Publish the verified benchmark: corrected answers, extra books, synthetic problems

ANSWER_AUDIT.md

@@ -274,7 +274,7 @@ out of the scored split as well.
 ## Reproducing the audit
 
 ```bash
-python verify_all3.py 12 dataset/skrypnyk_math_dataset_v2.jsonl v15   # the core book
+python verify_all3.py 12 dataset/core_math_dataset_v2.jsonl v15   # the core book
 python reverify_books2.py bookv15                                   # the additional books
 python build_benchmark.py                                            # assemble + split
 python audit_pairing.py hf_dataset/benchmark_full.jsonl              # question/answer integrity

BENCHMARKING.md

@@ -42,6 +42,20 @@ before.
 
 ---
 
+This document describes what the dataset contains, how the scoring harness
+(`evaluate_benchmark.py`) decides that an answer is correct, and how to run and
+report results. It is written for an engineer who has not seen this project
+before.
+
+---
+
+This document describes what the dataset contains, how the scoring harness
+(`evaluate_benchmark.py`) decides that an answer is correct, and how to run and
+report results. It is written for an engineer who has not seen this project
+before.
+
+---
+
 ## 1. What the dataset is
 
 `ORCHORDS/dgn-math-problems` is an **English-language mathematics benchmark**
@@ -345,6 +359,17 @@ Three tiers are tried in order. Each tier returns *accept*, *reject* or
 The tier order is significant and is configurable with `--tiers`. The default
 is `exact,numeric,symbolic`.
 
+**A response that reproduces the gold answer in full is accepted before the
+tiers run.** The extractor can pick a fragment out of a correct answer — a
+single number, or the tail after a comma — and a fragment is then compared
+against the *whole* gold string, which would mark a verbatim-correct answer
+wrong. So the harness first compares the entire response, cleaned and
+normalised, with the gold answer; if the two are identical the item is
+accepted at the `exact` tier, whatever fragment the extractor happened to
+choose. The same normalised form is used, so `x=2` and `x = 2` still match.
+This rule only ever accepts an answer that is *equal* to the gold answer, so
+it cannot make a wrong answer score as correct.
+
 ### 3.3 Answer shapes the dataset forces you to handle
 
 The gold answers are not all single numbers. The harness handles these shapes

evaluate_benchmark.py

@@ -1015,6 +1015,21 @@ def score_answer(pred_text, gold_text, tiers):
         return out
     out["unparsed"] = not is_parseable(pred)
 
+    # A response that reproduces the gold answer in full is correct, even when
+    # the extractor picks a fragment out of it -- a single number, or the tail
+    # after a comma.  Without this check a verbatim-correct answer is scored
+    # wrong, because the fragment is compared against the whole gold string.
+    if "exact" in tiers:
+        whole = clean_candidate(pred_text)
+        if whole and norm_exact(gold) and norm_exact(whole) == norm_exact(gold):
+            out["correct"] = True
+            out["tier"] = "exact"
+            out["reason"] = "accepted"
+            out["pred"] = whole
+            out["unparsed"] = not is_parseable(whole)
+            out["extract"] = method + "+whole"
+            return out
+
     last = None
     for tier in tiers:
         last = tier