Of all the questions I get asked as a player researcher, the one that comes up most is also the simplest: why is level 47 so much harder than level 46? The honest answer is almost always "because we measured that you needed a wall there." Game difficulty is not an aesthetic judgment. It is a series of empirically tuned decisions designed to keep players inside a narrow channel where the game is neither boring nor punishing. Every level number on a difficulty curve is the visible artifact of dozens of playtests, A/B runs, and telemetry signals you never see.
This article is the long version of that answer. I want to walk through the actual science behind how we balance puzzle levels at Suntongames — the theory we lean on, the player skill data we actually look at, the A/B method we use, the fraught tradeoffs of dynamic difficulty adjustment, and the way we use all of it to design a 120-level progression that does not collapse in the back half.
The Flow Channel: The Theory Behind Every Curve
Almost everything we do in difficulty design rests on a single piece of cognitive psychology: Mihaly Csikszentmihalyi's concept of flow. Flow is the state of optimal experience where challenge and skill are matched closely enough that the person is fully absorbed but not overwhelmed. The useful part of the theory for game designers is not the romantic description of flow itself — it is the diagnostic. Csikszentmihalyi identified two failure modes around flow, and they are the only two failure modes that matter in puzzle design.
When challenge is too high for the player's current skill, the result is anxiety. The player feels frustrated, punished, and — most importantly for us — likely to leave. When challenge is too low for the player's current skill, the result is boredom. The player feels patronized, the game feels pointless, and churn follows just as reliably. The narrow corridor between anxiety and boredom is what is commonly called the flow channel. The entire job of a difficulty curve is to keep the player inside that channel across hundreds of levels.
The phenomenological experience of flow is what we are designing for, but the diagnostic value of the theory is what we are designing against. We do not try to make players feel flow. We try to keep them out of the two states that destroy it.— Adapted from Mihaly Csikszentmihalyi, "Flow and the psychology of discovery and invention"
The theory is older than the casual puzzle genre itself, but it has held up remarkably well as empirical scaffolding. Recent game-studies work has repeatedly validated that self-reported flow in puzzle players correlates with retention and with measured session length, as documented in Computers in Human Behavior and surveyed in ACM's journal of games research. We treat the flow channel as a hypothesis we design around, then test it against telemetry every release.
The Player Skill Distribution — And Why a Single Curve Will Not Fit It
The first practical problem with flow theory is that "skill" is not a single number. In any puzzle game we have shipped, the player skill distribution is wide, multimodal, and shifts over the player's lifecycle. A level that sits perfectly in the flow channel for the median player is boring for the top quartile and anxiety-inducing for the bottom quartile. Designing a single difficulty curve that serves all three is impossible. The question is not whether to compromise — it is which compromise hurts the fewest players.
Our approach is a layered compromise. The base difficulty curve is tuned to the median player — that is the curve the level numbers represent. On top of it, we layer two soft adaptations: a low-skill safety net that offers a small assistive bonus (an extra move, a free undo) after two consecutive failed attempts, and a high-skill fast-track that unlocks optional hard-mode variants for players clearing levels on the first try. The base curve never changes for any player. The soft adaptations do. This is not DDA, which we will get to in a moment — it is a static curve with two visible, opt-in safety rails.
How We Actually Measure Difficulty: A/B Testing Levels
Tuning a level without measuring is just opinionating. Our standard practice is to ship every new level as an A/B test against a baseline variant for the first 5,000 unique attempts, with a defined success metric before launch. The metric is almost always one of three things: first-try win rate, mean attempts to clear, and 24-hour return rate for players who failed the level. The target bands are also defined up front — typically a 45–60% first-try win rate for median-difficulty levels, dropping to 25–35% for "wall" levels designed as checkpoint challenges.
The A/B method matters because it kills the most common tuning mistake: mistaking "hard for me" or "hard for my favorite playtester" for "hard for the cohort." Internal designers are almost always in the top decile of skill on their own game, which means their gut instinct on difficulty is systematically wrong. The A/B test is the empirical correction. If a level I personally cleared in three attempts has a 14% first-try win rate in production, the level is too hard, regardless of how it felt to me. This is a hard discipline to maintain — designers, including me, are emotionally attached to levels we authored — but it is the difference between a curve that flows and one that gates players out.
If your difficulty curve is the designer's opinion, you are designing for an audience of one. The A/B test is not bureaucracy. It is the only honest signal.
