{
 "meta": {
  "id": "cost",
  "title": "Every token, time and cost figure, reconciled",
  "sources": [
   "recommendations.html#1",
   "improvements.html#6",
   "cross-tool.html (Cost of a scan)",
   "cascade.html#5",
   "methodology.html#2",
   "index.html (KPI)",
   {
    "repo": "Vineethsain/skill-scanner-vs-skillspector",
    "path": "results/large-scale.json",
    "sha256": "e48f9cb342c05d2d13286c6b91c165411994d6febd71e14d1fe54c6f339d32b5"
   },
   {
    "repo": "Vineethsain/cisco-skill-scanner-judge-and-system-one-results",
    "path": "RESULTS.md",
    "sha256": "2503c65d7bda785a6b4ddcc7bb0e1bb56437a3a23c8f332ffab04b7f6414daa3"
   }
  ],
  "extracted_at": "2026-10-16T14:34:41+00:00",
  "notes": [
   "tokens_per_skill: 4,838/300 (improvements.html#6) is current; 3,800/90 (recommendations) is kept with superseded: true (owner decision 2); the full-corpus average and the Haiku arm totals are further variants.",
   "results_md_cost is the dataset's cost table (backed by RESULTS.md); the value ratios derived from it are dropped everywhere (owner decision 3).",
   "seconds_per_skill: cross-tool wall clock, instrumented per subprocess (covers SkillSpector; the in-process Gemma arm is not instrumented).",
   "context_input_tokens: cascade.html#5 recipe totals for the System One context ablation."
  ],
  "page_sha256": {
   "index.html": "ea2c1d462effa0d10c18016d0285ae5002a82e61b9d12aa0aa095343cc82cede",
   "cascade.html": "c1a4d8157eba904912e05da37e55dd3c9a4abc313064bd6dc0180ecce526bdc1",
   "cross-tool.html": "e7f8433087a55e382dc51895d350412a45ce19124d12beda376a47a1d55509aa",
   "improvements.html": "dbb68812c9eca9d368773fb520761af6c33836e65c5b7bcc35d84a0e642fcfb6",
   "recommendations.html": "cbb0c3dbc9b80000a67a18f13ae85a94a88827748e65758fc04f17fb9a8ce2da",
   "methodology.html": "a7a87af52e9ed8b91f308bf430c8247a235812bb6673c06ac189580cca8d2d3e"
  }
 },
 "data": {
  "tokens_per_skill": [
   {
    "variant": "improvements.html#6 single judge pass (120 source-disjoint records, Gemma 4 26B)",
    "input": 4838,
    "output": 300,
    "page": "improvements.html#6",
    "current": true,
    "superseded": false
   },
   {
    "variant": "recommendations.html stage table (LLM judge, new prompt)",
    "input": 3800,
    "output": 90,
    "page": "recommendations.html#1",
    "superseded": true,
    "current": false,
    "note": "'about 3,800 input and 90 output tokens per skill'; superseded by 4,838/300 (owner decision 2)"
   },
   {
    "variant": "improvements.html#6 five specialized passes",
    "input": 24703,
    "output": 1548,
    "page": "improvements.html#6",
    "ratio_to_single": 5.106035551880942
   },
   {
    "variant": "full gitskills judge run (shipped prompt), total / 1,876,206 skills",
    "input": 3779.5695259475774,
    "output": 84.71474241101457,
    "page": "results/large-scale.json full.judge.tokens"
   },
   {
    "variant": "cross-tool.html Haiku arm, msb-source-disjoint (totals / records scored on both)",
    "input": 10111.937972768532,
    "output": 1593.9417549167927,
    "page": "cross-tool.html#35",
    "totals": {
     "input": 13367982,
     "output": 2107191
    }
   },
   {
    "variant": "cross-tool.html Haiku arm, msb-balanced-800 (totals / records scored on both)",
    "input": 11293.035248041775,
    "output": 2368.476501305483,
    "page": "cross-tool.html#42",
