{
  "model": "claude-haiku-5-5",
  "category": "build",
  "scenarios": [
    {
      "id": "a1-static-site",
      "status": "pass",
      "product_defects": [],
      "task_outcome": "Built all five bakery pages; independently passed 26 validation tests, 62 local links and nine HTTP checks; verified BrowserVerify reports, artifacts and hashes.",
      "notes": "38 executed calls exceeded 25; TaskStop stopped the server. All 17 source assets have absent:true baselines; all 53 output snapshots match. Four invalid BrowserPlan inputs correctly settled refused; a model-supplied missing gallery selector produced failed, then a corrected plan passed 11 assertions. Four final browser plans passed. Three heartbeats stopped when output began. Cost $0.05310502. Recorded CLI exit 0, one turn-complete, valid paired sessions/protocol, 30 tools with six bundled Skill catalog entries, positive cache reads, exact usageTotals cost recomputation, empty .err files and recorded leftovers; current process/port checks clear. No DSH/GPCI/DeferredToolCall effects. Independent checks used scene-local scratch copies and fresh HOME; evidence: review-build-audit.json and review-build-location.txt."
    },
    {
      "id": "a2-vite-react",
      "status": "pass",
      "product_defects": [],
      "task_outcome": "Scaffolded React/TypeScript app and served Hello from G-Code; independent npm run build passed with existing dependencies under Node 24.22.0.",
      "notes": "Saved curl results show HTTP 200 and the changed heading in /src/App.tsx; tasks remained running at turn end as requested and shutdown removed the server. Runner samples leftovers five seconds after exit, so the approximate three-second timing is not directly measured. Original App.tsx scaffold remains in version 1. Cost $0.00337062. Recorded CLI exit 0, one turn-complete, valid paired sessions/protocol, 30 tools with six bundled Skill catalog entries, positive cache reads, exact usageTotals cost recomputation, empty .err files and recorded leftovers; current process/port checks clear. No DSH/GPCI/DeferredToolCall effects. Independent checks used scene-local scratch copies and fresh HOME; evidence: review-build-audit.json and review-build-location.txt."
    },
    {
      "id": "a3-vite-build-preview",
      "status": "pass",
      "product_defects": [],
      "task_outcome": "Added three coffee origins; independent production build passed and rebuilt HTML/JS both returned HTTP 200; local asset links resolved.",
      "notes": "TaskStop stopped the background server, including its shell child. The stale scenario wording expects a scaffold baseline, but outputs.ts:185 and :467 preserve absent:true for files created earlier in this session; original main.ts scaffold is retained as version 1, followed by the edited version. Both hashes match, and no original content was lost. Cost $0.004094495. Recorded CLI exit 0, one turn-complete, valid paired sessions/protocol, 30 tools with six bundled Skill catalog entries, positive cache reads, exact usageTotals cost recomputation, empty .err files and recorded leftovers; current process/port checks clear. No DSH/GPCI/DeferredToolCall effects. Independent checks used scene-local scratch copies and fresh HOME; evidence: review-build-audit.json and review-build-location.txt."
    },
    {
      "id": "a4-canvas-game",
      "status": "pass",
      "product_defects": [],
      "task_outcome": "Built Snake with injectable RNG; independently passed all 81 logic tests and verified the saved 10-assertion browser report and evidence.",
      "notes": "The model corrected test assumptions and a Node test-directory invocation, without a failure-guard stop. Four malformed FileEdit argument streams and one malformed BrowserPlan stream correctly settled refused and never executed; all canonical calls are paired. Skill browser-verify loaded successfully without approval. Six heartbeats stopped on visible output. Browser coverage limitations are candidly stated in the saved final. Cost $0.053402545. Recorded CLI exit 0, one turn-complete, valid paired sessions/protocol, 30 tools with six bundled Skill catalog entries, positive cache reads, exact usageTotals cost recomputation, empty .err files and recorded leftovers; current process/port checks clear. No DSH/GPCI/DeferredToolCall effects. Independent checks used scene-local scratch copies and fresh HOME; evidence: review-build-audit.json and review-build-location.txt."
    },
    {
      "id": "a5-md-docs",
      "status": "pass",
      "product_defects": [],
      "task_outcome": "Built the zero-dependency Markdown converter; independently passed 43 tests, rebuilt six HTML pages and checked all 36 internal links.",
      "notes": "All eight generated outputs and their versions/baselines are complete and hash-valid. Model-authored test/build bugs were corrected during the turn. The harness seed contains duplicated headings and no intended fenced code because shell backticks were consumed; this is a fixture issue, disclosed by the model. Six heartbeats stopped on output. Cost $0.03382571. Recorded CLI exit 0, one turn-complete, valid paired sessions/protocol, 30 tools with six bundled Skill catalog entries, positive cache reads, exact usageTotals cost recomputation, empty .err files and recorded leftovers; current process/port checks clear. No DSH/GPCI/DeferredToolCall effects. Independent checks used scene-local scratch copies and fresh HOME; evidence: review-build-audit.json and review-build-location.txt."
