Death by a Thousand Cuts
The per-dimension compliance threshold from yesterday’s behavioral contracts work had a known blind spot: an agent with 2 violations in each of 10 dimensions — 20 total, all individually below the threshold of 3 — looked perfectly compliant. The design review caught the gap and filed it as #88.
The fix is an aggregate safety valve. Sum all VIOLATED counts across all compliance dimensions; if the total exceeds a second threshold (default 5), fire. The key insight: the aggregate only fires when per-dimension hasn’t — every dimension’s count is below the per-dimension threshold. So summing counts directly measures “total behavioral drift below the per-dimension radar.”
The design needed a discriminator on BehavioralViolation — ViolationKind { PER_DIMENSION, AGGREGATE } — because the violations map has different semantics per kind. PER_DIMENSION contains only the exceeding dimensions. AGGREGATE contains all dimensions with any violations. Without the discriminator, consumers can’t interpret the map correctly.
The co-tuning guidance between the two thresholds was the trickiest part. The design review caught two mathematical errors in my initial formula. The correct general formula for ensuring at least K dimensions must contribute: aggregate >= (K-1) * (per_dimension - 1) + 1. At the defaults (per-dimension=3, K=3), that gives aggregate >= 5.
Also closed #86 as an ADR: no compliance status in A2A_CARD. The card represents what an agent declares; compliance is runtime state about whether it lives up to those declarations. Different lifecycles, caching would break, and probe() already gives consumers the same information. The right architectural answer is that these are separate concerns accessed through separate APIs.
The rest was cleanup from the #85 design review: unified InMemoryAgentRegistry.matchesCapability() to delegate to CapabilityResolver.match() (#83), renamed stale test inner classes (#91), and updated ARC42STORIES.MD and PLATFORM.md for the behavioral contracts framework.