{
  "_about": "Stats for the blocks and catalog-routing tests: curated agent, routers, cost and tokens. Every number was recomputed from the raw results in results/blocks/ and checked by an independent audit. Label per number: measured, computed, sourced. Small tests, one writer model (Sonnet 5.5), briefs and keys written by us, 3 repeats.",
  "prices_usd": {
    "sonnet_5_5_in_per_mtok": 2.0,
    "sonnet_5_5_out_per_mtok": 10.0,
    "sonnet_5_5_cache_hit_per_mtok": 0.1,
    "source": "https://platform.claude.com/docs/en/about-claude/pricing (fetched 2026-10-29)",
    "label": "sourced",
    "jev_per_billion_tokens": 42,
    "jev_label": "sourced (provider price page, checked 2026-10-29)",
    "gemini_3_8_flash_in_out_cache_per_mtok": [
      0.75,
      3.75,
      0.075
    ],
    "gemini_3_5_flash_lite_in_out_cache_per_mtok": [
      0.3,
      2.5,
      0.03
    ],
    "gemini_source": "https://ai.google.dev/gemini-api/docs/pricing (last updated 2026-10-28; 3.8 Flash prices double on 2027-01-01)"
  },
  "tokens_per_request": {
    "full_catalog_input": 122832,
    "full_catalog_input_B2": 122833,
    "label": "measured (L3 pilot, 2 calls, Sonnet 5.5, 573 atoms)",
    "source": "results/blocks/level3-full-catalog-pilot/",
    "agent_slice_input": 3601.5,
    "agent_slice_output_mean": 52,
    "agent_label": "measured, 6 calls (P B1/B2 x3)",
    "source_agent": "results/blocks/level2-with-profile/",
    "sonnet_router_input": 1768.5,
    "sonnet_router_output": 43,
    "jev_router_input": 1456.65,
    "jev_router_output": 216.2,
    "router_label": "measured means (Sonnet R 48 calls; Jev 120 calls)",
    "full_output_mean": 67.5,
    "routed_total_input_sonnet_router": 5370,
    "routed_total_input_jev_router": 5058,
    "input_token_ratio_sonnet_router": 22.9,
    "input_token_ratio_jev_router": 24.3,
    "ratio_label": "computed, uncached"
  },
  "latency_s": {
    "full_catalog": 4.505,
    "full_label": "measured, 2 calls (4.74, 4.27)",
    "sonnet_router": 1.656,
    "jev_router": 0.1916,
    "agent_slice": 1.69,
    "routed_sonnet_router_sum": 3.35,
    "routed_jev_router_sum": 1.88,
    "quicker_sonnet_router": 1.35,
    "quicker_jev_router": 2.39,
    "sum_label": "computed: sums of separately measured means; routed path not run end to end"
  },
  "cost_usd_per_request": {
    "uncached": {
      "full_catalog": 0.24634,
      "jev_routed": 0.00779,
      "sonnet_routed": 0.01169,
      "jev_ratio": 31.6,
      "sonnet_ratio": 21.1
    },
    "cached_steady_state": {
      "full_catalog": 0.01296,
      "jev_routed": 0.00095,
      "sonnet_routed": 0.00148,
      "jev_ratio": 13.6,
      "sonnet_ratio": 8.7
    },
    "router_step_only": {
      "sonnet_uncached": 0.00397,
      "jev_uncached": 7.02e-05,
      "ratio_uncached": 56.4,
      "sonnet_cached": 0.00061,
      "ratio_cached": 8.57
    },
    "formula": "input_tokens x price_in + output_tokens x price_out; Jev tokens at Jev price; cached = 100 percent cache hits at 0.10 USD/MTok on both sides; cache writes (1.25x) and 5-minute TTL ignored",
    "label": "computed"
  },
  "cost_usd_per_1000_requests": {
    "uncached": {
      "full_catalog": 246.34,
      "jev_routed": 7.79,
      "sonnet_routed": 11.69
    },
    "cached": {
      "full_catalog": 12.96,
      "jev_routed": 0.95,
      "sonnet_routed": 1.48
    }
  },
  "gemini_writer_estimate": {
    "label": "ESTIMATE: Gemini prices applied to Claude-tokenizer counts",
    "gemini_3_8_flash": {
      "uncached_full": 0.0924,
      "uncached_jev_routed": 0.003,
      "ratio_uncached": 31.1,
      "cached_ratio": 17.7
    },
    "gemini_3_5_flash_lite": {
      "uncached_full": 0.037,
      "uncached_jev_routed": 0.0013,
      "ratio_uncached": 28.9,
      "cached_ratio": 12.5
    }
  },
  "routing_accuracy": {
    "set_R_16_briefs_x3": {
      "sonnet": {
        "block": "18/18",
        "cross": "21/21",
        "null": "3/3",
        "judgement": "6/6",
        "clear_total": "42/42"
      },
      "jev": {
        "block": "15/18 (B4 spaceship -> none x3)",
        "cross": "21/21",
        "null": "3/3",
        "judgement": "6/6",
