A difficulty curve is not monotonic in any single resource, so a test must assert the trend

Fact

Authored difficulty is tempting to guard with a one-line invariant: total enemy health, or total obstacle count, must rise with every level. It fails immediately on good design, because a well-built curve periodically trades one resource for a different threat — the first level that introduces a new mechanic usually costs less of the old resource, and reads as a dip.

A finale that spikes creates the mirror problem: whatever follows it, including the first level of an endless or new-game-plus mode, is lower than its predecessor by design.

Why it matters

The assertion fails on exactly the content it was written to protect: the interesting levels. The natural response is to inflate the dipping level back over its neighbour, which flattens the variety the dip existed to create.

How to apply

• Assert the trend over a window, not wave on wave: resource(n) > resource(n - k), with k at least as long as the design's own repeating cycle. If bosses land every fifth level, compare across five, so each level is compared with the same phase of the cycle.
• Assert separately that nothing collapses — resource(n) > resource(n - 1) * 0.5 — which is the real failure that the monotonic rule was reaching for.
• Assert the shape you actually intend: that the last authored level is the peak, and that a procedural tail eventually surpasses it.
• If a dip is deliberate, the invariant should say which resource is being traded for what, rather than being loosened until it passes.