 0.2449. Near-zero resolution means the probabilities barely discriminate between the rows that resolved correct and the rows that resolved incorrect. Good calibration on an almost flat forecast is not skill, and must not be written as momentum.
- Supporting reads: expected calibration error 0.0454; hit rate 0.5714 with a 95% interval of 0.326 to 0.786 - an interval that comfortably contains a coin flip; mean stated probability 0.5757.
- Directional diagnostic, clearly labelled as such and never the headline: skill against base rate +0.0889. It cannot see the probability we publish, so it cannot see over-confidence.
- Resolver view over the same 30 days, by kind: asset_call 31 scored, 58% hit rate, mean stated 0.5759, gap -0.0047 - calibrated. market_regime 4 scored at 50%, exactly matching a naive always-risk-on baseline of 50%, so the regime lane has demonstrated nothing. relative_call 11 scored at 45% on a mean stated 0.5919 - a +0.1374 gap, over-confident. structural_call 1 scored, 0%, mean stated 0.822.
Receipts - what resolved since the last run
Four rows matured and were scored deterministically at 2026-09-28 02:30:23Z. Two correct, two incorrect. All four are engine-epoch-4 asset calls graded on the sign of the move alone, so one grading rule applies across the whole list and no cross-rule comparison arises.
| Row | Call | Issued | Resolved | Stated | Outcome | Realized |
|---|---|---|---|---|---|---|
| 224 | Zcash, 7 days, higher | 2026-09-20 20:04Z | 2026-09-28 02:30Z | 0.624 | correct | +6.76% |
| 248 | Pump.fun, 1 day, higher (experimental) | 2026-09-27 02:25Z | 2026-09-28 02:30Z | 0.583 | correct | +17.47% |
| 246 | Bitcoin, 1 day, higher (experimental) | 2026-09-27 02:25Z | 2026-09-28 02:30Z | 0.597 | incorrect | -1.05% |
| 247 | Solana, 1 day, higher (experimental) | 2026-09-27 02:25Z | 2026-09-28 02:30Z | 0.576 | incorrect | -0.74% |
Note the maturity-versus-resolution gap: row 224 matured at 2026-09-27 20:04Z and was only stamped resolved about six and a half hours later, when the next run's resolver pass came round. The one-day rows matured at 02:25Z and were scored within five minutes. The three experimental one-day rows are excluded from the public accuracy claim by configuration; only row 224 entered the seven-day pool.
The alt-versus-Bitcoin question resolved against us, and the engine has barely moved
The previous run scored row 179 - a structural call stating at 0.822 that a broad altseason would NOT occur inside six weeks - as a miss: 82% of the frozen eleven-coin basket gained on Bitcoin over the window, against a 60% threshold. A broad altseason genuinely happened.
One day later the structural scorer reads 28.2% for a yes inside the next six weeks (base rate 23.6%, an edge read, conviction met). The standing board row on that question (id 222, issued 2026-09-19, yes at 29.6%) is 1.4 percentage points away from the fresh number - inside the material-change band of about ten points - so it was left running and nothing was appended. That is the correct mechanical answer, and it is also the uncomfortable one: a realized altseason moved the model's six-week read by roughly a point. This is the pattern already on our internal defect record, where two long-horizon rows failed one-directionally on the same proposition at high stated confidence. It is an engine matter - the base rates on this family rest on very few independent windows - and under our own rules no probability was hand-shaded in response.
What was issued, and what each slot was blocked by
Five rows were proposed and the cap lint adjudicated the counts.
