The Plumber's Pressure Test: Finding Poker Leaks with AI (2026)

When a ceiling stain appears, good plumbers do not repaint the whole room. They isolate a section, pressurize the line, and listen for the hiss. Poker leaks work the same way. Your win rate is not bleeding from "bad luck"—it is dripping from repeatable joints: blind defense too wide, turn give-ups after floating, river thin-value misses. One joint at a time.

Most players review hands like homeowners with a wet sponge: they wipe every surface, fix nothing structural, and wonder why the stain returns next month. AI coaching tools change the workflow—not by replacing your judgment, but by pressurizing 20–50 hands in minutes and flagging which pipe section hisses loudest.

TryBluff's Leak Finder, PokerPro's aggregated leak reports, and similar AI coaches converge on the same threshold: pattern detection needs roughly 20–50 reviewed hands before recurring mistakes surface. Below that, you are guessing from one drip. This piece is a plumber's pressure-test manual—zone isolation, the three-valve protocol, joint signatures by street, one verified arithmetic spot, and when to distrust the gauge.

Why "Review Everything" Floods the Basement

Marathon reviews feel productive. They rarely fix leaks.

Approach	What happens	Leak outcome
Flood review (50+ hands, no filter)	Hand 38 gets skimmed; hand 9 repeats next session	Same joint keeps dripping
Pressure test (3–5 flagged hands + AI triage)	One joint isolated per cycle	Measurable BB/100 shift in 2–3 weeks
Vanity metrics (only biggest pots)	Cooler hands dominate; small-pot leaks ignored	Blind -45 BB/100 stays hidden

Sky Matsuhashi (Smart Poker Study) frames leak work as benchmark → diagnose → plug one hole. His blind-leak podcast uses BB/100 by position as the gauge: SB trending toward -15 BB/100 and BB toward -30 BB/100 over large samples is a healthy target band—not because those numbers are magic, but because they prove you are not paying full blind tax every orbit.

The plumber rule: depth beats flooding. Pressurize one zone per week.

Zone Map: Five Pipes Most Players Ignore

Before running AI, know which pipe you are testing. Leak tools aggregate tags; you choose the valve.

Zone	Gauge stat / filter	Hiss signature	First patch
Blind defense	BB/100 in SB and BB	Calling T6s vs UTG open; flatting instead of 3-betting TT	Tighten defend chart; 3-bet linear vs wide stealers
Flop c-bet defense	Fold-to-c-bet by position	Over-folding IP on dry boards; over-calling OOP with no plan	MDF check on one board class only
Turn give-up	Won $ at SD after flop call	Float flop → check-fold turn 70%+	Probe when PFR checks back turn
River value	W$SD vs WTSD gap	WTSD high, W$SD low—calling too wide, winning too little	Size up thin value; stop bluffing stations
3-bet pots	Filter 3-bet pots, sort by $ lost	Cold-calling 3-bets OOP with dominated broadways	Fold more OOP; 4-bet or fold pre

AI coaches excel at surfacing cross-hand patterns in these zones—"you folded to turn barrels 8/11 times after calling flop"—faster than manual filtering. They fail when the sample is 6 hands or when context (stack depth, opponent type) is stripped out. Treat AI output as a pressure gauge, not a demolition order.

The Three-Valve Pressure Test (15 Minutes)

Run this after each session or twice weekly. No FAQ module—just valves.

Valve 1 — Isolate (3 minutes)

While closing your client, tag hands where:

• You time-banked or felt "unsure"
• Pot exceeded ~30bb
• Result surprised you at showdown
• Structure was unfamiliar (3-bet pot, multiway, ICM)

Export or paste 20–50 hands from the session—not the whole database. PokerPro's leak report and similar tools note that 30+ hands minimum yields usable aggregation; 50–100 is better for position splits.

Pick one zone from the map above. Example: "This week = blind defense only." Everything else is out of scope—plumbers do not replumb the kitchen while testing the bathroom line.

Valve 2 — Pressurize with AI (8 minutes)

Paste the batch into your AI coach (TryBluff, DeepFold, FlopMentor, or any chat-based reviewer). Ask one focused prompt:

"Scan these hands for [zone] leaks only. List the top 2 recurring mistakes with hand counts, not solver frequencies."

