Context

The benchmark currently has 50 problems (10 per tier). Expanding to 75+ (15 per tier) would improve statistical power, reduce the impact of non-deterministic results on per-tier rates, and cover more Vera language features.

Guidelines for new problems

• Don't duplicate existing vera examples/ directory (in training data)
• Each problem should test something specific and be tagged accordingly
• Follow the existing problem structure (JSON + .vera + .py + .ts solutions)
• All canonical solutions must pass vera check + vera verify

Suggested additions per tier

Tier 1 (5 more): Pure arithmetic

• Integer overflow handling
• Multi-way branching (4+ conditions)
• Nested arithmetic with complex postconditions

Tier 2 (5 more): Builtin discovery

• string_split + string_join composition
• array_range + array_map pipeline
• json_parse + field extraction
• regex_match (if available in SKILL.md)
• Type conversion functions (int_to_string, float_to_int)

Tier 3 (5 more): ADTs + match

• Mutual ADTs (two types referencing each other)
• Generic ADTs (List<T> with forall)
• Three+ variant ADTs
• Nested pattern matching (match inside match)
• ADT with multiple type parameters

Tier 4 (5 more): Recursion + termination

• Lexicographic decreases (two measures)
• Accumulator pattern with non-trivial invariant
• Tree traversal with path tracking
• String recursion (if string indexing supports it)
• Mutual recursion with where blocks

Tier 5 (5 more): Effects

• Nested effect handlers
• Multiple effects in one function
• Effect polymorphism
• State + IO combination
• Error recovery with retry logic