Everything we tested, and what it actually showed

Does price respect a real swing high or low any more than it respects a random line the same distance away?

How it was tested. Every swing level was scored forward against a fake line pushed the same distance to either side, with that distance then mathematically extrapolated down to zero.

Real levels scored 47.6 / 49.2 / 47.5% against fake lines at 47.6 / 48.5 / 47.0%. Three independent runs put the average outcome at -0.033R to -0.036R. Both check groups agree, four calibration checks pass, and the profile is flat rather than growing with distance. An earlier +1.5pp reading was withdrawn once the fake line was fixed to differ in one respect only.

How much data

19,688 separate 1-hour trades across eight check symbols, plus a separate 24-minute run

Smallest effect it could spot

0.0115 to 0.0162R on 1-hour, about 0.6 to 0.8 percentage points; 0.0194R on 24-minute - the best-resolved test in the project

A real swing level is worth nothing over an arbitrary line the same distance from price, so use levels to decide where to pay attention, never as a reason to expect a bounce.

When price comes back into an order block, does it turn there more often than at any similar zone?

How it was tested. Every armed retest was scored against a matched fake zone - first shifted sideways, then, after that comparison proved rigged, matched instead on how deep the pullback was - and twenty combinations of the block settings were swept alongside.

The sideways-shifted comparison read +4.6 / +4.8pp and turned out to be an artefact: pullback depth on its own is worth 20pp, the fake running 41.2% in the shallowest tenth to 61.1% in the deepest while the real block sat flat near 55%. Matched properly on depth and timing: +1.6 / +1.2pp on 3-minute and -0.8 / -0.6pp on 1-hour, with a separate control saying the real block is 3.3pp worse than a matching non-swing zone. Not one of the 20 setting combinations cleared +2pp on both check symbols, and 0 of 12 quality groups cleared the smallest effect the test could spot. An archived 50.0% at 1,867 events was withdrawn as an overlapping-trades artefact, and the older "60% versus 33%" claim left nothing above about 6pp on retest.

How much data

3 symbols on 1-hour and 3-minute, 40,000 bars per symbol; 20 declared setting combinations; 12 quality groups

Smallest effect it could spot

0.040R, about 2 percentage points - still roughly 1.7 times too coarse to settle it

A block retest looks like a coin flip and the test was not sharp enough to see anything smaller, so treat any future block claim as false until it has been matched on how deep the pullback was.

Can any filter sort the order blocks that work from the ones that do not?

How it was tested. About 25 filter variants were tried across two halves of recent data on one symbol, and the two survivors were then re-run on a genuinely different symbol they had never touched.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

ATR, rising cleanly with its own threshold. On the fresh symbol the baseline was 50.0%, minimum height 50.4%, overshoot-6 45.3% and overshoot-7 42.3%. The overshoot rule did not merely fade, it inverted. Separately, ranking blocks by score instead of filtering them gave -0.01 / -0.48 / +0.40pp, and a model using all nine candle features captured only 1.6 to 2.9% of the room available.

How much data

1,101 events in the build data, 1,117 on the fresh symbol

A filter that survives both halves of one coin over one period, rises neatly with its own setting and comes with a good story can still be pure noise - it only counts once it survives a coin it has never seen.

When a broken block gets retested by a big-volume candle, does that retest actually work?

How it was tested. Every breaker retest was rebuilt offline across nine coins and three timeframes, then big-volume returns were compared with quiet ones four separate ways, including pairs taken inside the same zone.

Paired inside the same zone: -7.89 / -6.31 / -5.27pp. Not one of the 12 cells cleared its own floor, and the deepest samples lean the wrong way. On a separate run the filter read +1.15pp on the coin it was built on and inverted to -2.26pp and -0.91pp on two unseen coins. The original claim was 60% versus 33% - a 27-point gap.

How much data

2,369 resolved retests on 1-hour, 1,153 on 24-minute, 11,473 on 3-minute; nine coins; 25,345 blocks became 7,162 breakers

Smallest effect it could spot

0.0676R, about 3.4 percentage points - 2.8 times the project's best-ever finding

The 27-point gap is not in the data, so treat a breaker as a place on the chart rather than a reason to take the trade - though waiting for the return does cut your risk from about 3 ATR to about 1.73 ATR.

Is there anything special about the 0.618, 0.5 or 0.786 retracement?

How it was tested. The same leg, same bar and same direction were scored with only the ratio changed, measured from the close of the touch bar and compared against a 62-rung ladder running from 0.30 to 0.95.

Against all 62 rungs the three fib ratios ranked 8th / 19th / 34th of 62 on 3-minute and 9th / 43rd / 49th on 1-hour. The whole triple against the same triple nudged 0.01 to 0.06 came to -0.000pp. The only real pattern is depth: +1.7 to +1.8pp at a retracement of 0.44 to 0.53, sliding to -0.77pp at 0.95 - a smooth curve the fib rungs sit on unremarkably. Hand-drawn fibs scored -1.60 and -0.23 against their fake levels, both worse than the fake.

How much data

nine coins on 3-minute and 1-hour, both check groups

Smallest effect it could spot

0.32pp (0.0064R) on 3-minute; 0.009 to 0.015R on 1-hour - three to five times finer than the project's best-ever positive result

Price retraces but the golden ratio adds nothing, and if you ever backtest a fib with a resting order sitting at the level you will get a fake winner, because the fill itself only picks the moves that carried on.

Does price get pulled back to an untouched high-volume price more than to any other price?

How it was tested. Each untouched point of control was tracked over a 240-bar horizon against a fake level whose distance was drawn from a different event on the same side, so the two sat the same distance from price on average.

Corrected, the naked point of control was reached 88.4 / 88.1 / 87.5% of the time against the fake at 88.2 / 88.1 / 87.8% - a difference of +0.2 / +0.0 / -0.3pp, with the mean distances matching to two decimal places.

How much data

3 coins on 1-hour, 240-bar horizon of about ten days

Smallest effect it could spot

about 0.12R, roughly 6 percentage points - and it was the wrong measure anyway

Price reaches a naked point of control 88% of the time inside ten days, and it reaches any old price at that distance 88% of the time too, so it is a price, not a magnet.

Does a wider order block make the trade harder to win?

How it was tested. Straight arithmetic - your risk is the zone width plus a 1 ATR buffer, and risk is what R is measured against - checked against how far price actually travelled after each retest.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

The median move after a retest is 2.34 ATR from the zone edge, 75th percentile 3.92 and 90th percentile 4.56. On 3 ATR of risk the median move is 0.78R - it does not even reach 1R. Drawing the zone to the wicks rather than the bodies changed the outcome in only 2 of 15 paired setups.

How much data

22 retests; observed widths median 0.62 ATR, 90th percentile 1.16 ATR, maximum 1.60 ATR

Smallest effect it could spot

not applicable - this is arithmetic, not a measurement

A fat zone demands a bigger move for the same reward and no sample size can argue with that, so past roughly 1 ATR of width the maths stops closing whatever the zone is called - worry about width, not about wick versus body.

When the yellow dot prints, does price go my way?

