Entries are added weekly and never polished after the fact. If a finding later turns out to be wrong, the entry stays and the correction appears below under “Corrected”.
19 Oct 2026 – 27 Oct 2026DOES NOT HOLD
Queue position: visible, but without predictive power
PRE-REGISTERED
Fixed on Oct 13 and 14, before an evaluator existed: we measure how many contracts stand ahead of the largest resting order (20 lots and up, within 2 ticks of the best price) in its queue — from market-by-order data only. Two questions. (1) Does the position improve the direction forecast on top of the 61 CORTEX features? Hurdle +0.005 AUC. (2) Does it predict whether the order disappears within 60 s without a trade — beyond size, distance, age, level size and activity? Hurdle +0.01 AUC. Both additionally against 199 random permutations and per volatility layer.
WHAT THE DATA SAID
Question 1, 10 trading days (Oct 13 to 26), 22,923 out-of-sample samples: AUC 0.527 with position against 0.528 without — −0.0009, not above the chance bound. Question 2, 10 trading days (Oct 12 to 23), 7,949 out-of-sample orders: AUC 0.744 against 0.739 — +0.005, real and not a volatility echo, but only half the hurdle.
CONSEQUENCE
Does not hold — in both questions, by the rule fixed in advance. The position stays a display (“this many contracts ahead of it”) without interpretation, no new feature, no restart of the learning state. The idea arose at a single price level on screen; ten days of measurement give it no edge.
Pre-registration: s526-queue-position · 13 Oct 2026 | s528-order-schicksal · 14 Oct 2026
16 Oct 2026 – 27 Oct 2026OPEN
The learning verdict is collecting: 8 of 10 sessions of one learning state
PRE-REGISTERED
Fixed on Oct 15 (revision 3): only sessions of the same engine with no restart in between count; the learned and the traded question are measured against their base rate. The verdict comes at 10 sessions — until then the tool prints no result figure.
WHAT THE DATA SAID
The learning state restarted on Oct 16, and so did the count. Since then 8 sessions have been counted (Oct 16 to 27), all from the same cohort; both endpoints read “collecting”. The raw direction quoted in week 42 describes the state before this restart and is not part of the verdict.
CONSEQUENCE
The first verdict comes after the 10th session — later than announced in week 42, because the restart reset the count. Until then we publish the count only.
Pre-registration: s509b-lern-urteil · 15 Oct 2026
13 Oct 2026 – 16 Oct 2026OPEN
The diagnosis: CORTEX learns a different question from the one that is traded
PRE-REGISTERED
Newly fixed on Oct 15 (revision 3, stamped): a learning verdict counts only sessions of the same engine with no restart in between, and it measures the traded question against its base rate alongside the learned one. Verdict at 10 sessions.
WHAT THE DATA SAID
Raw direction on Oct 15: 48.9% (47.0 to 50.7, n = 2,781) against a cost line of 59.4%; at 70% confidence and above only 47.2%. The deciding head effectively learns a ±3-tick target, while +8 against −10 to −14 ticks is traded — even a perfect oracle on it would hit the traded question only about 53% of the time; break-even is 55.6%. The most promising of twelve challengers was 22.6 points below the champion after 9 of 10 days.
CONSEQUENCE
Status 3 of 10 sessions, first verdict around Oct 27. The learning check runs every night. The next rebuild thus has a precise brief: align the learned question with the traded one — pre-registered, with its own restart. A lead from order size (small orders sell every day, blocks buy) has been pre-registered since Oct 15 and is not a finding.
Pre-registration: s509b-lern-urteil · 15 Oct 2026
9 Oct 2026 – 13 Oct 2026HOLDS
Market-by-order: the first pre-registered gain from a new data block
PRE-REGISTERED
Fixed before computing: 49 against 61 features on the same 10 days, out of sample. The gain has to exceed 0.01 AUC, lie above the chance bound and pass the volatility gate (at least five strata per day).
WHAT THE DATA SAID
21,221 samples: AUC 0.5365 → 0.5552, gain +0.0187, chance bound −0.0146; stratified 0.524, positive on 80% of days. Along the way: four features with “MBO” in their name had until then been fed by an estimator from the old book. In the first census 35–45% of change messages arrive twice; a rebuilt book was crossed in 36% of seconds without protection.
CONSEQUENCE
Since Oct 13 CORTEX computes with 61 features (engine 13); market-by-order is the first input, the previous book the fallback layer — with a deliberate restart of the learned state. The crossing repair is live (36.46% → 0.03%). Important: 0.555 measures separation, not profit. Four further MBO questions are pre-registered.
