Receipts · forward-tested record
Every claim here was logged before the outcome was knowable.
IntellaZone is a trading-intelligence platform: long and short calls logged every evening with the factors behind them, casefiles, a trading desk, a synced portfolio, and a trading academy with a simulator on real bars. This page is the record of how those calls turned out.
Each call is written down the moment it is made, scored against SPY at fixed windows, and tested under hypotheses stamped in advance. Verdicts change only at pre-registered reads — including the ones that went against the product. Nothing on this page is a backtest, a pick, or advice.
Record through the night of Oct 25 · ledger updated Oct 22 · page built October 26, 2026 at 10:18 AM ET
How to read this page
- The big number
- How far the live long book’s picks beat or trailed SPY on average, seven days after each call. n is the number of picks scored; the net figure takes off an assumed round-trip cost in bps (basis points: 100 bps is one percentage point).
- Sealed, then read
- SEALED: the hypotheses were stamped and the read date fixed in advance, and no result is shown before the read. READ: the verdicts are recorded in the verdict trail, and they change only at pre-registered reads.
- Where the failures are
- Nothing is removed: failed and retired results stay in the scoreboard, the record by product and the verdict trail.
Live long book · avg excess vs SPY per pick at 7 days
0.79 pts
n=925 · -0.99 pts net of 20 bps round-trip · 42.4% of picks beat SPY (any margin, ties counted)
Live scanner (v14 core + universe-x1) · since July 20, 2026 · metric frozen Aug 12, 2026
Live scanner (v14 core + universe-x1) · since July 20, 2026 · metric frozen Aug 12, 2026
Short book, as-called: 0.00 pts (n=927) — a research instrument, not the product, shown however it looks.
FORWARD-TESTED
3,643 calls logged · never edited — deletions and re-bases are listed below
verdicts change only at pre-registered reads
Scored @ 7d
2,695
Scored @ 30d
2,580
Scored @ 90d
1,259
Verdicts reversed
0
since June 2026
Headline = the live long book (the product the platform surfaces) in average excess vs SPY, the figure the headline reports. The all-time blended record for every cohort ever logged — retired products and their failures included — is below. Nothing is removed.
Experiments running now
4 running · 3 readday 0 of 36 · read in 38 days
Day 0 Oct 28 · registered read Dec 3 · last pick night Nov 23
core scanner + universe-x1 · every pick logged Oct 27 – Nov 23, 2026, sorted by the premarket tone of the next morning (decided before 9:30 AM ET) and walked to its stop or target on daily prices up to Dec 1
- H1 (PRIMARY): picks on the side of the tone reach their target first at least 10 pts more often than picks against it, n ≥ 100 in each, past the level corrected for the 9 hypotheses live at the stamp
Stamped Oct 22, 2026. The running result is hidden from everyone, the owner included, until the read: the stop/target diagnostic stops at the picks of Oct 26 until E5-PM-H1 is recorded. The tone is logged each morning before the open and is not a result. Measured, never executed — the platform does not trade.
day 14 of 183 · read in 169 days
Day 0 Oct 12 · registered read Apr 13 · night 120 Apr 5
the live scanner (10 longs and 10 shorts a night) against the relative-strength lane B-RS (10 and 10), on the same 120 trading nights, Oct 12, 2026 – Apr 5, 2027
- H1 (PRIMARY): the live scanner's calls beat the lane's by ≥ 0.5 pt at 7 days, as called (each short's result flipped, longs and shorts pooled); the lane ahead by ≥ 0.5 pt is the negative; within 0.5 pt, indistinguishable; n ≥ 150 scored outcomes in the lane
Stamped Sep 27, 2026; dates fixed Oct 1, 2026. The lane's running result is hidden from everyone, the owner included, until the read: from Day 0 the owner's statistics route leaves out the lane's new rows until E5-RS-H1 is recorded. Measured, never executed — the platform does not trade, and the lane's list is never shown as picks.
day 21 of 65 · read in 44 days
Day 0 Oct 5 · registered read Dec 9 · last night of the first window Nov 30
core scanner + universe-x1 · every long call logged Oct 5 – Nov 30, 2026, banded upper / middle / lower by its own list's scores
- W1, on each list separately: the upper band beats the lower by ≥ 0.5 pt at 7 days, with the middle between them, n ≥ 30 per cell
Stamped Sep 26, 2026; filled Oct 2, 2026. The bands were drawn from logged scores alone (Sep 14–25) before any window call existed, and their cut-offs stay in the stamped charter, never on a page. The calibration record takes in no call from Day 0 until the last read, so the running answer is hidden from everyone. A second window (Dec 1, 2026 – Jan 28, 2027, read Feb 8, 2027) is read only if the first confirms; only that read could return the letters.
day 27 of 36 · read in 9 days
Day 0 Sep 29 · registered read Nov 4 · last pick night Oct 26
core scanner + universe-x1 · every pick logged Sep 29 – Oct 26, walked to its stop or target on daily prices up to Nov 2
- H1 (PRIMARY): among picks whose first day does not close in their favour, the stop rate beats the target rate by ≥ 15 pts, n ≥ 100
Stamped Sep 26, 2026. The running result is hidden from everyone, the owner included, until the read: the E4 page counts only E4's own picks, and the stop/target diagnostic stops at the night before Day 0. Measured, never executed — the platform does not trade.
registered read Sep 20 · recorded Sep 25 · 5 of 5 verdicts recorded below
Day 0 Aug 21 · registered read Sep 20 · recorded Sep 25 · 7d cutoff Sep 13
core scanner + universe-x1 · 30 picks/night (core 20, universe-x1 10)
- H1 range-state conditioning
- H2 ranking-inversion persistence
- H3 grade monotonicity
- X1 day-1 kill rule
- X2 day-5 time stop
Verdicts come from the pre-registered bars at the Sep 20 read. Live flags are measured, never executed — the platform does not trade.
registered read Sep 23 · recorded Sep 25 · 3 of 3 verdicts recorded below
Day 0 Aug 24 · registered read Sep 23 · recorded Sep 25 · 7d cutoff Sep 16
three ghost lanes (equal-weight, trend-momentum, relative-strength) race the composite · 20 each/night
- each lane vs the composite at 7d, ±0.5 pt bars, n ≥ 150 per lane
Read Sep 25, 2026; the verdict rows are in the ledger below.
registered read Sep 24 · recorded Sep 25 · 2 of 2 verdicts recorded below
Day 0 Aug 25 · registered read Sep 24 · recorded Sep 25 · 7d cutoff Sep 17
an independent model opinion logged beside every scanner pick, model recorded per row
- H1 disagreement as a filter: endorsed beats rejected by ≥ 1.0 pt at 7d, n ≥ 100 / 60
- H2 calibration: Brier < 0.25 and below base rate; overconfidence gap ≤ 15 pts
The verdict is the one AI the charter names (Fable, with Opus as fallback), all its rows together; each model's rows are shown beside it as description and never compared. Corrected Sep 23, before the read: this line used to say the read would slice by model.
Scoreboard
as of Oct 22Confirmed
5
Failed or retired
14
Context / untestable
4
Sealed
4
Repairs disclosed
10
Reads scheduled
13
Confirmed 5 of 27Failed or retired 14 of 27Context 4 of 27Sealed 4 of 27
Tallied from the ledger below, never hand-counted. 2 products have been retired by their own scoreboard; one suspension executed by pre-committed rule.
Nightly integrity
15 clean · 0 warn · 15 graded PASS WARN · a manual lane short (none since Sep 16 2026) FAIL SKIPPED · non-trading day
Every night at 11:45 PM ET a check counts what each logging lane wrote and grades the night. Last graded: 2026-10-25 · SKIPPED.
Is the confidence number worth anything?
