Datasets:
Languages:
English
Size:
n<1K
ArXiv:
Tags:
physical-reasoning
benchmark
vision-language-models
physics-simulation
minimal-pairs
evaluation
License:
Harness update: Antigravity CLI backend (tools denied, web search forbidden, transcript audit), Claude run tags, timeout fix; card documents the Antigravity safeguards (part 2)
1bdbac8 verified Download scripts/agy_tool_audit.py from emilioferrara/DominoBench: direct link, hf CLI and curl.
- Browser
- Download file 3.19 kB
-
https://huggingface.co/datasets/emilioferrara/DominoBench/resolve/main/scripts/agy_tool_audit.py
- Command line
-
hf download hf://datasets/emilioferrara/DominoBench/scripts/agy_tool_audit.py
-
curl -L -o agy_tool_audit.py https://huggingface.co/datasets/emilioferrara/DominoBench/resolve/main/scripts/agy_tool_audit.py
3.19 kB
| """Audit Antigravity evaluation runs for tool use (2026-11-14 consistency check). | |
| Each evaluation conversation leaves a transcript under eval/agy_home/.gemini/antigravity-cli/brain/<conversation_id>/ | |
| .system_generated/logs/transcript_full.jsonl, and each answer file records meta.conversation_id. For every answer file | |
| of every antigravity-cli model, this finds its transcript (if kept), lists the tool calls the model made and whether | |
| each one executed or was refused, and reports answers whose conversation executed any tool other than reading the | |
| picture. Writes the list of affected answer files to results/agy_tool_audit.json.""" | |
| import collections | |
| import glob | |
| import json | |
| import os | |
| ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__))) | |
| B = os.path.join(ROOT, "eval/agy_home/.gemini/antigravity-cli/brain") | |
| ALLOWED = {"read_file", "view_file"} | |
| def conversation_tools(cid): | |
| """[(tool, executed?, detail)] for one conversation, or None if its transcript was not kept.""" | |
| f = os.path.join(B, cid, ".system_generated", "logs", "transcript_full.jsonl") | |
| if not os.path.exists(f): | |
| return None | |
| steps = [] | |
| for line in open(f): | |
| try: | |
| steps.append(json.loads(line)) | |
| except Exception: | |
| pass | |
| out = [] | |
| for i, s in enumerate(steps): | |
| for call in s.get("tool_calls") or []: | |
| name = call.get("name") | |
| # the step(s) after a planner call carry its outcome; a refusal names the deny rule or permission | |
| nxt = json.dumps(steps[i + 1:i + 3]) | |
| refused = any(w in nxt for w in ("deny rule", "Permission denied", "permission check failed", | |
| "auto-denied", "not allowed")) | |
| out.append((name, not refused, json.dumps(call.get("args"))[:160])) | |
| return out | |
| summary = collections.Counter() | |
| affected, missing, total = [], 0, 0 | |
| for d in sorted(glob.glob(os.path.join(glob.escape(ROOT), "results", "m*", "antigravity-cli__*"))): | |
| for f in sorted(glob.glob(os.path.join(glob.escape(d), "*.json"))): | |
| j = json.load(open(f)) | |
| cid = (j.get("meta") or {}).get("conversation_id") | |
| total += 1 | |
| tl = conversation_tools(cid) if cid else None | |
| if tl is None: | |
| missing += 1 | |
| continue | |
| executed = sorted({t for t, ok, _ in tl if ok and t not in ALLOWED}) | |
| for t, ok, _ in tl: | |
| summary[(t, "executed" if ok else "refused")] += 1 | |
| if executed: | |
| affected.append({"file": os.path.relpath(f, ROOT), "tools": executed, | |
| "correct_item": j.get("correct_item")}) | |
| print(f"answer files: {total}; transcript found for {total - missing}; missing {missing}") | |
| print("tool calls (name, outcome): ", dict(sorted(summary.items()))) | |
| by_model = collections.Counter(a["file"].split("/")[2] for a in affected) | |
| print(f"answers whose conversation EXECUTED a tool other than reading the picture: {len(affected)}", dict(by_model)) | |
| print("tools executed in those:", collections.Counter(t for a in affected for t in a["tools"])) | |
| json.dump(affected, open(os.path.join(ROOT, "results", "agy_tool_audit.json"), "w"), indent=1) | |