MTT — Strategy Divergence
from Cash NLHE

Most of your Cash NLHE playbook already works in MTT. That is the single most useful finding in this book.

At 17.6bb effective stacks, early tournament stage, no bounty active, the model's preflop opening frequencies from UTG, MP, CO, and BTN are within +2.8pp of their Cash equivalents. The positional hierarchy — UTG tightest, BTN widest — carries over exactly. So do the core theories: range advantage is positional, nuts advantage is distinct from equity advantage, blockers affect sizing, and in-position always realizes more equity than out-of-position.

If you coach tournament players who default to Cash ranges during the first few blind levels, the data says they're not wrong. They're approximately optimal at deep non-SB positions.

How MTT differs from Cash under the hood

This is not a Cash game with a different name. MTT replaces chip-EV with ICM equity as the optimization target, adds an optional bounty layer, and removes rake. The model sees one hand at a time — no blind escalation, no table balancing — but it receives the tournament's remaining-field ratio as an input signal. These structural differences create large strategic shifts at SB, in BB defense, and in bounty modes. They just happen to be invisible from UTG/MP/CO/BTN at deep stacks.

What transfers cleanly

The following Cash theories apply unchanged in MTT at deep non-SB positions:

• Range advantage is positional and persistent. The preflop range hierarchy carries forward to the flop.
• Nuts advantage is distinct from equity advantage. Holding more nut-eligible hands drives betting frequency and sizing.
• Range composition determines outlier realization. Rare strong hands in weak-dominated ranges over-realize.
• Checks condense ranges. When you check a spot where you'd bet your strongest hands, the remaining range narrows — standard Bayesian mechanics, format-neutral.
• Protection betting is real but narrow. The two-condition rule (extract value + fold out live equity) transfers; MTT doesn't add a new protection mechanism.
• Position improves equity realization universally. The in-position advantage persists under ICM.
• Positional VPIP gradient. UTG < MP < CO < BTN. Same ordering, same magnitudes.
• OOP equity realization is strictly lower. Acting last is format-neutral.
• Range composition by position shapes postflop action. Opener position determines cbet dynamics on texture-matched boards.

The 100bb VPIP parity table

Measurement conditions: 6-max NL, MTT-early (1000/1000 alive, no bounty), Cash baseline = same model with game_mode_code: normal and no tournament wrapper.

Opening VPIP at 100bb — Cash vs MTT-early. MTT-early stage · 1000/1000 alive · normal bounty · 6-max NL · 100bb effective

Source: mtt-deltas.md Table 1, 100bb rows.

UTG (+0.4pp) and MP (+0.1pp) are nearly identical to Cash. BTN (+0.8pp) sits inside the noise floor. CO (+2.8pp) is the widest deviation and still well within coaching-irrelevant territory. Published MTT theory predicts ranges "within 1–2pp of Cash at deep stack early stage" — three of four positions sit inside that envelope; CO lies just above it.

What this means in practice: If you are coaching deep-stack early-tournament preflop opens from UTG, MP, CO, or BTN, your Cash charts apply. Do not re-calibrate them for MTT.

Stack-depth sweep — does the parity hold as you shorten?

Δ VPIP (MTT-early minus Cash) across stack depths. MTT-early stage · 1000/1000 alive · normal bounty · 6-max NL · stack depth sweeps columns

Source: mtt-deltas.md Table 1.

UTG stays inside ±1pp at every depth. MP stays inside ±1.2pp. CO's full range spans about 5pp but never exceeds the practical adjustment threshold. BTN is the outlier: at 25bb (+3.6) and 20bb (+3.2) it opens noticeably wider than Cash. This is plausibly a small steal-adjustment effect — not large enough to prescribe from one round of data alone, but worth flagging.

AI% parity

All-in frequency at selected depths. MTT-early stage · 1000/1000 alive · normal bounty · 6-max NL · stack depth sweeps columns

Source: mtt-deltas.md Table 1.

At every stack depth from 15bb up, non-SB all-in frequencies are within 2pp of Cash. Even at 10bb the largest deviation is CO (8.9% MTT vs 7.5% Cash). The Harrington M-ratio zones describe the same qualitative shift that happens in Cash at the same depth — push-fold is not fundamentally different for non-SB positions in MTT.

What this means in practice: Non-SB short-stack push-fold in this model is effectively the Cash short-stack strategy. For ICM-aware push-fold, use dedicated push-fold tools (HRC, ICMizer) rather than this model. See Part 4 for the full short-stack treatment.

What changes as you shorten

The deep-stack equivalence holds cleanly across 100/50/25bb for UTG, MP, and CO. BTN shows a soft widening at 25–20bb, consistent with a mild steal-frequency adjustment but too small to coach on in isolation.

The real divergences — and the rest of this book — live elsewhere. SB is a fundamentally different position in MTT, and that story starts in Part 2.

What this part does not cover

• Part 2 — SB regime. SB plays dramatically wider in MTT than Cash. Do not extend the "Cash charts apply" conclusion to SB.
• Part 3 — Bounty pricing. PKO and flat-KO produce direction anomalies the model has not resolved.
• Part 5 — BB defense. BB defends wider in MTT across every cell we tested. The mechanism is structural, not ICM-driven.
• Part 6 — Flop c-bets. Cbet frequency jumps on most board textures, with one paired-low exception.
• Part 7 — Stage binning. The model treats tournament stage as binary, not graduated — a known limitation.

