A game built from a closed vocabulary can be verified; one built from per-unit free text can only be playtested. The escape from the first to the second is measurable — and it is where balance stops being checkable.
Links: HeroClix (the dated specimen — 0 → 434 characters of bespoke rules text per figure, measured), BattleTech Simulator (the sibling worry: rules drifting toward OP and absurd), BattleValue, D&D Monster Tournament — Exact Markov Chains, Gaming
Combinatorial. Units are combinations drawn from a fixed vocabulary of abilities. The space is finite and enumerable, so interactions can be reasoned about exhaustively — with 28 abilities there are only 28 things that can interact, and a designer can hold the whole lattice in view.
Generative. Each unit carries arbitrary rules text unique to itself. The space is unbounded. No enumeration exists, so balance stops being a property you can verify and becomes an empirical question settled by play.
Every long-running unit-collection game faces the same pressure: a combinatorial space exhausts. After enough releases the interesting combinations are used up. There are only two exits — grow the vocabulary, or let each unit write its own rules — and the second is far cheaper.
| 2002 (Infinity Challenge, 176 figures) | 2024 (43 figures) | |
|---|---|---|
| standard vocabulary | 28 abilities, exactly 7 per stat slot | ~50 (≈1.8×) |
| ability slots that are free text | 0.0% | 19.8% |
| units carrying ≥1 free-text ability | 0% | 86% |
| bespoke rules text per unit | 0 characters | 434 characters |
Zero across 176 units, twenty-two years later 434 characters on 86% of the roster. The vocabulary did grow, but that is a rounding error beside the escape hatch. The exit is not vocabulary growth; it is the shift of rules authorship from the system to the unit.
A second measurement separates two things that are easy to conflate. Fitting unit cost on stat lines alone, 2002 vs 2024: R² falls 0.889 → 0.682, but mean absolute error is 10.1 points in both eras and residual SD is 13.4 vs 13.3. The pricing function did not get worse; modern units just cluster in a narrower cost band (SD 40 → 24), so the same error explains less variance. R² measures explained variance, not accuracy — here they come apart cleanly.
Yet a cost-matched cross-era test shows the baseline moved: a 67-point 2002 unit beats cost-matched contemporaries 85% of the time and cost-matched 2024 units 37.5% — and that test is one-sided in the old unit’s favour, since ~20% of the modern units’ abilities are free text the model cannot read.
So the unit cost is an intra-era currency, not an inter-era one. Power creep in its cleanest form is not sloppy pricing — it is a stable pricing function applied to a rising baseline. Which means internal balance metrics cannot detect it: every set looks well costed against itself. Detecting creep requires cross-era play, which is exactly what a collectible game rarely tests.
The transition is not a single step. HeroClix went through a middle stage worth naming, because each stage needs a different analytic tool:
| Era | Rules regime | Right instrument |
|---|---|---|
| Early | simple rules; each universe has its own ability set | exact enumeration, per universe — small closed vocabulary |
| Middle | abilities unified into one chart and expanded; cross-universe play legal | same tool, wider vocabulary. Best ratio of tractability to design richness |
| Modern | abilities become per-unit free text | needs a parser before a solver is applicable at all |
The early stage is verifiable in the source documents: a 2002 colour-coded Marvel/DC rulebook diff shows Impervious as a DC-only ability, and the Marvel set’s entire defense vocabulary lacks it. Two universes, two vocabularies — so the “closed vocabulary” of the early era is closed per universe, which makes within-universe analysis easy and cross-universe analysis meaningless.
The middle era is the sweet spot for any analyst, and it is worth recognising while a system is in it: one vocabulary, still enumerable, maximum expressive range. That window closes without announcement.
Verifiability is the thing traded away. In the combinatorial regime, “is this balanced?” is answerable in principle by enumeration. In the generative regime it is answerable only by sampling actual play — so mispricing is no longer prevented, only detected, and only after release. Power creep is not inevitable under that regime, but it becomes very hard to detect early, which is the same failure mode by a slower route.
It degrades the analyst, not just the designer. Any model that reads the vocabulary — a simulator, a solver, a ranking — sees 100% of a combinatorial set and progressively less of a generative one. The HeroClix Markov model reads 100% of the 2002 vocabulary and ~80% of a 2024 set’s ability slots. A method’s coverage can decay with the age of its target even though the method never changed. Any long-horizon analysis should therefore report coverage per release, not once.
The escape is a ratchet. Once units carry bespoke text, returning to a closed vocabulary would invalidate the existing catalogue. So the transition is effectively irreversible, which means the decision to take it is far more consequential than it looks at the time — it is a one-way door disguised as a content decision.