        "clear_total": "39/42"
      },
      "note": "easy set we wrote; non-discriminating"
    },
    "set_R2_24_briefs_x3": {
      "sonnet": {
        "clear": "36/36",
        "null": "9/9",
        "two_catalog": "10/12",
        "judgement": "11/15",
        "empty_replies": "6 of 72 (token cap, not bad JSON)",
        "ms_per_route": 1953
      },
      "jev": {
        "clear": "36/36",
        "null": "9/9",
        "two_catalog": "12/12",
        "judgement": "12/15 (Y6 toy plane -> none x3)",
        "ms_per_route": 192
      },
      "non_llm": {
        "lexical_bm25_clear": "7/12",
        "tags_clear": "7/12",
        "hybrid_clear": "8/12",
        "tags_hybrid_cannot_say_none": true
      }
    },
    "label": "measured",
    "source": "results/blocks/scores/routers-non-llm/summary.md",
    "accuracy_not_compared_claim": "Accuracy not compared on screen: small test, we wrote the briefs"
  },
  "curated_agent": {
    "level1_B4_slice_stage1": {
      "calls": 48,
      "express_compile": "24/24",
      "json_compile": "21/24",
      "matching_briefs_B1_B2_B4": "9/9 both formats",
      "express_out_tokens_matching": "66-113",
      "json_out_tokens_matching": "266-591",
      "b8_refused": "0/3 both formats",
      "post_hoc_rescore_express_F1_B1_B7": "0.57 -> 1.0",
      "post_hoc_rescore_json": "0.57 -> 0.86",
      "note": "post-hoc: key error found after seeing results (house/cat accepted only motion_bricks); pinned scores kept"
    },
    "level2_P_profile_slice_P3": {
      "calls": 24,
      "express_only": true,
      "compile": "24/24",
      "matching_F1": 1.0,
      "refusals_B7_B8": "6/6"
    },
    "guardrail_G": {
      "calls": 24,
      "clear_refused": "6/6",
      "injection_refused": "6/6",
      "controls": "6/6",
      "borderline_built_bricks": "6/6 (judgement)",
      "prompt": "byte-identical to P (profile + P3 slice)"
    },
    "refusals_total_with_profile": "18/18 (G 12 + P 6), all instructed by the profile; unprofiled B8 0/3",
    "level3_full_catalog_pilot": {
      "calls": 2,
      "picked_right_block": "2/2",
      "input_tokens": 122832,
      "other_22_calls": "not run"
    }
  },
  "google_claims": {
    "source": "https://developers.googleblog.com/fast-reliable-a2ui/",
    "express_latency": "up to 3x",
    "express_tokens": "up to 3x fewer vs Direct JSON",
    "decision_models_like_jev_latency": "up to 80x (baseline not stated)",
    "jev_questions_per_screen": "about 30 (a count, not a speed)",
    "their_run": "11 prompts: tree on screen 92-298 ms, screens complete 0.8-1.7 s"
  },
  "our_express_vs_json": {
    "sonnet_5_5": {
      "output_tokens": "2.1x fewer",
      "time": "2.0x faster",
      "compiled": "18/18 (JSON 15/18)"
    },
    "gemini_3_5_flash_thinking_off": {
      "tokens": "1.7x",
      "time_paired": "1.24x (n=15)",
      "compiled": "15/18 vs 18/18"
    },
    "gemini_3_8_flash": {
      "tokens": "1.55x",
      "time": "no reliable gain (1.06x, n=10)",
      "compiled": "16/18 vs 10/18"
    }
  },
  "jev_output_free_correction_2026_10_30": {
    "source": "typesafe.ai/blog/introducing-system-one-models-and-jev, Cost row: input $0.042/MTok ($42 per billion), output FREE",
    "label": "computed; supersedes the Jev rows above that priced output at the input rate",
    "jev_router_cost_per_route_usd": 6.12e-05,
    "router_step_ratio_uncached": 64.9,
    "router_step_ratio_cached": 10.0,
    "jev_routed_per_request_usd": {
      "uncached": 0.00778,
      "cached": 0.00094
    },
    "jev_routed_per_1000_usd": {
      "uncached": 7.78,
      "cached": 0.94
    },
    "ratio_vs_full_catalog": {
      "uncached": 31.7,
      "cached": 13.8
    },
    "cached_precision": "derived from rounded 0.00095; margin about 0.1x",
    "unchanged": "all Sonnet-router rows, token ratios, latency, accuracy"
  },
  "not_claimed": [
    "80x confirmed or refuted",
    "our 24x (writer input tokens) compared with their 80x (a latency claim)",
    "Jev better than Sonnet in general",
    "routing improves accuracy",
    "any model ranking",
    "cheaper without saying uncached or cached",
    "routed path run end to end"
  ]
}