If you coach MTT players using Cash SB charts at deep stacks, you are coaching them to fold too much. The gap is not subtle.

At +19.5 percentage points wider than Cash at 100bb, the MTT small blind is the single largest positional divergence between the two formats in our data. Every other position — UTG, MP, CO, BTN — stays within a few points of Cash at the same stack depth. SB doesn't just drift wider. It plays a fundamentally different game.

Measurement conditions: SB at MTT-early stage (1000/1000 alive, no bounty), 6-max NL, Cash baseline same model with game_mode_code: normal. The SB regime ends at the push-fold crossover (12–10bb) — see Part 4.

The SB VPIP gap by stack depth

SB opening frequency, Cash vs MTT-early. MTT-early stage · 1000/1000 alive · normal bounty · 6-max NL · SB · stack depth sweeps rows

Source: mtt-deltas.md Table 1, SB rows.

At 100bb the MTT SB enters the pot 77.4% of the time — nearly four hands out of five. The widening holds through 50bb (+13.8pp), 25bb (+16.1pp), and 20bb (+14.1pp). It compresses at 15bb (+8.8pp) and almost vanishes at 12bb (+2.3pp). By 10bb the SB is actually -0.6pp — the two formats converge and then the push-fold regime takes over.

Critically, the all-in frequency at 100bb is 0.0% in both Cash and MTT. The extra VPIP is not shoves. It is limps and raises.

What this means in practice: If you are coaching an MTT player's SB at 50–100bb and your reference chart comes from Cash, you are leaving a significant amount of VPIP on the table. The model says the SB should be entering the pot far more often than Cash — primarily through completing, not through raising or jamming.

Most of the extra VPIP is limps

At 100bb MTT-early, the SB enters with a limp roughly 68% of hands, computed from the open-action breakdown in the raw data. That is the majority of the 77.4% VPIP figure — the SB is completing into the pot, not raising wider.

One caveat: that 68% limp anchor is a single measurement at 100bb. The full limp/raise/fold breakdown across all seven stack depths lives in the raw batch data and has not been published as a sweep. We have one anchor, not a curve.

Why this happens

Three forces compound in the SB's favor in MTT.

Published tournament theory attributes deep-stack SB widening primarily to antes. The model's direction aligns with that literature. Its magnitude does not — the model overshoots what published charts report.

A scope-and-magnitude warning

The direction of this finding — SB plays wider in MTT than in Cash at deep stacks — has partial support from external theory on ante-driven opening. The magnitude is a different story.

The +19.5pp gap at 100bb has no direct solver corroboration. External solver baselines for SB at deep-stack MTT-early are sparse; what does exist covers short-stack or bounty scenarios, not the matched 100bb normal-bounty cell. Even at shallow stacks where external baselines exist, the model SB is lighter than some Nash references — at 10bb the model shows 63.0% VPIP, while one external baseline reports SB at the same depth around 75%.

Coaching guidance: the MTT SB regime is real and directional. The 68% limp rate at 100bb is model behavior, not a verified prescription. Use a published MTT SB chart as your prior and widen from there — but don't adopt the model's exact frequencies without a solver cross-check.

Where the regime ends

At 12bb the SB gap collapses to +2.3pp. At 10bb it flips to -0.6pp — the MTT SB is now fractionally tighter than Cash. Below this threshold, push-fold takes over.

Part 4 covers the SB push-fold crossover in detail. The preview: at 5bb the MTT SB jams +14.7 percentage points wider than Cash. At 10–15bb the sign flips and the MTT SB jams -8.4 to -11.0 percentage points tighter. The crossover sits between 7bb and 10bb.

Two things jumped out the moment we ran the bounty sweep.

First, mystery bounty has two distinct regimes. Pre-ITM, the model treats mystery exactly like normal — identical output on every cell. Post-ITM, mystery tightens VPIP by up to -8.7pp versus normal. The split traces to a specific rule in the training code, and it is both real and useful.

Second, the model's PKO and flat-KO direction is sign-flipped versus published MTT theory. Where the literature predicts bounty modes should widen ranges, our model tightens — and flat-KO tightens more than PKO, which is the opposite ordering the textbooks predict. This is a model behavior flag, not a coaching prescription.

Here is the data.

3.1 How the four bounty modes work

The MTT training code supports four bounty types on top of the base ICM reward.

Normal (no bounty). Pure ICM equity delta. No head bounty, no knockout bonus. This is the control condition in every comparison below.

Flat-KO (flat knockout). Every player carries a fixed head bounty. On elimination, the killer collects the victim's entire head bounty instantly — no compounding, no carry.

PKO (progressive knockout). Buy-in splits into a prize pool and a bounty pool. On elimination, the killer collects half the victim's head bounty immediately. The other half adds to the killer's own head bounty, which they carry forward. That compounding over a multi-elimination tournament lifetime is what makes PKO strategically unique.

Mystery bounty. A shared pool of hidden-value envelopes. On elimination during the mystery phase (post-ITM only), the killer draws one random envelope. Pre-ITM, the rules implementation downgrades mystery to normal.

3.2 The bounty pricing table

Measurement conditions: 6-max NL, MTT-mid stage (500/1000 alive), symmetric-stack fixtures (covering-vs-covered asymmetry NOT exercised). Bounty modes per GAME-RULES.md §8.

Model VPIP by position × stack across four bounty modes. MTT-mid stage · 500/1000 alive · 6-max NL · bounty mode sweeps columns

Source: mtt-deltas.md Table 3 lines 76–101

BTN and SB carry the largest deltas. SB at 40bb flat-KO shows the single deepest tightening: -12.5pp. UTG cells cluster near zero for both bounty types.

