I pulled the logs for all 17,084 hands I have personally been dealt and compared them against a monte carlo simulation that ran over 2.9 million sets of 17,084 hands to determine whether certain cards are dealt more often than expected, certain suits are dealt more often than expected, or certain ranks were dealt more often than expected.
My findings are as follows:
-
Raw-card level: broadly normal, with one real isolated tail. All four suits are ordinary, and the overall Broadway-card count is ordinary. Aces are the standout: 2,735 Aces versus fair P99 = 2,733, so that is a genuine P99-ish high-tail observation. Fives are mildly scarce at 2,526, P5-ish low. This does not look like a general high-card excess; it is specifically an Ace excess.
-
Two-card grouping level: broadly normal. I had 1,004 pocket pairs, exactly the fair median. TT+ totaled 400, versus median 387, but remained below P95 = 419. TT+ were 39.84% of all pairs, also below the composition P95 of 40.98%. Every broad hand family, suited hands, connectors, one-gappers, broadways, etc., remained ordinary. The clearest exception is 55: only 53, below the fair P1 cutoff of 56. KK+ is P95-ish high at 176, but AA and KK individually are both ordinary.
-
Temporal spacing/clustering: overall normal. This is one of the stronger reassuring results. All eight explicit clustering tests were ordinary, and every Any-Pair and TT+ block VMR from 50 through 1,000 hands was ordinary. Overall pair droughts, TT+ gaps, TT+ cold periods, and broad overdispersion do not show abnormal hot/cold cycling. There are isolated tails, particularly a 2,134-hand 55 drought, P99-ish, but that is closely related to the already-low lifetime 55 count rather than independent evidence.
-
Overall: this archive shows a few expected tail observations, not affirmative evidence against fair dealing. The anomalies are real enough to record, especially Aces high, 55 low/long-drought, and the recent TT+ burst. But they do not line up into a coherent pattern: suits are normal, broadways overall are normal, total pairs are exactly normal, premium-pair composition is normal, all broad starting-hand families are normal, explicit clustering is normal, and overdispersion is normal. Nothing predeclared went beyond the observed fair-reference range. My overall read is that the multi-year one-seat dealer history looks substantially more like ordinary fair variance with a handful of tails than a persistent dealing distortion.
The raw report can be found here.
I also ran a test for what I nicknamed the "MU Hypothesis". MU meaning "manufactured upset."
This test occurred twice, once at 665 observed hands and once at 1,503 observed hands. All data was collected on the Cat's Chance table.
The question the MU Hypothesis is trying to answer is:
Do mathematically unlikely upsets occur more frequently in Torn Poker than the exact remaining-card probabilities predict?
I ran two different tests, MU-R and MU-E.
MU-R is the statistical likelihood of the post-flop leader being passed on the turn or on the river.
MU-E determines whether the pot equity favorite ended up with the correct share of the pot at the time of the next card. To do this, I took the total pot at the start of the turn before the turn card was revealed but after all flop betting had concluded. I calculated the equity favorite at that point, and by how much they were favored. I then determined if over the sample size, their share of the pot at the state equity was calculated was consistent with the equity they held. I then did the same for the turn to the river.
MU-Clean was used as my control. The sample size is tiny, but involves All In automatic runnouts, in an effort to prevent decisions made after the observed point from contaminating the observation.
These tests are subject to selection bias due to them having reached a showdown. In order to know the amount of equity each player held at a given point, their hole cards had to be known, which can have an influence on the data.
Here are my findings:
At 665 observed hands:
MU verdict: weak deviation worth tracking. There is one interesting signal, but the clean sample is nowhere near large enough to call it a real inconsistency.
Strongest apparent MU signal: MU-R flop→turn reversals. There were 41 actual upsets vs 29.21 exact expected, about 1.40× expectation. The one-sided exact tail is about 1.2%.
What looks most normal: turn→river is almost absurdly ordinary at 32 upsets vs 32.48 expected. Turn equity favorites lost 32 times vs 32.17 expected. Flop equity favorites also ultimately lost at essentially normal rates across the strength bands.
The clean subset points weakly in the same direction, but it is tiny. Only 8 clean flop→turn states had a unique current leader: 3 upsets vs 1.13 expected. That is elevated, but still inside the exact 95% fair range and has a one-sided tail around 8.4%. Clean turn→river is 2 vs 2.18 expected.
MU-E currently looks normal. Event-weighted frozen-pot realization is close to exact equity both before the turn and before the river. The clean frozen-turn sample is especially ordinary. Dollar-weighted results move around more because a handful of large pots dominate them, but none is remotely a strong statistical inconsistency.
Nothing currently justifies a strong suspicion. The flop→turn result is interesting enough that I would absolutely keep tracking it. But the fact that turn→river, final equity-favorite losses, persistence, and frozen-pot realization are normal keeps this firmly in the “weak deviation worth tracking” category.
At 1,503 Observed Hands:
MU verdict: weak deviation worth tracking, but less suspicious than it looked at 665 hands. The main reason is replication.
The strongest apparent signal is still MU-R flop→turn upsets. At 665 hands it was 41 observed vs 29.21 expected, 1.40×. At 1,503 hands it is now 79 vs 60.74, 1.30×, with a one-sided exact tail around 0.68%. So the pooled result remains elevated, but the effect size decreased.
More importantly, the new 838 hands alone do not independently reproduce a strong anomaly: 38 flop→turn upsets vs 31.53 expected, 1.21×, with a tail probability of about 12.2%. Directionally high, yes; statistically remarkable, no.
The clean subset moved even more strongly toward normality:
Old 665 clean flop→turn: 3 vs 1.13 expected
New 838 alone: 1 vs 1.43 expected
Cumulative 1,503: 4 vs 2.56 expected, tail ≈ 24.5%
That is probably the single most important development. The bias-resistant sample did not replicate the apparent upset excess.
Everything downstream is also calm. Observed turn→river reversals are actually below expectation at 58 vs 68.29, while clean turn→river is 4 vs 4.69. Flop leaders ultimately losing showdown are 124 vs 113.06 expected, which is ordinary.
Frozen-pot MU-E is also very close to expectation, including the now-larger clean subsets: 28 clean pre-turn frozen-pot states and 45 clean pre-river states, with favorite realization slightly above expected rather than showing excess upset losses.
So compared with the 665-hand checkpoint, my interpretation shifts slightly toward ordinary variance/selection bias.
The MU-R flop→turn signal is still real enough numerically to keep watching, but it is being driven substantially by showdown-conditioned observations.
The new independent batch weakens the effect, and the clean sample actively regresses toward expectation.
Here are the tables with the observations.
TL;DR: The algorithm is fair and random. Your hand selection is not, and any losses or upsets are attributable to user error and variance, not the algorithm screwing you. That's the math.