Update README.md

README.md

@@ -46,8 +46,9 @@ Each task's `rubric` field is a JSON array:
 ]
 ```
 
-- `positive hli verifier` — a factual claim the answer must contain. If the answer includes this claim, the rubric item is a PASS.
-- `negative hli verifier` — something the answer must *not* claim. If the answer includes this claim, the rubric item is a FAIL.
+- `positive hli verifier` — a factual claim the answer must contain. If the claim is met my the agent's answer, the rubric item result is a PASS.
+- `negative hli verifier` — something the answer must *not* claim. If the claim is met my the agent's answer, the rubric item result is a FAIL.
+
 
 ## Environments
 
@@ -57,4 +58,17 @@ Each task includes a `docker_image` field pointing to a pre-built Docker Hub ima
 docker pull andrewparkscaleai/coding-agent:kovidgoyal__kitty__815df1e210e0a9ab4622f5c7f2d6891d7dbeddf1
 docker run -it <image> bash
 # repo is at /app
-```
+```
+
+## Inference and Eval
+
+We follow the standard SWE-Agent scaffold, and we provide a sample config (with the prompts) in  [default_code_qa.yaml](default_code_qa.yaml)
+
+
+Evaluation is performed by an LLM judge (Claude Opus 4.5) that scores the agent's answer against each rubric criterion independently. Each criterion receives a binary score (met or not met) indicating and is then aggregated.
+
+The primary metric is the Task Resolve Rate: the percentage of tasks for which the agent's answer is comprehensive (i.e. passes all rubric items and scores 1.0), as graded by a set of task-specific rubrics. 
+
+The agents are also instructed to avoid modifying source-code files, and clean up any temporary scripts made. So we add a programmatic check that fails a task that has any code changes after submission. 
+
+Our rubric evaluation prompt and other relevant details are in [rubric_evaluation_config.yaml](rubric_evaluation_config.yaml)