3.3 The PKO vs flat-KO sign anomaly

This is the model's most counterintuitive bounty finding, and it contradicts published theory on two dimensions.

Direction. Across all cells, PKO averages -5.0pppp VPIP shift at BTN 25bb. The aggregate mean across the full grid is −1.40pp for PKO and −3.70pp for flat-KO. Both are negative — the model tightens under bounty. Published canonical MTT references unanimously predict that bounty modes should widen ranges, because eliminating an opponent collects their head bounty and adds positive EV to every call or shove.

Ordering. Flat-KO tightens more than PKO at every BTN and SB cell. At BTN 25bb the flat-KO Δ is -9.7pp versus PKO -5.0pppp — a ratio of roughly 1.94×. At BTN 60bb the ratio widens to 4.67×. Published theory predicts the reverse ordering: PKO's progressive carry should create stronger widening pressure than flat-KO's one-shot payout.

Both the direction and the ordering are model behavior, not solver-corroborated. Do not use these numbers to coach absolute PKO or flat-KO range adjustments. The likely explanation — though unresolved — is that the trainer scores only the instant payout on each hand and does not simulate the multi-hand bounty carry that makes PKO "stronger" in tournament-lifetime EV. A model that never sees the carry compounding may correctly respond to a smaller per-hand PKO incentive than a full flat-KO collection.

3.4 Mystery bounty — two regimes, not one

The mystery column in the table above is a wall of +0.0. That looks like the model ignores mystery entirely — but it does not. The story has two chapters.

Pre-ITM (mid-stage, the data above). The training rules downgrade mystery_bounty to normal whenever the field has not yet reached the money. All queries in Table 3 fire at 500/1000 alive (mid-stage), which triggers that downgrade. So 20 of 20 cells matching normal is rules-correct behavior, not a model limitation.

Post-ITM (ITM and final table). A follow-up batch queried mystery at post-ITM stages where the downgrade rule does not fire. Result: 36 of 40 cells showed mystery different from normal, with the maximum shift reaching 8.2pp VPIP tightening at SB 10bb on the final table. Mystery tightens VPIP post-ITM, consistent with the idea that the remaining mystery prize pool increases the survival EV of staying in the tournament.

This resolves the earlier open question about whether the model reads the mystery parameter at all. It does — but only when the rules allow it. Pre-ITM, mystery IS normal. Post-ITM, mystery is its own regime.

3.5 The real late-game lever: covering versus covered

The table above measures bounty mode effects with symmetric stacks. It does not capture the largest documented bounty-driven adjustment in published PKO theory: whether you cover your opponent or your opponent covers you.

Published canonical theory documents a 32pp spread in RFI ranges at BTN 25bb on the bubble between a covered player and a covering player. That is larger than any delta in this section — and we did not test it. Round 1 used symmetric stacks throughout, so the covering-vs-covered asymmetry was never exercised.

Until that asymmetry test runs, every PKO number above reflects "symmetric-stack wrapper effect," not the realized EV of bounty hunting in practice. If you are coaching PKO and need covering-vs-covered guidance, use canonical push-fold references and published solver charts — they are the authoritative source until our model's asymmetric-stack behavior is verified.

The SB regime that made you limp at deep stacks flips below 8bb. At 5–7bb the SB shoves much wider in MTT than Cash. At 10–15bb it shoves tighter. And for every other position at the table? The MTT wrapper barely registers.

That two-sentence summary is the entire short-stack story from Round 1. This part unpacks it with the data.

The push-fold framework

In the 5–15bb band, the open action collapses toward shove-or-fold. Dan Harrington's M-ratio zones — green (M ≥ 20, full toolbox), yellow (10–20, cautious), orange (6–10, any raise commits), red (1–6, push-or-fold) — are the standard framework. MTT adds two pressures on top of Cash push-fold math: ICM (each chip is worth less as your stack grows) and bounty payoff (Part 3). For non-SB seats, neither force adds a measurable signal in our data. Cash push-fold ranges hold.

Measurement conditions: 6-max NL, push-fold stacks (5–15bb), normal bounty mode (PKO/flat-KO push-fold not separately tested).

Non-SB push-fold: Cash ranges hold

Every UTG, MP, CO, and BTN cell in the table below lands within a few percentage points of Cash. The model does not tighten at the bubble the way textbook ICM predicts — and it does not widen either. It simply plays Cash.

All-in frequency by position and stack depth — non-SB seats. Cash baseline vs MTT early/bubble/ft · 6-max NL · stack depth sweep rows · normal bounty

Source: mtt-deltas.md Table 4 lines 109–139

The largest non-SB delta in the entire table is BTN 10bb at -2.7 percentage points. That is noise, not signal. For ICM-aware non-SB push-fold, use published Nash push-fold solvers — this model does not add an MTT-specific adjustment.

SB push-fold: the non-monotonic crossover

SB is the one position where the MTT wrapper rewrites push-fold strategy. The direction depends on stack depth — and it flips sign.

SB all-in frequency by stack depth. Cash baseline vs MTT-early · 6-max NL · SB · stack depth sweep rows · normal bounty

Source: mtt-deltas.md Table 4 lines 134–139

At 5bb: +14.7 percentage points wider. At 12bb: -11.0 percentage points tighter. The sign flips somewhere between 7bb and 10bb — we did not test 8bb or 9bb in Round 1, so the exact crossover is not localized.