Good AI output looks like:

• "7/22 BB defend spots: called with hands below chart vs early-position opens"
• "4/9 3-bet pots: check-folded turn after flop float with equity"

Bad output looks like:

• "You should bluff more" (no denominator)
• "Solver says 67.3% c-bet" (fabricated precision)
• "Fold preflop more" (no position or stack context)

Cross-check one flagged hand manually. If the AI says you over-folded river, verify pot odds: facing a $45 bet into $60, you need $45 / ($60 + $45) ≈ 43% equity to call—not a vibe.

Valve 3 — Patch One Joint (4 minutes)

Write one testable rule for next session:

• Not "play better in blinds" → "BB vs UTG open at 100bb: fold everything outside 22–99, AQs+, KQs, suited connectors 98s+"
• Not "barrel more" → "When flop float succeeds and villain checks turn, probe 55–66% pot with any pair+ or 8+ outs"

Track the joint for one week. Re-run the pressure test on the same zone. Matsuhashi's blind-leak fix sequence applies: know your BB/100 benchmark, defend deliberately, filter one hour daily on that zone until the gauge moves.

Joint Signature: Blind Defense (Worked Example)

No solver percentages—just arithmetic you can verify.

Setup: $1/$2 online, 100bb effective. You are in the BB. BTN opens to $5 (2.5x), SB folds. You hold J8s.

Pot before your decision: $1 + $2 + $5 = $8. Call costs $3 more.

Implied odds check: You need roughly 11:1 implied to profitably set-mine with a small pair; J8s is not set-mining—it is a dominated speculative call OOP. Against a typical BTN steal range (~45% opens), J8s realizes poorly: you will check-fold many flops, check-call one street and fold, or win a small pot when you hit.

Leak math: If you make this call 40 times per 1,000 hands at -$4 average outcome vs folding for $0, that is -$160 per 1,000 hands = -16 BB/100 from one joint—before postflop mistakes compound.

AI might flag "wide BB defend vs steals." Your job: pull three flagged hands, confirm the pattern, then remove J8s, T8s, Q8s from BB defend vs BTN at 100bb this week. One joint patched.

Compare to benchmark: folding every BB loses -100 BB/100 in that seat. Defending intelligently moves you toward -30 BB/100. The gap between -60 and -30 is almost always joint work, not hero calls.

When the Gauge Lies

AI leak detection fails in predictable ways:

Failure mode	Symptom	Fix
Thin sample	"Leak" from 4 hands	Wait for 20+ in-zone hands or merge two sessions
Missing context	AI ignores 20bb vs 100bb	Re-prompt with stack depth and stake
Result bias	Flags winning hands as mistakes	Ask "decision quality, not outcome"
Generic GTO noise	Solver-sounding % with no pot math	Discard; verify with pot odds only
Platform mismatch	Mis-parsed hand history	Fix format; one hand manually before batch

TryBluff notes Leak Finder is most useful after 20+ reviewed hands—same threshold. Treat sub-20-hand "leaks" as yellow alerts, not red demolition.

Live Table Translation

Pressure testing is post-session work. On the felt:

1. One sticky note per patched joint — "BB vs BTN: no J8s this week"
2. Pre-session gauge check — Glance at BB/100 in tracker; is the joint still dripping?
3. Showdown core sample — One blind defend showdown per session: was the hand in chart?

Live players without HUD can log seat + hand shorthand on phone, run the same 15-minute protocol weekly. The pipe does not care whether the gauge is digital or paper.

Maintenance Schedule

Frequency	Task	Minutes
After session	Tag 3–5 hands, pick one zone	2
2× weekly	Three-valve pressure test	15
Monthly	Compare BB/100 or overall bb/100 to prior month	5
Quarterly	Rotate zone map; do not patch same joint forever	10

Sites like PokerPro offer ~3 free leak reports per month on aggregated stats—enough for weekly cadence if you batch sessions. Heavier grinders may need unlimited tiers; the protocol stays identical.

If you want one integrated bench for import + AI chat + calculators, PokerJudge bundles those in a single post-session kit—but the pressure-test logic above works with any coach that accepts hand histories.

Closing Gauge

Leaks are joints, not curses. Isolate the zone, pressurize 20–50 hands, patch one rule, read the gauge next week. AI is the pressure pump; your arithmetic is the leak detector. Repaint when the pipes are dry—not before.