Part of #20 (slice I). The gap, not the matrix: an 'Experimental protocol' section in docs/fase5-metrica.md (models >=2, N runs per scenario with a power justification, ablation axes as existing levers: INSTRUCTIONS_PROFILES, --sampling, tool surface) plus a runbook in bench/README.md with the exact commands the matrix will execute (a320-bench run ... && a320-bench score runs/ --json). Zero new code; nothing in CI touches the network. Unblocks when an Anthropic API key exists (Vertex express already validated).
Part of #20 (slice I). The gap, not the matrix: an 'Experimental protocol' section in
docs/fase5-metrica.md(models >=2, N runs per scenario with a power justification, ablation axes as existing levers:INSTRUCTIONS_PROFILES,--sampling, tool surface) plus a runbook inbench/README.mdwith the exact commands the matrix will execute (a320-bench run ... && a320-bench score runs/ --json). Zero new code; nothing in CI touches the network. Unblocks when an Anthropic API key exists (Vertex express already validated).