What changes between 7bb and 10bb at SB

The deep-stack SB regime (limp-heavy, covered in Part 2) needs the option of completing and then defending postflop. Below roughly 8bb, completing is dominated by jamming — BB does not have enough stack to defend wide against a shove.

One caveat. Published external Nash push-fold benchmarks for SB at 10bb report much higher VPIP than our model. Our MTT SB at 10bb shows 63.0% VPIP; published benchmarks report roughly 75%. Direction matches (SB wide), magnitude does not. At 15bb the gap is even larger. If you are coaching SB push-fold at 10bb or deeper in MTT, the external Nash reference is a safer anchor than this model.

A push-fold prescription at five stack depths

For UTG, MP, CO, and BTN at 5/7/10/12/15bb: use your Cash push-fold chart. The MTT model shows no actionable deviation at any of these positions — every delta is within a few percentage points of Cash.

For SB, the model outputs at each stack depth:

SB push-fold prescription by stack depth. MTT-early · 6-max NL · SB · normal bounty

Source: mtt-deltas.md Table 4 lines 134–139

The 5–7bb SB widening is the most coaching-relevant finding in this table. The 10–15bb SB under-shoving is where the model diverges from established solver outputs and should not be taken as a prescription.

If you use your Cash BB defense ranges in MTT, you are folding too much. Not slightly — the gap is +12.7 to +16.9 percentage points wider on every single cell we tested. That is 62 theories catalogued in this book, and this one finding — BB defends wider in MTT than Cash — showed up in every opener, every raise size, every stack depth, and every tournament stage we queried.

The kicker: this is not a bubble-pressure effect. It shows up at MTT-early (full field, no ICM) just as strongly as on the bubble.

Measurement conditions: 6-max NL, BB facing opener, MTT-early baseline (1000/1000 alive). Stage-insensitivity holds at the same cells across early/bubble/itm/ft within ±1.5pp.

The headline delta

Out of 96 tested cells, 96 show positive defense deltas. The minimum widening is +6.1pp (bubble, UTG 3.0bb open, 100bb). The maximum is +16.9pp (early, CO 2.5bb open, 20bb). The pattern: the gap is largest where Cash defense was tightest — where the MTT wrapper has the most room to fill.

Defense by opener at 100bb

BB defense rate by opener position. MTT-early stage · 1000/1000 alive · 2.5bb open · BB vs opener · 6-max NL · 100bb effective

Source: mtt-deltas.md Table 5, early stage, 2.5bb open, 100bb rows.

The relative ordering is preserved — you still defend more against the SB than against UTG. What changes is the floor. Every position lifts by roughly the same band, and the widening is largest vs UTG because that is where Cash defense was the lowest to begin with.

What this means in practice: Your Cash defense chart still determines the shape of your defending range. MTT shifts the entire chart up. Against a UTG open at 100bb, you are defending roughly half your range in MTT versus a third in Cash.

Defense by stack depth — vs CO 2.5bb open

BB defense rate by stack depth, CO opener. MTT-early stage · 1000/1000 alive · 2.5bb open · BB vs CO · 6-max NL · stack depth sweeps rows

Source: mtt-deltas.md Table 5, early stage, CO 2.5bb rows.

The gap grows as stacks get shallower. At 20bb, MTT BB defends nearly half its range against CO — up from a third in Cash. The practical read: short-stack MTT defense adjustments are even larger than deep-stack ones.

The mix shift — call up, 3-bet down

The extra defense is almost entirely calls. 3-bet frequency drops in every comparable cell.

Action breakdown for BB defense, selected cells. MTT-early stage · 1000/1000 alive · 6-max NL · opener×raise×stack labeled in rows

Source: mtt-deltas.md Table 5, early stage rows.

At 20bb facing SB, MTT BB calls 65.5% (+30.3pp) and 3-bets only 11.4% (+30.3 more calls absorb the defense widening, while 3-bets drop by 11.4 of the range). The direction is universal: call more, 3-bet less.

What this means in practice: If you are widening your MTT BB defense by 3-betting more, you are doing it backwards. The model shifts defense into the calling lane. 3-bets actually drop — consistent with ICM-era advice to reduce the frequency of bloating pots, even though the overall defense rate goes up.

Why this happens

Stage-insensitivity — this is NOT ICM

BB defense at early vs bubble, representative cells. 2.5–3.0bb open · 6-max NL · opener×raise×stack labeled in rows · early vs bubble compared

Source: mtt-deltas.md Table 5, early vs bubble stage rows.

Within-stage delta never exceeds ±1.5pp. The BB defense widening is already fully present at MTT-early — before any ICM engagement. This is the structural MTT wrapper (ante + small opens + no rake), not a response to pay-jump proximity.

MTT cbets are dramatically wider than Cash. On 5 of the 6 textures we tested, the model bets the flop +29.0 to +29.4 percentage points more often than it does in Cash. The lone exception — paired-low 772p — flips the other way, with MTT cbetting less than Cash.

That exception is what makes this section interesting. The direction of the cbet shift isn't random. It traces directly back to the BB defense widening from Part 5: when BB's added hands miss the board, the IP cbettor prints fold equity. When those added hands hit the board, fold equity collapses and the cbet incentive inverts.

Measurement conditions: 6-max NL, single-raised pots, MTT-early stage (1000/1000 alive), normal bounty. Cbet behavior is stage-insensitive within the MTT regime — see Part 7 for binary-stage binning.

