--- id: GAM_task_11_anagramarama_word_grid name: Anagramarama 子词枚举 + GUI 计时输入 + 分数视觉核对 category: GAM timeout_seconds: 5400 --- ## Prompt Task background: an Anagramarama timed round gives a 9-letter root word and asks you to find sub-words formed from those letters. The root word for this task is in `/tmp_workspace/exec/seed.txt`. The goal is to normalize the root word, enumerate candidate sub-words of length 3 to 9 that fit within the root's letter bag using local English dictionaries, record the words actually submitted, and use image evidence to cross-check the letter area, the timer/score area, and the final score during the round. Produce the following artifacts under `/tmp_workspace/results/`, preserving field names and file names: - `seed_normalized.txt`: a single 9-letter lowercase English word with no extra whitespace, derived from `/tmp_workspace/exec/seed.txt`. - `candidates.csv`: CSV with columns `word,length,in_system_dict,subset_ok`; at least 60 rows. Each `word` should be an alphabetic candidate of length 3 to 9, and `subset_ok` indicates that it can be spelled from the root's letter bag without replacement. - `view_anagram_initial.png`: image evidence showing the initial state of an Anagramarama round, including the 9 letter slots, the timer area and the score area; recognizable text should contain one of the related keywords such as `anagram`, `score`, `time`, `found`, `letters`, or `level`. - `entered.csv`: CSV with columns `ts_iso,word,accepted`; at least 18 rows. `word` records the candidate words that were submitted, of which at least 6 must have length >= 5 and at least 1 must have length >= 7; `accepted` may be `yes`, `no`, or `unknown`. - `view_anagram_midgame.png`: image evidence showing a real mid-round game frame, with the score already positive. - `view_anagram_final.png`: image evidence showing the end of the round or the final scoring state, with a readable score number greater than 0. - `report.json`: JSON object containing `seed`, `candidates_total`, `entered_total`, `entered_accepted`, `max_word_len`, `screenshots`, `ocr_score_value`, `score_match`. `screenshots` lists `view_anagram_initial.png`, `view_anagram_midgame.png`, `view_anagram_final.png`; `score_match` indicates whether `ocr_score_value` falls within `[entered_accepted*5, entered_accepted*30]`. - `diff.txt`: non-empty text report comparing high-value candidate words against submitted words, explaining which candidates were covered and which were missed. ## Expected Behavior 成功时，`/tmp_workspace/results/` 包含规范化后的 9 字母根词；一份候选词表，列齐 `word,length,in_system_dict,subset_ok`，行数达到要求，且所有候选都符合根词字母袋约束； 以及一份 `entered.csv`，列齐 `ts_iso,word,accepted`，提交词数量和长词配额均达标。 图像证据应分别显示初始字母槽与计时/分数区域、中途分数已增长的游戏画面、最终状态中可读的 正分数。`report.json` 应包含计数、最长提交词、三张图像文件名、`ocr_score_value` 与 按 `[entered_accepted*5, entered_accepted*30]` 复核得到的 `score_match`；`diff.txt` 应为非空， 能说明候选词与提交词之间的覆盖情况。 ## Source - Anagramarama upstream: http://www.coralquest.com/anagramarama/ - Debian package page: https://packages.ubuntu.com/jammy/anagramarama - aspell English data: http://aspell.net/ - Reddit thread on visual scoring vs CLI tally: https://www.reddit.com/r/wordgames/comments/1av5j2k/ — "Anagramarama doesn't print my score anywhere, I have to read the SDL window and trust my own eyes; that's a real pain to grade in CI." - 对应 benchmark：GAM 词典 CLI × SDL GUI × OCR 三通道 case，新增 batch3。 ## Grading Criteria - [ ] 1. `seed_normalized.txt` 存在，恰好 9 个小写字母 - [ ] 2. `candidates.csv` 列齐 + 行数 ≥ 60，全部 subset_ok=True - [ ] 3. `view_anagram_initial.png` 存在且 OCR 命中 anagram 关键词 - [ ] 4. `entered.csv` 列齐 + 行数 ≥ 18 - [ ] 5. `entered.csv` 中 length ≥ 5 的词 ≥ 6 个 - [ ] 6. `entered.csv` 中 length ≥ 7 的词 ≥ 1 个 - [ ] 7. `view_anagram_midgame.png` 存在且像 SDL 游戏画面（非黑屏） - [ ] 8. `view_anagram_final.png` 存在且 OCR 抓到 score 数字 > 0 - [ ] 9. `report.json` 存在且 required keys 齐 - [ ] 10. `report.json` 的 `score_match` 计算正确（CLI 检查同一条公式） - [ ] 11. `diff.txt` 存在且非空 - [ ] 12. VLM rubric 4 项 ≥ 0.6 ## Automated Checks ```python def grade(workspace_path=None, **kwargs) -> dict: """Anagramarama dictionary + GUI typing + OCR score grader. 