{
  "schema_version": 1,
  "builder_plan": {
    "scenario": "An agent reads external documents or tool responses that may contain malicious instructions.",
    "change": "Replay attacks and legitimate lookalikes in an isolated copy of your agent workflow.",
    "baseline": "Your existing checks with the same tasks, permissions, and tool responses.",
    "steps": [
      "Locate the paper's implementation and check its license, model support, data needs, and hardware requirements. Stop here if you cannot obtain the required artifacts.",
      "Save a small set of representative product tasks and expected outcomes. Keep evaluation tasks out of any training or tuning, and choose quality and cost limits before running.",
      "Run the baseline and the proposed change on identical inputs. Keep other settings fixed and repeat paired runs to expose variation.",
      "Save per-task outputs, failures, timings, and all usage costs. Review regressions before expanding the test."
    ],
    "measure": "Successful attacks missed, legitimate actions blocked, detection delay, and review cost.",
    "decision": "Keep the change only if it meets your preselected quality and latency limits and improves the metric you care about after setup and operating costs. Otherwise retain the baseline. This is an exploration plan, not a validated recommendation."
  },
  "builder_review": null,
  "arxiv_id": "2610.30742",
  "title": "UniGuard: Multi-View Guardrail Learning for Cross-Interface Safety Generalization",
  "summary": "UniGuard replaces interface-specific guardrail training (fixed prompt/response formats) with multi-view training over semantically aligned prompt, conversation, agent-trace, and multi-agent representations of the same safety case, using classification, contrastive, and calibration losses. The cost is a fine-tuning run on a 0.6B model (feasible on a single 24 GB GPU) plus constructing or adopting multi-view training data; inference cost remains that of a small classifier. It can fail because the 1,000-case synthetic dataset is small, gains are reported against its own pretrained baseline rather than standard production guardrails, and synthetic multi-view transformations may not capture real agent-trace distributions in the reader's deployment.",
  "rationale": "A 0.6B fine-tuned guardrail is deployable within the reader's hardware and budget, but the small synthetic training set and interface-generalization claims require validation against the reader's actual traffic before production adoption.",
  "technique": "Multi-view guardrail training for cross-interface safety",
  "family": "seguranca_e_guardrails",
  "recommendation": "testar",
  "published": "2026-10-22",
  "as_of": "2026-10-24",
  "editorial_status": "indexed",
  "source_url": "https://arxiv.org/abs/2610.30742",
  "independent_implementations": 3,
  "total_implementations": 3,
  "stars": 1906,
  "citations": 0,
  "claims": [
    {
      "claim_id": "claim-01",
      "statement": "UniGuard replaces interface-specific guardrail training (fixed prompt/response formats) with multi-view training over semantically aligned prompt, conversation, agent-trace, and multi-agent representations of the same safety case, using classification, contrastive, and calibration losses. The cost is a fine-tuning run on a 0.6B model (feasible on a single 24 GB GPU) plus constructing or adopting multi-view training data; inference cost remains that of a small classifier. It can fail because the 1,000-case synthetic dataset is small, gains are reported against its own pretrained baseline rather than standard production guardrails, and synthetic multi-view transformations may not capture real agent-trace distributions in the reader's deployment.",
      "basis": "inferred",
      "result": "",
      "baseline": "",
      "conditions": "",
      "source_url": "https://arxiv.org/abs/2610.30742",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "claim_id": "claim-02",
      "statement": "UniGuard-0.6B improves F1 by 45.12 points on average over its pretrained guardrail on unseen-interface benchmarks and reduces expected calibration error by 35.41 points, remaining competitive with much larger guardrails; these are point improvements, not a multiplicative factor.",
      "basis": "inferred",
      "result": "",
      "baseline": "",
      "conditions": "",
      "source_url": "https://arxiv.org/abs/2610.30742",
      "source_page": null,
      "source_excerpt": ""
    }
  ],
  "exposure_map": [
    {
      "dimension": "quality",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "compute",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "latency",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "operations",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "compatibility",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "security",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "data_and_training",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "reproducibility",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    }
  ],
  "risks": [],
  "minimum_test": [],
  "open_questions": [],
  "independent_tests": [],
  "lab_reproductions": [],
  "report_available": false,
  "validation_tier": "unknown",
  "evidence_tier": "unknown",
  "infrastructure_basis": "unknown",
  "equations": [],
  "equations_status": "not_fetched",
  "core_kind": "",
  "equations_fetched_at": "",
  "equations_source_url": "",
  "equations_source_sha256": "",
  "exposures_reviewed_at": "",
  "exposures_source_sha256": ""
}