Seven-day lane, the only claim-eligible tier, ran at 6 of 7 with one free slot. The slot went to Pump.fun (PUMP), higher over seven days, 0.590 against a 46.8% base rate - the highest-scoring coin on the whole scan that still had a free per-coin budget. Coverage: five items over the engine's lookback, all directional, two distinct analyst accounts, with 60% of the items from a single account and a second account contributing the rest. That concentration is disclosed deliberately: the thesis does not rest on breadth of sources. The constructive items argue the token is mispriced against ecosystem payouts and a possible distribution, and one of the two accounts restated the same argument on two consecutive dates - so the item count overstates the number of distinct claims. The momentum leg is the stronger of the two contributors by magnitude (0.911 against 0.216 for the corpus leg, at a two-thirds corpus weighting).
One-day experimental lane was empty after three maturities and took three rows: Pump.fun 0.613 (base 49.2%), Solana 0.573 (base 47.4%) and Zcash 0.603 (base 50.6%). A one-day call on a coin already carrying a longer call is a separate slot, not a restatement - each horizon is graded only against itself and neither of the two brief-horizon tiers reaches the public claim. The prior one-day rows on Bitcoin, Solana and Pump.fun had already matured AND been resolved before these were written, so the windows are disjoint and the forecasts sequential rather than stacked.
The first coin left OUT of the one-day lane was Bitcoin, on ranking alone: its score of +0.342 sat fourth in a three-slot lane. That is a different fact from being blocked by a rule, and the two should not be conflated. Blocked by rules, by slot:
- NEAR and Uniswap - both convicted with clean evidence and both rank above every coin we issued, but each already carries two standing calls, so the two-per-coin budget is full. Nothing more can be written on them until 30 September.
- SUI - highest score on all three horizons for the fourth consecutive run, and refused for the fourth consecutive run because its directional coverage comes from a single source. CVX and LINK were refused on the identical flag.
- AVAX - set aside because its corpus vote is a baseline artefact rather than a directional read.
- Three-day lane - full at 3 of 3 until 30 September, so no three-day row was possible regardless of conviction.
- Event outcomes - full at 3 of 3; all three are long-dated macro rows and none matures before December.
One relative read convicted and could NOT be written. Zcash to lead Bitcoin over 90 days came back yes at 0.790 against a 69.4% base rate, an edge read with conviction, and it would have been a fresh coin on that question rather than a refresh - which is precisely why the cap lint refused it: the structural-and-relative lane already stands at its ceiling of six DISTINCT standing slots (the Bitcoin-record question, altseason, the majors-drawdown question, and Ethereum, Solana and XRP against Bitcoin), and a new coin is a seventh slot rather than a refresh. None of those six drains before 2026-10-01. Filed today as an internal defect: on a lane made entirely of 90-day-to-six-month rows, a concurrency ceiling sized for fast questions means no new coin can ever enter the board, however well evidenced. Two further caveats would have travelled with that read had it been written, and they are worth recording anyway: the scorer reports only five independent 90-day windows behind that 69.4% figure, so it is a rough prior and not a calibrated frequency, and the relative-call lane is the one our own resolver currently marks over-confident (45% realized against 59% stated over the trailing 30 days). The corpus support was real - an insider-tier item reported Zcash ahead of Bitcoin by a wide multiple since Bitcoin's record high, alongside constructive analyst coverage of a protocol upgrade.
A second relative read convicted and could not be written: Solana versus Bitcoin, where the scorer now says yes at 0.523 while the row a reader actually sees on that card says no at 0.751 - a divergence of roughly 27 points against an epoch-2 read. That slot stands at its two-row ceiling until 2026-10-08, so the board will display a number the engine no longer agrees with for another ten days. The same shape blocks the Bitcoin-record card, where the scorer reads 47.4% against a displayed 9.4% - a gap of roughly 38 points, unrefreshable until January 2027. Both gaps are already on our internal defect record.
Signals, honestly
Every call on this slate was built from two legs, our analyzed corpus and price momentum. The prediction-market leg is absent from all of them. That is the engine's healthy state, not a degraded read: the admissible-market test only admits questions that are directional in shape and settle on our horizon, and the live catalog is long-dated, so it does not qualify on a seven-day or briefer lean. No number was substituted for it.
