{
 "meta": {
  "id": "glossary",
  "title": "Glossary: every tooltip definition and every term defined in the methodology",
  "sources": [
   "index.html",
   "deterministic.html",
   "judged.html",
   "cascade.html",
   "large-scale.html",
   "cross-tool.html",
   "improvements.html",
   "experiments.html",
   "recommendations.html",
   "methodology.html",
   "reproduce.html"
  ],
  "extracted_at": "2026-10-16T14:34:41+00:00",
  "notes": [
   "tooltips: every distinct (term, definition) pair found in a `title` attribute, with the pages it appears on.",
   "terms: definitions written on the pages in prose (methodology, cascade, cross-tool), lightly normalised."
  ],
  "page_sha256": {
   "index.html": "ea2c1d462effa0d10c18016d0285ae5002a82e61b9d12aa0aa095343cc82cede",
   "deterministic.html": "752550a10cd6b5f5a0c465f1de4da1168521cefc519bedba7cd5c006f67e4cc4",
   "judged.html": "c8a6f6c0912a7791db4b80f3195dd316bc5603fbae3ecc0ad9297232724f0ebd",
   "cascade.html": "c1a4d8157eba904912e05da37e55dd3c9a4abc313064bd6dc0180ecce526bdc1",
   "large-scale.html": "2101600e4bc7833d4222bc8f066a548661443464a1ee4981cac5d07372d25a62",
   "cross-tool.html": "e7f8433087a55e382dc51895d350412a45ce19124d12beda376a47a1d55509aa",
   "improvements.html": "dbb68812c9eca9d368773fb520761af6c33836e65c5b7bcc35d84a0e642fcfb6",
   "experiments.html": "6b3dcd0ea2e5ff7b991df0410c5c929e998bee5fa4fc2865fcfa6ff3e486d42d",
   "recommendations.html": "cbb0c3dbc9b80000a67a18f13ae85a94a88827748e65758fc04f17fb9a8ce2da",
   "methodology.html": "a7a87af52e9ed8b91f308bf430c8247a235812bb6673c06ac189580cca8d2d3e",
   "reproduce.html": "95e1be0a878f1f106476790d86c38e29a1e0ec5a836c0f8434f9fb6ff5c81c99"
  }
 },
 "data": {
  "tooltips": [
   {
    "term": "95% interval",
    "definition": "The range the true value would fall in 95 times out of 100, given this sample size.",
    "pages": [
     "experiments.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "Added",
    "definition": "Findings the meta-judge introduced that no other stage had reported.",
    "pages": [
     "judged.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "Agreement",
    "definition": "Chance-corrected agreement between two reviewers. Zero means they agree no more than chance; one means perfectly.",
    "pages": [
     "cross-tool.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "Agrees",
    "definition": "Whether the two match closely enough that the harness can be trusted with a new configuration.",
    "pages": [
     "deterministic.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "any-intervention",
    "definition": "Scoring where asking a human to review also counts as catching something. The right lens when flags go to a review queue.",
    "pages": [
     "deterministic.html",
     "judged.html",
     "cascade.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "Any-intervention",
    "definition": "Scoring where asking a human to review also counts as catching something. The right lens when flags go to a review queue.",
    "pages": [
     "cascade.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "Arm",
    "definition": "One configuration of the scanner, measured on the same packages as the others.",
    "pages": [
     "judged.html",
     "improvements.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "Block-only",
    "definition": "Scoring where only a hard block counts as catching something. The right lens when a flag automatically stops the skill.",
    "pages": [
     "cascade.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "Chain",
    "definition": "A sequence of stages. A later stage can raise a decision but never lower it.",
    "pages": [
     "cascade.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "Cost per skill",
    "definition": "Wall-clock time and, where a model is charged for, money per skill scanned.",
    "pages": [
     "recommendations.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "F1",
    "definition": "The balance of precision and recall. High only when the scanner both finds most bad skills and is usually right when it flags one.",
    "pages": [
     "deterministic.html",
     "judged.html",
     "cascade.html",
     "large-scale.html",
     "cross-tool.html",
     "improvements.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "False negative",
    "definition": "A bad skill that was missed.",
    "pages": [
     "deterministic.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "False positive",
    "definition": "A harmless skill that was wrongly flagged.",
    "pages": [
     "deterministic.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "False-positive rate",
    "definition": "The share of harmless skills that were flagged.",
    "pages": [
     "cascade.html",
     "cross-tool.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "Fits a 32k-token window whole",