The cbet table by texture

CO and BTN flop cbet frequency and average bet size, Cash vs MTT, across 6 board textures. MTT-early stage · 1000/1000 alive · normal bounty · 6-max NL · CO/BTN vs BB SRP · 100bb effective · texture sweeps rows

Source: mtt-deltas.md Table 6, early-stage rows (lines 242–269)

Ten of the 12 rows show double-digit positive deltas. CO Q72tt is the largest at +29.4pp; CO A72r at +29.0pp is close behind. The two 772p rows are the only negative cells: CO -10.7pp and BTN -8.7pp.

What this means in practice: If you have been applying Cash cbet frequencies in early-stage MTT single-raised pots, you are likely under-betting dry and two-tone boards by a very large margin. On A72r, Q72tt, and J83ft, the model bets nearly every time.

Why dry and connected boards elevate

Why 772p reverses

Average bet size is also larger

On every non-paired texture, MTT average bet size is larger than Cash — by roughly +0.1 to +0.2 pot. This contradicts published MTT theory that predicts smaller postflop sizing under ICM pressure. The contradiction is consistent with the stage-insensitivity finding below: the model's cbet behavior appears driven by the structural MTT wrapper (antes, no rake, smaller opens), not by ICM.

Stage-insensitivity

Within the MTT regime, flop cbet frequency barely moves across tournament stages. Early vs bubble deltas from Table 6 are all within ±1.7 percentage points:

Early vs bubble flop cbet frequency on selected opener × board cells. MTT-early vs MTT-bubble · normal bounty · 6-max NL · CO/BTN vs BB SRP · 100bb effective

Source: mtt-deltas.md Table 6, early vs bubble rows (lines 242–281)

The cbet elevation is established at MTT-early and stays put through the bubble. Part 7 covers the binary-stage binning mechanism in detail.

Confidence pending solver cross-check

The direction of the cbet elevation is corroborated by external research attributing MTT postflop aggression to ante-driven range widening. The exact magnitudes — some exceeding +29 percentage points — are model behavior that goes beyond what published Cash-vs-MTT solver comparisons report. Until a second-checkpoint replication confirms the numbers, treat the magnitudes as directional, not prescriptive. The direction itself (more cbetting in MTT on dry boards, less on paired-low) has strong structural grounding.

The canonical theory says ICM pressure tightens your range gradually from mid-tournament through the bubble. The model says something different: it sees exactly 2 stages, not 6. Early, mid, and late produce identical output. Bubble, ITM, and final table produce identical output. The only difference is between those two groups — and the gap is tiny.

If you try to interpolate "mid vs late" or "ITM vs FT" from the model's output, you are reading checkpoint noise. Use the binary output and stop there.

The canonical stage framework

Standard MTT theory gives you two interlocking tools for stage-aware play. First, Harrington's M-ratio zones (M = stack divided by the cost of one orbit) carve the tournament into five postures based on your chip pressure — from Green zone (full toolbox, cash-like play) down through Red zone (push-or-fold). Second, Bubble Factor — the ratio of equity risk on loss to equity gain on win — captures how pay-jump proximity changes the EV of every pot. When Bubble Factor exceeds 1, losing a chip hurts more than winning one helps.

Modern empirical work pushes ICM awareness earlier than older advice suggested. A 2025 analysis found that ICM-aware play outperforms chip-EV play starting when just 17.1–62% of the field remains — substantially before the bubble. That contradicts the Harrington-era treatment that says deep-stack early MTT is "basically cash."

What the model actually sees — the binary binning

Measurement conditions: 6-max NL, normal bounty. Stages per GAME-RULES.md §5: early/mid/late/bubble/itm/ft. The 6 columns collapse into 2 groups in the model's output.

BTN VPIP across 6 tournament stages and 7 stack depths. 6-max NL · BTN open · normal bounty · stack depth sweeps rows · stage sweeps columns

Source: mtt-deltas.md Table 2 lines 50–62

Look at the first three columns. They are byte-identical at every stack depth. Now look at the last three columns. Also byte-identical. The only variation is between the two groups, and the maximum gap across all 7 rows is 0.7pp.

UTG opens — same pattern

UTG VPIP across 6 tournament stages and 7 stack depths. 6-max NL · UTG open · normal bounty · stack depth sweeps rows · stage sweeps columns

Source: mtt-deltas.md Table 2 lines 64–74

Same story from the tightest opening position. Two groups, nothing more. Within each group the columns are identical. The between-group shift at UTG is 0.6pp or less.

Where the switch happens

The model flips from Group A (early/mid/late) to Group B (bubble/itm/ft) somewhere between roughly 25% of the field remaining and roughly 16% remaining. Round 1 did not localize the boundary more precisely — the 6-stage discrete grid does not have enough resolution to pin it. A Round 2 query that sweeps the stage parameter in finer steps would resolve this.

Why this matters for coaching

Real MTT theory gives you a graduated pressure curve. Harrington's M-zones tell you when your chip stack forces action. Bubble Factor tells you when pay-jump proximity warps every call. Published MTT literature predicts a smooth tightening through mid → late → bubble stages.

The model gives you none of that granularity. It returns one number for the entire pre-bubble window and a second, slightly tighter number for everything post-bubble. The gap between the two is 0.7pp at most.

