fix: ground Work reports in measurements

agent_platform/runtime/loop.py

@@ -39,6 +39,10 @@ For a code-and-report task, follow this order: write a real source file, read it
 file with a complete command, write the report using only the measured output, then read the report.
 The executable source and the report must be separate files: run the `.py`, `.js`, or other real
 source file, never a Markdown, text, JSON, CSV, YAML, or other documentation/data file.
+For comparisons and benchmarks, implement every compared subject, give each one equivalent independent
+inputs, validate results against a trusted reference, use a suitable high-resolution timer and meaningful
+sizes/repeats, and print concise numeric measurements instead of raw input arrays. Include exact measured
+values from successful execution output in the report; never substitute estimated or invented numbers.
 Write source files with real line breaks rather than literal `\\n` escape sequences. If any command
 fails, inspect the failure, repair the underlying problem, and successfully rerun the relevant check;
 never replace failed output with invented numbers, placeholders, or a hand-written success report.
@@ -82,6 +86,7 @@ EXECUTION_REQUEST = re.compile(
 )
 SOURCE_FILE_REQUEST = re.compile(r"(脚本|程序|代码|\b(?:script|program|code)\b)", re.IGNORECASE)
 REPORT_FILE_REQUEST = re.compile(r"(报告|\breport\b)", re.IGNORECASE)
+COMPARISON_REQUEST = re.compile(r"(比较|对比|基准|\b(?:compare|comparison|benchmark)\b)", re.IGNORECASE)
 MUTATION_TOOLS = {"write_file", "apply_patch"}
 VERIFICATION_TOOLS = {"read_file", "git_status", "git_diff", "exec", "poll_process"}
 EXECUTION_TOOLS = {"exec", "start_process", "poll_process"}
@@ -228,7 +233,8 @@ def deliverable_evidence_failure(
     written_source_paths: set[str],
     written_report_paths: set[str],
     reports_written_after_execution: set[str],
-    verified_paths: set[str],
+    report_grounding_required: bool,
+    grounded_report_paths: set[str],
 ) -> str | None:
     reasons: list[str] = []
     if source_required and not written_source_paths:
@@ -237,11 +243,19 @@ def deliverable_evidence_failure(
         reasons.append("Create a separate report file such as `report.md`.")
     elif report_required and not reports_written_after_execution:
         reasons.append("Write or update the report only after a successful execution, using its measured output.")
-    elif report_required and not (reports_written_after_execution & verified_paths):
+    elif report_grounding_required and not grounded_report_paths:
-        reasons.append("Read the generated report file after writing it to verify the delivered content.")
+        reasons.append(
+            "Update the report with at least one exact decimal or scientific-notation measurement from the "
+            "latest successful execution output; estimated or invented values are not accepted."
+        )
     return " ".join(reasons) or None
 