    "definition": "The share of skills small enough to send to a model in one piece without dropping anything.",
    "pages": [
     "methodology.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "FPR",
    "definition": "False-positive rate: the share of harmless skills that were flagged. This is what makes a scanner tiring to live with.",
    "pages": [
     "deterministic.html",
     "judged.html",
     "cascade.html",
     "large-scale.html",
     "improvements.html",
     "recommendations.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "Input tokens",
    "definition": "Total text sent to the model, which is what the cost is charged on.",
    "pages": [
     "cascade.html",
     "cross-tool.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "Judge",
    "definition": "The full LLM analyzer, which reads the skill and writes structured findings.",
    "pages": [
     "index.html",
     "judged.html",
     "large-scale.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "Judged real",
    "definition": "The share of sampled findings a reviewing model called a genuine security risk.",
    "pages": [
     "experiments.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "Median files",
    "definition": "The middle file count, which says more about a typical skill than an average would.",
    "pages": [
     "methodology.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "Median scanner-relevant bytes",
    "definition": "Text the scanner actually reads, after excluding images, fonts, archives and vendor directories.",
    "pages": [
     "methodology.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "Meta invoked",
    "definition": "The share of packages where the meta-judge actually ran. It skips packages whose findings are unambiguous.",
    "pages": [
     "judged.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "Population",
    "definition": "How many skills the figure was computed over, and their split between bad and harmless where both exist.",
    "pages": [
     "index.html",
     "deterministic.html",
     "judged.html",
     "cascade.html",
     "methodology.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "Precision",
    "definition": "Of the skills flagged, the share that really were bad. Low precision wastes reviewer time.",
    "pages": [
     "deterministic.html",
     "judged.html",
     "cascade.html",
     "large-scale.html",
     "cross-tool.html",
     "improvements.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "Previously published",
    "definition": "The figure reported for the deterministic rules before this work, which we treated as the number to reproduce.",
    "pages": [
     "deterministic.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "Re-measured",
    "definition": "What the same corpus produced when we ran it again from its pinned revision.",
    "pages": [
     "deterministic.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "Recall",
    "definition": "Of the bad skills present, the share that got flagged. Low recall means bad skills ship.",
    "pages": [
     "deterministic.html",
     "judged.html",
     "cascade.html",
     "large-scale.html",
     "cross-tool.html",
     "improvements.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "Recipe",
    "definition": "How much surrounding context was included in the request sent to the model.",
    "pages": [
     "cascade.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "Rules",
    "definition": "The deterministic stage: regular expressions, YARA signatures, syntax trees and dataflow. No model involved.",
    "pages": [
     "index.html",
     "judged.html",
     "cascade.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "Scan errors",
    "definition": "Packages the scanner could not read at all. These are excluded rather than counted as clean.",
    "pages": [
     "deterministic.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "Skills reaching each stage",
    "definition": "Each stage in order, showing the share of skills that still needed it. A stage sees fewer skills when earlier ones already settled them, which is where the saving comes from.",
    "pages": [
     "cascade.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "Suppressed",
    "definition": "Findings the meta-judge marked as false positives and hid.",
    "pages": [
     "judged.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "System One model",
    "definition": "A 9-billion-parameter decision model that answers fixed questions with calibrated probabilities rather than writing text.",
    "pages": [
     "index.html",
     "cascade.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "True negative",
    "definition": "A harmless skill that was correctly left alone.",
    "pages": [
     "deterministic.html"