1202 scored closed calls · 7d windowFrozen for a test: this record takes in no call logged on or after October 5, 2026. A pre-registered test (G-R1) is checking whether this score ranks the calls logged from that night on, and a record that kept updating here would show its answer before it is read. New calls come back into this record after the test's last read. The first read is on December 9, 2026, and a second on February 8, 2027 if the first passes.
Scores run overconfident by about 21.2 points on average; the widest gap is in the 70–80 band, which claims 74 and delivers 41.
The highest-conviction 302 closed predictions beat SPY by more than 1% at seven days in 38.7% of cases, against 59.7% for the lowest-conviction 300 — the ordering is inverted: the scores the scanner is most confident about are doing worse, by more than the intervals allow for chance.
Every pick carries a conviction score. This asks the question nobody asks of their own: does that number behave like a probability, and does the ordering it implies actually order outcomes? A hit means the call beat SPY by more than 1% at seven days. When the test cannot be passed, the statistics are withheld rather than published with a caveat — a rigorous-looking figure computed from an input that does not satisfy its assumptions is worse than no figure at all.
All-time record by product
win rate vs SPY at 7d / 30d / 90d · n scoredAI watchlist alerts (retired)
177 logged
7d
45.5%
n=134
30d
40.6%
n=160
90d
56.5%
n=131
Daily Brief v2 watchlist calls (live since Aug 12, 2026)
110 logged
7d
44.6%
n=74
30d
39.6%
n=53
90d
—
n=0
Daily Brief v1 (retired)
110 logged
7d
38.8%
n=85
30d
31.6%
n=98
90d
40.2%
n=87
Daily Brief v2 (live since Aug 12, 2026)
54 logged
7d
47.7%
n=44
30d
32.3%
n=31
90d
—
n=0
Scanner long picks
1,463 logged
7d
46.1%
n=1162
30d
41.1%
n=1066
90d
36.5%
n=513
Scanner short picks
1,472 logged
7d
49.1%
n=1124
30d
52%
n=1083
90d
51.7%
n=520
TSP allocation picks
257 logged
7d
50%
n=72
30d
43.8%
n=89
90d
4 of 8
n=8 · low
WIN = beat SPY by more than 1% over the window (short calls inverted); LOSS = trailed by more than 1%; PUSH = within ±1%. The hairline on each bar marks 50% — the coin flip. Under n=30 a cell gives its counts (calls that beat SPY, of those graded) and no rate. Retired products stay on the board.
The verdict trail
50 entries · including the failuresE5-PM — sealed (Day 0: Oct 28, 2026 · read Dec 3, 2026)
Charter stamped Oct 22, 2026 (docs/research/e5pm-charter.md). Before each open the morning read already reads the premarket move of SPY; this test writes it down once a trading morning before 9:30 AM ET as a tone (positive, negative or flat at a tenth of a percent), then asks whether picks on the side of that tone reach their target before their stop clearly more often than picks against it, with the walk of E5-X1, on the picks of Oct 27 – Nov 23, 2026, the running result hidden from everyone until the read on Dec 3, 2026. Its clock started at its own Day 0: the cap of three was lifted on Oct 22, 2026, by Steve. One pass would be a candidate: under Article 2 nothing on the platform changes until a second blind window passes too.
E5-PM-H1
Premarket tone and the side of the pick, forward: among core and universe-x1 picks logged on the 20 trading nights before the mornings of the window, a pick whose side agrees with the premarket tone recorded the next morning before 9:30 AM ET (a long on a positive morning, a short on a negative one) is WITH the tone, the opposite AGAINST; flat and unknown mornings count for neither. Pass if WITH reaches its target first at least 10 pts more often than AGAINST, n ≥ 100 in each, with the morning-clustered t past the level corrected for every hypothesis live at the stamp (Bonferroni); open picks kept in the denominator; walked on daily prices by the same rules as E5-X1.
PRIMARY for its read week, named at the stamp (Oct 22, 2026) before any of its data existed; k = 9 hypotheses live at the stamp, so the bar is one-sided at 0.05 ÷ 9 (Bonferroni). Day 0 Oct 28, 2026; pick nights Oct 27 – Nov 23; prices through Dec 1; read Dec 3, 2026 (docs/research/e5pm-charter.md). The tone is logged by the morning-read cron, dormant until a yes by Steve (PREMARKET_TONE_OPEN and the migration of its table, applied on that yes). Sealed: no person or agent computes a stop-or-target result or an R figure for these picks, or joins a tone to a pick, before the read. The 7-day results against SPY stay public call by call on the track record, as for every call; that is a different measure. Longs and shorts are also compared with themselves across positive and negative mornings, beside the verdict, because a trending window can make the tone mostly one side. A secondary question (did WITH end up better in R) is printed beside the verdict and can never be a finding on its own.
SEALED
E5-RS — sealed (Day 0: Oct 12, 2026 · read Apr 13, 2027)
Charter stamped Sep 27, 2026; dates fixed Oct 1, 2026. In B1 the relative-strength lane finished 1.6 pt a week ahead of the live scanner on a score that counted a short like a long; on the fair score registered on Sep 23, before anyone had seen a B1 figure (a long earns what the stock did against SPY, a short the opposite), the two finished within half a point. This asks the question again on 120 fresh nights, with that score and B1's bar. Whatever it shows, nothing on the platform changes: one result is a candidate, and the lane's list is never shown as picks.
E5-RS-H1
The live composite beats the relative-strength lane (B-RS) by at least 0.5 pt at 7 days, as called (each short's result flipped, longs and shorts pooled), n ≥ 150 per lane, on the counted nights of Oct 12, 2026 – Apr 5, 2027 (the as-called measure registered on Sep 23, 2026).
PRIMARY, named before any of its data existed: the as-called measure registered on Sep 23, 2026, before the B1 read, applied on fresh nights. If the lane beats the live composite by the same margin the row is recorded NEGATIVE; within 0.5 pt, INDISTINGUISHABLE; with fewer than 150 scored outcomes in the lane, OBSERVATIONAL. Counted in 2027, the year it is read. The long book alone is printed beside it as description only and can never become a second verdict. Sealed: nobody, the owner included, sees the lane's running figure before the read. Wording set on Oct 2, 2026, ten days before Day 0, to the stamped charter's own sentence; the row first posted that morning named no direction and put the 150 floor on both arms.
SEALED
G-R1 — sealed (Day 0: Oct 5, 2026 · reads Dec 9, 2026 and Feb 8, 2027)
Charter stamped Sep 26, 2026 and filled Oct 2, 2026 (docs/research/g-r1-charter.md). E4-H3 found the old letters did not rank outcomes, with a flaw the read named: most A picks came from one list and most B picks from the other. This gives the score one fair second look: bands drawn separately for each list from scores alone, before any window call existed, then two fresh windows. The charter's own expectation is failure. The letters return only if both lists hold the bar in both windows.
G-R1-W1
G-R1, first window (Oct 5 – Nov 30, 2026): on the core list and on universe-x1 separately, long calls in the upper score band beat those in the lower band by at least 0.5 pt at 7 days, with the middle band between them, n ≥ 30 per cell.
Named before any of its data existed. Cut-offs as stamped, in the charter and never on a page. Read Dec 9, 2026. If it does not confirm, the letters stay internal, the score alone is shown, and the second window is not read.
SEALED
E5-X1 — sealed (Day 0: Sep 29, 2026 · read Nov 4)
Charter stamped Sep 26, 2026 (docs/research/e5x1-charter.md). The second try of E4-X1's day-1 warning, the same rule, walk, bar and minimum, on picks made after the stamp, with the running result hidden from everyone until the read. E4-X1 was not blind, so it is the candidate finding that prompted this test, not a first pass: a pass here is the first blind pass, and under Article 2 nothing on the platform changes until a second blind window passes too.
