docs: publish audited baseline prompt protocol

baseline_prompts.md

@@ -0,0 +1,107 @@
+# Audited baseline prompts and run protocol
+
+Source: `general_agent_judge/graders/baselines.py`
+
+Source SHA-256: `385e33c73a4c371876945380ce8bb1871cbd01ab1b584f2a41e14aed601fc9af`
+
+## Scope
+
+The prompts below are extracted from the executable grader classes. The
+configuration in this document defines the **new audited rerun protocol**. It
+does not retroactively establish the model identifiers or sampling parameters
+of the historical reported runs. In particular, the archived code does not
+preserve a complete historical 8-pass harness, so no same-backbone or
+prompt-invariant interpretation is claimed for the previously reported margin.
+
+## LLM-as-Judge
+
+- Model ID for audited reruns: `qwen3.7-plus`
+- Temperature: `0.1`
+- Maximum output tokens: `256`
+- Passes: `1`
+
+System prompt:
+
+```text
+You are an evaluator. Given a task description and an agent's response, determine whether the agent successfully completed the task.
+
+Output ONLY a JSON object: {"verdict": "pass" or "fail", "confidence": 0.0-1.0, "reason": "brief explanation"}
+```
+
+User prompt:
+
+```text
+## Task
+{question}
+
+## Reference Answer
+{reference_answer}
+
+## Agent Response
+{agent_response}
+
+## Agent Status: success
+
+Did the agent successfully complete the task? Respond with JSON: {"verdict": "pass"/"fail", "confidence": 0.0-1.0, "reason": "..."}
+```
+
+## Agent-as-Judge
+
+- Model ID for audited reruns: `qwen3.7-plus`
+- Temperature: `0.1`
+- Maximum output tokens: `512`
+- Passes: `1`
+
+System prompt:
+
+```text
+You are an expert evaluator assessing whether an AI agent successfully completed a given task. You will be provided with:
+1. The task description
+2. Reference answer / expected outcome
+3. Checkpoints that should be satisfied
+4. The agent's response and execution details
+
+Evaluate thoroughly and output ONLY a JSON object:
+{"verdict": "pass" or "fail", "score": 0.0-1.0, "confidence": 0.0-1.0, "checkpoints_met": [true/false for each], "reason": "detailed explanation"}
+```
+
+User prompt:
+
+```text
+## Task
+{question}
+
+## Reference Answer
+{reference_answer}
+
+## Checkpoints (2 items)
+  - {checkpoint_1}
+  - {checkpoint_2}
+## Agent Response
+{agent_response}
+
+## Execution Details
+- Status: success
+- Steps: 12
+- Tokens: 3456
+- Wall time: 7.8s
+
+## Tool Calls (1 total)
+  ✓ {tool_name}(123ms)
+
+Evaluate whether the agent successfully completed the task. Respond with JSON.
+```
+
+## 8-pass ensemble
+
+The ensemble repeats the LLM-as-Judge prompt exactly eight times.
+
+- Model ID for audited reruns: `qwen3.7-plus`
+- Temperature: `0.7`
+- Maximum output tokens per pass: `256`
+- Passes: `8`
+- Aggregation: strict majority vote; a `4–4` tie is failure
+
+This setting is implemented by `EightPassEnsembleBaselineGrader`. It is a
+prospective, reproducible configuration and does not retroactively establish
+the missing temperature or harness metadata for the historical reported runs.