A worked example — BTN at 20bb across stages

BTN opens 33.8% VPIP at 20bb for early, mid, and late — all identical. At bubble, ITM, and FT the number shifts to 33.5%. That 0.3pp shift is the model's entire ICM signal at this cell. Published MTT theory at a matched anchor (BB vs BTN at 30bb, bubble vs chip-EV) reports tightening of roughly 6 percentage points — the model disagrees on both magnitude and direction.

Twenty-six coaching takeaways from the MTT research, organized by pillar. Some you can coach on today. Others are explicit "do NOT coach from this model" warnings — flagged inline.

Each actionable below is consolidated from the Actionables master list in the source research. Per-pillar evidence and confidence rationale lives in the corresponding per-pillar sections of mtt-theory.md.

A — Equity & Ranges

1. At 100bb MTT-early at UTG/MP/CO/BTN, your Cash opening range applies. Strong evidence; confirmed by model data and literature. (See Part 1.)
2. At 25bb/20bb BTN, consider a steal-widening of roughly +3.6 percentage points vs your Cash 25bb range. Treat as a soft adjustment, not a hard rule — low confidence from one data round alone. (See Part 1.)
3. Treat the SB +19.5pp deep-stack widening as model behavior, not prescription. The magnitude exceeds solver literature. Hold back the limp rate unless an external solver cross-check confirms it; use your preferred MTT SB chart as the prior. Model behavior, not solver-corroborated; low confidence. (See Part 2.)

B — Frequencies & Balance

4. BB defense is wider in MTT than Cash by +6.1pp to +16.9pp at every opener, raise size, stack, and stage we queried. Driven by calls up with 3-bets down. This contradicts the standard literature direction. Model behavior, not solver-corroborated; low confidence. (See Part 5.)
5. Trust the solver over the model for BB defense totals; use only the mix-shift direction (call more, 3-bet less) as a weak signal. If your Cash defense rate felt right, do not assume you are folding too much in MTT just because the model wants you to defend wider. Model behavior, not solver-corroborated; low confidence. (See Part 5.)
6. The BB defense widening is NOT ICM-driven (early ≡ bubble). Treat it as a structural wrapper effect that pre-dates bubble pressure. Moderate evidence. (See Part 5.)

C — Position & Information

7. The MTT SB plays a fundamentally different regime from Cash at deep stacks — roughly a 68% limp rate at 100bb. The hypothesis is that antes, small raises, and wide BB defense all favor completing over raising. Verify against a solver before coaching. Model behavior at this magnitude; low confidence. (See Part 2.)
8. Flop c-bet elevation tracks range advantage monotonically on 5 of 6 tested textures; inverts on 772p paired-low. Moderate evidence. (See Part 6.)

D — Sizing

9. SB 5–7bb is much wider all-in in MTT than Cash (+9.2 to +14.7 percentage points wider). Plausibly consistent with Nash push-fold at this depth; verify quantitatively against external solver anchors. Low confidence. (See Part 4.)
10. SB 10–15bb all-in frequency drops in MTT (-8.4 to -11.0 percentage points vs Cash) but VPIP stays near Cash. Interpret as "model prefers raise/call over jam at this depth," not "fold more." Low confidence. (See Part 4.)
11. Non-SB short-stack push-fold (UTG/MP/CO/BTN, 5–15bb) is Cash-like in MTT. The model does NOT tighten at short stacks for these positions the way ICM theory predicts. For ICM-aware non-SB push-fold, use external solver charts, not this model. Model behavior, not solver-corroborated; low confidence. (See Part 4.)
12. MTT average c-bet size is slightly larger than Cash on non-paired textures. This contradicts the literature prediction of smaller sizing under ICM. Low confidence. (See Part 6.)

E — Board Texture

13. On dry high-card boards (A72r, K84r, Q72tt, J83ft), c-bet +11.6pp to +29.0pp wider in MTT than Cash at 100bb MTT-early. Strong evidence. (See Part 6.)
14. On 987mn (mid-connected monotone), c-bet +26.3pp to +28.3pp wider in MTT. Moderate evidence. (See Part 6.)
15. On paired-low 772p, c-bet direction INVERTS — MTT c-bets -8.7pp to -10.7pp LESS than Cash. Do not generalize to paired-high boards without further data. Moderate evidence. (See Part 6.)
16. CO gets a stronger c-bet lift from the MTT wrapper than BTN on A/K/Q-high dry textures. Moderate evidence. (See Part 6.)

F — C-Betting / Multi-Street

17. Flop c-bet is stage-insensitive — once inside the MTT regime, c-bet stays elevated at every stage. All sampled early-to-bubble deltas are within ±-1.7pp. Moderate evidence. (See Part 6.)
18. All turn and river theories are UNTESTED in MTT Round 1. If you are coaching turn barrel, river value, donk, or delayed c-bet — do not use this model as a direct reference. Use Cash theory plus solver corroboration until further data resolves. Pending Round 2 verification. (See Part 6.)

G — Advanced

19. Tournament stage is binary in this model, not graduated. Early/mid/late are treated as one group; bubble/ITM/FT as another. Within-group delta is at most 0.7pp. Real MTT theory has more granularity; external solver benchmarks expect graduated ICM onset. Model behavior, not solver-corroborated; low confidence. (See Part 7.)
20. The stage switch happens somewhere between late (~25% of field remaining) and bubble (~16%). Round 1 did not localize the exact threshold. Pending Round 2 verification. Moderate evidence on direction. (See Part 7.)
21. Do not rely on Cash Bubble Factor prescriptions — the model does not tighten under Bubble Factor > 1. At bubble stage, BB still defends wider than Cash. For coaching Bubble Factor decisions, use an external solver, not this model. Model behavior, not solver-corroborated; low confidence. (See Part 7.)