How it was tested. Every dot was scored on whether price travelled one average bar's range in its favour before travelling the same distance against it, on three coins and five chart speeds, and was also compared against simply noticing the wave was already deep.

Median 19 bars before the real turn, with price running a median 3.0 ATR against you first. As a top/bottom finder a random bar marks a pivot 2.9% of the time, the dot 6.4%, 'the wave is simply deep' 9.8%, and 'deep but no dot' 10.5% — the dot loses to plain depth in 29 of 30 cells.

How much data

22,656 dots on ETH, BTC and SOL across five timeframes; 22,500 scored for direction; 18,000 oversold episodes

Smallest effect it could spot

0.021R on the race; per-cell error bars of 0.72 to 2.35 percentage points

The dot only tells you momentum has crossed; it is a coin flip that usually fires around 19 bars too early, so never treat it as an entry.

Is there a filter, a slower chart, or a waiting rule that turns the dot into something worth trading?

How it was tested. About 1,100 rule variants across six angles were screened, the seven that passed were handed to an independent re-tester who ran the control the first screen had skipped, and separately every fast-chart-to-slow-chart pairing, every waiting length from 1 to 21 bars, and resting limit orders were tested against matched controls.

On data it had never seen, the volume filter reversed: the quietest 2% of dots won 56.46% against the loudest 2% at 52.64%. Waiting failed at every length from 1 to 21 bars; limit orders lifted the win rate to 51-53%, but the same order placed from a random bar did 50.5-52.1%. Lift found on the tuning coin carried over to fresh coins at a slope of 0.20 — about 80% of it was noise.

How much data

~1,100 variants screened and 7 re-tested, over 22,656 dots; 9 fast-to-slow chart pairs x 16 agreement rules across 174 comparison cells

Smallest effect it could spot

±0.90 percentage points on the best surviving filter

Stop hunting for the setting that fixes the dot — filtering it, confirming it on a slower chart, and waiting for a better price have all been tried and none of them work.

If I simply buy when the fast wave crosses up and sell when it crosses down, what does that cost me?

How it was tested. Every crossover was scored offline on nine major coins against a control matched on six things, so the comparison bars were as alike as possible apart from the crossover itself.

How much data

Nine major coins; the signal fires about 13 times a day on the 15-minute chart

Smallest effect it could spot

0.0219R

This one is not a nothing — trading raw wave crossovers loses a small, steady amount every time, before you even pay fees.

Does the 9 crossing the 21 tell me which way price is going?

How it was tested. The forward outcome after every cross was measured on nine coins and three chart speeds, then 25 other fast/slow pairs chosen in advance were scored on the same machinery.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

on the 3-minute -0.55pp, negative on 9 of 9 coins, against a floor of about 0.012R. Across the 25-pair grid the entire spread is smaller than one cell's own measuring error, and 9/21 ranks 10th, 6th, 20th and 21st of 25. The earlier evidence for 9/21 was withdrawn — its best figure, 50.24%, sits inside its own 0.44 unit of noise.

How much data

Nine coins on 1-hour, 24-minute and 3-minute, including 313,133 events on the 3-minute; ETH, BTC and SOL re-checked separately

Smallest effect it could spot

about 0.024R on the hour and 0.012R on the 3-minute; 0.37 to 0.94 percentage points per archived cell

The cross is a coin flip on every coin tested and 9/21 is in no way a special pair, so keep it as a picture of trend, never as a reason to click.

At a level I already like, am I better off waiting for the moving averages to cross before I get in?

How it was tested. The same level, stop and target were traded with only the trigger changed, comparing entering straight away against ten different crossover pairs.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

Slow pairs including 9/21: 27-42% and -0.11 to -0.28R. Every crossover lost to entering immediately, and the ranking tracks how slow the pair is.

How much data

39 pivots on the 3-minute over 365 bars; 38-39 trades per condition

Smallest effect it could spot

not stated; the spread among neighbouring fast pairs is itself the noise band, and each cell holds only 38-39 trades

Your stop does not move while you wait, so every bar you spend waiting for confirmation buys a worse price for the same risk — thin evidence at 38-39 trades a cell, but it all points one way.

When momentum makes a higher low while price makes a lower low, does the turn actually come?

How it was tested. Five ways of spotting a divergence were measured on one coin and re-checked on two coins that had not been used to pick them, then sliced by trend, depth and confirmation, and finally used as a filter on dot entries against the control of taking every dot.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

Slices: against the trend 50.3/50.6%, with the trend 52.7% where it was fitted but 49.9% on the check, deep turn 50.1/50.7%, confirmed 50.6/50.5%. Used as a filter on dots it made things worse — every dot with no filter 40% and +0.002R, classic divergence 38% and -0.065R, trend-aligned hidden divergence 34% and -0.237R.

How much data

Cells from 480 to 7,862 trades on ETH, BTC and SOL from 15-minute to 4-hour; 820-910 divergences per coin; the filter arm had 32 classic and 29-37 hidden against a control of 269

Divergences describe momentum fading, not a turn arriving — and filtering other signals with them made those signals worse, not better.

Does an overbought or oversold RSI reading tell me where price goes next?

How it was tested. The bands were first set at 40 and 60, where the top and bottom fifth of readings actually fall on this market over about eleven months rather than the textbook 70/30, and then the move 12 bars later was measured from each band in units of a typical bar's range, counting each move only once.

On calibration, median RSI is 49.7 (1H), 50.3 (24M) and 49.7 (4H), and cut points of 40 and 60 give a 21.2 / 58.7 / 20.1 split on the hour, where the old 55 line had been calling a third of all bars 'extended'.

How much data

About 8,000 bars on 1-hour and 24-minute; 124 contracted, 338 neutral and 203 extended hourly readings, plus 314 neutral on the 24-minute

Smallest effect it could spot

0.375 to 0.797 ATR depending on the band

Read RSI as how far a move has already stretched, not as where it goes next — and keep the lines at 40 and 60, because at 55 a third of all bars looked 'extended'.

When a big-volume bar fires on the fast chart and the hour above it is also big, does price run further?

How it was tested. Every 24-minute bar trading at least twice its 30-bar average volume was followed for the next 24 bars, and the ones whose last completed hourly bar was also 2x were compared against the rest, using events that shared no bars and rules written down before any number was looked at.

The peeking version (82 vs 155 events) gave 0.497 ATR against a floor of 3.022 ATR. Control of every bar: -0.032 ATR.

How much data

8,000 x 24M bars and 8,000 x 1H bars of ETHUSDC.P; 237 qualifying events; 2,494-bar control

Smallest effect it could spot

2.449 ATR (peeking version 3.022 ATR)

The most trusted rule in the framework has never been shown to work, and the usable version fired 17 times in 133 days, so settling it on one coin would take decades.

The ladder says the hour above must also be twice normal volume, so can you actually read that number when the signal fires?

How it was tested. Worked out how the bars nest inside each other, then measured what share of the confirming hour's volume was the signal bar's own volume.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

On the two events measured the signal bar was 53% and 54% of the whole hour's volume.

