{
 "model": "gpt-6-1-terra",
 "modelName": "GPT-6.1 Terra",
 "modelId": "gpt-6.1-terra",
 "suite": "Superbench 2.1",
 "promptEdition": "2.1",
 "date": "2026-10-28",
 "status": "integrated",
 "authorization": "The repository owner asked on 2026-10-27 for GPT-6.1 Terra to build the four Superbench 2.1 briefs on the uAssol host using parallel subagents, to publish the results on the site, and to place the model in Tier B of the editorial ranking. On 2026-10-28 the owner pointed out that Superbench 2.1 also includes two returning 3D games, and the two returning briefs were run under their original repository-run protocol.",
 "expectedRuns": 6,
 "completedRuns": 6,
 "protocol": "One isolated Codex CLI 0.163.0 session per frozen brief. All four ran in parallel on uAssol with --model gpt-6.1-terra and model_reasoning_effort=medium. Each session started in its own empty directory, which held only a byte-identical copy of the prompt file (SHA-256 checked against benchmarks/prompts). It received a short fixed wrapper: read ./PROMPT.md, write the requested single HTML document to ./index.html, do not browse, run or test anything, and stop. Session logs show each agent ran exactly two shell commands: one to read PROMPT.md and one to write index.html. The operator copied the artifacts unchanged, added only publication metadata, and ran the browser checks afterwards. Returning briefs (2026-10-28): Ember Glider and Low-Poly Tower Defense keep the original 1.x repository-run protocol. Two parallel Codex CLI 0.163.0 sessions on uAssol (--model gpt-6.1-terra, model_reasoning_effort=medium, workspace-write sandbox with network access for the Three.js CDN). Each ran in its own detached checkout of the repository with the other models' runs for that brief removed. Each received the exact prompt file followed by benchmarks/BENCHMARK-PROMPT.md. Shell tools and headless Chrome were allowed for the Loop/Validation rule. The operator copied each index.html unchanged and finalized it with benchmarks/scripts/run.mjs.",
 "isolation": "Each session's working directory contained only its prompt file. Separation from the rest of the filesystem relied on the instructions and the workspace-write sandbox, not on an access-control jail. The logs show no read of any other file, no network fetch, no MCP call and no browser use. For the two returning briefs, each checkout held the repository but not the other models' artifacts for its brief, and no reference implementation was supplied.",
 "modelVersionEvidence": "Codex CLI reported model: gpt-6.1-terra, provider: openai for every session. OpenAI announced GPT-6.1 Terra on 2026-10-26 and documents the API model ID as gpt-6.1-terra. The Codex model cache on the host lists the same slug with display name GPT-6.1 Terra.",
 "timing": [
  {
   "benchmark": "skyhook-cargo",
   "wallSeconds": 76,
   "tokensReported": 24669,
   "artifactBytes": 17782,
   "artifactSha256": "58a72e77b887277ebea12aefe514ce2eb40ba528e4bc75112365c0d9e41e10f7"
  },
  {
   "benchmark": "concourse-divide",
   "wallSeconds": 94,
   "tokensReported": 18164,
   "artifactBytes": 15511,
   "artifactSha256": "8cb0e3d83cf1f90d057cbbeba938f9064117a1d6aeb11b26c1d52404aebe3977"
  },
  {
   "benchmark": "lightkeepers-atelier",
   "wallSeconds": 72,
   "tokensReported": 18134,
   "artifactBytes": 16500,
   "artifactSha256": "e805d712cf63aa3e9878ee84a1fc04d1aa4f849296c3494a5206bc177ddb4452"
  },
  {
   "benchmark": "wrecklight-salvage",
   "wallSeconds": 67,
   "tokensReported": 28512,
   "artifactBytes": 14442,
   "artifactSha256": "e7cd02df97050459d519243562b1c8b29e7c987cf158cad26231d730aad712ed"
  },
  {
   "benchmark": "ember-glider",
   "wallSeconds": 135,
   "tokensReported": 94178,
   "artifactBytes": 21275,
   "artifactSha256": "83fe9cc09aa1cb7abf11380ad7983cb7832d12439180f01b538cf55bf18f6576"
  },
  {
   "benchmark": "low-poly-tower-defense",
   "wallSeconds": 122,
   "tokensReported": 77730,
   "artifactBytes": 22022,
   "artifactSha256": "e578c23ca53630d50de3a6e927c7ea67d8f74c99ce277d988831ffead91f5285"
  }
 ],
 "review": "The operator loaded each artifact over file:// in headless Chrome 156 at 1280x720 and 390x844. Console errors, page exceptions and outside requests were recorded, and the artifacts were driven with real CDP key, pointer and button events. Each run has its own VALIDATION.md listing the checks that passed, failed or were not observed. No numeric rubric score, rank or pass badge is stored. The two returning 3D games were served over local HTTP in headless Chrome 156 at 1280x800. Each was driven through its window.__glider or window.__defense hook, plus real key and click events, against the prompt's twelve validation checks.",
