{
  "testDate": "2026-08-02",
  "model": "glm-5.2",
  "endpoint": "https://api.z.ai/api/paas/v4/chat/completions",
  "method": {
    "fixtures": 8,
    "conditions": 2,
    "requests": 16,
    "runsPerFixtureCondition": 1,
    "thinking": "disabled",
    "sampling": "do_sample=false",
    "conditionOrder": "alternated by fixture"
  },
  "aggregate": {
    "genericRejection": {
      "exactMatches": 6,
      "fixtures": 8,
      "totalTokens": 1811,
      "medianLatencyMs": 1897
    },
    "alignedFeedback": {
      "exactMatches": 8,
      "fixtures": 8,
      "totalTokens": 1923,
      "medianLatencyMs": 1596
    },
    "observedExactMatchDelta": 2,
    "percentagePointDeltaOnThisFixtureSet": 25,
    "additionalTotalTokens": 112
  },
  "fixtures": [
    {"id":"archive-invoice","expected":"validate_totals","generic":"validate_totals","aligned":"validate_totals"},
    {"id":"close-incident","expected":"attach_evidence","generic":"export_timeline","aligned":"attach_evidence"},
    {"id":"publish-report","expected":"redact_pii","generic":"generate_pdf","aligned":"redact_pii"},
    {"id":"deploy-release","expected":"approve_change","generic":"approve_change","aligned":"approve_change"},
    {"id":"refund-payment","expected":"lookup_payment","generic":"lookup_payment","aligned":"lookup_payment"},
    {"id":"rotate-key","expected":"create_backup_key","generic":"create_backup_key","aligned":"create_backup_key"},
    {"id":"merge-dependency","expected":"run_security_scan","generic":"run_security_scan","aligned":"run_security_scan"},
    {"id":"update-shared-document","expected":"acquire_lock","generic":"acquire_lock","aligned":"acquire_lock"}
  ],
  "sources": {
    "paper": "https://arxiv.org/abs/2505.21055",
    "repository": "https://github.com/THUNLP-MT/ALIGN",
    "repositoryCommit": "4a853dd462f08eb1fbad35ab6d5b9207d860ddcf",
    "glmGuide": "https://docs.z.ai/guides/llm/glm-5.2"
  },
  "limitations": [
    "This is a one-step recovery probe, not an ALFWorld, ScienceWorld, WebShop or M3ToolEval reproduction.",
    "Each condition was observed once on one model route, so the percentages do not estimate a population success rate.",
    "Aligned feedback names the missing prerequisite; the test measures information supplied by the interface, not general intelligence.",
    "Sequential latency is retained for auditability but is not a provider performance claim."
  ]
}