{
  "campaign": "layout-skill-bench",
  "description": "Four opencode sessions, one verbatim brief, three of them on one local model (Qwen3.8 Flash Next MLX) each loading a different design skill, plus a no-skill baseline (`vanilla`) that ran a hosted model and is not like-for-like — see its comparabilityNote, each producing a five-iteration landing-page design for the same note-taking app. Source vendored under demos/<arm>; per-arm self-reports (recovered from the opencode session database) under results/<arm>.md; this file is the single source of numbers for the dashboard.",
  "brief": "I want you to design the landing page for a note-taking application as essentially a second brain. You should design five iterations and each of them should be accessible within the slash one, slash two, slash three like pages directory. And then you should add a little button that lets me switch between them easily.",
  "briefNote": "Byte-identical across the three skill arms; the vanilla baseline adds a two-line no-skills preface and one trailing space, which its own briefNote records. Otherwise byte-identical across all four (SHA eaebe847..., 318 characters), confirmed against the opencode session database in Task 2. It is carried per-arm below (rather than only once at the top) so each arm record is self-contained for the dashboard, but it is the same string three times, not three different briefs. The prompt originates from whichai.dev (\"Which AI Made This?\"), which runs the same exercise on frontier models with and without a design skill; this campaign runs it on one local model.",
  "arms": [
    {
      "id": "design-skill",
      "harness": "opencode",
      "model": "qwen3.8-flash-next",
      "hosted": false,
      "label": "opencode + Qwen3.8 Flash Next + frontend-design",
      "color": "#2a9d8f",
      "sourceFolder": "~/LocalProjects/qwen38-next-layout-bench/design-skill",
      "skillConfig": "frontend-design",
      "skillConfigNote": "Loaded from the user's global skill directory (~/.agents/skills/frontend-design/SKILL.md), unrelated to the Leonxlnx/taste-skill package the other two arms pull from. This arm has no skills-lock.json at all — its skill was identified by reading the session transcript, not a lock file. It is NOT a no-skill baseline.",
      "brief": "I want you to design the landing page for a note-taking application as essentially a second brain. You should design five iterations and each of them should be accessible within the slash one, slash two, slash three like pages directory. And then you should add a little button that lets me switch between them easily.",
      "form": "Five static HTML directories, no build step, no framework (index.html plus 1/ 2/ 3/ 4/ 5/, each with its own index.html)",
      "config": {
        "harness": "opencode",
        "harnessVersion": null,
        "modelId": "ddalcu-Qwen3.8-Flash-Next-MLX-Serve-mixed-4-8bit",
        "runtime": "served locally on the rig",
        "effort": null,
        "source": "bench/layout-skill-bench/results/design-skill.md"
      },
      "short": "frontend-design"
    },
    {
      "id": "taste-skill",
      "harness": "opencode",
      "model": "qwen3.8-flash-next",
      "hosted": false,
      "label": "opencode + Qwen3.8 Flash Next + design-taste-frontend",
      "color": "#e76f51",
      "sourceFolder": "~/LocalProjects/qwen38-next-layout-bench/taste-skill",
      "skillConfig": "design-taste-frontend",
      "skillConfigNote": "One of three skills pinned by taste-skill/skills-lock.json (brandkit, design-taste-frontend, high-end-visual-design), all from Leonxlnx/taste-skill and vendored locally under .agents/skills/. The lock records what was AVAILABLE, not what ran: reading the session transcript shows only design-taste-frontend was actually invoked (87,308-character skill message matching design-taste-frontend/SKILL.md's 87,253 bytes on disk). taste2's lock file is byte-identical to this one, but taste2 invoked a different skill (high-end-visual-design) — the two arms are NOT a repeat run of the same configuration.",