Pre-registration: s525-weglass-61 · 13 Oct 2026
1 Oct 2026 – 13 Oct 2026REBUILT
The model that stood still — faults no alarm reported
PRE-REGISTERED
There is no pre-registration for this: it is a finding about the measurement basis, made while checking whether the model learns at all. We report it anyway because it invalidates every learning statistic for the affected days.
WHAT THE DATA SAID
The trainer's memory grew by about 25 MB per learning round; after roughly 25 minutes learning stopped — with no error message, while the acceptance test reported 10 of 10 green. On Oct 1 twelve features had also been changed without a restart: the model learned two definitions at once. And on days with a feed reconnect a context feature counted every tick twice.
CONSEQUENCE
Training on the CPU (660–760 MB instead of 3,505 MB, shown about 358 learning steps per trading hour), restart on Oct 9, a fixed rule “feature changed ⇒ restart”, a guard for the double count. Learning now uses strategy trades only. A “freshness witness” checks whether the content of the book moves — not whether data arrives.
6 Oct 2026 – 9 Oct 2026DOES NOT HOLD
Costs eat the information: holding horizon, limit entry and target width
PRE-REGISTERED
Fixed in advance per study: only what lies entirely above zero after costs in the 95% interval gets built — in every cell of the grid, not in the best one. Before that, two independent reviewers had assessed 15 trading days: −0.96 ticks per trade, profit factor 0.76, stated confidence 74.4% against a 49.1% hit rate.
WHAT THE DATA SAID
Holding horizon (49 days, 265,668 samples): at 90 s the top decile stands at +0.07 ticks (−0.68 to +0.65) — about 1.1 to 1.5 ticks above every null model, but costs of about 1.5 ticks eat exactly that. Limit instead of market: −1.94 ticks; fills come at 80–87% exactly where the price runs against the trade. Tighter target in the replay (258 trades, 46 days): no variant better.
CONSEQUENCE
Do not build in 12 of 12 cells, for holding horizon and for limit alike. What we learned is where the loss happens: the features carry information, and it is lost between information and execution. Every further idea has to beat the cost line, not the coin flip.
Pre-registration: s512-horizont · 9 Oct 2026 | s512-limit-entry · 9 Oct 2026 | s515-scalp-bracket · 9 Oct 2026
8 Oct 2026 – 9 Oct 2026DOES NOT HOLD
A deep network with two time scales sees the flow — not the direction
PRE-REGISTERED
Fixed before the first run: a causal network with a short and a long time scale (275,593 parameters, pre-trained on 49 days) has to improve the out-of-sample prediction — against a random encoder, in at least 3 of 4 time blocks and beyond a permutation bound.
WHAT THE DATA SAID
Improvement −0.0001 (95%: −0.0020 to +0.0017), a random encoder does just as well, 2 of 4 blocks positive, 24 test days, three seeds. What the network learns is the persistence of the aggressor side (explained variance 0.119 over 15 minutes) — not price direction (≤ 0.02).
CONSEQUENCE
Verdict no, in revision 3. Before the verdict we found two data errors (null quotes on 9 days, 154 seconds of a different contract on one day); the corrected version reached the same result. A repeat comes at 150 days at the earliest.
Pre-registration: s513-dl-kontext · 9 Oct 2026
23 Sept 2026 – 7 Oct 2026DOES NOT HOLD
CORTEX does not choose the first target better than a fixed width
PRE-REGISTERED
Fixed on Sep 23: the T1 head may choose the first target per market state if the 2.5% bound of its AUC lies above 0.55 and the deciles rise monotonically — for 8 ticks and for a width of 12 ticks or more.
WHAT THE DATA SAID
10 days, 41,367 rows. AUC at 6 ticks 0.511, at 8 ticks 0.525 (lower bound 0.490), at 12 ticks 0.541 (0.490), at 14 ticks 0.546 (0.487); the deciles are not monotonic. A sweep over 183 trades and 34 days found no fixed width significantly better than 8 ticks.
CONSEQUENCE
T1 stays fixed at 8 ticks. The head keeps learning alongside but decides nothing. A re-measurement with finer volatility strata found a small, stable residual (AUC 0.548) — below the bar, so no abort head.