How much data

arithmetic on bar nesting, plus the two largest events of 2026-08-28

Smallest effect it could spot

not applicable — arithmetic, not a measurement

Over half of the 'confirmation' is the same trades counted twice, so the ladder has never been tested the way it is actually used, and the unbroken 5-for-5 live record is anecdote against 237 measured events that show nothing.

Is a bar at twice its average volume actually a rare event worth reacting to?

How it was tested. Divided every bar's volume by its own 20-bar average and counted how often each level came up, with no other conditions attached.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

How much data

every bar on 9 crypto majors, 1-hour, after a 300-bar warm-up

Smallest effect it could spot

no floor applies — this is a count, not a test

Twice-average volume turns up roughly once every twelve bars, so a big volume number by itself does not make a setup rare or special — whatever rarity a setup has comes from where the bar sits, not how loud it is.

Does only taking sweeps that arrive with heavy volume make them better trades?

How it was tested. Scored swing-failure entries with and without a rule that the sweep candle trade at least twice normal volume, then stacked that filter on top of the existing entry rule, with a live forward log checked the same way.

Stacked on the entry rule it made things worse: +0.131R against +0.227R for the entry rule alone. Median volume on sweeps was already 2.07x normal.

How much data

the offline swing-failure sample, plus 21 live sweeps logged forward

Smallest effect it could spot

0.041R

A sweep is a heavy-volume event by definition, so calling one unconfirmed for low volume is not supported — the best sweep in the live log (6.17 ATR) fired on 0.41x volume, the quietest of all 21.

Does our money-flow reading tell you anything about where price goes next?

How it was tested. Bars were sorted into fifths by their money-flow reading across three coins and four timeframes, the top-minus-bottom gap was measured, and the whole thing was rechecked on coins the test had never seen.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

How much data

453,176 resolved bars (the original claim came from 6,344)

Smallest effect it could spot

not stated; the run-to-run wobble was 0.20-0.29pp per cell

The eye-catching lead was a two-month accident on one coin; money flow has no shown ability to predict direction, and agreement across horizons on the same bars proves nothing — only a fresh coin does.

Does the well-known paid indicator's money flow do any better than ours?

How it was tested. Pulled 6,344 bars of the paid indicator's own plotted output straight off the chart, sorted each reading into fifths and measured the top-minus-bottom gap at three forward horizons.

Its wave matched ours at a correlation of 1.0000.

How much data

6,344 bars of 15-minute ETHUSDC.P

Smallest effect it could spot

1.44pp per cell

The paid tool has no mechanical edge over ours, and its popular free clone reads a different quantity entirely (-6.35 against -1.11 on the same bar), so never quote the clone as if it were the paid number.

Is the green-and-red money flow cloud really weighted by volume, and does the darker inner band add anything?

How it was tested. Read the shipped formula line by line, checked a copy of it against the live chart bar for bar, then sorted every bar into blank, an exact copy of the outer cloud, or something new.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

The inner pressure band is blank on 36.8% of 1-hour bars and an exact copy on 19.2%, leaving 44.0% that says anything new (15-minute: 36.9 / 17.3 / 45.8). The code matches its own formula to 5.2e-14.

How much data

298,200 bars across 3 coins and 2 timeframes, plus 201 bars checked against the code

The name is inherited, not earned: the cloud is a slow read of candle bodies, and on about 55% of bars the volume band is blank or just repeating the cloud — with blank meaning both 'neutral volume' and 'volume disagrees'.

If I wait for price to poke through a level and close back inside, then enter with my stop tucked behind the wick, does it make money?

How it was tested. Every sweep was traded exactly as specified — enter at the close of the sweep bar, stop 1 ATR beyond the wick, risk capped at 2 ATR, target the nearest untaken level, real exchange fees charged — and scored against entering at a random bar with the identical stop and target.

Moving the target does not rescue it: nearest level -0.0362R, 1.0R target -0.0725R, 1.5R -0.0807R, 2.0R -0.0856R, 3.0R -0.0857R, and the sweep is worse than random at every one. Average win 0.155R against a full 1R loss.

How much data

18,243 trades on six fresh symbols, 1-hour

Smallest effect it could spot

0.010R — the sharpest test in the whole project

You can win 69% of the time and still lose money, because the wins are small and the losses are full size — and no choice of target fixes it.

Does a wick that takes out a high or low and closes back inside tell you which way price goes next?

How it was tested. Fading the sweep was scored as a race from the sweep bar's close — does price travel 1 ATR the reversal way before it travels 1 ATR against — on nine symbols, with the long/short mix matched, checked in all four calendar quarters and on data it had never seen.

Re-run against a matched comparison group it comes out at -1.64pp, which is -0.033R in the reversal direction, and that repeats on both check groups, in all four quarters, and at every sweep size from 0.5 to 2.0. Every slice leaves it there: wick size 49.1-50.3%, swing age 45.3-49.8%, direction 48.8-50.1%.

How much data

31,710 sweeps, nine symbols, four timeframes

Smallest effect it could spot

0.022R

Traded mechanically as a reversal the pattern is not even a coin flip — it leans slightly against you, and the lean is smaller than what the fees cost, so it is information about location, not money.

Is a sweep with a long wick and a decisive close back inside a better trade than a scrappy one?

How it was tested. 73 sweeps were split into thirds by how deeply price closed back inside, by wick length and by swing size, with the grading rule written down in advance, and each third was compared against simply taking every sweep.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

Strong reclaim was negative on all three timeframes: -0.02 / -0.30 / -0.84R. Graded by swing size instead: strong -0.095R, weak -0.179R, both worse than doing nothing.

How much data

73 sweeps across 30-minute, 1-hour and 4-hour

The most impressive-looking sweep is the worse trade, because a big reclaim means the move already happened and you are entering late.

Does insisting on heavy volume in the sweep candle pick out the better sweeps?

How it was tested. Sweeps were scored with and without a rule demanding at least twice normal volume on the sweep candle, that filter was then stacked on top of the entry rule, and the same idea was recorded live on every sweep as it formed.

Stacked on the entry rule it made things worse: +0.131R against +0.227R for the entry rule alone. Median volume on a sweep is already 2.07x normal, so the filter screens out almost nothing. In the live log the best sweep recorded (best move 6.17 ATR) fired on 0.41x volume, the lowest reading of any logged sweep. The opposite claim — that heavy-volume sweeps are worse, at 47.5% and 48.5% — did not reproduce when re-tested properly.

How much data

the 73-sweep offline set, plus 21 independent sweeps logged live

Smallest effect it could spot

0.041R

A sweep is a high-volume event by definition, so calling one 'unconfirmed' because volume looks low has nothing behind it.

The indicator scores every untaken high and low by how much volume sits behind it — does a big score mean price gets there faster, or that the level holds when price arrives?

How it was tested. Every scored level was tracked for how long price took to reach it, how often price closed straight through it within 20 bars, and how a fade of the sweep did, with each level's distance at birth held constant so a nearer level cannot look magnetic for the wrong reason.

from 4-plus ATR, 43 / 60 / 52 / 77 — a high-scoring level takes about twice as long to be reached, and the check symbols agree almost exactly. Once touched, price closed through within 20 bars 86.2 / 84.7 / 83.4 / 84.8% of the time, flat at every score, and fat levels produced smaller pullbacks first, not bigger (2.39 / 2.18 / 2.06 / 1.83 ATR). About 99% of levels get swept eventually whatever their size, and fading the sweep wins 48-49% across every pool size.