    ],
    "kind": "tooltip"
   },
   {
    "term": "True positive",
    "definition": "A bad skill that was correctly flagged.",
    "pages": [
     "deterministic.html"
    ],
    "kind": "tooltip"
   }
  ],
  "terms": [
   {
    "term": "block-only lens",
    "definition": "Scoring where only a hard block counts as catching something. The right lens when a flag automatically stops the skill.",
    "page": "cascade.html"
   },
   {
    "term": "any-intervention lens",
    "definition": "Scoring where asking a human to review also counts as catching something. The right lens when flags go to a review queue.",
    "page": "cascade.html"
   },
   {
    "term": "detection",
    "definition": "The share of skills in a group that got flagged at all, for review or for blocking.",
    "page": "index.html"
   },
   {
    "term": "signal recall",
    "definition": "The share of malicious packages where the scanner produced any relevant signal, including below the threshold that would flag the package.",
    "page": "RESULTS.md"
   },
   {
    "term": "flag rate",
    "definition": "On an unlabelled population, the share of records flagged; it bounds the false-positive rate from above rather than measuring it.",
    "page": "large-scale.html"
   },
   {
    "term": "MEDIUM+ / HIGH+",
    "definition": "Detection at a severity threshold: any finding at MEDIUM or above (review), or HIGH or above (the shipped block gate).",
    "page": "recommendations.html"
   },
   {
    "term": "shipped install gate",
    "definition": "Each tool's own default decision: this scanner gates at HIGH; SkillSpector gates an install above a 0–100 risk score of 50.",
    "page": "cross-tool.html"
   },
   {
    "term": "paired recall difference",
    "definition": "Both tools resampled together on the same records so shared corpus difficulty cancels; an interval excluding zero means the tools are separable on that lens.",
    "page": "cross-tool.html"
   },
   {
    "term": "complementarity",
    "definition": "The share of records exactly one tool flags.",
    "page": "cross-tool.html"
   },
   {
    "term": "verdict repair",
    "definition": "Opt-in, escalate-only rewrite of a SAFE package verdict that contradicts its own findings to SUSPICIOUS; findings are never changed.",
    "page": "judged.html"
   },
   {
    "term": "meta-judge / meta-analyzer",
    "definition": "A second model pass that can only demote findings; it has a routing gate that returns early when no finding is ambiguous.",
    "page": "judged.html"
   },
   {
    "term": "System One model",
    "definition": "A decision model (OpenJev, jev-1.13.0) that answers fixed questions with calibrated probabilities rather than writing text; the eight-probe format is used as a screen.",
    "page": "index.html"
   },
   {
    "term": "OpenJev screen",
    "definition": "OpenJev's prompt_injection probability (or a logistic regression over all eight probe logits) decides which skills reach the judge.",
    "page": "large-scale.html"
   },
   {
    "term": "decomposed judge",
    "definition": "One judge pass per focus (specialized prompts), findings unioned; --llm-decompose.",
    "page": "recommendations.html"
   },
   {
    "term": "policy pack / preset",
    "definition": "balanced, low-noise (11 rules reported at LOW) and quiet (19 rules at LOW, judge low-confidence and contextual findings capped) presets.",
    "page": "large-scale.html"
   },
   {
    "term": "sole driver",
    "definition": "Records a rule alone drives to MEDIUM or above; removing the rule would clear them.",
    "page": "large-scale.html"
   },
   {
    "term": "judge clears",
    "definition": "The share of a rule's flagged real skills the judge rates clean; the judge is not ground truth.",
    "page": "large-scale.html"
   },
   {
    "term": "packing ladder",
    "definition": "Rung A sends the whole record when it fits; rung B fills the budget by priority (SKILL.md first); rung C sends one request per candidate finding.",
    "page": "methodology.html"
   },
   {
    "term": "noise floor",
    "definition": "The share of package verdicts that change when one judged configuration is repeated over the same packages (14%).",
    "page": "experiments.html"
   },
   {
    "term": "Wilson 95% interval",
    "definition": "The range the true value would fall in 95 times out of 100, given this sample size.",
    "page": "experiments.html"
   },
   {
    "term": "source-disjoint split",
    "definition": "MaliciousSkillBench's frozen test split whose sources do not overlap the train/validation split; used to design nothing.",
    "page": "deterministic.html"
   },
   {
    "term": "superseded",
    "definition": "A figure whose method was later replaced (the held-out-half screen threshold chosen on test members).",
    "page": "large-scale.html"
   }
  ]
 }
}