| # GCA Agentic Trace Benchmark (91 cases) | |
| This directory contains 91-case benchmark aligned to the benchmark design described in the GCA paper. | |
| Each benchmark instance is written in an **agentic trace** format with: | |
| - `case_id` | |
| - `query` | |
| - `tool_sequence` | |
| - `gold_trace` (ordered tools with arguments and normalized outputs) | |
| - `final_answer` | |
| - `meta` | |
| Files: | |
| - `gca_benchmark_91_full_traces.jsonl`: full trace benchmark with per-step outputs and final answer | |
| - `gca_benchmark_91_runner_compatible.jsonl`: minimal version compatible with the earlier `case_id/query/gold_trace` loader | |
| - `manifest.json`: schema, topic counts, and tool catalog | |
| Important note: | |
| These are **benchmark stubs** for evaluating agent planning and answer composition. The per-step outputs are synthetic, | |
| normalized placeholders meant to support regression-style testing and are not live environmental measurements. | |