How much data

34,951 levels price later reached; 43,076 levels scored; about 27,000 sweeps across three symbols and four timeframes

Smallest effect it could spot

1.16 percentage points on the fade comparison

A big score is not a stronger level, it is a slower target — use it to see where price might eventually go, never to decide whether a level will hold.

Does price get dragged back to an untouched high-volume price more than to any other price the same distance away?

How it was tested. How often price reached the untouched shelf was compared against how often it reached a fake target drawn the same distance away on the same side, after a first attempt that used a fake target sitting at a different distance had to be thrown out.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

Corrected: the shelf was reached 88.4 / 88.1 / 87.5% of the time against 88.2 / 88.1 / 87.8% for the matched fake target — a difference of +0.2 / +0.0 / -0.3pp, with the two average distances matching to two decimal places.

How much data

3 symbols on 1-hour, 240-bar window (about ten days)

Price reaches an untouched volume shelf 88% of the time within ten days — and it reaches any price at that distance 88% of the time. It is a price, not a magnet.

How much room does my stop need to survive a stop hunt?

How it was tested. For every sweep, the distance the wick travelled past the level before price turned back was measured as a multiple of that chart's average range, on four timeframes and re-run on a second independent sample.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

The old claim that 1.5x ATR clears every sweep is dead: the largest observed sweep grew every time a sample was added, 0.89 to 1.22 to 1.33 to 1.65, while the median and 90th percentile barely moved. In dollars the same medians read $2.98 on 30-minute but $9.13 on 4-hour, which is why the number must be kept in ATR.

How much data

85 sweeps across 15-minute, 30-minute, 1-hour and 4-hour, spans of 3.1 to 83.2 days

Smallest effect it could spot

not applicable — a direct measurement

Stop hunts are shallow: about 1 ATR beyond the swept wick clears roughly nine in ten of them, so surviving one is cheap — but this says how far price pokes, never which way it goes next.

When the chart says a range is accumulation or distribution, how often does it call the eventual break correctly?

How it was tested. Scored each range's first label against the side price actually broke out of, then repeated the scoring at every conviction level from 1 to 7 to see if being pickier helps.

How much data

4,894 range phases across nine major coins

Smallest effect it could spot

4.72pp on the sealed-range version, 2.52pp on the faster version

It is right about half the time no matter how picky you are, so read it as a description of where price has been, not as a call on where it goes next.

Does the accumulation-versus-distribution reading tell you which way a range will break before it breaks?

How it was tested. Used only ranges that actually broke, froze the reading at the last bar before the break and then further and further back in time, and scored each freeze against how often that side broke anyway.

twenty bars back -3.3 / +0.2 / +0.8pp. A plain count of up bars against down bars scores the same (78.3 vs 77.7).

How much data

about 500 ranges per coin on three coins

Smallest effect it could spot

0.130R, and 0.16-0.20R for the rows the verdict rested on (roughly 6.5pp)

The reading was watching the break form, not forecasting it — any number that shrinks the earlier you take the snapshot was never a forecast.

Does the volume part of the phase engine actually know anything about volume, or is it just counting up-closes?

How it was tested. Replaced volume with a flat 1.0 on every bar and counted how often the vote changed, then ran a version with volume shuffled against price as a fake check, keeping the phases, ranges and every gate identical.

How much data

nine major coins, 40,000 1-hour bars each (2022-01 to 2026-08); 4,894 range phases, 4,773 resolved, 167,623 bar readings

Smallest effect it could spot

2.46-5.45pp (about 0.050-0.110R)

About 71% of the volume term is a disguised up-close count, and taking volume out altogether leaves how often the label is right unchanged (+0.13 / +0.06pp).

The volume part of the engine votes short more often than long by accident — does correcting that make it better at calling direction?

How it was tested. Measured the built-in lean, re-centred it, and re-ran four scoring methods across three splits of the data, two settings and two read points, with a deliberately fake arm run alongside as a check.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

Direction-calling ability does not move — every cell sits inside its own floor, and the fake arm threw up numbers the same size as the real ones.

How much data

four scoring methods x three data splits x two settings x two read points, plus a fake arm; repeated on 24-minute bars

Smallest effect it could spot

not stated as one number; every cell was inside its own

Changing how often a tool leans long or short moves the scoreboard without changing whether it knows anything — the third time in this project a result turned out to be that in disguise.

A failed push above the top of a range was flagged as being counted the wrong way round — does that hold up?

How it was tested. Reproduced the original number exactly, re-scored the term against a benchmark that leans the same way it does, compared it against the mirror-image spring term, and tried flipping the sign.

Flipping the sign gives +0.18pp and destroys 86% of the distribution calls (2,354 down to 330).

How much data

nine coins, three data splits; the original flag came from about 412 pooled ranges on three coins

Smallest effect it could spot

4.3-5.0pp; 4.33pp for the flip

No code bug here — the alarm came from comparing the term against the wrong benchmark, and flipping it would silence most of the labels for nothing in return.

If proper Wyckoff ideas are added — effort versus result, climax, dry-up, shakeout — does the phase label get better?

How it was tested. Built four textbook constructs with thresholds fixed in advance and no tuning, then measured both how quickly the label arrives and how often it is right.

Deleting the volume term and adding nothing does the same damage (-2.99 / -3.15) as the constructs (-2.72 / -4.82).

How much data

same nine coins, 40,000 bars each, 4,894 range phases

Smallest effect it could spot

0.11-0.39 bars on speed; 3.3-5.0pp (0.067-0.100R) on accuracy

The textbook shapes make the label arrive later without making it any more accurate — the delay is big enough to measure, the improvement is not.

Does the phase drawing only name a range after price has genuinely left it, or does it quietly rewrite the past so it looks like it knew?

How it was tested. Replayed the rules in Python from the chart's own bars and compared them against what the indicator actually drew, across the whole timeframe stack, plus a side-by-side check of the offline copy against the live chart.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

Price alignment error 0.0000 on every timeframe, and phases tile the chart with no gaps or overlaps. The offline copy matched the live logic on 40,000 of 40,000 bars per coin with 0 mismatches, and 5 of 5 phase names matched the live chart in order.

How much data

about 39,000 bars (4H 3,114 / 1H 12,453 / 24M 10,515 / 3M 13,087), plus 40,000 bars each on three coins for the copy check

Smallest effect it could spot

not stated — a correctness check, not an edge test

What you saw live is what you see later, which is more than most phase indicators can say — but drawing honestly is not the same as predicting anything.

When price closes outside the band, should I bet on a snap back?

How it was tested. Every hourly close outside the band was traded in the fade direction, each trade counted once, against a control matched on the long/short mix and against a scrambled version of the tape, with all 88 cells written down before any number was computed.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

But the band threshold itself adds only +0.72pp ± 0.38 over the same rule on ordinary bars, and only +0.13pp on the second check group. The response is not orderly: the moderate bucket reads higher than the more extreme one. It is +1.59pp before Oct 2025 and +0.19pp after. 24-minute reads +0.29pp and 3-minute reverses sign. Gross +0.027R against 0.081R of round-trip fees.