E5-X1-H1
Day-1 kill rule, forward: among core and universe-x1 picks logged Sep 29 – Oct 26, 2026, a pick whose first full trading day does not close in its favour (a flat close included) is marked KILLED-D1. Pass if that cohort's stop rate beats its target rate by ≥ 15 pts, n ≥ 100, open picks kept in the denominator; walked on daily prices up to Nov 2 by the same rules as E4-X1.
PRIMARY for its read week, named at the stamp before any of its data existed. Exploratory in the year's count: it adds one test (k = 20 at the read if nothing is recorded first). Sealed: no person or agent computes a day-1 flag, a stop-or-target result or an R figure for these picks before Nov 4, 2026. The 7-day results against SPY stay public call by call on the track record, as for every call; that is a different measure. A secondary question (would selling at the day-1 close have done better, in R) is printed beside the verdict and can never be a finding on its own.
SEALED
E4 — live (Day 0: Aug 21, 2026 · read Sep 20)
Charter stamped Aug 21. Standing caveat, pre-committed: E3 matured in a single bullish regime, so long-side results were flattered and shorts penalized by drift. E4 verdicts carry whatever regime E4 delivers; none may be excused or upgraded by regime narrative after the fact. Kill criteria: pooled 7d n < 300 at the read downgrades every verdict to observational; the Best Available suspension lifts only if H1 or H2 passes with a deployable positive signal.
E4-H1
Range-state conditioning (lead): COILED picks beat TRENDING-WEAK picks by ≥ 1.0 pt at 7d, COILED positive with win rate > 50%, n ≥ 100 per cell.
COILED did not hold its own bar out of sample (−0.6020, 38.5%, raw), so no TRENDING-WEAK result could have passed the test. E3's range-state split is stamped as in-sample noise. COILED n 288; TRENDING-WEAK n 0; pooled n 438. Left out as no_data: 0. Graded more than 7 days 6 hours after the pick: 0. Recorded 2026-09-25 from read-files/E4-2-range-state-H1.sql (revision 3), printed as run.
FAILS
E4-H2
Ranking-inversion persistence: long book minus short book at 7d, n ≥ 150 per side.
Long book minus short book: −0.1807 pt, inside ±1.0. No cross-sectional signal. Longs −0.8604, shorts −0.6797, raw; n 219/219, pooled 438. Breakdown, not a verdict: core −0.2135, universe-x1 −0.1110. UVXY-class left out: 0. Left out as no_data: 0. Graded more than 7 days 6 hours after the pick: 0. Recorded 2026-09-25 from read-files/E4-1-ranking-H2.sql (revision 3), printed as run.
NO SIGNAL
E4-H3
Grade monotonicity: A ≥ B+ ≥ B at 7d (longs) with A−B separation ≥ 0.5 pt, n ≥ 30 per cell.
The grades did not rank: A −2.1742, B+ −0.7766, B 0.3288. Grades are demoted to internal-only pending a recalibration charter, and the public grade-based win-rate context is removed. n 45/101/51; A minus B −2.5030 pt at 7 days. Caveat: the A cell is 98% core list and the B cell 71% universe-x1, so an A-versus-B gap also compares the two lists. Per-list breakdown: A / core −2.2959 (n 44), A / x1 3.1800 (n 1), B+ / core −0.7306 (n 90), B+ / x1 −1.1536 (n 11), B / core 1.4947 (n 15), B / x1 −0.1569 (n 36). Left out as no_data: 0. Graded more than 7 days 6 hours after the pick: 0. Recorded 2026-09-25 from read-files/E4-3-grades-H3.sql (revision 3), printed as run.
FAILS
E4-X1
Day-1 kill rule: a pick whose first full trading day closes against entry is marked KILLED-D1. Pass if that cohort's eventual-stop rate exceeds its eventual-target rate by ≥ 15 pts, n ≥ 100.
E4-X1 PASS: the stop rate beat the target rate by 15 points or more (one pass = a candidate finding, Article 2). Of the 239, 13.39% were still open at the Sep 18 cutoff (not yet resolved then, which is not the same as never resolving); they stay in the denominator, as in the design basis. KILLED-D1 239 (a first day not in the pick's favour, a flat close included, as the launch-night tracker counted it): stopped 64.44%, target 22.18%, still open at the Sep 18 cutoff 13.39%, gap of record 42.2594 pts. Counted strictly, without the 1 flat close, as description only: 238, gap 42.8571 pts. By side (description, not a verdict): longs 136 — stopped 75%, target 13.97%; shorts 103 — stopped 50.49%, target 33.01%. Read at Sep 25, 2026, 2:32 PM ET. Not blind: the E4 page has shown a live preview of this cohort since launch night. One pass is a candidate finding; nothing changes on the platform until it repeats on a fresh window. Exploratory: one of 19 tests this year; the chance of a false positive rises with k. The only hypothesis designated PRIMARY before its data existed was E4-H1. Recorded 2026-09-25 from the /stop-target read of record (?read=e4), printed as run.
CONFIRMS
E4-X2
Day-5 time stop: picks unresolved at the day-5 close are marked EXPIRED-D5. Pass if their forward expectancy from day 5 is ≤ +0.10R, n ≥ 80.
E4-X2 UNTESTABLE: its stamped measure, forward expectancy from day 5 to resolution or day 90 (pass at or below +0.10R), was defined on 2026-09-24 before any exit-rule number (R is the stop distance; forward R runs from the day-5 close to the walk's exit or the session-90 close), and the read of record needs every EXPIRED-D5 pick resolved or at session 90 on one fixed price history, so it runs on or after 2027-01-22. Also, 64 of the 154 EXPIRED-D5 picks were still open at the Sep 18 cutoff, 0 to 14 sessions after day 5. The measure's definition was written on the night of Sep 23, before any exit-rule number (E4-X2-DEFINITION-2026-09-24.md: R is the stop distance, from the day-5 close to the walk's exit or the session-90 close). The read of record runs on or after 2027-01-22, once every EXPIRED-D5 pick has resolved or reached session 90. Not blind: the E4 page and the ordinary stop/target diagnostic have shown this cohort's day-5 figures to signed-in accounts since launch night. Recorded 2026-09-25 from the /stop-target read of record (?read=e4), printed as run.
UNTESTABLE
B1 and A1 — read (Day 0: Aug 24 / Aug 25, 2026 · read Sep 25)
Two of the three concurrent slots. Both scoreboards are read; the rows below carry the verdicts.
B1-H1
The live composite beats the equal-weight lane (B-EQ) by at least 0.5 pt at 7 days (raw convention, longs and shorts together), n ≥ 150 per lane.
The live composite and B-EQ finished within 0.5 pt at 7 days (−0.4427; live −0.6269, B-EQ −0.1843). The composite did not beat this benchmark by the registered margin. The registered measure is direction-blind: symbol minus SPY with shorts not flipped, longs and shorts pooled. With every arm half short, an arm scores higher when the names it shorted rose, and symmetric skill cancels; the platform said so before the read (track-record API note, Aug 12: averaging raw relatives across both directions cancels to ~0 by construction). As-called, shorts flipped (not the registered measure): live +0.4531, B-EQ +0.5378, the same reading as the registered measure. With exactly 50% shorts, as-called equals long book minus pooled raw. Any follow-up pre-registers the as-called measure (B1-AS-CALLED-MEASURE-PREREGISTERED-2026-09-23.md). Shorts: live 50.0%, B-EQ 50.0%. Long book (the stamped secondary): live −0.1739, B-EQ +0.3535. Day 0 was Aug 24 as stamped; the lanes logged no rows that night, so live's 20 Aug 24 picks are left out to keep the arms matched. Left out as no_data: 0. Graded more than 7 days 6 hours after the pick: 1. Recorded 2026-09-25 from read-files/B1-benchmark-lanes.sql (revision 3), printed as run. Disclosed 2026-09-26 (Fable): from Aug 24 to the Sep 25 read, /api/read-stats returned each B1 lane's and the live lane's 7-day aggregate (scored n, share beating SPY, average excess, by direction) to any signed-in account, and the members' dual track-record page listed the newest twenty scored outcomes of any lane; the code cannot say whether anyone looked. The lane rows themselves were never on a public page. Both routes were narrowed on Sep 26. Disclosed 2026-09-27 (Fable): in the same weeks the members' dual track-record page's summary cards showed each lane's 7-day aggregate by type, a third way; narrowed on Sep 26 (70e4cd2).
