v1.1: correct 16 answer keys and document the judge system prompt

README.md

@@ -11,6 +11,24 @@ configs:
 
 AA-LCR includes 100 hard text-based questions that require reasoning across multiple real-world documents, with each document set averaging ~100k input tokens. Questions are designed such that answers cannot be directly retrieved from documents and must instead be reasoned from multiple information sources.
 
+## New in Version 1.1 (September 2026)
+
+- **Sixteen corrected answer keys.** Each one was re-verified against the question's own source documents. Most were values stored as a decimal fraction where the question asked for a percentage (question 28, `0.658`, is now `65.8%`), or a date stored as a spreadsheet serial number (question 40, `45444`, is now `June 2024`). Four were corrections to the answer itself rather than its notation: questions 10, 25, 30 and 67.
+- **A judge system prompt.** The equality checker previously ran with no system prompt, which left it to improvise on unit conversion, notation and multi-value answers one call at a time. See Scoring Approach below.
+
+The questions and the document sets are unchanged. Scores measured on version 1.0.0 are not directly comparable with scores measured on version 1.1.
+
+Version 1.0.0 remains available at revision [`bdae010`](https://huggingface.co/datasets/ArtificialAnalysis/AA-LCR/tree/bdae010bbce259820c0e34c1d7cce210d966fb75), so earlier results stay reproducible:
+
+```python
+csv_path = hf_hub_download(
+    repo_id="ArtificialAnalysis/AA-LCR",
+    filename="AA-LCR_Dataset.csv",
+    repo_type="dataset",
+    revision="bdae010bbce259820c0e34c1d7cce210d966fb75",
+)
+```
+
 ## Dataset Development
 
 AA-LCR was created through a rigorous multi-phase process involving several members of the Artificial Analysis research team and more than a dozen undergraduate students who were engaged on a short-term contract basis to write and/or validate questions.
@@ -200,21 +218,85 @@ def get_document_set(
 
 ## Scoring Approach
 
-We use an LLM-based equality checker to evaluate responses:
+We use an LLM-based equality checker to evaluate responses. The checker receives a system prompt and a user prompt.
+
+System prompt:
+
+```
+Decide whether the CANDIDATE ANSWER is correct or incorrect against the OFFICIAL ANSWER.
+Note the following points when assessing correctness:
+
+- Numbers should still match when they are the same value written differently, e.g., a
+  percentage, a count of percentage points, and the equivalent decimal fraction are the same
+  value: 0.675, "67.5%" and "67.5 percentage points" all match. So do different scales
+  (thousand, million, bn) and different notations (thousands separators, currency symbols,
+  LaTeX markup, and numbers written as words).
+- Where the question asks for a particular format (e.g., a percentage, a number of decimal
+  places, a unit, a rounding, or an ordering) the CANDIDATE ANSWER must meet it. If the
+  question asks for no particular format, accept any equivalent form.
+- In cases where the question asks for an ordered list, a title, honorific or article added
+  to an entry in the CANDIDATE ANSWER can change where that entry sorts. Accept the ordering
+  if it is correct either with those additions or without them.
+- Grade the value the CANDIDATE ANSWER finally commits to, and it must commit to one. Values
+  reached while working, and alternatives it considers and sets aside, do not count. If it
+  offers several values without selecting one, it is incorrect even if one of them is right.
+  Hedging is fine as long as one clearly definitive answer is given.
+```
+
+User prompt:
 
 ```
 Assess whether the following CANDIDATE ANSWER is CORRECT or INCORRECT.
 For the CANDIDATE ANSWER to be correct, it must be consistent with the OFFICIAL ANSWER.
 
-The question, for reference only: {question}
+The question, for reference only: START QUESTION {question}
+
+END QUESTION
+
 The OFFICIAL ANSWER: {official_answer}
-CANDIDATE ANSWER TO ASSESS: {candidate_answer}
 
-Reply only with CORRECT or INCORRECT.
+END OFFICIAL ANSWER
+
+BEGIN CANDIDATE ANSWER TO ASSESS
+
+{candidate_answer}
 
+END CANDIDATE ANSWER TO ASSESS
+
+Reply as JSON, with a verdict of CORRECT or INCORRECT.
+```
+
+GPT-5.6 Luna (medium) is used as the equality checker model.
+
+The prompts above are those used from version 1.1 onwards.
+
+<details>
+<summary>Version 1.0.0 scoring prompt</summary>
+
+Version 1.0.0 runs without a system prompt, and with the following user prompt:
+
+```
+Assess whether the following CANDIDATE ANSWER is CORRECT or INCORRECT.
+For the CANDIDATE ANSWER to be correct, it must be consistent with the OFFICIAL ANSWER.
+
+The question, for reference only: START QUESTION {question}
+
+END QUESTION
+
+The OFFICIAL ANSWER: {official_answer}
+
+END OFFICIAL ANSWER
+
+BEGIN CANDIDATE ANSWER TO ASSESS
+
+{candidate_answer}
+
+END CANDIDATE ANSWER TO ASSESS
+
+Reply only with CORRECT or INCORRECT.
 ```
 
-Qwen3 235B A22B 2507 Non-reasoning is used as the equality checker model.
+</details>
 
 ## Access and Citation