EqualifyEverything / benchmarks-ai-alt

First project: build the corpus of functional images and gold standard alt text the benchmark scores against. Two agents in a loop, a seeking agent and an adversarial reviewer, with a stop condition that no agent can declare. - directives/00 states the spec: item definition, coverage, difficulty and diversity targets, collection constraints, acceptance criteria. - directives/01 and 02 are the seeking and adversarial review directives. - directives/03 and run.sh sequence the rounds and stop on two consecutive quiet review rounds. - tools/validate.mjs enforces the schema and computes the stop condition, with a fixture self-test. No dependencies. - run.sh --selftest exercises the whole loop against a stub agent, no API calls. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>