INDISTINGUISHABLE
B1-H2
The live composite beats the trend-and-momentum lane (B-TM) by at least 0.5 pt at 7 days, same terms.
The live composite and B-TM finished within 0.5 pt at 7 days (+0.1603; live −0.6269, B-TM −0.7873). The composite did not beat this benchmark by the registered margin. The registered measure is direction-blind: symbol minus SPY with shorts not flipped, longs and shorts pooled. With every arm half short, an arm scores higher when the names it shorted rose, and symmetric skill cancels; the platform said so before the read (track-record API note, Aug 12: averaging raw relatives across both directions cancels to ~0 by construction). As-called, shorts flipped (not the registered measure): live +0.4531, B-TM −0.0363, the same reading as the registered measure. With exactly 50% shorts, as-called equals long book minus pooled raw. Any follow-up pre-registers the as-called measure (B1-AS-CALLED-MEASURE-PREREGISTERED-2026-09-23.md). Shorts: live 50.0%, B-TM 50.0%. Long book (the stamped secondary): live −0.1739, B-TM −0.8236. Day 0 was Aug 24 as stamped; the lanes logged no rows that night, so live's 20 Aug 24 picks are left out to keep the arms matched. Left out as no_data: 0. Graded more than 7 days 6 hours after the pick: 1. Recorded 2026-09-25 from read-files/B1-benchmark-lanes.sql (revision 3), printed as run. Disclosed 2026-09-26 (Fable): from Aug 24 to the Sep 25 read, /api/read-stats returned each B1 lane's and the live lane's 7-day aggregate (scored n, share beating SPY, average excess, by direction) to any signed-in account, and the members' dual track-record page listed the newest twenty scored outcomes of any lane; the code cannot say whether anyone looked. The lane rows themselves were never on a public page. Both routes were narrowed on Sep 26. Disclosed 2026-09-27 (Fable): in the same weeks the members' dual track-record page's summary cards showed each lane's 7-day aggregate by type, a third way; narrowed on Sep 26 (70e4cd2).
INDISTINGUISHABLE
B1-H3
The live composite beats the relative-strength lane (B-RS) by at least 0.5 pt at 7 days, same terms.
B-RS beat the live composite by 1.6118 pt at 7 days (live −0.6269, B-RS +0.9848; raw convention, n 318/320). The registered measure is direction-blind: symbol minus SPY with shorts not flipped, longs and shorts pooled. With every arm half short, an arm scores higher when the names it shorted rose, and symmetric skill cancels; the platform said so before the read (track-record API note, Aug 12: averaging raw relatives across both directions cancels to ~0 by construction). As-called, shorts flipped (not the registered measure): live +0.4531, B-RS +0.8290, and on that measure the reading differs: the two arms finish within 0.5 pt. With exactly 50% shorts, as-called equals long book minus pooled raw. Any follow-up pre-registers the as-called measure (B1-AS-CALLED-MEASURE-PREREGISTERED-2026-09-23.md). Shorts: live 50.0%, B-RS 50.0%. Long book (the stamped secondary): live −0.1739, B-RS +1.8138. Day 0 was Aug 24 as stamped; the lanes logged no rows that night, so live's 20 Aug 24 picks are left out to keep the arms matched. Exploratory: one of 16 tests this year; the chance of a false positive rises with k. The only hypothesis designated PRIMARY before its data existed was E4-H1. Left out as no_data: 0. Graded more than 7 days 6 hours after the pick: 1. Recorded 2026-09-25 from read-files/B1-benchmark-lanes.sql (revision 3), printed as run. Disclosed 2026-09-26 (Fable): from Aug 24 to the Sep 25 read, /api/read-stats returned each B1 lane's and the live lane's 7-day aggregate (scored n, share beating SPY, average excess, by direction) to any signed-in account, and the members' dual track-record page listed the newest twenty scored outcomes of any lane; the code cannot say whether anyone looked. The lane rows themselves were never on a public page. Both routes were narrowed on Sep 26. Disclosed 2026-09-27 (Fable): in the same weeks the members' dual track-record page's summary cards showed each lane's 7-day aggregate by type, a third way; narrowed on Sep 26 (70e4cd2).
NEGATIVE
A1-H1
Disagreement as a filter: scanner picks the second-opinion model endorses beat the ones it rejects by ≥ 1.0 pt at 7d, n ≥ 100 endorsed / 60 rejected.
The picks the AI rejected did better than the ones it endorsed (1.0217 against 0.2992). Separation −0.7225 pt at 7 days, read as called; endorsed n 394, rejected n 94. Within-night separation −1.1460; within-direction −0.5876; raw sign −0.4561; rejected arm 57.4% short, endorsed 48.2% short. Per model, description only, never compared: claude-fable-5 endorsed 0.7377 (n 292), rejected 1.9026 (n 46); claude-opus-4-7 endorsed −0.9559 (n 102), rejected 0.1775 (n 48). Left out as no_data: 0. Graded more than 7 days 6 hours after the pick: 33. Opus answered on five nights (Aug 25–26, Sep 1–3) and Fable on twelve; no night had both. On Sep 19 an audit saw A1's overall outcome counts by side (hit rate near 50%) but not the endorsed/rejected split. Recorded 2026-09-25 from read-files/A1-1-disagreement-H1.sql (revision 3), printed as run.
FAILS
A1-H2
Calibration: Brier score < 0.25 and below the base rate; overconfidence gap ≤ 15 pts against the prior 53.5.
Calibration failed: condition 1 missed, condition 2 missed. Brier 0.2506 against 0.25 and the base-rate 0.2488; band-by-band gap 2.74 pts. n 488, base rate 0.5348. Log loss 0.6943. Overall-average gap 2.74 pts (context; the Sep 19 audit had already seen the overall hit rate). Calibration curve, stated probability: share that came good (n): 40: 0.0% (3); 41: 100.0% (1); 42: 50.0% (6); 43: 50.0% (2); 44: 33.3% (9); 45: 57.1% (21); 46: 75.0% (20); 47: 48.0% (25); 48: 37.5% (8); 49: 30.8% (13); 50: 56.8% (74); 51: 58.5% (65); 52: 54.8% (84); 53: 60.8% (79); 54: 42.9% (56); 55: 40.0% (20); 56: 100.0% (1); 57: 0.0% (1). Per model, description only, never compared: claude-fable-5 Brier 0.2499, base rate 0.5799, band gap 6.71 pts (n 338); claude-opus-4-7 Brier 0.2521, base rate 0.4333, band gap 6.21 pts (n 150). Left out as no_data: 0. Graded more than 7 days 6 hours after the pick: 33. Opus answered on five nights (Aug 25–26, Sep 1–3) and Fable on twelve; no night had both. On Sep 19 an audit saw A1's overall outcome counts by side (hit rate near 50%) but not the endorsed/rejected split. Recorded 2026-09-25 from read-files/A1-2-calibration-H2.sql (revision 3), printed as run.
