docs: explain how to aggregate inference results

README.md

@@ -362,6 +362,7 @@ Deeper references for specific topics live under [`docs/`](docs/):
 - [`test_cases_deep_dive.md`](docs/test_cases_deep_dive.md) — Deeper examples and patterns for test-case authoring.
 - [`history_and_metadata.md`](docs/history_and_metadata.md) — Multi-turn test cases with prior conversation history loaded from a companion `.meta.yaml`.
 - [`debugging_tests.md`](docs/debugging_tests.md) — How to debug a failing test (VSCode launch configs and friends).
+- [`analyzing_results.md`](docs/analyzing_results.md) — Use `tb agg` to compute pass@1, pass@20, and pass^20 from a `tb infer` result JSONL.
 
 **Fixtures and judges**
 - [`fixtures.md`](docs/fixtures.md) — How fixtures are wired up (dependency injection via `conftest.yaml` and scenario overrides).

docs/analyzing_results.md

@@ -0,0 +1,19 @@
+# Analyzing inference results
+
+## Result file
+
+Run `tb infer` as described in the [ThinkingBox README](../README.md) to
+produce a batch result JSONL file. Each line represents one task attempt. To
+report pass@20 and pass^20, run 20 attempts per task.
+
+## Compute the reported metrics
+
+```bash
+uv run tb agg output_thinkingbox_bench_v1.jsonl
+```
+
+For a JSONL containing 20 attempts for every task, the aggregate output
+includes the paper's pass@1, pass@20, and pass^20 values. pass@20 measures
+whether at least one of 20 attempts succeeds; pass^20 estimates whether all 20
+attempts succeed. These metrics are omitted if tasks have unequal attempt
+counts.