{
  "paper": {
    "num": 10,
    "title": "TRACE: Temporal Risk Assessment and Concept Editing for Safer Language Generation",
    "authors": "Ankur Garg, Xuemin Yu, Hassan Sajjad, Samira Ebrahimi Kahou",
    "affiliation": ""
  },
  "inputs": {
    "codebook": 48,
    "prior a": "5",
    "\u03c1": "0.5",
    "region weight": "0.5",
    "\u03bb": "1",
    "hard negatives": "30% of safe",
    "harm onset": "0.4",
    "your sequence": "benign:20, caution:8, benign:6, harmful:14, benign:6",
    "seed": 3
  },
  "tiles": [
    {
      "label": "Streaming F1 (test)",
      "value": "0.96 vs 0.61",
      "sub": "TRACE vs DiffMean; FPR 2% vs 33%",
      "tone": "good"
    },
    {
      "label": "ASR analog",
      "value": "0.94 \u2192 0.18",
      "sub": "unsteered \u2192 TRACE edit at \u03bb = 1 (global: 0.00)",
      "tone": "accent"
    },
    {
      "label": "\u0394Ref analog",
      "value": "+0.00 vs +0.46",
      "sub": "safe responses disrupted: TRACE vs global direction",
      "tone": "warn"
    },
    {
      "label": "Your sequence stops at",
      "value": "token 39",
      "sub": "harm starts at token 35",
      "tone": "neutral"
    }
  ],
  "figures": [
    {
      "path": "fig_stream.png",
      "title": "When would generation stop?",
      "caption": "Cumulative streaming risk R_t = 1 \u2212 \u220f(1 \u2212 q_i) for your concept sequence and two test responses. Dotted lines mark where harmful content truly starts. The strip shows your sequence's true concept families and the tokens TRACE would edit."
    },
    {
      "path": "fig_steering.png",
      "title": "Selective edits vs one global direction",
      "caption": "Sweeping the steering strength \u03bb. TRACE edits only tokens assigned to harmful concepts (Eq. 7); the global direction shifts every token. Lower left is better."
    },
    {
      "path": "fig_concepts.png",
      "title": "Did the codebook find the harmful concepts?",
      "caption": "Each bar is a learned concept: its true composition (simulated ground truth) and its smoothed harmfulness score s_c (Eq. 3) computed only from response labels."
    },
    {
      "path": "fig_latency.png",
      "title": "Earlier detection costs false positives",
      "caption": "Median detection position vs false-positive rate across thresholds (the paper's Figure 2 analog)."
    }
  ],
  "compare": [
    {
      "metric": "Harmful-class F1, response / streaming (TRACE)",
      "paper": "WildGuard 0.835 / 0.837; S-Eval 0.933 / 0.935",
      "yours": "0.940 / 0.959",
      "note": "synthetic activations"
    },
    {
      "metric": "DiffMean F1, response / streaming",
      "paper": "WildGuard 0.491 / 0.491; S-Eval 0.820 / 0.820",
      "yours": "0.780 / 0.606",
      "note": ""
    },
    {
      "metric": "Steering ASR (unsteered \u2192 TRACE)",
      "paper": "WildGuard 0.873 \u2192 0.306; S-Eval 0.634 \u2192 0.062",
      "yours": "0.939 \u2192 0.184",
      "note": "activation-level analog"
    },
    {
      "metric": "\u0394Ref, TRACE vs refusal direction",
      "paper": "+0.006 vs +0.717 (WildGuard); \u22120.002 vs +0.439 (S-Eval)",
      "yours": "+0.000 vs +0.458",
      "note": "analog"
    },
    {
      "metric": "Refusal direction ASR",
      "paper": "0.000 on both datasets",
      "yours": "0.000 at \u03bb = 1",
      "note": ""
    },
    {
      "metric": "Stop threshold",
      "paper": "threshold = 0.605 (Figure 1 example)",
      "yours": "0.550",
      "note": "validation-selected"
    }
  ],
  "runbook": {
    "name": "smash-10-trace-simulator",
    "url": ""
  }
}