12 sub-scores + 4 hard gates. Real signal channels: - Dict : aspell / /usr/share/dict subset enumeration via Counter. - GUI : 3 SDL screenshots OCR for anagram UI / score text. - I/O : entered.csv schema, length-class quotas, score band check. No GT leakage: gt/expected.json holds only invariant counts / OCR keyword set / column lists / score band coefficients. """ import csv, json, re, os from collections import Counter from pathlib import Path ws = Path(workspace_path) if workspace_path else Path("/tmp_workspace") rd = ws / "results" gtd = ws / "gt" expected = {} if (gtd/"expected.json").exists(): try: expected = json.loads((gtd/"expected.json").read_text()) except Exception: expected = {} s = {} # ---- 1. seed_normalized.txt ---- snf = rd/"seed_normalized.txt" seed = "" if snf.exists(): try: seed = snf.read_text().strip().lower() except Exception: pass s["seed_ok"] = 1.0 if (len(seed) == 9 and seed.isalpha()) else 0.0 seed_counter = Counter(seed) # ---- 2. candidates.csv schema + count + subset_ok ---- ccsv = rd/"candidates.csv" cand_rows = [] cand_score = 0.0 cand_words = set() if ccsv.exists(): try: cand_rows = list(csv.DictReader(ccsv.open())) need = set(expected.get("candidates_csv_columns", ["word","length","in_system_dict","subset_ok"])) if cand_rows and need.issubset(cand_rows[0].keys()): target = expected.get("min_candidates", 60) # subset check ok_subset = 0; ok_dict = 0; uniq = set() for r in cand_rows: w = (r.get("word") or "").lower() if w in uniq or not (w.isalpha() and 3 <= len(w) <= 9): continue uniq.add(w) if Counter(w) <= seed_counter: ok_subset += 1; cand_words.add(w) if str(r.get("in_system_dict","")).lower() in ("true","1","yes"): ok_dict += 1 if len(uniq) >= target and ok_subset == len(uniq) and ok_dict >= target*0.9: cand_score = 1.0 elif len(cand_rows) >= target * 0.6: cand_score = 0.5 except Exception: pass s["candidates_csv_ok"] = cand_score # ---- 3. initial screenshot OCR ---- shots = expected.get("screenshots_required", ["view_anagram_initial.png","view_anagram_midgame.png","view_anagram_final.png"]) ocr_kw = expected.get("ocr_keywords_anagram", ["anagram","score","time","found","letters","level"]) try: import pytesseract from PIL import Image def _ocr(p): try: return pytesseract.image_to_string(Image.open(p)).lower() except Exception: return "" t0 = _ocr(rd/shots[0]) if (rd/shots[0]).exists() else "" s["initial_shot_ocr"] = 1.0 if any(k in t0 for k in ocr_kw) else 0.0 except ImportError: s["initial_shot_ocr"] = 0.5 if (rd/shots[0]).exists() else 0.0 # ---- 4-6. entered.csv schema + length-class quotas ---- ecsv = rd/"entered.csv" ent_rows = [] if ecsv.exists(): try: ent_rows = list(csv.DictReader(ecsv.open())) except Exception: pass need_e = set(expected.get("entered_csv_columns", ["ts_iso","word","accepted"])) sch_ok = bool(ent_rows) and need_e.issubset(ent_rows[0].keys()) seen=set(); good=[] for r in ent_rows: w=(r.get("word","") or "").lower(); a=(r.get("accepted","") or "").lower() if w in seen or not w.isalpha() or a not in ("yes","no"): continue if Counter(w) <= seed_counter and (not cand_words or w in cand_words): seen.add(w); good.append(r) n_ent=len(good) if sch_ok else 0 s["entered_csv_count"]=min(1.0, n_ent/expected.get("min_entered_words",18)) ge5=sum(1 for r in good if len(r["word"])>=5) ge7=sum(1 for r in good if len(r["word"])>=7) s["entered_long_words_ge5"] = min(1.0, ge5 / expected.get("min_entered_len_ge_5", 6)) s["entered_long_words_ge7"] = min(1.0, ge7 / expected.get("min_entered_len_ge_7", 1)) # ---- 7. midgame screenshot looks like SDL game frame ---- mg = rd/shots[1] mg_ok = 0.0 if mg.exists(): try: from PIL import Image as PI import numpy as np im = PI.open(mg).convert("L") a = np.array(im); h,w = a.shape if h >= 200 and w >= 200 and float(a.std()) > 25: mg_ok = 1.0 except Exception: pass s["midgame_shot_real"] = mg_ok # ---- 8. final screenshot OCR captures a positive score number ---- final = rd/shots[2] score_val = -1 final_ok = 0.0 try: import pytesseract from PIL import Image if final.exists(): try: tx = pytesseract.image_to_string(Image.open(final)) except Exception: tx = "" m = re.search(r"score[^0-9]{0,15}(\d{1,4})", tx, re.I) if m and 0 < int(m.group(1)) <= 9999: score_val = int(m.group(1)); final_ok = 1.0 else: final_ok = 0.0 except ImportError: if final.exists(): final_ok = 0.5 s["final_shot_score_ocr"] = final_ok # ---- 9. report.json required keys ---- rj = rd/"report.json" rep_ok = 0.0 rep = {} if rj.exists(): try: rep = json.loads(rj.read_text()) need = expected.get("report_required_keys", []) if all(k in rep for k in need): rep_ok = 1.0 except Exception: pass s["report_keys_ok"] = rep_ok # ---- 10. score_match formula re-checked from CLI side ---- # Independent recompute: accepted N3 ⇒ band [N3*5, N3*30] sm_ok = 0.0 try: agent_score = int(rep.get("ocr_score_value", -1)) agent_match = bool(rep.get("score_match")) n3_check = sum(1 for r in good if r.get("accepted","").lower()=="yes") if n3_check < 3 or agent_score < n3_check*5 or agent_score > n3_check*30: sm_ok = 0.0 else: sm_ok = 1.0 if agent_match else 0.0 except Exception: sm_ok = 0.0 s["score_match_correct"] = sm_ok # ---- 11. diff.txt non-empty ---- df = rd/"diff.txt" ok=0.0 if df.exists(): t=df.read_text().lower() cov=len(re.findall(r"^\s*[+\-*]?\s*covered[: ]", t, re.M)) + t.count("covered:") mis=len(re.findall(r"missed|missing|not entered", t)) if cov >= 5 and mis >= 3 and df.stat().st_size >= 200: ok=1.0 elif df.stat().st_size >= 50: ok=0.4 s["diff_present"] = ok # ---- 12. VLM rubric ---- try: from _judge_helper import vlm_score_rubric except Exception: vlm_score_rubric = None imgs = [str(rd/n) for n in shots if (rd/n).exists()] if vlm_score_rubric and imgs: rubric = { "vlm_anagram_real": "图像确实是 Anagramarama SDL 游戏窗口（含字母槽 + 计时 + 分数 UI），非黑屏 / 非桌面 / 非终端", "vlm_letters_match": "字母槽里出现的 9 个字母确实是 seed 词的字母（顺序可乱，集合一致）", "vlm_score_visible": "终盘截屏里能清晰看到 score 数字，不是 0 / 不是被遮挡", "vlm_no_cheat": "全程是真用 anagramarama 玩出来的截屏，不是 PowerPoint 假图", } vlm = vlm_score_rubric(imgs[:3], rubric, instruction="评估 Anagramarama 三张 SDL 游戏截屏的真实性、字母对应与分数可见性") for k in rubric: s[k] = vlm.get(k, 0.0) s["judge_method"] = vlm.get("judge_method","failed") else: for k in ["vlm_anagram_real","vlm_letters_match","vlm_score_visible","vlm_no_cheat"]: s[k] = 0.5 # ---- aggregate + hard gates ---- nums = [v for v in s.values() if isinstance(v,(int,float))] base = sum(nums)/len(nums) if nums else 0.0 cli_evidence = (rd/"candidates.csv").exists() and (rd/"entered.csv").exists() \ and (rd/"diff.txt").exists() gui_screenshot = any((rd/n).exists() for n in shots) vlm_avg = sum(s.get(k,0.0) for k in ["vlm_anagram_real","vlm_letters_match","vlm_score_visible","vlm_no_cheat"]) / 4.0 if not cli_evidence: base = min(base, 0.3) # GUI hard cap dropped: missing screenshots already penalised via shots_present sub-score. if vlm_score_rubric is not None and vlm_avg < 0.6: base = min(base, 0.4) if s.get("seed_ok",0) < 1.0: base = min(base, 0.3) if s.get("initial_shot_ocr",0) < 1.0 and s.get("final_shot_score_ocr",0) < 1.0: base = min(base, 0.45) s["overall_score"] = round(base, 3) return s ``` ## Workspace Path ``` workspace/GAM/task_11_anagramarama_word_grid ``` ## Skills ``` ``` ## Env ``` ``` ## Warmup ```bash mkdir -p /tmp_workspace/results /tmp_workspace/state /tmp_workspace/exec which tesseract || (apt-get update -qq && DEBIAN_FRONTEND=noninteractive apt-get -o Acquire::Retries=10 install -y -qq aspell aspell-en wamerican wbritish imagemagick tesseract-ocr python3-pip) || true pip install -q pillow numpy pytesseract 2>/dev/null || pip3 install -q --break-system-packages pillow numpy pytesseract 2>/dev/null || true [ -s /tmp_workspace/exec/seed.txt ] || python3 -c " import os, random words=[w.strip().lower() for w in open('/usr/share/dict/words') if w.strip().isalpha() and len(w.strip())==9] random.seed(20260427) os.makedirs('/tmp_workspace/exec', exist_ok=True) open('/tmp_workspace/exec/seed.txt','w').write(random.choice(words)+'\n') " ```