FAILS
Mirror Backtest — run Aug 22, 2026 (diagnostic)
The one backtest this platform runs, and only as a check on itself: the frozen E3 configuration replayed over the E3 window, single run, published regardless of outcome. The platform does not backtest to find strategies.
MIRROR
Replay of the frozen scanner over the E3 window lands within ±1.5 pts of the forward result.
Replay +0.726 vs forward −0.13 — inside the pre-registered band, so the forward record is not a logging artifact. The long-short spread flipped sign versus live E3; that finding was handed to E4-H2 rather than interpreted here.
CONSISTENT
E3 — read complete (Day 0: Jul 20 · 7-day checkpoint Aug 3 · 30-day read Aug 20, 2026)
Pre-registered Jul 19. Cohorts: core scanner, universe-x1 (a zero-overlap defensive/dividend list), TSP allocation, Daily Brief. The 30-day read executed a pre-committed suspension: Best Available moved to Parked. That consequence was written before E3 began.
E3-H3
The scanner's 7-day edge persists out-of-sample (confidence interval above 50, n ≥ 30).
By pick-week, 7d avg excess vs SPY: Jul 20 +2.28 (n=122) → Jul 27 −1.93 (n=150) → Aug 3 +0.03 (n=150) → Aug 10 −0.66 (n=90). Pooled −0.13 (n=512, 47.9% win rate). The launch-week edge seen at the Aug 3 checkpoint (+1.49, n=81) did not persist. SUSPEND-to-Parked fired as registered.
FAILS
E3-H1
The regime gate discriminates out-of-sample: the approved side beats the blocked side by ≥ 2.0 pts.
Inverted. The flag read "bull" every night of the matured window; approved longs returned −1.28 vs SPY (7d, n=256) while the blocked short-ranked names returned +1.01 raw (n=256). Single-regime caveat stamped: the gate never faced a non-bull flag in the window.
FAILS
E3-H2
A+ conviction picks outperform A picks by ≥ 1.5 pts (pooled n ≥ 30).
Zero A+ grades fired in 30 days (n=0). Logged calibration finding, not a verdict: the grade ladder was non-monotonic on longs at 7d — A −0.83 (n=52), B −1.28 (n=59), B+ −1.59 (n=141). Became an E4-H3 design input.
UNTESTABLE
E3-H4
The universe-x1 expansion list performs within 1 pt of the core list.
Core −0.12 (n=341) vs x1 −0.16 (n=171); delta 0.04. The machine generalizes across a zero-overlap universe — and the signal it generalized was not working in this window. Both halves stated.
CONFIRMS
Q-R1/R2
Range-state questions: does the edge differ across TRENDING / RANGING / COILED tape?
COILED +0.83 (7d, n=186, 55.9% win) vs TRENDING-WEAK −0.77 (7d, n=420, 43.8%). The sharpest split in the dataset, found post-hoc — so it cannot be claimed as validated. Pre-registered as E4-H1 instead.
OBSERVATIONAL
E3-BRIEF
Daily Brief v1 (AI free-picks from the watchlist) adds value vs SPY.
Negative at the Aug 3 checkpoint; −1.26 (7d, n=94) at the 30-day read. Product retired Aug 12 — see Retirements.
NEGATIVE
E2 — closed (Jun 1 – Jul 17, 2026; addendum Jul 19)
Ran across three sub-regimes: melt-up tail, geopolitical whipsaw (strikes, FOMC), and the relief rally. Closed with an addendum after a scoring-corruption repair; the addendum had power to confirm or mark low-confidence only, never to reverse. Zero verdict reversals.
E2-GATE
Regime-gated long picks beat SPY at 7 days.
+2.93 avg excess, n=158 at close. The addendum's rescored master table landed on the same number.
CONFIRMS
E2-H4
The edge is a 7-day edge; it decays by 30 days.
Completed 30-day cohort −1.41 — the picks that won at 7d faded by 30d, reinforcing the pre-registered horizon.
REINFORCES
E2-RELIEF
Relief-rally cohort at 30 days (journal cell, not a hypothesis).
25% beat rate / −8.97 avg, n=40 — matured directly into the July semiconductor rout. Stamped as hostile-tape context per the pre-committed journal rule, and displayed, not excused.
HOSTILE MARKET
WHIP-1 — one historical run of Rule A's exit (Sep 25, 2026)
The TSP rule's one-close exit against confirmation rules, on SPY's history split before the run into a discovery window (to 2009) and a replication window (2010 on), run once together. Measurement only: a better arm is an E5 candidate, to be tested forward under its own charter, never adopted from history.
WHIP-1
WHIP-1 — does a confirmation rule cut Rule A's false exits without costing more than it saves? Primary: A2 (two closes) against A0 (as built), on a discovery window to 2009 and a replication window from 2010.
One historical run on 2026-09-25, published as run: SPY's closes, Rule A's own entry, five exit arms; the primary comparison A2 (two consecutive closes) against A0 (as built, one close), named before any data. Discovery 1993-11-11 to 2009-12-31: A0 38 of 45 exits whipsawed (84.4%), A2 26 of 32 (81.3%); maximum drawdown A0 24.8%, A2 21.8%. Replication 2010-01-01 to 2026-09-24: A0 24 of 32 exits whipsawed (75.0%), A2 18 of 26 (69.2%); maximum drawdown A0 18.3%, A2 18.4%. Secondary, exploratory: B2 clears the charter's bar — by §5 an E5 candidate, to be tested forward under its own charter. Not adopted. Exploratory: one of 19 tests this year; the chance of a false positive rises with k. The only hypothesis designated PRIMARY before its data existed was E4-H1. Recorded 2026-09-25 from docs/research/whip-1-result-2026-09-25.md (scripts/whip-1.mjs at commit 8f47f34), printed as run.
FAILS
Retirements
Products killed by their own scoreboard. Records preserved in full on the Track Record page; v1 and v2 numbers are never blended.
BRIEF-V1
Daily Brief v1 — AI-selected picks from the user watchlist.
Failed the Aug 3 read (negative excess vs SPY), confirmed unhealthy at the Aug 20 read. Retired Aug 12, 2026. Replaced by v2, whose picks are restricted to the scanner's own validated candidate pool and logged as a fresh cohort under its own tag, starting at n=0.
RETIRED
SCAN-V2
Legacy scanner-v2 demo logging.
Disabled Jul 2, 2026. Its era's corrupted outcome rows (677) were deleted in the all-time cleanup with the deletion documented — the only rows ever removed, removed for being unscoreable, not for being losses.
RETIRED
Data integrity — caught and repaired
A ledger that only listed wins would be marketing. These are the bookkeeping failures found and disclosed, each fixed or waiting on a stated date. Since Sep 3 a nightly integrity check grades every lane at 11:45 PM ET and writes the grade to a log the public page reads.
OCT-2
The scan compares a stock's and SPY's returns over the same number of bars, so a name missing a day is set against different days, and it scores a name whose latest bar is a day older than SPY's as if it were current. Found by the Oct 1, 2026 figure audit; not yet corrected.
The correction (Instrumentation Amendment 6) is built and held back. Three registered tests (E5-X1, G-R1 and E5-RS) freeze what the scan reads and which calls it logs until their reads, and by their charters a correction going live inside a window would end that window. It starts on a date stamped after the E5-RS read (Apr 13, 2027), unless a later charter freezes the scan again. Until then every logged call records the date of its own last bar and of SPY's, and since Oct 2 how many names were a day behind that night, so each read can count the calls a late bar touched. The casefile's weekly and monthly trend cells, which no test reads, were corrected on Oct 2, 2026.