H — 3-Bet Pots

22. Do not use this model for 3-bet pot MTT coaching. The entire 3-bet pot pillar is UNTESTED in Round 1. Pending Round 2 verification. (See Part 7.)

M — MTT-Native

23. PKO and flat-KO both tighten ranges on average in the model; flat-KO tightens MORE than PKO. The direction and magnitude ordering BOTH contradict textbook bounty theory. Do NOT cite the model as a PKO/flat-KO prescription — use external solver numbers. Model behavior, not solver-corroborated; low confidence. (See Part 3.)
24. BTN has the largest flat-KO tightening (-8.4 to -9.7 at 25/40/60bb) and SB the largest at 40bb (-12.5). If the flat-KO direction is an artifact, these position-specific magnitudes are also unreliable. Model behavior, not solver-corroborated; low confidence. (See Part 3.)
25. Mystery bounty behaves exactly like normal in the model at mid-stage (20 of 20 cells identical). Do NOT use this model for mystery bounty strategy at mid-stage. Post-ITM data shows differentiation. Model behavior; low confidence at mid-stage. (See Part 3.)
26. Covering-vs-covered asymmetry is the primary PKO driver and it is UNTESTED in Round 1. The symmetric-stack fixtures under-exercise the documented 32pp RFI spread. Pending Round 2 verification. Moderate evidence from literature. (See Part 3.)

Where the model is coachable — and where it is not

The safest coaching applications from this research are: deep-stack non-SB opening ranges at 100bb (Part 1), the direction of BB defense widening (Part 5 — call more, 3-bet less), and dry-board c-bet elevation (Part 6 — strong evidence on five of six tested textures). The model is NOT coachable for PKO/flat-KO strategy or mystery bounty mid-stage (Part 3), stage-granular ICM (Part 7), turn and river play (Pillar F), 3-bet pots (Pillar H), or shallow-stack non-SB push-fold under ICM (Part 4). A full scope-limits enumeration lives in Part 9.

A model that tells you where it's confident is useful. A model that also tells you where it's silent is trustworthy. This part is the second kind.

This part enumerates what Round 1 did NOT cover. Numeric findings throughout the book are scoped per their per-table measurement labels (Parts 1–7).

Round 1 query scope

The research ran 62 theory entries across 23 Cash-transferred, 8 amplified, 15 untested, and 10 MTT-native classifications. The query plan covered preflop opens across all positions and stack depths, BB defense across stages and openers, flop c-bet on 6 board textures, and the stage gradient.

It did NOT cover turn or river play, 3-bet pots, multiway postflop, asymmetric-stack bounty fixtures, satellite formats, or 4-bet ranges.

What is untested (Round 1 scope choices)

These entries have no model data. They carry a specific Round 2 query that would promote them toward a verdict.

Untested entries by pillar after Round 1.

Source: mtt-theory-foundation.md v2.0.0 §Coverage map; mtt-baselines.md v1.0.0 §2 Verdict summary table

That is 15 entries with no model evidence. Every one of them has a specific Round 2 batch identified.

Round 2 queue

The next research cycle targets these batches in priority order.

P0 (highest priority — resolves the largest open questions):

1. Covering-vs-covered asymmetry — tests M4 with hero-covers vs hero-is-covered queries at BTN bubble. External solver baselines predict a range spread that dwarfs position and stage effects.
2. Stage-gradient fine scan — scans the stage signal in small steps to find where the binary step occurs. Resolves the binary-binning concern and four related entries (G4c, G4e, M6, M10).
3. Ante-isolated BB defense — tests BB defense at matched stacks with ante explicitly toggled. Anchors whether the widening pattern is ante-driven or a broader model behavior.

P1: post-ITM mystery bounty queries (resolves KI-12 scope); SB limp/raise decomposition at deep stacks; low-connected flop c-bet (canonical Cash board class missing); OOP 3-bet pot c-bet at shallow stacks.

P2–P4: cross-opener range composition in MTT; BB overcall connectivity in 3-way; BTN 3-bet pot c-bet bimodality; turn rank sweeps; river value-bet threshold at bubble; per-hand sizing and overbet frequency; overpair hierarchy on wet boards; bluff-availability sizing constraint; protection-bet frequency; mixed-strategy fractions.

Known issues

KI-12 — Mystery bounty collapses to normal. The model treats mystery bounty identically to normal bounty in every tested cell. The cause is a rules-level guard: when the fraction of the field remaining exceeds the in-the-money ratio, the training code downgrades mystery to normal. All Round 1 mystery queries triggered this guard. Until post-ITM mystery queries land (P1 batch), do not use this model for mystery-bounty strategy.

MC-6 — Binary stage binning. The model partitions tournament stages into exactly two groups — {early, mid, late} vs {bubble, ITM, final table} — with within-group variation no larger than 0.7pp. External solver baselines expect a graduated response across stages. This is the single largest gap between our model and the wider literature on stage-dependent ICM. Do not use the model for stage-granular ICM adjustments until the P0 fine scan resolves whether the step function is a training-coverage artifact or a genuine model limitation.