How much data

23,886 trades, nine symbols, five calendar years, 1-hour

Smallest effect it could spot

not stated as one number

Do not fade a close outside the band on the hourly — but the work is being done by the 20-bar average, not the band, and what is left costs three times more to trade than it pays.

Would switching to the other common band type be better?

How it was tested. Two teams ran it blind to each other, each matching how often the two band types fire so both were scored on the same number of trades, with no width setting ever tuned against results.

The fitting symbol read -1.74pp, the opposite sign, favouring Keltner. 14 of 15 cells sat below the floor and none cleared on both check groups with the same sign.

How much data

23,886 Bollinger trades and 18,714 Keltner trades on 1-hour; 48 cells declared for one design, 18 for the other

Smallest effect it could spot

1.27pp; the finest either design could resolve was 0.025R on 1-hour and 0.0145R on 3-minute

Keep the band type you already have — not because it won, but because the gap between them is three times smaller than the smallest effect the test could spot.

If the two band types differ at all, is it the centre average or the way the width is measured?

How it was tested. One change at a time — swap only the centre average and hold the width measure fixed, then swap only the width measure and hold the centre average fixed.

Changing only the width measure gives +0.579pp ± 0.498 against a 1.39pp floor, and the fitting symbol flips sign at -2.22pp.

How much data

1-hour, nine symbols; trade counts matched to within 0.6% on the centre-line swap

Smallest effect it could spot

0.73pp for the centre average, 1.39pp for the width measure

The centre average is not where anything lives, so arguing over a 20-bar simple versus a 21 exponential is arguing over nothing.

Would a different length or width setting work better than the shipped one?

How it was tested. All 20 combinations of two lengths by two width multipliers by five average types, every cell written down before running.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

The shipped 20-bar / 2.0-width / simple-average setting sits mid-pack, and the single cell that leaned the other way on the fitting symbol reversed on both check groups.

How much data

20 declared cells, nine symbols, 1-hour

Leave the length, the multiplier and the average type exactly where they are — the whole grid is one flat field.

When the bands pinch together, does a real move follow in the direction they point?

How it was tested. Five separate rule families were fitted on one coin and then scored by different people on two coins the fitter never saw, after which the squeeze machinery was ripped out and replaced with "every 16 bars, just bet the direction rule".

Stripping the squeeze out entirely and using a plain timer scored 50.30% against the full rule's 51.90%. The detector fires on 8.6-9.1% of bars in all nine cells, near-identical across three coins and three timeframes.

How much data

2,534 fitted events and 5,075 scored on unseen data; 19 agents

Smallest effect it could spot

52.5% was the bar set in advance — 2.5pp above a coin flip

The squeeze is an alarm clock, not a signal — a timer that ignores the bands completely does the same job.

Is the standing rule "only fade the band in a range, never in a trend" actually supported?

How it was tested. The rule was traced back to the study it came from and re-run against the bands that are actually on the chart.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

Its own numbers were inconsistent anyway: top of band while trending read +0.44 on 30-minute and +1.46 on 1-hour, while ranging read -1.20 and -0.19, and the plain top-of-band reading flipped sign across timeframes at +0.05, +1.18 and -0.63 average-range units.

How much data

the original study's key ranging buckets held 13 and 19 observations, with samples reused, no control and no second coin

Smallest effect it could spot

not stated; ranging buckets of 13 and 19

This rule was being issued live every hour for weeks and was measured on the wrong indicator — it is gone, and nothing replaced it.

One reading in the band study was enormous. Was it real?

How it was tested. The single largest cell — the most extreme 1% of 24-minute bars — was checked against the matching 1-hour cell, split by era, and cross-checked against 3-minute data covering the same days.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

It collapses on an era split, -15.7 in the first half and -3.9 in the second, and is invisible in 3-minute data covering the same days. The second design found the mirror image, its two clearing cells pointing in opposite directions.

How much data

15 declared readings; 3-minute cross-check covering the same 312 days

When you look at fifteen readings, the biggest one is usually the noise — which is why the bar gets set before looking, not after.

Does the 15-part score tell you which way price is about to go?

How it was tested. The card was re-scored against a like-for-like benchmark that holds its own long/short mix fixed, because the original had compared a signal that trades both ways against simply buying and holding on a sample that drifted down.

How much data

32,809 trades across eight check symbols, 1-hour bars, 4.5 years

Smallest effect it could spot

2 percentage points, the size fixed in advance as worth having

Never quote the 2.7 figure again; the card's real reading is a third of a point, which is nothing you can trade.

Is the card worth more when the market is falling or going sideways?

How it was tested. One trade at a time, split by market phase, scored first against buy-and-hold and then against a benchmark that keeps the card's own long/short mix fixed.

Corrected, downtrends turn negative on both check groups at -0.24pp and -0.81pp, ranges fall to +0.99pp and +0.04pp, and uptrends read -0.08pp then +2.41pp.

How much data

eight check symbols, 1-hour bars, 4.5 years

Smallest effect it could spot

2 percentage points, set before the test

The phases where the card looked best are exactly the phases where it leans short — it was being paid for the lean, not for the timing.

Once you trade the card's calls with a real stop, a real target and fees, does it make money?

How it was tested. The calls were run through the trade harness at a 1 ATR stop for 1R and a 3 ATR stop for 3R, counted first on every qualifying bar and then one trade at a time.

The overlapping count inflated the result thirtyfold and made its error bars five times too small.

How much data

103,305 overlapping bars against 3,702 one-trade-at-a-time trades

Smallest effect it could spot

0.080R

It breaks even before fees are even in the picture, and any bar-by-bar result should be treated as inflated until it is re-run one trade at a time.

The card weights its parts from 2 to 14 points. Does the weighting change the answer?

How it was tested. The same card re-scored with all fifteen parts weighted equally, picking the same number of signals as the shipped version.

How much data

the same 32,809-trade universe, top 30% of readings

Smallest effect it could spot

2 percentage points

Stop tuning the weights — both versions read zero, so there is nothing to tune toward.

Is one bad part dragging the other fourteen down?

How it was tested. The card was re-summed fifteen times, each time with a different part removed; a removal had to beat the full card by 2 points on both check groups to count.

Removing the two heaviest parts, SR VWAP and Trend, hurts on both groups: -0.49/-0.24 and -0.10/-0.32.

How much data

the same 32,809-trade universe, both check groups

Smallest effect it could spot

2 percentage points on both check groups

There is no part to blame and no part to cut — the card is not one broken row away from working.

A 15-part score sounds like fifteen opinions agreeing. How many is it really?

How it was tested. A version of the indicator that publishes every part separately was dumped off the live chart, and how closely the parts move together was measured.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

The three heaviest parts hold 40 of the 100 points and move together at +0.52 to +0.73 (DI against Trend +0.73 and +0.61). The part holding the fourth-heaviest weight, 7 points, votes on only 2.4-2.6% of bars and contributes 0.2 of 100. Two parts agreed on 100% and 97.3% of bars, which is a fixed offset rather than a vote.