      "brief": "I want you to design the landing page for a note-taking application as essentially a second brain. You should design five iterations and each of them should be accessible within the slash one, slash two, slash three like pages directory. And then you should add a little button that lets me switch between them easily.",
      "form": "Vite + React + TypeScript app (react 19, motion, tailwind v4, vite 7) with five client-side routes /1..5, plus a fixed-position switcher nav",
      "config": {
        "harness": "opencode",
        "harnessVersion": null,
        "modelId": "ddalcu-Qwen3.8-Flash-Next-MLX-Serve-mixed-4-8bit",
        "runtime": "served locally on the rig",
        "effort": null,
        "source": "bench/layout-skill-bench/results/taste-skill.md"
      },
      "short": "design-taste"
    },
    {
      "id": "taste2",
      "harness": "opencode",
      "model": "qwen3.8-flash-next",
      "hosted": false,
      "label": "opencode + Qwen3.8 Flash Next + high-end-visual-design",
      "color": "#f4a261",
      "sourceFolder": "~/LocalProjects/qwen38-next-layout-bench/taste2",
      "skillConfig": "high-end-visual-design",
      "skillConfigNote": "One of three skills pinned by taste2/skills-lock.json, byte-for-byte identical to taste-skill's lock file (diffed; both pin brandkit, design-taste-frontend, high-end-visual-design from Leonxlnx/taste-skill). The lock records what was AVAILABLE, not what ran: reading the session transcript shows only high-end-visual-design was actually invoked (10,476-character skill message matching high-end-visual-design/SKILL.md's 10,561 bytes on disk). brandkit does not appear in any of the three arms' sessions.",
      "brief": "I want you to design the landing page for a note-taking application as essentially a second brain. You should design five iterations and each of them should be accessible within the slash one, slash two, slash three like pages directory. And then you should add a little button that lets me switch between them easily.",
      "form": "Five static HTML pages (pages/1.html .. pages/5.html) served by a small Node HTTP server (server.js, plain node:http, no framework), \"/\" redirects to \"/1\"",
      "config": {
        "harness": "opencode",
        "harnessVersion": null,
        "modelId": "ddalcu-Qwen3.8-Flash-Next-MLX-Serve-mixed-4-8bit",
        "runtime": "served locally on the rig",
        "effort": null,
        "source": "bench/layout-skill-bench/results/taste2.md"
      },
      "short": "high-end-visual"
    },
    {
      "id": "vanilla",
      "harness": "opencode",
      "model": "qwen3.8-flash",
      "hosted": true,
      "label": "opencode + Qwen3.8 Flash (hosted), no skill",
      "color": "#264653",
      "sourceFolder": "~/LocalProjects/qwen38-next-layout-bench/vanilla",
      "skillConfig": "no skill",
      "skillConfigNote": "The baseline the other three lacked: the prompt opens \"do not use any skills. do not look at other files.\" No skill directory was loaded and the session invoked none.",
      "brief": "do not use any skills.\ndo not look at other files.\n\nI want you to design the landing page for a note-taking application as essentially a second brain. You should design five iterations and each of them should be accessible within the slash one, slash two, slash three like pages directory. And then you should add a little button that lets me switch between them easily. ",
      "briefNote": "The body matches the campaign brief except for one trailing space (319 characters against 318); the two lines above it are what makes this arm the baseline.",
      "comparabilityNote": "Not like-for-like with the other three. They ran ddalcu-Qwen3.8-Flash-Next-MLX-Serve-mixed-4-8bit on the rig at variant xhigh for no cost; this arm ran qwen/qwen3.8-flash through openrouter at variant high and billed US$0.028931. Different model build, different thinking level, hosted rather than local: its wall time and token counts do not compare across that line.",
      "form": "Five static HTML directories, no build step, no framework (index.html plus 1/ 2/ 3/ 4/ 5/, each with its own index.html)",
      "config": {
        "harness": "opencode",
        "harnessVersion": null,
        "modelId": "qwen/qwen3.8-flash via openrouter",
        "runtime": "hosted by openrouter",