We also use a specific methodological guard: we never tune a level live based on its own telemetry alone, because first-try and second-try players self-select. A level that looks "too easy" by week three is being played mostly by players who survived the wall before it. The cohort changed, not the level. Our A/B framework holds the cohort constant by sampling random incoming players into variants, which is the only way the comparison is meaningful. The mechanics of running this are well-trodden in the games industry — GDC talks on A/B testing in casual games cover the operational details at length.
Dynamic Difficulty Adjustment: The Promise and the Trap
Dynamic Difficulty Adjustment (DDA) — the family of systems that adapt a level's parameters in real time based on the player's observed performance — is one of the most seductive ideas in modern game design, and also one of the most misused. The pitch is obvious: instead of one curve, give each player their own. The flow channel is, after all, defined per player. Why not tune per player too?
The problem is that DDA in casual puzzles carries three serious risks that the marketing pitch usually glosses over. First, it can feel like cheating. When players notice the game quietly giving them easier boards after a streak of failures, their perception of mastery collapses — the achievement feels unearned. Second, it can mask real difficulty spikes. If a level is badly tuned, DDA will quietly absorb the failure rather than flagging it for the designer. The level never gets fixed; the system just hides it. Third, it is almost impossible to A/B test fairly. Because the system adapts to the cohort, the cohort you are measuring is the one the system already reacted to.
The conclusion we drew from that study is the conclusion we still hold: DDA is useful as an invisible safety net for the bottom decile of players who would otherwise hard-churn, but it should never be the primary balancing mechanism. We use it sparingly — a single, conservative assist after a third consecutive failure, and only on levels flagged as wall checkpoints. The bulk of the difficulty work happens in the static curve, where it can be measured, A/B tested, and tuned deliberately.
Designing a 120-Level Progression Curve
The visible product of all of the above is a 120-level progression curve. The shape of that curve is not arbitrary, and it is not linear. It is a structured rhythm of difficulty steps, plateaus, and wall checkpoints, and the rhythm is the part that matters most. A linear ramp that simply gets harder every level is the worst curve we have ever tested — players churn steadily from level 30 onward, with no recovery. The curves that work look more like a heartbeat.
The 4-Level Rhythm
Our standard unit of progression is a four-level block. Within the block, levels follow a consistent pattern: an introduction level that introduces one new constraint at a low difficulty, a reinforcement level that repeats the constraint with slightly tighter parameters, a challenge level that combines the new constraint with a previous one, and a wall level that pushes the player close to their failure threshold before the next block resets the rhythm. The wall is followed by an easier opening to the next block, which gives the player a breather and a sense of progress even as the overall curve trends upward.
The Wall Checkpoints
Every 20 levels, we place a deliberate wall — a level tuned to roughly a 25% first-try win rate, designed to make most players retry three to five times before clearing. Walls look like accidents to a player who has just breezed through the previous ten levels. They are not. They are the curve's most important retention feature. Players who clear a wall report higher satisfaction than players who never encounter one, because the wall is what makes the previous nineteen levels feel earned. Without the wall, the game becomes a treadmill. With the wall, it becomes a journey. The psychology here — that overcoming artificial resistance increases valuation of the outcome — is well established in behavioral research on the effort heuristic.
The Endgame Compression
From level 90 onward, the curve changes shape. New mechanics are no longer introduced — players have seen the full toolkit by then. Instead, difficulty rises through compression: tighter move budgets, more constrained boards, less forgiving cascades. The result is a curve that feels hard in a different way. Early-game difficulty feels like learning. Late-game difficulty feels like mastery tested under pressure. The distinction matters because it is what lets a 120-level campaign feel coherent rather than padded.
What the Science Tells Us — And What It Does Not
The honest summary is this: the science of difficulty balancing is robust enough to tell us what to optimize for (flow, retention, clear rate), but it is not robust enough to tell us exactly what to ship. Every level curve is a craft decision that uses the science as scaffolding. We use Csikszentmihalyi's flow channel as a diagnostic. We use cohort telemetry to correct our designer-intuition bias. We use A/B tests to convert opinion into measurement. We use DDA sparingly because the long-term data is not kind to it. And we use the 4-level rhythm with deliberate wall checkpoints because that is the structure that, in our cohort, produces the longest sustained engagement.
If you take only one idea from this article, take this: difficulty is not a feeling. It is a measured curve, and the only honest way to design one is to measure it. Every level you ship is a hypothesis. The players who play it are the experiment. Everything else is ego.