    "totals": {
     "input": 8650465,
     "output": 1814253
    }
   }
  ],
  "results_md_cost": {
   "source": "RESULTS.md (dataset Vineethsain/cisco-skill-scanner-judge-and-system-one-results), 'What each stage costs, and what it buys'",
   "lens": "block_only",
   "population": "msb-source-disjoint (1,384 packages)",
   "rows": [
    {
     "configuration": "Rules only",
     "f1": 0.1374,
     "points_over_rules": 0.0,
     "usd_for_corpus": 0.0,
     "seconds_per_skill": 0.014,
     "f1_points_per_dollar": null
    },
    {
     "configuration": "Rules plus System One model",
     "f1": 0.245,
     "points_over_rules": 10.8,
     "usd_for_corpus": 0.1755,
     "seconds_per_skill": 0.044,
     "f1_points_per_dollar": 61.3
    },
    {
     "configuration": "Rules plus judge",
     "f1": 0.3768,
     "points_over_rules": 23.9,
     "usd_for_corpus": 1.0211,
     "seconds_per_skill": 2.377,
     "f1_points_per_dollar": 23.4
    },
    {
     "configuration": "Rules plus judge plus meta",
     "f1": 0.3724,
     "points_over_rules": 23.5,
     "usd_for_corpus": 1.8454,
     "seconds_per_skill": 2.377,
     "f1_points_per_dollar": 12.7
    }
   ],
   "prices_usd_per_million_tokens": {
    "judge_input": 0.13,
    "judge_output": 0.4,
    "system_one_input": 0.042
   },
   "measurement": "System One measured over all 1,384 packages; judge tokens measured on a 60-package sample then projected; batch inference (halves the judge rate) not assumed.",
   "note": "The value ratios the dataset derives from this table are dropped (owner decision 3); its 24.50% F1 for rules+System One matches no Space table."
  },
  "seconds_per_skill": [
   {
    "arm": "gemma-4-26b",
    "corpus": "msb-source-disjoint",
    "tool": "skillspector",
    "mean_seconds_per_skill": 26.018,
    "page": "cross-tool.html#7"
   },
   {
    "arm": "gemma-4-26b",
    "corpus": "msb-balanced-800",
    "tool": "skillspector",
    "mean_seconds_per_skill": 23.174,
    "page": "cross-tool.html#14"
   },
   {
    "arm": "gemma-4-26b",
    "corpus": "openskillrisk-corpus",
    "tool": "skillspector",
    "mean_seconds_per_skill": 23.519,
    "page": "cross-tool.html#21"
   },
   {
    "arm": "gemma-4-26b",
    "corpus": "harmfulskillbench-corpus",
    "tool": "skillspector",
    "mean_seconds_per_skill": 17.079,
    "page": "cross-tool.html#28"
   },
   {
    "arm": "claude-haiku-4.5",
    "corpus": "msb-source-disjoint",
    "tool": "skill-scanner",
    "mean_seconds_per_skill": 13.763,
    "page": "cross-tool.html#35"
   },
   {
    "arm": "claude-haiku-4.5",
    "corpus": "msb-source-disjoint",
    "tool": "skillspector",
    "mean_seconds_per_skill": 39.585,
    "page": "cross-tool.html#35"
   },
   {
    "arm": "claude-haiku-4.5",
    "corpus": "msb-balanced-800",
    "tool": "skill-scanner",
    "mean_seconds_per_skill": 19.865,
    "page": "cross-tool.html#42"
   },
   {
    "arm": "claude-haiku-4.5",
    "corpus": "msb-balanced-800",
    "tool": "skillspector",
    "mean_seconds_per_skill": 30.982,
    "page": "cross-tool.html#42"
   },
   {
    "arm": "static",
    "corpus": "msb-source-disjoint",
    "tool": "skill-scanner",
    "mean_seconds_per_skill": 0.015,
    "page": "cross-tool.html#49"
   },
   {
    "arm": "static",
    "corpus": "msb-source-disjoint",
    "tool": "skillspector",
    "mean_seconds_per_skill": 2.453,
