Spaces:
Running
Running
| """Development-only matched-date replay of the captured MCP schema failure.""" | |
| import datetime,json,time | |
| from pathlib import Path | |
| import needle | |
| from context_adapter import DecisionContext | |
| from decision_policy import validate_decision | |
| from native_schema import inference_tools | |
| def main(): | |
| record=json.loads(Path('benchmarks/workflow/results/mcp-failure-native.json').read_text())['records'][2] | |
| captured=record['native'][0] | |
| minute=datetime.datetime.fromtimestamp(record['response']['created'],datetime.timezone.utc).strftime('date: %Y-%m-%d %a %H:%M') | |
| system=minute+'; '+captured['system'];rows=[] | |
| for projected in (False,True): | |
| tools=inference_tools(captured['candidates']) if projected else captured['candidates'] | |
| agent=needle.Needle(tools=tools,system=system,generation=3) | |
| start=time.perf_counter();native=agent.complete(captured['query']) | |
| rows.append({'titles_removed':projected,'tools':tools,'system':system,'query':captured['query'],'latency_ms':(time.perf_counter()-start)*1000,'native':native,'validated':validate_decision(native,captured['candidates'],DecisionContext(**record['request']['decision_context']),system+'\n'+captured['query'])}) | |
| Path('benchmarks/workflow/results/native-schema-controlled-audit.json').write_text(json.dumps({'methodology':'Development causal probe, no tool execution. Both fresh calls use the same explicit date fact at the response minute of the captured failure; no SDK clock/source/weights/flags/threshold changes. Original internal constructor timestamp was not separately captured, so this is a matched replay, not a claim about hidden historical bytes.','rows':rows},indent=2)+'\n') | |
| if __name__=='__main__':main() | |