How much data

703 bars each of 15-minute and 1-hour

A high score is not fifteen things agreeing; it is four or five things said several ways, so read it as less confirmation than it looks like.

Is the copy used for all this testing really the same card that is on the screen?

How it was tested. All fifteen parts were rebuilt in Python by three modules working in parallel and compared bar by bar against dumps of the live indicator's own per-part output.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

A real bug was found in the trend-strength code along the way, affecting 11 bars of 13,398 on coins with coarse price steps.

How much data

13,398 scored bars across nine symbols and two timeframes; 6,745 bars for the assembled score

Smallest effect it could spot

not stated - this is a correctness check

Everything else measured here is about the real card, not a rough imitation of it.

If I pull my stop in closer, do I pay less in fees?

How it was tested. Every entry was traded both long and short with the same stop and target and the two averaged, so market drift cancels out exactly, then the whole thing was repeated on a shuffled price tape where no trend effect can exist.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

Widening from 1 ATR to 3 ATR saves +0.0333R per trade on the real tape and +0.0332R on the shuffled tape, the same number. The harness matched the fee-only line to 4 decimal places at every stop width.

How much data

nine symbols on 1-hour bars; the original grid used seven symbols and 3,000 entries each

Smallest effect it could spot

not applicable - this is arithmetic, matched to 4 decimal places

A 0.5 ATR stop costs six times what a 3 ATR stop costs for the identical dollars at risk, because halving the stop doubles your position size and fees are charged on the position, not on the risk.

Before I am right or wrong about anything, what does a single trade cost me?

How it was tested. Measured straight off the random-entry comparison arm, which loses money for one reason only: fees.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

How much data

18,243 trades at a 2 ATR risk cap on the cheapest fee tier

Any idea has to clear 0.03 to 0.05R before it is worth taking, and the largest effect this project has ever measured is smaller than that.

On each of my charts, how much of the money I am risking goes to the exchange?

How it was tested. Worked the fee out as a share of risk - fee rate times price, divided by twice the ATR, since the method risks about two ATR - then ran it at each timeframe's current ATR and checked the formula reproduced every row of the original fee table.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

At a 0.12% round trip the same formula gives ATR 25 -> 5%, ATR 12 -> 9-10%, ATR 7.50 -> 15%, ATR 5 -> 23%, ATR 3 -> 38%.

How much data

4 timeframes at one price snapshot (ETH 2,519.40); formula cross-checked against all 5 rows of the source fee table

Smallest effect it could spot

not applicable - this is arithmetic, not a measurement

Fees are a rounding error on the 4-hour and nearly a seventh of your risk on the 3-minute, so the quieter the chart the bigger the exchange's cut of the same trade.

How quiet does the market have to get before a trade is not worth taking?

How it was tested. Three fixed dollar thresholds had been written down at different times and each treated as a constant; re-deriving the formula showed price sits on top of the fraction, so the figure has to move with the price of ETH.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

The floor is 0.20% of price: $3.76 at ETH 1,880 but $5.04 at ETH 2,519.40, so quoting $3.76 today understates it by about a third. The 15% cut-off the floor is solved from was inherited convention and was never tested, and the fee gate built on it was about three times stricter than a limit-order trader actually faces (23% assumed, versus about 15% limit-in/market-out and about 8% limit both sides at ATR 4.86). It was deleted as a rule on 2026-08-06.

How much data

3 previously recorded dollar figures, re-derived at current price

Smallest effect it could spot

not applicable - this is arithmetic, not a measurement

Never skip a setup because of a fixed dollar volatility figure: the real threshold is 0.20% of price and moves as price moves, and the cut-off behind it was a convention nobody ever measured.

Do moves that get going keep running, so that a bigger target picks up more than a coin flip would pay?

How it was tested. The same both-ways grid, compared against what a coin-flip market would pay after fees, then re-priced once trades sharing the same market move stopped being counted as separate ones, and re-checked on two separate groups of symbols.

Re-priced, the range runs about -0.014R to +0.064R and includes zero. The estimate later rises to +0.0312R with 8 of 9 symbols leaning positive, but it fails on both checking groups, it lives disproportionately in one calendar year (2025), and the only slice sharp enough to see the target effect reads negative. That cell's actual net expectancy is +0.0051R.

How much data

344,763 trades across nine symbols on the dense re-run; 3,702 trades sharing no bars in the earlier version

Smallest effect it could spot

0.0676R originally; 0.0535R on 1-hour; 0.0308R on the 3-minute cut

Keep wide stops for the fee saving, not because trends were shown to carry - the carrying-on half of the story is the part that died.

For an order block trade, should my stop sit in the middle of the zone or past its far edge?

How it was tested. Identical entry and identical target on every trade with only the stop moved, then the answer was required to hold for longs and shorts and for the first and second half of the sample.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

In 22 of 62 cases the midpoint stop was taken out by a move the wider stop would have survived to reach target, and zero cases ran the other way. Sweeping the multiple, the sign only holds steady across every split from 1.5x ATR (76% win, +0.087R) through 1.75x (78%, +0.081R) to 2.0x (80%, +0.077R); 1.0x (+0.044R) and 1.25x (+0.053R) both flip sign.

How much data

65 zones, 62 of them retested, over 360 hourly bars across 15 days

Smallest effect it could spot

not stated; 62 trades over 15 days on one timeframe

Put the stop past the far edge of the zone rather than at its midpoint - the 22-to-0 count never inverted - but 15 days of data cannot pin down the exact multiple beyond a rough 1.5x to 2x ATR band.

Does banking half the position early and pulling the stop to breakeven make me more money?

How it was tested. The same trades run four different ways, with the result then broken down into win rate, average win and average loss.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

Half off at 0.5R plus a breakeven stop: 67.0% win, -0.032R, with the average win collapsing to +0.47R while the average loss does not move at all. Reproduced from win rate times average win and loss to four decimal places.

How much data

a 360,000-bar panel across nine symbols

Smallest effect it could spot

not applicable - this is a breakdown of the same trades

Taking half off nearly doubles your win rate and turns a small profit into a loss - it relabels losers as winners instead of making the trades better.

If I write the trade down before it happens - trigger, target, give-up level, deadline - how often does the call come good?

How it was tested. Six calls were fixed in writing with a trigger price, a target, a give-up level, the exchange and a deadline, then scored on 24-minute closing prices, counting only bars strictly after the trigger bar.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

properly-shaped pass out of 6.

How much data

6 pre-registered calls, 2026-08-25 to 2026-08-29, ETHUSDC.P on Hyperliquid

Smallest effect it could spot

not stated as an effect size; the log's own target for a meaningful record is roughly 90 calls

Four out of six reads like a system that works; the honest count is one clean result out of six, which tells you nothing yet.

With the target and the give-up level set at sensible distances apart, does the break actually pay?