Pre-registration: s473-t1-wahl · 23 Sept 2026
28 Sept 2026 – 29 Sept 2026DOES NOT HOLD
Profile lines, time of day, order flow: no edge at the line
PRE-REGISTERED
Three questions, each fixed before any calculation — on days already recorded, which we state openly; new days are checked separately and never pooled in. (1) Does price bounce more often at profile lines (POC, value area, prior-day and prior-week levels) than at placebo lines? (2) Does the bounce at neutral lines depend on time of day or volatility? (3) Does the order flow of the 5 seconds before the touch predict the bounce? Yardstick every time: ±8 ticks within 120 seconds.
WHAT THE DATA SAID
(1) Profile lines 51.5% bounce, placebo 53.0% — difference −1.5 points, 95% interval −6.5 to +3.4 (36 days). (2) Midday vs. open −1.2 points, quiet vs. fast +1.6 points, both intervals include zero (8,896 touches, 39 days). (3) Aggressor pressure 0.0 points (−3.4 to +3.5), order-flow imbalance −1.6 points (−5.3 to +2.1).
CONSEQUENCE
Three times no difference. A bounce of about 52% happens at any price, not specifically at lines. The lines stay on the chart — as orientation, not as a trading edge. What is ruled out is an advantage above roughly 3.5 points; smaller effects are not refuted, but not supported either.
Pre-registration: s486-level-reaktion · 29 Sept 2026 | s486b-rueckschwung-regime · 29 Sept 2026 | s486c-orderflow-linie · 29 Sept 2026
11 Sept 2026 – 26 Sept 2026DOES NOT HOLD
Gate 1, third round: the order book does not improve the prediction
PRE-REGISTERED
Fixed on Sept 12, before the days existed: only days under the new quote rule (tick and quote days are never pooled), target +8 ticks before −10 ticks — exactly the geometry the strategy trades —, verdict after 10 days with order book.
WHAT THE DATA SAID
10 book days (one day excluded), effectively 614 independent samples. The gain from the order book is negative out of sample in all 11 subsets (−0.001 to −0.081 AUC), none above costs. The base rate for +8 before −10 is about 56%, the break-even after costs about 64%.
CONSEQUENCE
Verdict not learnable, acknowledged on Sept 26 and confirmed the following night. We do not keep searching with a different target geometry until something turns up — the next question gets its own pre-registration with a rationale.
Pre-registration: s445-gate1-rev3 · 12 Sept 2026
22 Sept 2026 – 23 Sept 2026REBUILT
The CORTEX exit: back into shadow mode after one day
PRE-REGISTERED
The new chain of Sept 22 (entry “The new signal chain”) let CORTEX decide the exit as well — without a prior shadow run. That was the mistake. There is no pre-registered verdict for this: it was stopped after one day, far below its own minimum sample.
WHAT THE DATA SAID
On the first day with a freshly reset model the CORTEX exit ended 10 of 14 trades. Together they made +5 ticks per contract; with the fixed brackets alone the replay gives +39.5 — the exit cost 34.5 ticks. Its own exit statistic looked good, because it compared against the stop instead of against the trade without the rule.
CONSEQUENCE
Since Sept 23 the fixed brackets end the trade again. Every new exit rule runs in shadow mode first and is measured against “brackets only”; it goes live only when its advantage over at least 5 days lies entirely above zero in the 95% interval.
21 Sept 2026 – 23 Sept 2026OPEN
The new signal chain: CORTEX alone decides — and may end the trade
PRE-REGISTERED
Fixed on Sept 22, before the first data existed: more than 6.6 entries per day, a CORTEX-exit share between 10 and 30%, verdict only after 5 trading days and 30 trades. Old and new chain are never pooled.
WHAT THE DATA SAID
No verdict yet. Measured before the rebuild: 6 of the model's 49 inputs are practically silent, and no single feature separates up from down (AUC 0.47–0.51). Three throttles made the signal chain 3.3 s slow.
CONSEQUENCE
Throttles removed (≈ 0.2 s by calculation), training moved to its own process, fresh learning state for everyone with release 0.5.30. Not built, because it measured too weak beforehand: a bracket width taken from the bracket head (AUC 0.54).
Pre-registration: s469-kette-umbau · 22 Sept 2026
14 Sept 2026 – 20 Sept 2026DOES NOT HOLD
The weekend study's last candidate falls — despite a 65.6% hit rate
PRE-REGISTERED
In a daily forward test since Aug 29, frozen: “flush absorption” (price drops ≥ 12 ticks in 10 s under strong buying → long entry). Decision at n ≥ 50 and Wilson lower bound ≥ 55% — with positive expectancy.
