{
  "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": "2610.01170",
  "title": "HeadEdit: Calibrating Language Model Behavior Through the Frozen Unembedding Matrix",
  "summary": "HeadEdit replaces manual activation steering, target-token heuristics, or additional preference-tuning runs for correcting residual misbehaviors such as over-refusal, unnecessary tool calls, and sycophancy, by applying a prompt-adaptive vocabulary-wide logit correction derived from a low-rank subspace of the frozen unembedding matrix. Its cost is a one-time extraction of paired completions to fit the behavioral subspace plus a small inference-time projection, with negligible latency overhead and no parameter updates or retraining. It can fail if the paired-completion data does not represent the production behavior distribution, if the linear decodability assumption breaks for a given model or behavior, or if the correction shifts logits on benign prompts in ways the paper's limited task suite did not expose.",
  "rationale": "It is gradient-free, inference-time, and cheap enough for a 24 GB GPU or API-adjacent setup, but the behavioral gains are validated only on the paper's specific refusal, tool-use, and sycophancy tasks and must be re-verified on the reader's own prompts and models.",
  "technique": "Unembedding-based behavioral calibration (HeadEdit)",
  "family": "seguranca_e_guardrails",
  "recommendation": "testar",
  "published": "2026-10-01",
  "as_of": "2026-10-18",
  "editorial_status": "indexed",
  "source_url": "https://arxiv.org/abs/2610.01170",
  "independent_implementations": 0,
  "total_implementations": 1,
  "stars": 121,
  "citations": null,
  "claims": [
    {
      "claim_id": "claim-01",
      "statement": "HeadEdit replaces manual activation steering, target-token heuristics, or additional preference-tuning runs for correcting residual misbehaviors such as over-refusal, unnecessary tool calls, and sycophancy, by applying a prompt-adaptive vocabulary-wide logit correction derived from a low-rank subspace of the frozen unembedding matrix. Its cost is a one-time extraction of paired completions to fit the behavioral subspace plus a small inference-time projection, with negligible latency overhead and no parameter updates or retraining. It can fail if the paired-completion data does not represent the production behavior distribution, if the linear decodability assumption breaks for a given model or behavior, or if the correction shifts logits on benign prompts in ways the paper's limited task suite did not expose.",
      "basis": "inferred",
      "result": "",
      "baseline": "",
      "conditions": "",
      "source_url": "https://arxiv.org/abs/2610.01170",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "claim_id": "claim-02",
      "statement": "HeadEdit improves all nine experimental settings across three behavioral tasks and three model families, with negligible inference overhead and no systematic loss of general capabilities; no multiplicative factor is reported in the abstract.",
      "basis": "inferred",
      "result": "",
      "baseline": "",
      "conditions": "",
      "source_url": "https://arxiv.org/abs/2610.01170",
      "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": ""
}