 
+def measurement_tokens(value: Any) -> set[str]:
+    text = str(value or "").lower()
+    return set(re.findall(r"(?<![\d.])(?:\d+\.\d+(?:e[+-]?\d+)?|\d+e[+-]?\d+)(?![\d.])", text))
+
+
 def recovery_message(failure: str) -> str:
     return (
         "Your attempted checkpoint was rejected by the execution policy. Continue the original task now. "
@@ -553,6 +567,7 @@ class AgentLoop:
         execution_required = depth == 0 and bool(EXECUTION_REQUEST.search(request_text))
         source_required = depth == 0 and bool(SOURCE_FILE_REQUEST.search(request_text))
         report_required = depth == 0 and bool(REPORT_FILE_REQUEST.search(request_text))
+        report_grounding_required = report_required and bool(COMPARISON_REQUEST.search(request_text))
         mutation_count = 0
         verified_mutation = -1
         successful_tools = 0
@@ -560,7 +575,8 @@ class AgentLoop:
         written_source_paths: set[str] = set()
         written_report_paths: set[str] = set()
         reports_written_after_execution: set[str] = set()
-        verified_paths: set[str] = set()
+        grounded_report_paths: set[str] = set()
+        latest_execution_measurements: set[str] = set()
         unresolved_execution_failure: str | None = None
         failed_execution_revisions: dict[str, int] = {}
         successful_write_fingerprints: set[str] = set()
@@ -604,7 +620,8 @@ class AgentLoop:
                         written_source_paths=written_source_paths,
                         written_report_paths=written_report_paths,
                         reports_written_after_execution=reports_written_after_execution,
-                        verified_paths=verified_paths,
+                        report_grounding_required=report_grounding_required,
+                        grounded_report_paths=grounded_report_paths,
                     ),
                 )
                 if not failure and action_required and successful_tools == 0:
@@ -723,8 +740,11 @@ class AgentLoop:
                             written_report_paths.add(artifact_path)
                             if successful_executions or not execution_required:
                                 reports_written_after_execution.add(artifact_path)
-                    elif name == "read_file" and artifact_path:
+                                report_content = str(arguments.get("content", ""))
-                        verified_paths.add(artifact_path)
+                                if not report_grounding_required or (
+                                    latest_execution_measurements & measurement_tokens(report_content)
+                                ):
+                                    grounded_report_paths.add(artifact_path)
                     fingerprint = write_call_fingerprint(name, arguments) if isinstance(arguments, dict) else None
                     if fingerprint:
                         successful_write_fingerprints.add(fingerprint)
@@ -733,6 +753,7 @@ class AgentLoop:
                     if is_substantive_execution(name, arguments):
                         successful_executions += 1
                         unresolved_execution_failure = None
+                        latest_execution_measurements = measurement_tokens(result.get("output"))
                 elif is_substantive_execution(name, arguments) and not result.get("policy_duplicate", False):
                     unresolved_execution_failure = f"{name}: {public_tool_summary(result, 1000)}"
                     execution_failures.append(unresolved_execution_failure)
@@ -780,7 +801,8 @@ class AgentLoop:
                 written_source_paths=written_source_paths,
                 written_report_paths=written_report_paths,
                 reports_written_after_execution=reports_written_after_execution,
-                verified_paths=verified_paths,
+                report_grounding_required=report_grounding_required,
+                grounded_report_paths=grounded_report_paths,
             ),
         ):
             return (

tests/test_agent_loop.py

@@ -11,6 +11,7 @@ from agent_platform.runtime.loop import (
     AgentLoop,
     deliverable_evidence_failure,
     is_substantive_execution,
+    measurement_tokens,
     non_source_execution_error,
     normalize_write_file_content,
     python_source_error,
@@ -77,7 +78,8 @@ def test_script_and_report_require_distinct_verified_artifacts() -> None:
             written_source_paths={"benchmark.py"},
             written_report_paths=set(),
             reports_written_after_execution=set(),
-            verified_paths={"benchmark.py"},
+            report_grounding_required=True,
+            grounded_report_paths=set(),
         )
         == "Create a separate report file such as `report.md`."
     )
@@ -88,12 +90,18 @@ def test_script_and_report_require_distinct_verified_artifacts() -> None:
             written_source_paths={"benchmark.py"},
             written_report_paths={"report.md"},
             reports_written_after_execution={"report.md"},
-            verified_paths={"benchmark.py", "report.md"},
+            report_grounding_required=True,
+            grounded_report_paths={"report.md"},
         )
         is None
     )
 
 
+def test_measurement_tokens_only_capture_exact_numeric_measurements() -> None:
+    assert measurement_tokens("n=1000 elapsed=0.012345s score=4.7e-06") == {"0.012345", "4.7e-06"}
+    assert measurement_tokens("O(n²), version 3") == set()
+
+
 def test_read_only_shell_commands_do_not_satisfy_execution_evidence() -> None:
     assert not is_substantive_execution("exec", {"command": "cat report.txt"})
     assert not is_substantive_execution("exec", {"command": "sed -n '1,20p' script.py"})
@@ -461,7 +469,7 @@ async def test_script_and_report_completion_requires_both_files(settings) -> Non
             max_tool_output_chars=10_000,
         ).run(
             spec=get_model_spec("work-light"),
-            messages=[{"role": "user", "content": "写脚本对比算法并给一个报告"}],
+            messages=[{"role": "user", "content": "写一个脚本并给一个报告"}],
             identity=UserIdentity("u1", "", "", "user"),
             raw_user_jwt="jwt",
             chat_id="script-report",