WHAT THE DATA SAID
At n = 90 the candidate hits 65.6% (lower bound 55.3%) — and loses $5.17 per trade. On Sept 14 (n = 52) the threshold was briefly met; the rule has no stopping at the first crossing.
CONSEQUENCE
Discarded by rule since Sept 18, as were the other two candidates. Lesson for every future pre-registration: a hit rate without expectancy is not an edge.
14 Sept 2026 – 20 Sept 2026DOES NOT HOLD
The “fireworks” in the HUD: spectacular, but not predictive
PRE-REGISTERED
Four displays in the app each got a baseline that simply bets against the trend or with the last move. They only hold if the Wilson lower bound beats that baseline — one verdict after 30 resolved events.
WHAT THE DATA SAID
Reversal signal 57.0% (lower bound 46.8%) vs 48.2%. Takeover at the inside 58.5% (n = 41). A price where pressure was absorbed holds 48.6% vs 46.5%. Strongest HUD events 45.7% vs 50.1%. None holds.
CONSEQUENCE
All stay displays — with honest legends (“pressure absorbed here — no proven barrier”). Found and fixed on the way: the wall tracker counted cancellations as absorption.
Pre-registration: s449-cortex-rev · 14 Sept 2026 | s449g-absorptions-konto · 14 Sept 2026 | s449i-hud-ereignisse · 14 Sept 2026
14 Sept 2026 – 20 Sept 2026DOES NOT HOLD
Highly significant, but too small: repeating prints as an iceberg hint
PRE-REGISTERED
Do ≥ 5 equal-size prints at one price within 60 s indicate real reloading in the order book? Required in advance: at least 10 percentage points difference to the control and p < 0.05, once, on the first 5 days.
WHAT THE DATA SAID
20,170 candidates vs 18,649 controls: 11.4% vs 8.5% — p ≈ 10⁻²², but only 2.9 percentage points.
CONSEQUENCE
No display in the app, no parameter search on the same days. The example shows what a pre-set effect size is for: significant does not mean relevant.
Pre-registration: s447d-clip-wiederkehr · 12 Sept 2026
7 Sept 2026 – 13 Sept 2026HOLDS
The exchange as referee: our delta was 30% wrong
PRE-REGISTERED
For the first time with real CME order data as ground truth: the rule holds if it assigns ≥ 90% of volume correctly on ≥ 6 of 7 days. Switch only with a confirmation day after deployment (≥ 95%, per-minute r ≥ 0.95).
WHAT THE DATA SAID
Quote rule 97.5% of volume correct, the tick rule used until then only 69.9% — with the wrong daily sign on 4 of 7 days. With a locked quote every neighbour rule is below chance. Confirmation day Sept 11: 97.0%, per-minute r = 0.985.
CONSEQUENCE
Delta switched in indicator, app and AI; locked prints count as “side not determinable” instead of guessed. Old and new days are never pooled — the learning state was restarted.
Pre-registration: s441-je-trade-join · 10 Sept 2026 | s441-gesperrt-naechst · 10 Sept 2026 | s443-quote-regel-umsetzung · 10 Sept 2026
7 Sept 2026 – 13 Sept 2026DOES NOT HOLD
Absorption and exhaustion: largely an artefact of classification
PRE-REGISTERED
“Just a label” if the tick version and exchange truth agree on the sign ≥ 90% of the time; “noise” below 75%. Second question: does the true exhaustion predict the price target better (AUC ≥ 0.53)?
WHAT THE DATA SAID
Signs agree only 74.1% of the time; the tick version reports exhaustion in 41–49% of seconds, the truth in 34%. And the true exhaustion does not hold either: AUC 0.510 vs 0.510.
CONSEQUENCE
The correct rule makes the measure true, but not useful. For the model restart that means: it costs no signal, because there was none.
Pre-registration: s441-buch-wahrheit · 10 Sept 2026 | s441-exh-wahrheit-ziel · 10 Sept 2026
7 Sept 2026 – 13 Sept 2026DOES NOT HOLD
The replay machine: would the strategy have won the other way round?
PRE-REGISTERED
105 replayable trades, tick by tick on our own tape, including costs. A variant only holds with a positive net result and a Wilson lower bound above its own break-even line. One look per variant.
WHAT THE DATA SAID
Original −1.76 ticks per contract. Reversed direction −1.73. No time stop, T1 = stop, no break-even lock: −1.36 to −1.72. Larger scale: −2.01 to −0.89. Brackets that grow with speed: −1.86 / −1.87. No variant holds.