    "page": "cross-tool.html#49"
   },
   {
    "arm": "static",
    "corpus": "msb-balanced-800",
    "tool": "skill-scanner",
    "mean_seconds_per_skill": 0.017,
    "page": "cross-tool.html#55"
   },
   {
    "arm": "static",
    "corpus": "msb-balanced-800",
    "tool": "skill-scanner-all-packs",
    "mean_seconds_per_skill": 0.28,
    "page": "cross-tool.html#55"
   },
   {
    "arm": "static",
    "corpus": "msb-balanced-800",
    "tool": "skillspector",
    "mean_seconds_per_skill": 2.399,
    "page": "cross-tool.html#55"
   },
   {
    "arm": "static",
    "corpus": "openskillrisk-corpus",
    "tool": "skill-scanner",
    "mean_seconds_per_skill": 0.018,
    "page": "cross-tool.html#61"
   },
   {
    "arm": "static",
    "corpus": "openskillrisk-corpus",
    "tool": "skill-scanner-all-packs",
    "mean_seconds_per_skill": 0.266,
    "page": "cross-tool.html#61"
   },
   {
    "arm": "static",
    "corpus": "openskillrisk-corpus",
    "tool": "skillspector",
    "mean_seconds_per_skill": 2.457,
    "page": "cross-tool.html#61"
   },
   {
    "arm": "static",
    "corpus": "harmfulskillbench-corpus",
    "tool": "skill-scanner",
    "mean_seconds_per_skill": 0.017,
    "page": "cross-tool.html#67"
   },
   {
    "arm": "static",
    "corpus": "harmfulskillbench-corpus",
    "tool": "skill-scanner-all-packs",
    "mean_seconds_per_skill": 0.325,
    "page": "cross-tool.html#67"
   },
   {
    "arm": "static",
    "corpus": "harmfulskillbench-corpus",
    "tool": "skillspector",
    "mean_seconds_per_skill": 2.4,
    "page": "cross-tool.html#67"
   }
  ],
  "wall_clock": [
   {
    "what": "System One (OpenJev) over the 1,384-package track",
    "seconds": 42,
    "usd": 0.18,
    "backed_by": "RESULTS.md stage table (System One model row)"
   },
   {
    "what": "System One per skill",
    "milliseconds": 30,
    "usd_cents": 0.013,
    "backed_by": "RESULTS.md stage table (Cost per skill cell)"
   },
   {
    "what": "Judge per skill",
    "seconds": 4,
    "backed_by": "RESULTS.md stage table ('about four seconds per skill')"
   },
   {
    "what": "Judge over 1.88M skills, total input tokens",
    "tokens": 7091251022,
    "backed_by": "results/large-scale.json full.judge.tokens"
   },
   {
    "what": "Full-corpus GPU run",
    "hardware": "four NVIDIA H200 GPUs (150 GB each), 96 CPU cores",
    "backed_by": "results/large-scale.json hardware"
   }
  ],
  "context_input_tokens": [
   {
    "recipe": "K0",
    "input_tokens": 2062360
   },
   {
    "recipe": "K1",
    "input_tokens": 2495808
   },
   {
    "recipe": "K3",
    "input_tokens": 2544772
   },
   {
    "recipe": "K7",
    "input_tokens": 2552672
   }
  ],
  "tokenizer": {
   "bytes_per_token_code": 3.06,
   "naive_estimate": 4.0,
   "budget_bytes_per_token": 2,
   "largest_request": {
    "bytes": 124989,
    "prompt_tokens": 40855
   },
   "character_budget": 100000,
   "prompt_cache": {
    "cached_tokens": 0,
    "cache_write_tokens": 0,
    "request_tokens": 4859,
    "note": "no caching on the Bedrock mantle route"
   },
   "page": "methodology.html#2"
  },
  "usage_counter_bug": "the analyzer's usage counter reported the most recent call rather than the sum, so no cost is quoted for three passes",
  "decomposed_cost_note": "Three passes cost three times the model calls; no cost figure quoted (usage counter bug)."
 }
}