DISCLOSED
OCT-1
From Aug 12 to Oct 1, 2026 the public Track Record's 'Daily Brief v1 (retired)' row counted the live v2 brief's calls with it (37 of 147 by Oct 1), because the page groups calls by type and both briefs log type 'brief'. The ledger's promise that v1 and v2 numbers are never blended was broken on that page for 50 days. Found by the Oct 1 figure audit; split the same week, route-side; no row changed, nothing re-graded.
Split on Oct 1, 2026 by the source each call was logged under (Fable, afternoon review T1-1). Measured on Oct 1: the retired v1 brief alone, 110 calls, 85 graded at 7 days, 38.8% wins (the blended row showed 41.6%); the live v2 brief, 37 calls, 28 graded at 7 days, 50.0% wins, now its own row. The resolved-call appendix labels each v2 call as v2. Found the same evening and split the same way: the brief's watchlist calls log as type 'ai_commentary', so the 'AI watchlist alerts (retired)' row had counted the live v2 brief's 65 watchlist calls (Aug 12 – Sep 30) with v1's 177; v1 alone 134 graded at 7 days, 45.5% wins (the blended row showed 45.3%); v2 38, 44.7%, now its own row. Corrected Oct 2, 2026: this row first said Aug 13 and 49 days, the UTC date of the first v2 brief. The platform dates by New York, where that brief was logged at 8:01 PM on Aug 12.
DISCLOSED
OCT-5
The scorer graded a call by the clock: once it was 7, 30 or 90 times 24 hours old. A call logged after the 11 PM scoring run had begun (a catch-up or a late repair) missed its seventh night and was priced a night late, yet stored as its 7-day result; from the spring clock change, 61 trading nights carried the same exposure for one window.
Corrected under Instrumentation Amendment 4 (grade by night), stamped Oct 1, 2026: from calls logged on New York night Oct 5, 2026, a call is graded at the first scoring run on or after its logging night plus the window, whatever minute it was logged (Scorer Fix 1, lib/grade-night.js). Calls logged before Oct 5 keep the clock rule, and no row graded before the stamp is re-graded. A read-only check of the record since Aug 20 found one night affected: 34 calls logged at 11:10 PM on Aug 28, 2026 (30 AI second-opinion, 4 Daily Brief). Their 7-day results fell on a weekend and are priced at the same Friday close either way; their 30-day results were taken at Monday Sep 28's close instead of Friday Sep 25's, and stand as recorded. The core scan, the B1 lanes, universe-x1 and the TSP lane logged no call after 11 PM. Read notes print a graded-late count for any window that includes nights before the start.
DISCLOSED
SEP-29
The casefile's earnings section measured each report's price reaction and the +5d and +30d drift from the day the fiscal quarter ended, about a month before the company reported, from May 8 to Sep 29, 2026.
The EPS history feed dates each row by the quarter's last day, and the route measured from that day. Found Sep 29 by a read of the feed's dates; corrected the same day (e96856a): the report date now comes from the reports feed, each move runs from the close before the report date to the close of the session after it, and a quarter with no report date found shows no price figures. No stored record or experiment read these figures; the casefile displayed them.
DISCLOSED
SEP-27
The calibration record counted a call logged with no conviction score as a score of 0, from the panel's first day to Sep 27, 2026.
Found by the G-R1 review on Sep 26 and corrected Sep 27 (846fca7): a missing, empty or blank score now counts as missing, and a real 0.0 stays a score. The buckets on /receipts, /track-record and each casefile's "This score's record" move by exactly the calls that had no score; nothing else changes. Disclosed here because the panel is public.
DISCLOSED
SEP-11
The core scan logged 9 longs and 9 shorts on Sep 11, 2026, two short of its 10 a side, and no cause was recorded.
Found by the nightly integrity check that night (graded FAIL) and written here on Sep 30 after a read-only audit of the record, which found it the one short night not yet disclosed. The 18 calls stand as that night's picks; the E4 exit read's fixed set (438 picks, counted on Sep 23 before the verdict) includes them as logged. The nightly guards that would now refuse a thin night, retry the names that did not score, and keep the same name off both sides came on Sep 22.
DISCLOSED
SEP-3
Production stuck on an Aug 30 build for four days: a truncated integrity-check file failed every deploy.
Rebuilt Sep 3; the check now writes integrity_log nightly. A holiday weekday graded FAIL until Sep 5 — it now grades SKIPPED from the same trading calendar every logging lane uses.
DISCLOSED
AUG-4
Benchmark-snapshot gap: the TSP logger wrote no SPY snapshot after Jul 21; the Daily Brief snapshot was context-dependent; a zero-value bug ate exactly-flat returns.
Found because the Aug 3 read's cohort and gate averages disagreed by 0.74 pts. Root-caused to three bugs within 24 hours, repaired with official EOD closes (63 predictions stamped, 30 outcomes re-based, zero scoring windows contaminated). A snapshot-health watchdog now runs in the read tooling.
DISCLOSED
JUL-20
Regime-gate auth bug: server-to-server calls carried no session, silently neutralizing the macro tilt.
Caught hours before E3 Day 0; fixed by threading authentication through every caller. Disclosed because a silently-neutral gate would have invalidated E3-H1.
DISCLOSED
JUL-19
E2 scoring corruption: a scorer zero-fill wrote fake outcomes (a −84.59% "loss" on a real +9.66% win).
Rescored against real window-end closes under a pre-committed addendum with confirm/low-confidence power only. True numbers for the Jun 15 cohort were worse than the corrupted ones and were recorded as-is.
DISCLOSED
Next gates
Nothing about the live methodology changes between reads. The deploy freeze of Sep 19–24 ended with the three reads recorded on Sep 25; the reads still to come are dated below, fixed at registration.
E5-X1-READ
E5-X1 read, Nov 4, 2026 (Day 0 Sep 29; last pick night Oct 26; prices up to Nov 2).
Stamped Sep 26, 2026. The read opens at 00:00 New York time on Nov 4, owner-only, on the /stop-target read of record (?read=e5x1); the set is counted on Oct 27 (rows only, no prices, no outcomes). If the builds had not been live by the evening of Sep 28, Day 0 and every date would have moved by the charter's rule.
SCHEDULED
E5-PM-READ
E5-PM read, Dec 3, 2026 (Day 0 Oct 28; pick nights Oct 27 – Nov 23; prices through Dec 1).
Stamped Oct 22, 2026 (docs/research/e5pm-charter.md). The read opens at 00:00 New York time on Dec 3, owner-only (?read=e5pm on the /stop-target read of record); the set is counted on Nov 24 after 9:30 AM ET (rows only, no prices, no outcomes). Until E5-PM-H1 is recorded the ordinary stop/target diagnostic stops at the picks of Oct 26, for everyone, even after E5-X1 is recorded.
SCHEDULED
G-R1-W1-READ
G-R1 first read, Dec 9, 2026 (Day 0 Oct 5; window Oct 5 – Nov 30, 2026): do re-drawn score bands rank 7-day outcomes on long calls, on the core list and on universe-x1 separately?
Stamped Sep 26, 2026; dates fixed at the fill. The bands were drawn from logged scores alone (Sep 14–25) before any window call existed, and their cut-offs stay in the stamped charter, never on a page. The calibration panel is frozen from Day 0 to the last read. A second read, Feb 8, 2027, runs only if this one confirms; only that one could return the letters.
SCHEDULED
G-R1-W2-READ
G-R1 second read, Feb 8, 2027 (window Dec 1, 2026 – Jan 28, 2027), only if the first read is recorded CONFIRMS.
Stamped Sep 26, 2026; dates fixed at the fill, before any window call existed. The same test on a fresh window, and the only read that could return the letters. If the first read is anything but CONFIRMS, this read does not run.
SCHEDULED
E5-RS-READ
E5-RS read, Apr 13, 2027 (Day 0 Oct 12, 2026; night 120 Apr 5, 2027; rows still ungraded after Apr 20, 2027 left out and counted).