CONSEQUENCE
The entries themselves carry no edge — no management rescues them. Strategy not reversed, brackets not enlarged. The shortcut “flip the sign” (+0.95 ticks) was retracted before any code changed.
Pre-registration: s437-inversion · 9 Sept 2026 | s439-bracket-struktur · 9 Sept 2026 | s440-skala · 10 Sept 2026 | s440-vollgas-bracket · 10 Sept 2026
31 Aug 2026 – 6 Sept 2026DOES NOT HOLD
Order flow describes the move — but does not predict the next one
PRE-REGISTERED
Fixed since July 29: the sign of the order-flow imbalance must hit the next mid-price move with a skill ≥ 1.25 against the chance line. If this stage falls, the whole cascade above it falls.
WHAT THE DATA SAID
19 days, 77,972 windows: hit rate 49.2–50.2%, skill 0.995 on all four horizons. The simultaneous move, however, is explained almost completely by the same order flow (median R² 0.893).
CONSEQUENCE
On Sept 1 stages 5 to 8 were shut down — exactly as fixed in advance. A planned order-flow entry gate was never built.
Pre-registration: s394 · 29 Jul 2026
31 Aug 2026 – 6 Sept 2026HOLDS
A well-known iceberg trader's theses against 33 days of data
PRE-REGISTERED
Exploratory, but with a fixed hurdle: candidate only above the 95th percentile of a corrected control, not worse on ≥ 60% of days and an effect of ≥ 2 percentage points.
WHAT THE DATA SAID
Sweep zones pull price back (on 25 of 33 days), sweeps through a 60-minute extreme reverse more often (55.4% vs 51.8%). Falling short: iceberg zones as anchors and the entry rule “ATR + 15%”.
CONSEQUENCE
Two candidates for a later pre-registration on new days — no code in strategy or CORTEX. Experience supplies hypotheses, not the yardstick.
Pre-registration: s429-ereigniszone · 31 Aug 2026
31 Aug 2026 – 6 Sept 2026DOES NOT HOLD
“Open road after full throttle”: the market picture does not survive the data
PRE-REGISTERED
Seven states of an intuitive market picture (full throttle, refuel stop, jam, wall) against time-matched control moments: holds above the control's 95th percentile and on ≥ 60% of days.
WHAT THE DATA SAID
436,088 seconds, 19 days: after a refuel stop the move continues less often than chance (46.5% vs 58.8%, better on 0 of 19 days). Large walls, on the other hand, measurably hold (break-through 42.9% vs 49.5%).
CONSEQUENCE
No phase module built. The opposite signature became a new, separate pre-registration on fresh days — instead of turning thresholds until the picture fits.
Pre-registration: s421-autobahn · 31 Aug 2026 | s421-autobahn-rev2 · 31 Aug 2026
24 Aug 2026 – 30 Aug 2026DOES NOT HOLD
The only positive learnability trace does not replicate
PRE-REGISTERED
A sequence model (GRU) had shown a weak trace in August (AUC 0.548 on 2 test days). Replicated only at p ≤ 0.01 on days that analysis never saw — with a working positive control.
WHAT THE DATA SAID
6 fresh days: AUC 0.5045, p = 0.366. Positive control sharp (0.796), placebo silent — an edge from AUC ≈ 0.605 would have been found. New feature sets failed as well (also in the repeat on Sept 11).
CONSEQUENCE
The August trace counts as small-sample chance. Four model classes are empty on this question; no model restart for new features.
Pre-registration: s415-sequenz-oos · 26 Aug 2026 | featrev-s415 · 26 Aug 2026
17 Aug 2026 – 23 Aug 2026DOES NOT HOLD
Filtering out small prints does not improve the delta
PRE-REGISTERED
Paired test against the raw delta: confirmed only at ≥ 3 percentage points difference and p ≤ 0.005. The exploration day on which the idea arose was excluded.
WHAT THE DATA SAID
Delta from clearly classifiable volume only: −0.55 percentage points. From prints of 5 contracts and more only: +0.63 percentage points (p = 0.40). Divergences under high rule agreement do not hold either (p = 0.19).
CONSEQUENCE
The raw delta stays — a month later the exchange truth showed that the problem lay elsewhere: in the rule itself (week 37).
Pre-registration: s407h-identifizierbarkeit · 18 Aug 2026 | s407h-divergenz · 18 Aug 2026