    },
    {
      "id": "a6-express-api",
      "status": "pass",
      "product_defects": [],
      "task_outcome": "Built Express notes CRUD and static frontend; independently asserted 14 local HTTP responses covering create/list/get/update/delete, validation and 404s, then reaped the test server.",
      "notes": "The saved run installed dependencies, exercised endpoints and stopped its server with TaskStop. One model FileEdit string mismatch correctly failed with closest-line context and was corrected. Reviewer used the already-installed Express 5.3.0 without installing or upgrading dependencies. Cost $0.006489245. Recorded CLI exit 0, one turn-complete, valid paired sessions/protocol, 30 tools with six bundled Skill catalog entries, positive cache reads, exact usageTotals cost recomputation, empty .err files and recorded leftovers; current process/port checks clear. No DSH/GPCI/DeferredToolCall effects. Independent checks used scene-local scratch copies and fresh HOME; evidence: review-build-audit.json and review-build-location.txt."
    },
    {
      "id": "a7-chinese-site",
      "status": "pass",
      "product_defects": [],
      "task_outcome": "Built four Chinese pages; independently passed all lang/title checks and 26 local links, and verified the seven-assertion browser report.",
      "notes": "Strict UTF-8 decoding and replacement-character checks passed for 13 original source/session/protocol files; the final summary is readable Chinese. Menu data has four categories and 11 dishes, and browser evidence confirms RMB rendering. Browser report hashes and artifacts match; its preview port is closed. Cost $0.013564805. Recorded CLI exit 0, one turn-complete, valid paired sessions/protocol, 30 tools with six bundled Skill catalog entries, positive cache reads, exact usageTotals cost recomputation, empty .err files and recorded leftovers; current process/port checks clear. No DSH/GPCI/DeferredToolCall effects. Independent checks used scene-local scratch copies and fresh HOME; evidence: review-build-audit.json and review-build-location.txt."
    },
    {
      "id": "j1-todo-fullstack",
      "status": "pass",
      "product_defects": [],
      "task_outcome": "Built JSON-backed todo API and vanilla frontend; independently passed 59 store/HTTP tests with ephemeral servers and verified the final 13-assertion browser report.",
      "notes": "The model corrected its initial Node test invocation. BrowserPlan actually ran and failed on a nonexistent entry file; BrowserVerify actually ran and failed on an inappropriate button-value assertion. Both correctly remain failed, and corrected plans pass; failed evidence is retained. Saved curl results cover CRUD/filter/error responses; TaskStop stopped the app and reviewer test servers closed. Cost $0.03758046. Recorded CLI exit 0, one turn-complete, valid paired sessions/protocol, 30 tools with six bundled Skill catalog entries, positive cache reads, exact usageTotals cost recomputation, empty .err files and recorded leftovers; current process/port checks clear. No DSH/GPCI/DeferredToolCall effects. Independent checks used scene-local scratch copies and fresh HOME; evidence: review-build-audit.json and review-build-location.txt."
    },
    {
      "id": "j2-cli-tool",
      "status": "pass",
      "product_defects": [],
      "task_outcome": "Built quoted-field CSV statistics CLI; independently passed 33 tests and ran it successfully on the 2,000-row sample.",
      "notes": "Initial inspection failed because package.json was absent, then the model proceeded normally. The model corrected its Node test-directory invocation; shell pipelines ending in tail correctly returned the pipeline status while showing test failures, followed by a successful retest. One heartbeat stopped on output. Cost $0.015156115. Recorded CLI exit 0, one turn-complete, valid paired sessions/protocol, 30 tools with six bundled Skill catalog entries, positive cache reads, exact usageTotals cost recomputation, empty .err files and recorded leftovers; current process/port checks clear. No DSH/GPCI/DeferredToolCall effects. Independent checks used scene-local scratch copies and fresh HOME; evidence: review-build-audit.json and review-build-location.txt."
    },
    {
      "id": "j7-sonnet-sanity",
      "status": "pass",
      "product_defects": [],
      "task_outcome": "Created TABLE.md with eight planets; independently ran verify-table.mjs and confirmed eight table data rows.",
      "notes": "Baseline run used claude-haiku-5-5 despite the historical scenario name: three requests, 39,297 input tokens, 36,523 cache-read tokens, and recomputed cost $0.00167978. Both output snapshots match. Table row structure was checked; changing astronomical moon counts were not fact-checked. Recorded CLI exit 0, one turn-complete, valid paired sessions/protocol, 30 tools with six bundled Skill catalog entries, positive cache reads, exact usageTotals cost recomputation, empty .err files and recorded leftovers; current process/port checks clear. No DSH/GPCI/DeferredToolCall effects. Independent checks used scene-local scratch copies and fresh HOME; evidence: review-build-audit.json and review-build-location.txt."
    }
  ]
}