Stamped Sep 27, 2026; dates fixed Oct 1, 2026 by the charter's rule from Day 0. If Day 0 moves, every date moves by the same rule and is written here before any Day 0 row exists.
SCHEDULED
C1-CHARTER
Universe-C1 (crypto), revised charter: to be stamped after the E5-X1 read on Nov 4, 2026.
Stamped Aug 21, 2026 for three coins and never started. Revised before its Day 0: a rule-drawn list of coins judged against Bitcoin, its own cohort and its own read, never blended with the stock record. Rows logged before the revised stamp are a rehearsal of the pipeline and enter no verdict; their results are sealed from everyone.
SCHEDULED
E4-X2-READ
E4-X2 read of record, on or after Jan 22, 2027.
The measure was defined on Sep 24, 2026, before any exit data (parameters in lib/stop-target-walk.js); the Sep 25 read records E4-X2 as UNTESTABLE, and this later read is the one of record.
SCHEDULED
E-A-READ
E-A first read, Jan 11, 2027 (nightly shadow cross-sections; Day 0′ Sep 18, 2026).
Registered Sep 14, 2026 under Instrumentation Amendment 3; a confirmatory second read follows. Dates are held in lib/experiments/registry.js and shown on /experiments.
SCHEDULED
E-D-READ
E-D first read, about Mar 8, 2027, after 120 nightly rows (the regime shadow; Day 0 Sep 15, 2026).
Registered Sep 14, 2026. The read waits for 120 nightly rows: Mar 8 is trading day 120 from Day 0, so a missed night moves it later. A pass is not adopted on this read; the labels become a candidate gate input only after a second, confirmatory read. The date and the rule are held in lib/read-notify.js.
SCHEDULED
E-B-READ
E-B read, Mar 11, 2027.
Registered Sep 14, 2026; the read date was fixed at registration.
SCHEDULED
E-A-CONFIRM
E-A confirmatory read, Apr 12, 2027 (a fresh window of picks, Dec 14, 2026 – Mar 11, 2027).
Registered Sep 14, 2026 with the first read: the same test at a stricter bar, on picks the Jan 11 read never saw. Only a pass on both reads changes the ranking key. The date is held in lib/experiments/registry.js.
SCHEDULED
E-D-CONFIRM
E-D confirmatory read, about Aug 27, 2027, after 120 further nightly rows (240 in all from Day 0, Sep 15, 2026).
Registered Sep 14, 2026 with the first read: the same test on the regime shadow, and the only read that can make the HMM labels a candidate gate input — their 30-night separation must beat the live labels' with t ≥ 3. Aug 27 is trading day 240 from Day 0 on the shared market calendar (the charter estimated about Aug 26), so a missed night moves it later. The rule is held in lib/experiments/hmm-read.js.
SCHEDULED
E-C-READ
E-C read, Sep 13, 2027.
Registered Sep 14, 2026; the read date was fixed at registration.
SCHEDULED
The read is coming
4 reads are registered with their dates fixed in advance: E-A on January 11, 2027, E-B on March 11, 2027, E-C on September 13, 2027 and E-D on March 8, 2027. Ask for the one that interests you and the day it lands you get the pre-registered question, the verdict as scored, and this page.
Each method, and every test it has faced
from the verdict trail aboveThe nightly scanner (core list)
Each trading night it scores about 120 stocks and funds and logs its ten strongest longs and ten strongest shorts before the outcome.
Its first test passed in July: long picks the regime gate allowed beat SPY at 7 days (E2-GATE). Out of sample it has not shown an edge (E3-H3), its grades did not rank outcomes (E4-H3), and on B1’s registered score a plain relative-strength ranking beat it (B1-H3); on the fairer score, which counts a correct short as a gain, the two finished within half a point, and that comparison is being repeated sealed (E5-RS). One rule about the day after a pick passed (E4-X1) and is being repeated blind (E5-X1).
What it reads trend, momentum, where the price sits in its recent range, the quality of the set-up, trading volume, confluence of signals, supply and demand zones, strength against the S&P 500, chart and candlestick patterns, the market regime.
20 tests, with their verdicts
E5-X1-H1
Day-1 kill rule, forward: among core and universe-x1 picks logged Sep 29 – Oct 26, 2026, a pick whose first full trading day does not close in its favour (a flat close included) is marked KILLED-D1. Pass if that cohort's stop rate beats its target rate by ≥ 15 pts, n ≥ 100, open picks kept in the denominator; walked on daily prices up to Nov 2 by the same rules as E4-X1.
E5-X1 — sealed (Day 0: Sep 29, 2026 · read Nov 4)
SEALED
G-R1-W1
G-R1, first window (Oct 5 – Nov 30, 2026): on the core list and on universe-x1 separately, long calls in the upper score band beat those in the lower band by at least 0.5 pt at 7 days, with the middle band between them, n ≥ 30 per cell.
G-R1 — sealed (Day 0: Oct 5, 2026 · reads Dec 9, 2026 and Feb 8, 2027)
SEALED
E5-RS-H1
The live composite beats the relative-strength lane (B-RS) by at least 0.5 pt at 7 days, as called (each short's result flipped, longs and shorts pooled), n ≥ 150 per lane, on the counted nights of Oct 12, 2026 – Apr 5, 2027 (the as-called measure registered on Sep 23, 2026).
E5-RS — sealed (Day 0: Oct 12, 2026 · read Apr 13, 2027)
SEALED
E4-H1
Range-state conditioning (lead): COILED picks beat TRENDING-WEAK picks by ≥ 1.0 pt at 7d, COILED positive with win rate > 50%, n ≥ 100 per cell.
E4 — live (Day 0: Aug 21, 2026 · read Sep 20)
FAILS
E4-H2
Ranking-inversion persistence: long book minus short book at 7d, n ≥ 150 per side.
E4 — live (Day 0: Aug 21, 2026 · read Sep 20)
NO SIGNAL
E4-H3
Grade monotonicity: A ≥ B+ ≥ B at 7d (longs) with A−B separation ≥ 0.5 pt, n ≥ 30 per cell.
E4 — live (Day 0: Aug 21, 2026 · read Sep 20)
FAILS
E4-X1
Day-1 kill rule: a pick whose first full trading day closes against entry is marked KILLED-D1. Pass if that cohort's eventual-stop rate exceeds its eventual-target rate by ≥ 15 pts, n ≥ 100.
E4 — live (Day 0: Aug 21, 2026 · read Sep 20)
CONFIRMS
E4-X2
Day-5 time stop: picks unresolved at the day-5 close are marked EXPIRED-D5. Pass if their forward expectancy from day 5 is ≤ +0.10R, n ≥ 80.
E4 — live (Day 0: Aug 21, 2026 · read Sep 20)
UNTESTABLE
B1-H1
The live composite beats the equal-weight lane (B-EQ) by at least 0.5 pt at 7 days (raw convention, longs and shorts together), n ≥ 150 per lane.
B1 and A1 — read (Day 0: Aug 24 / Aug 25, 2026 · read Sep 25)
INDISTINGUISHABLE
B1-H2
The live composite beats the trend-and-momentum lane (B-TM) by at least 0.5 pt at 7 days, same terms.
B1 and A1 — read (Day 0: Aug 24 / Aug 25, 2026 · read Sep 25)
INDISTINGUISHABLE
B1-H3
The live composite beats the relative-strength lane (B-RS) by at least 0.5 pt at 7 days, same terms.
B1 and A1 — read (Day 0: Aug 24 / Aug 25, 2026 · read Sep 25)
NEGATIVE
A1-H1
Disagreement as a filter: scanner picks the second-opinion model endorses beat the ones it rejects by ≥ 1.0 pt at 7d, n ≥ 100 endorsed / 60 rejected.