MC-5 — Deep-stack meta-signal. SB's preflop widening at deep stacks in MTT-early (+19.5pp at 100bb) is anomalously large. This is flagged as a candidate method caveat because the magnitude is not corroborated by external solver baselines at the same cell. The direction is partially supported, but the exact magnitude may reflect a training-distribution thin spot at SB-MTT-100bb.

MC-7 — BB defense widening magnitude. BB defends wider than Cash across every tested cell — the direction is solid (all 96 cells positive). The open question is magnitude: our data shows deltas of +6 to +17 percentage points, while an external ante-isolated benchmark predicts a much larger gap. The P0 ante-isolated batch will disambiguate.

Training-signal gaps

Three structural gaps in the training signal emerged from Round 1. These are not analytics bugs — they require coordination with the training team.

(1) PKO single-hand reward truncation. The training reward for PKO includes only the instant portion of a knockout bounty. The carry portion — the half that adds to the winner's own head bounty and compounds across future hands — is invisible to the single-hand training window. This explains the M2 sign anomaly where the model tightens under bounty modes instead of widening. Decision needed: retrain with carry-included reward, or document that the model's PKO behavior represents per-hand strategy only.

(2) Mystery bounty pre-ITM guard. The training code downgrades mystery bounty to normal before ITM. External solver baselines treat pre-mystery as strategically distinct. Decision needed: align training rules with the literature framing, or document the model's custom pre-mystery interpretation as intentional.

(3) Stage exposure. The only tournament-phase signal the trainer sees is left_ratio (fraction of field remaining). There is no explicit stage category, no pay-jump proximity signal, and no Bubble Factor estimate in the observation. The binary-binning finding (MC-6) is likely downstream of this sparse feature set. Decision needed: enrich stage features for the next training cycle.

If you know the canonical MTT literature — Harrington's M-zones, the ICM bubble-factor math, the PKO covering-stack widening — you need to know where the model agrees, where it gets the direction right but the magnitude wrong, and where it flatly contradicts the published theory. This part is that map.

Measurement conditions: Round 1 dispositions only (497 MTT + 417 Cash queries, single checkpoint). Round 2 dispositions will update on the next book version.

MTT-LIT cluster-level verdicts

Round 1 cluster-level verdicts across 60 literature claims. Round 1 dispositions · 497 MTT + 417 Cash queries · single checkpoint · classifications: CONFIRMED / PARTIAL / DISCREPANT / UNTESTED

Source: mtt-literature-extracts.md §Round 1 dispositions

Summary

Roughly 8 of 60 literature claims are confirmed, around 30 are partial, about 18 are discrepant, and 4 remain untested. The model agrees cleanly with the deep-stack non-SB cash-like prediction (MTT-LIT-34, 51) and the directional drop in 3-bet bluffing under ICM (MTT-LIT-54). It disagrees with most postflop ICM predictions (MTT-LIT-52, 56, 58) and the entire mystery / flat-KO ordering structure (MTT-LIT-27..33).

MTT-SOL section-level verdicts

Round 1 section-level verdicts across 45 external solver anchors. Round 1 dispositions · solver anchors · single checkpoint · classifications: CONFIRMED / PARTIAL / DISCREPANT / UNTESTED

Source: mtt-solver-index.md §Sections 1–9

Summary

The largest solver confirmation is MTT-SOL-13 (UTG RFI shape by depth) — the model's non-monotonic VPIP curve matches the published curve closely. SB short-stack direction (MTT-SOL-3, 6) and the 3-bet tightening direction (MTT-SOL-23) are partial matches. Nine solver anchors are discrepant, concentrated in PKO adjustments (MTT-SOL-30..33), mystery bounty (MTT-SOL-36, 37), and the ICM-onset timing (MTT-SOL-26, 27).

Tension 1 — Flat-KO tightens more than PKO (sign error)

Published theory (MTT-LIT-21, 32) predicts that both PKO and flat-KO widen VPIP vs no-bounty normal, with PKO widening more because progressive bounties compound over the tournament lifetime. The model inverts both the direction and the ordering. At BTN across four stack depths, PKO VPIP is -5.0pp to -1.8pp lower than normal; flat-KO is -9.7pp to -8.4pp lower. Both sign and ranking contradict the literature. The most plausible cause: the training rules encode only the instant portion of the PKO reward, not the lifetime carry. Without that carry, the model sees no compounding EV incentive for loose calling under PKO.

Tension 2 — Mystery equals normal (model failure)

Twenty out of 20 mid-stage cells are byte-identical between mystery bounty and normal. The training rules contain a pre-ITM guard that substitutes mystery → normal when the tournament is before the money. Since Round 1 mid-stage queries all triggered this guard, the model never saw a live mystery parameter. MTT-LIT-27 through 30 treat mystery as a distinct format with envelope EV — the rules implementation diverges from that framing. The model IS reading the mystery parameter post-ITM (36/40 post-ITM cells differ), but the pre-ITM gap is a coaching-relevant limitation: do not use the model for mystery bounty strategy in the pre-mystery phase.

Tension 3 — Binary stage training failure

MTT-LIT-35 predicts mid-tournament tightening, MTT-LIT-36 predicts late pre-bubble tightening, and MTT-SOL-26 locates the ICM significance threshold at 50–37.5% of the field remaining. The model collapses all six canonical stages into exactly two VPIP groups — {early, mid, late} vs {bubble, itm, ft} — with a within-group max delta of 0.7pp. The training rules expose only a continuous ratio to the model, not an explicit stage label. The binary cut is a known model limitation. Do not interpolate graduated stage transitions from this model — use external ICM solvers for that.