How it was tested. Named 2026-08-26 with the geometry stated first - trigger a 24-minute close above 2514.90, target 2534.00 (+18.00), give-up 2505.50 (-10.50), 1 to 1.71, break-even 36.8%, deadline 2026-08-28 23:00 UTC - and checked that the target had not already been touched when the trigger fired.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

The bar that did it ran 29.20 points on 3.23x average volume; the trigger bar itself ran 1.77x.

How much data

1 call; fired 2026-08-27 08:00, target hit 24 minutes later

The one call out of six that was built properly, and it is still a single trade decided by one 29-point bar.

When a call passes, was there actually anything at risk?

How it was tested. Two of the six passing calls were re-measured afterwards: how far the target already sat when the trigger fired, and what the risk-to-reward really was at the price the trigger actually filled at.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

TEST 4 declared 1 to 1.18 and a 45.9% break-even, but the trigger bar's own low of 2483.40 was already 4.90 points through the target, and the close-basis fill landed 14 points past the named level, making the real geometry 1 to 0.38 and the real break-even 72.6%.

How much data

2 of the 6 pre-registered calls (TEST 2 and TEST 4)

Two of the four wins were handed over by the shape of the trade, not the idea - measure the risk-to-reward at the fill, not at the level you named.

If you name both edges of the range a day in advance, does price actually close through either one?

How it was tested. Named 2026-08-25 23:20 UTC - up side a 24-minute close above 2469.00 (target 2483.50), down side a close below 2437.80 (target 2419.70), deadline 2026-08-26 15:20 UTC, closing prices only, so a wick through the level does not count.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

Both edges were wicked through (2474.40 and 2431.60) across a 42.80-point range and neither closed through.

How much data

39 closed 24-minute bars inside the 40-bar window

Whether you judge a level on closes or on wicks changed the answer more than the market did - on wicks this same call would have fired both ways and been counted twice.

Would my call have scored the same if I had used a different exchange's prices?

How it was tested. The first named-level call - a 24-minute close below 2477.60, target 2453.40, give-up 2488.20, 40-bar limit, all fixed in public at 09:00 UTC on 2026-08-25 - was scored on Hyperliquid as written, then re-scored on Binance prices over the identical window.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

FAIL on Binance - it printed 2489.00 three minutes after the trigger, clearing the give-up by 0.80. A 3.40-point gap between venues flipped the answer.

How much data

1 call scored twice; resolved 11 bars into a 40-bar window

Smallest effect it could spot

not stated as a number; the log said roughly 90 such calls would be needed to show a large effect

Name the exchange before the call - a few points of difference between venues can turn the same trade from a win into a loss.

When I draw a support or resistance zone myself, does it hold?

How it was tested. Three order blocks drawn on the live chart, each with its give-up rule - a body close outside the zone - stated at the moment of drawing so it could not be moved afterwards, then left alone and scored when price arrived.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

The 2496.20-2503.30 zone broke within two cycles (close 2508.60, 5.30 above the top). The 2483.50-2495.80 zone survived four tests including an 18-point wick and a close 1.20 above its base, then died 96 minutes later on a close of 2471.20.

How much data

3 zones drawn, 2026-08-27 to 2026-08-29

On the only three zones scored under a rule fixed before the fact, hand-drawn levels went nought for three - and the cleanest-looking one broke fastest.

The level-plus-volume-plus-structure checklist had a 6-out-of-8 record on live breaks - was that real?

How it was tested. The checklist was graded on live breaks over roughly 20 monitoring cycles with a running record of correct and incorrect calls, then the record itself was audited for whether any pass rule had been written down before the outcomes existed.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

No success rule was ever set beforehand, so the time horizon and the threshold for each 'correct decline' were chosen once the outcome was already known.

How much data

8 graded instances

If you decide after the move what counts as being right, you will be right most of the time - which is exactly why the trigger, target and deadline have to be written down first.

Was the one number in this project that said something actually real?

How it was tested. The card leans short and the test window drifted downwards, so scoring it against simply buying and holding paid it for its mix of longs and shorts rather than its timing; it was re-scored against a comparison that holds that mix fixed and tests only the timing.

Individual symbols run from -0.80pp to +1.66pp. The phase pattern that made it look strongest in ranges and downtrends was the mistake itself: corrected, downtrends read -0.24pp and -0.81pp.

How much data

32,809 trades across eight check symbols, 1-hour bars, 4.5 years; the offline copy matches the live card on 13,398 bars

Smallest effect it could spot

+2pp, set before the test

Never quote the +2.7pp figure again — a signal that trades both ways has to be scored against something that also trades both ways, or it collects free points for the market's drift.

When a test here says it found nothing, how small a thing could it really have spotted?

How it was tested. The original check wobbled a fake reaction by only 0.03 ATR inside a 0.10 ATR counting band, so every fake landed dead centre and looked easy to detect; it was redone using the real height of an actual order block.

Recalibrated, the smallest detectable effect moves from 1.03% to 4.02% — from about 1 reaction in 100 to about 1 in 25, roughly 0.18R per trade. Fees are 0.035R and the largest effect the project ever measured was +0.024R.

How much data

22,687 blocks, median height 0.800 ATR, 90th percentile 1.121 ATR

Smallest effect it could spot

about 1 reaction in 25, or 0.18R per trade

Read every 'nothing found' here as 'no large effect' — these tests could not have seen an edge the size of the biggest one this project ever measured.

The volume rule says the hour above must also be heavy — can you actually know that at the moment you would press the button?

How it was tested. Plain arithmetic on how bars nest inside each other, plus measuring how much of the confirming hour's volume was the signal bar's own volume.

On the two big events the signal bar was 51,621 of 97,579 (53%) and 59,690 of 110,159 (54%) of the whole hour's volume — over half the confirmation is the same trades counted twice. Tested using only what was visible at the time, confirmed minus partial is -0.855 ATR against a detection floor of 2.449 ATR, and the usable version fires 17 times in 133 days.

How much data

Arithmetic plus the two largest live events of 2026-08-28; the measured version ran on 8,000 bars and 237 separate events

Smallest effect it could spot

2.449 ATR

You cannot read a confirming hourly volume figure until the hour has closed, and by then the entry is long gone — the most trusted rule in the framework had never been tested the way it is used.

If I test a fib the way I would actually trade it — an order sitting at the level waiting to be filled — does it work?

How it was tested. The identical test was re-run with entry taken at the level, as a resting order would fill, instead of at the close of the bar that touched it.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

It is entirely construction: a bar that reaches a level closes on the favourable side of it 57-58% of the time, and at deep levels the fact that the move is still going guarantees it. This would have been the eleventh false positive.

How much data

The same nine-symbol panel

Any backtest that fills a resting order at a level is quietly picking the moves that carried on — the fill does the selecting for you, so test at the close of the touch bar instead.

Does price get pulled back to an untouched volume shelf more than to any other price?

How it was tested. The first fake comparison level was built from the day's open and sat a different distance from price, so it was harder to reach for reasons that had nothing to do with being a real level; it was rebuilt at a matching distance on the same side.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

Corrected, the untouched point of control was reached 88.4 / 88.1 / 87.5% of the time against 88.2 / 88.1 / 87.8% for the matched fake — a difference of +0.2 / +0.0 / -0.3pp, with average distances matching to two decimal places.