B1 and A1 — read (Day 0: Aug 24 / Aug 25, 2026 · read Sep 25)
FAILS
A1-H2
Calibration: Brier score < 0.25 and below the base rate; overconfidence gap ≤ 15 pts against the prior 53.5.
B1 and A1 — read (Day 0: Aug 24 / Aug 25, 2026 · read Sep 25)
FAILS
WHIP-1
WHIP-1 — does a confirmation rule cut Rule A's false exits without costing more than it saves? Primary: A2 (two closes) against A0 (as built), on a discovery window to 2009 and a replication window from 2010.
WHIP-1 — one historical run of Rule A's exit (Sep 25, 2026)
FAILS
MIRROR
Replay of the frozen scanner over the E3 window lands within ±1.5 pts of the forward result.
Mirror Backtest — run Aug 22, 2026 (diagnostic)
CONSISTENT
E3-H3
The scanner's 7-day edge persists out-of-sample (confidence interval above 50, n ≥ 30).
E3 — read complete (Day 0: Jul 20 · 7-day checkpoint Aug 3 · 30-day read Aug 20, 2026)
FAILS
E3-H1
The regime gate discriminates out-of-sample: the approved side beats the blocked side by ≥ 2.0 pts.
E3 — read complete (Day 0: Jul 20 · 7-day checkpoint Aug 3 · 30-day read Aug 20, 2026)
FAILS
E3-H2
A+ conviction picks outperform A picks by ≥ 1.5 pts (pooled n ≥ 30).
E3 — read complete (Day 0: Jul 20 · 7-day checkpoint Aug 3 · 30-day read Aug 20, 2026)
UNTESTABLE
E2-GATE
Regime-gated long picks beat SPY at 7 days.
E2 — closed (Jun 1 – Jul 17, 2026; addendum Jul 19)
CONFIRMS
E2-H4
The edge is a 7-day edge; it decays by 30 days.
E2 — closed (Jun 1 – Jul 17, 2026; addendum Jul 19)
REINFORCES
It does not place trades, size positions or know anything about a member’s account, and its calls are descriptions, not advice.
The second list (universe-x1)
The same nightly method applied to a second list of names, logged beside the core list and never blended with it.
It was tested once against the core list and performed within a point of it (E3-H4). It is part of two sealed tests now running.
What it reads trend, momentum, where the price sits in its recent range, the quality of the set-up, trading volume, confluence of signals, supply and demand zones, strength against the S&P 500, chart and candlestick patterns, the market regime.
3 tests, with their verdicts
E3-H4
The universe-x1 expansion list performs within 1 pt of the core list.
E3 — read complete (Day 0: Jul 20 · 7-day checkpoint Aug 3 · 30-day read Aug 20, 2026)
CONFIRMS
E5-X1-H1
Day-1 kill rule, forward: among core and universe-x1 picks logged Sep 29 – Oct 26, 2026, a pick whose first full trading day does not close in its favour (a flat close included) is marked KILLED-D1. Pass if that cohort's stop rate beats its target rate by ≥ 15 pts, n ≥ 100, open picks kept in the denominator; walked on daily prices up to Nov 2 by the same rules as E4-X1.
E5-X1 — sealed (Day 0: Sep 29, 2026 · read Nov 4)
SEALED
G-R1-W1
G-R1, first window (Oct 5 – Nov 30, 2026): on the core list and on universe-x1 separately, long calls in the upper score band beat those in the lower band by at least 0.5 pt at 7 days, with the middle band between them, n ≥ 30 per cell.
G-R1 — sealed (Day 0: Oct 5, 2026 · reads Dec 9, 2026 and Feb 8, 2027)
SEALED
It does not place trades, size positions or know anything about a member’s account, and its calls are descriptions, not advice.
The daily brief
An AI-written daily summary; its calls are logged before the outcome and graded like the scanner’s.
Its first version, which picked from members’ watchlists, did not add value against SPY (E3-BRIEF) and was retired. The current version has no registered test yet; its calls are graded on the public record like every other.
What it reads the night’s strongest long calls from the core scan, the only names whose calls it logs, when a member opens it, also their watchlist, the market regime reading and the risk limits they set.
1 test, with its verdict
E3-BRIEF
Daily Brief v1 (AI free-picks from the watchlist) adds value vs SPY.
E3 — read complete (Day 0: Jul 20 · 7-day checkpoint Aug 3 · 30-day read Aug 20, 2026)
NEGATIVE
It does not place trades, size positions or know anything about a member’s account, and its calls are descriptions, not advice.
The TSP fund calls
Daily allocation calls on the Thrift Savings Plan funds, graded at 7, 30 and 90 days against market-traded stand-ins for each fund.
No registered test has been read for it yet. Its calls are logged and graded on the public record.
What it reads each fund’s market-traded stand-in price alone: where it sits against its own moving averages, and whether it has risen over recent months.
It does not place trades, size positions or know anything about a member’s account, and its calls are descriptions, not advice.
How the record is kept
- Every directional call is written to the database at the moment it is made, with the symbol's price and SPY's price locked in.
- Calls log only in a fixed evening window after the daily bar has closed; a daytime visit logs nothing.
- A scheduled job scores each call at 7, 30 and 90 days against SPY. Nothing is edited retroactively; no loss is deleted.
- The first run of each trading day is canonical; re-runs are flagged as duplicates and excluded from every read.
- Prices are compared to official end-of-day closes; a snapshot-health watchdog runs before every read.
How the methods are tested
- At most three experiments run at once. The three slots opened on Sep 25, 2026, when E4, B1 and A1 were read; the next charters are stamped one at a time: E5-X1 (Day 0 Sep 29, 2026), then G-R1 (Day 0 Oct 5, 2026), then E5-RS (Day 0 Oct 12, 2026), which held all three; the cap of three was lifted on Oct 22, 2026, by Steve (docs/research/experiments-registered-2026-09-14.md, the Article 1 amendment), and the verdict clock of a charter now starts at its own Day 0, each charter stating how many hypotheses are live at its stamp and correcting its pass rule for that count.
- When k charters run at once, each primary read is held to alpha ÷ k (Bonferroni), stated in the read — for charters stamped from Oct 1, 2026.
- No methodology changes mid-experiment. Weights, gate logic and universes change only through a charter, after a read.
- Incubation gate: no signal is displayed to users until it has ≥ 30 sealed forward days and a passing read.
- Replication before adoption: a result must repeat out-of-sample before it changes the product.
- Every read runs the six-question audit first: lookahead, leakage, outliers, regime dependence, chart-vs-number spot checks, n / confidence-interval honesty.
The gate's own receipt
30 nights recordedNights recorded
30
since Sep 14
Gate open
4
regime mapped to bull
Gate parked
26
any other mapped regime
30 nights recorded so far. The receipt is read at 60 recorded nights and again at 120, the first about January 21, 2027 if no night is missed. Nothing is shown before the first read.
Open means the recorded regime mapped to bull; parked means any other mapped regime; a reading the gate does not map is counted apart and is in neither cohort. Each figure is SPY’s log return from that night’s close over the next 1, 7 and 30 trading days; a night whose window has not closed has no figure and is not in n. The t is Welch’s, between the two cohorts, reported as a number. Reads are taken at 60 and 120 recorded nights and then every 60, on the first that many nights, once the last of them has a closed 30-day window; nothing is published between reads. This says what the index did on each kind of night. It does not score the gate, and it is not advice.
Win rate is not return: a 60% win rate says nothing about position sizing or the magnitude of wins versus losses. Win rates below n=30 are not conclusive. Past results do not guarantee future results. Educational information, not investment advice. Symbols, picks and sealed scoreboards are not shown on this page by design; the Track Record lists every call once you are signed in.