        "effort": "high",
        "source": "bench/layout-skill-bench/results/vanilla.md"
      },
      "short": "no skill"
    }
  ],
  "results": [
    {
      "arm": "design-skill",
      "metric": "wallMinutes",
      "value": 12.4,
      "note": "743.6 s, first message to last recorded token event"
    },
    {
      "arm": "design-skill",
      "metric": "tokensOutput",
      "value": 37742
    },
    {
      "arm": "design-skill",
      "metric": "tokensInputUncached",
      "value": 48917
    },
    {
      "arm": "design-skill",
      "metric": "tokensCacheRead",
      "value": 437595
    },
    {
      "arm": "design-skill",
      "metric": "toolCalls",
      "value": 12,
      "note": "6 bash + 6 write"
    },
    {
      "arm": "design-skill",
      "metric": "messages",
      "value": 14,
      "note": "1 user + 13 assistant; one user turn (combined skill + brief message), no follow-up prompts"
    },
    {
      "arm": "design-skill",
      "metric": "pageCount",
      "value": 5
    },
    {
      "arm": "taste-skill",
      "metric": "wallMinutes",
      "value": 24.7,
      "note": "1,483.4 s, first message to last recorded token event"
    },
    {
      "arm": "taste-skill",
      "metric": "tokensOutput",
      "value": 45018
    },
    {
      "arm": "taste-skill",
      "metric": "tokensInputUncached",
      "value": 354721
    },
    {
      "arm": "taste-skill",
      "metric": "tokensCacheRead",
      "value": 905050
    },
    {
      "arm": "taste-skill",
      "metric": "toolCalls",
      "value": 25,
      "note": "6 bash + 2 todowrite + 13 write + 4 edit"
    },
    {
      "arm": "taste-skill",
      "metric": "messages",
      "value": 23,
      "note": "2 user + 21 assistant; two user turns (skill-load message, then the brief)"
    },
    {
      "arm": "taste-skill",
      "metric": "pageCount",
      "value": 5
    },
    {
      "arm": "taste2",
      "metric": "wallMinutes",
      "value": 38.0,
      "note": "2,279.0 s, first message to last recorded token event"
    },
    {
      "arm": "taste2",
      "metric": "tokensOutput",
      "value": 89922
    },
    {
      "arm": "taste2",
      "metric": "tokensInputUncached",
      "value": 464702
    },
    {
      "arm": "taste2",
      "metric": "tokensCacheRead",
      "value": 412103
    },
    {
      "arm": "taste2",
      "metric": "toolCalls",
      "value": 15,
      "note": "6 bash + 1 read + 7 write + 1 edit"
    },
    {
      "arm": "taste2",
      "metric": "messages",
      "value": 19,
      "note": "4 user + 15 assistant; four user turns (skill-load message, the brief, an opencode auto-compaction bookkeeping event, and a \"continue if you have next steps\" nudge) — the only one of the three sessions with a compaction event"
    },
    {
      "arm": "taste2",
      "metric": "pageCount",
      "value": 5
    },
    {
      "arm": "vanilla",
      "metric": "wallMinutes",
      "value": 15.5,
      "note": "932.9 s, first message to last recorded token event"
    },
    {
      "arm": "vanilla",
      "metric": "tokensOutput",
      "value": 30166
    },
    {
      "arm": "vanilla",
      "metric": "tokensInputUncached",
      "value": 37981
    },
    {
      "arm": "vanilla",
      "metric": "tokensCacheRead",
      "value": 472064
    },
    {
      "arm": "vanilla",
      "metric": "toolCalls",
      "value": 24,
      "note": "8 write + 5 bash + 2 edit + 2 todowrite + 1 grep + 1 read, and 5 parts that aborted"
    },
    {
      "arm": "vanilla",
      "metric": "messages",
      "value": 24,
      "note": "3 user + 21 assistant; the two extra user turns are nudges after failures (\"continue, algo ralhou\", \"try again\"), not design direction"
    },
    {
      "arm": "vanilla",
      "metric": "pageCount",
      "value": 5
    }
  ],
  "configNote": "Run configuration for each arm, transcribed from that arm's own report named in `source`. A field the report does not state is null, never guessed; none of these reports records a reasoning effort or thinking level.",
  "shortNote": "A short display name for each arm, for chart rows and other narrow places. It names the same arm as `label`; it measures nothing."
}