How much data

3 symbols on 1-hour bars, 240-bar horizon (about ten days)

Smallest effect it could spot

not stated in the same units

Price reaches an untouched volume shelf 88% of the time within ten days, and it reaches any other price at that same distance just as often — it is a price, not a magnet.

How many real, separate swing-failure observations does the forward log actually hold?

How it was tested. Every row of the log was checked for exact duplicates and for repeated pokes at the same price shelf inside a single session.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

session. Nine separate shelves were each swept 2-3 times. The 30 raw rows become 16 independent observations against a stated target of about 30; the count has since grown to 21.

How much data

Every row logged 2026-08-26 to 2026-08-28, about 30 raw rows

Smallest effect it could spot

not applicable - a count

Poking the same shelf three times in one session is one observation, not three — count shelves, not rows, or the sample inflates and manufactures a result out of nothing.

Did the level-plus-volume-plus-structure checklist really call breaks correctly six times out of eight?

How it was tested. The claimed live record was re-examined for whether a rule saying what counts as a correct call had been written down before the calls were graded.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

Withdrawn as unverified on 2026-08-05: no success rule existed in advance, so both the time horizon and the threshold were chosen once the outcome was already known.

How much data

8 graded instances over roughly 20 monitoring cycles

Write down what counts as a win before the move happens — grade afterwards and you will be right almost every time, on paper only.

Does every reading on my chart actually compute what its label says it computes?

How it was tested. Four features were rebuilt in plain code from their written formulas and compared value-by-value against the live chart, with the checking rules fixed in writing before any test ran.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

Divergence markers: 1,004 comparisons, zero false fires and zero misses, all 5 events matched exactly. The scorecard code is byte-identical in both indicators, with 17 bull and 17 bear parts all correctly paired and none left unpaired. Two things do mislead, though: on the hourly the money-flow pressure band is flat on 36.8% of bars and simply retraces the outer cloud on another 19.2%, so it says something new on only 44.0%; and the divergence row stays lit for 10 bars after one event, so the hidden card counts each diamond 9.86 bars for every one you can see. The Dot row lights on 19.89% of all bars (19.76-20.06% across six series).

How much data

298,200 bars across 3 symbols and 2 timeframes for the money-flow check; 1,004 bar-by-bar comparisons for the divergence markers; 15,576 characters of shared scorecard code compared byte for byte

Smallest effect it could spot

not applicable - correctness check, no edge was measured

Nothing is broken, but the hidden scorecard and the chart in front of you do not always tell the same story, and the pressure band says nothing new on more than half of all bars.

Does the market-phase label only name a range after price has genuinely left it, or does it quietly redraw the past so it looks like it knew?

How it was tested. The phase rules were replayed in plain code from the chart's own bars and compared against what the indicator actually drew, across the whole timeframe stack.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

Price alignment error 0.0000 on every timeframe, and the phases tile the chart with no gaps and no overlaps. Switching to smoothed candles changes nothing (identical to 1e-9), and the labels read the same on the 4H, 24M and 3M charts when anchored to the 1-hour (identical to 1e-8). Against the live chart the phase names matched 5 of 5, in order.

How much data

about 39,000 bars - 4H 3,114 / 1H 12,453 / 24M 10,515 / 3M 13,087

Smallest effect it could spot

not applicable - correctness check; this checks the names and their order, not the exact bar each phase begins on

What you saw live is what you see when you scroll back later, which is more than most phase tools can say - though drawing it honestly is not the same as it being worth trading.

When the scorecard is tested over years of history, is it the same card that is actually on my screen?

How it was tested. All 15 parts of the card were rebuilt in plain code and compared bar by bar against 18 dumps of the real card's own output pulled off the live chart.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

A real bug surfaced on the way: rounding dust of about 1e-17 made two mathematically equal one-tick moves compare as unequal, flipping the trend-direction reading on 11 of 13,398 bars (0.15%), with zero effect on BTC, ETH or BNB.

How much data

13,398 scored bars across nine symbols and two timeframes; 6,745 bars for the assembled score

Smallest effect it could spot

not applicable - correctness check, no edge was measured

The live chart only holds about 1,000 bars while the questions needed 40,000, so this copy is what made every later scorecard test possible - and it can be trusted.

Is every liquidity level drawn on my chart really on a swing high or low, or are some of them just lines?

How it was tested. Every drawn level was checked against the actual candle data over the full history, with the swing count per 1,000 bars compared across five timeframes and the score checked against the running median volume.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

Swings per 1,000 bars at the same setting: 3m 96.6, 15m 94.8, 30m 94.8, 1h 95.2, 4h 92.4. A score of 5 lands on exactly 1.00x the median volume, each step up about 1.23x. Two faults were found and fixed: 27.5% of levels used to pin at the top score of 10 (16.9% after), and brand-new levels were hidden 56% of the time against 34% for levels over 100 bars old.

How much data

3,809 levels on 1-hour ETH; density on 3m, 15m, 30m, 1h and 4h; 43,076 levels for the score calibration; 12.8 million level-bars for the age check

Smallest effect it could spot

not applicable - correctness check, no edge was measured

The lines are drawn exactly where they claim to be and one setting works on every timeframe - but whether a bigger score means a stronger level is a separate question, and no edge was found there.

The green and red cloud is called money flow - is it measuring money, or volume, at all?

How it was tested. The shipped formula was read straight out of the code and the darker inner pressure band was traced through its arithmetic, then checked against live readings.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

The inner pressure band is forced to sit between zero and that volume-free cloud, so it can never exceed it or take the opposite sign; when volume actually opposes the cloud, pressure draws a flat zero, which looks identical to neutral volume. The same pass found that the plot labelled VWAP inside the oscillator panel is simply the fast wave minus the slow wave, residual exactly 0.0 - true of the paid script and of ours. Both the guidebook and the caption were corrected.

How much data

source code check plus live 15-minute readings and the audit's bar-by-bar count over 298,200 bars

Smallest effect it could spot

not applicable - correctness check, no edge was measured

A row you may have been reading as a volume signal is a price-shape average carrying an inherited name, and it can never warn you that volume disagrees because it never looks at volume.

A ribbon of eight moving averages looks like a lot of information - is it?

How it was tested. The ribbon's direction and its width were compared against just the 20 and 55 moving averages on every bar where the ribbon was stacked in order.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

How much data

275,265 bars - the 76.6% of bars on which the ribbon is stacked

Smallest effect it could spot

not applicable - correctness check, no edge was measured

The six middle lines cannot carry anything the outer two do not, so it is one reading dressed up as eight.

If I switch my chart to smoothed candles, are my indicators reading the real price or the smoothed one?

How it was tested. Every price read in the three scripts was routed through the real candles, then every numeric reading on smoothed candles was compared against standard candles on the same bar, matched by timestamp.

This test’s result and its floor were measured in different units, so the two cannot be placed on the same scale.

How much data

54 readings across the three scripts (12 + 13 + 29)

Smallest effect it could spot

not applicable - correctness check, no edge was measured

The indicators were quietly reading prices that never traded - it is fixed now, but any reading taken off a smoothed chart before the fix described something that did not happen.