 "disclosure": "All four artifacts are published exactly as delivered. One brief works: Wrecklight Salvage completed an earned Phase 1 to Phase 3 run ending in a two-second dock settle. Three are failed runs. In Skyhook Cargo the pad is unreachable because the crane's maximum cable leaves the crate 0.11 units above it. In Concourse Divide every traveler starts inside the other team's destination block and the crowd deadlocks on the first tick. In Lightkeeper's Atelier the WebGL program never links because of a uniform precision mismatch. Both returning 3D games work and are published exactly as delivered. Ember Glider finished an earned 14-ring flight at 1:51 against a 3:00 par, with the medal; its off-screen ring indicator ignores rings above or below the frame. Low-Poly Tower Defense finished an earned ten-wave victory with 17 of 20 lives; it has no hover inspection and no muzzle flash. Across the six briefs: three working games and three failed runs.",
 "limits": [
  "New 2.1 briefs: Codex agent authoring, not a hosted-API single-message completion. The model could read a prompt file and write a file, and Codex owns the context and step structure. Returning briefs: full repository-run agent authoring with tools, under the 1.x protocol.",
  "One pass per brief at medium reasoning effort, with no retries, so nothing here measures pass rate or variance. Both returning-game agents stopped before finishing their own twelve-point validation loop.",
  "The only browser available renders through SwiftShader on the CPU, so no frame-rate figure can be claimed.",
  "The Wrecklight Salvage earned run was driven by a scripted controller that read positions through a read-only state probe in a private test copy; the published file is byte-identical to the model output.",
  "The returning-game earned runs were driven by scripted controllers through each prompt's required test hook. The boss leak cost and the reduced-motion camera behaviour were confirmed from the code rather than seen.",
  "The Tier B placement is the owner's editorial routing choice, not a score derived from these six runs."
 ],
 "runSets": [
  {
   "date": "2026-10-27",
   "track": "v21",
   "benchmarks": [
    "skyhook-cargo",
    "concourse-divide",
    "lightkeepers-atelier",
    "wrecklight-salvage"
   ],
   "method": "Four parallel Codex CLI sessions on gpt-6.1-terra, one per frozen Superbench 2.1 brief, each isolated in an empty directory holding only that prompt file and writing one index.html. No browsing, no other files, no browser, no repair turn.",
   "evaluation": "No numeric score or aggregate pass claim is assigned. The per-artifact operator reports record the standalone, browser and behaviour checks: one working run and three failed runs."
  },
  {
   "date": "2026-10-28",
   "track": "core",
   "benchmarks": [
    "ember-glider",
    "low-poly-tower-defense"
   ],
   "method": "Two parallel Codex CLI repository-run sessions on gpt-6.1-terra under the original 1.x protocol (exact prompt plus BENCHMARK-PROMPT.md, with tools, headless Chrome and the Three.js CDN), each finalized with benchmarks/scripts/run.mjs.",
   "evaluation": "No numeric score or aggregate pass claim is assigned. The per-artifact operator reports record twelve checks each: both games work, with the defects listed."
  }
 ],
 "runs": [
  {
   "benchmark": "skyhook-cargo",
   "prompt": "benchmarks/prompts/21-skyhook-cargo.md",
   "entry": "/benchmarks/skyhook-cargo/gpt-6-1-terra/index.html",
   "manifest": "/benchmarks/skyhook-cargo/gpt-6-1-terra/manifest.json",
   "validationReport": "/benchmarks/skyhook-cargo/gpt-6-1-terra/VALIDATION.md",
   "reviewerConclusion": "Boots offline with no console error, page exception or external request. Swing, braking, reversal, cable limits, pause and reset all behave, and the HUD matches the drawn state. Failed run: with the cable fully out the crate still hangs 0.11 units above the pad, so it can never land and the settle ring and delivery can never trigger. A long-cable push against the tower also freezes the simulation until Reset."
  },
  {
   "benchmark": "concourse-divide",
   "prompt": "benchmarks/prompts/22-concourse-divide.md",
   "entry": "/benchmarks/concourse-divide/gpt-6-1-terra/index.html",
   "manifest": "/benchmarks/concourse-divide/gpt-6-1-terra/manifest.json",
   "validationReport": "/benchmarks/concourse-divide/gpt-6-1-terra/VALIDATION.md",
   "reviewerConclusion": "Boots offline with no console error, page exception or external request. It renders 48 travelers, three gates with honest Open, Closing and Closed states, selection with route and waiting reasons, and counts that hold active + arrived = 48. Failed run: each team starts inside the other's destination block, so the planner finds no free destination slot and all 48 travelers wait forever. 0 of 48 arrived in 60 simulation seconds with every gate open."
  },
  {
   "benchmark": "lightkeepers-atelier",
   "prompt": "benchmarks/prompts/23-lightkeepers-atelier.md",
   "entry": "/benchmarks/lightkeepers-atelier/gpt-6-1-terra/index.html",
   "manifest": "/benchmarks/lightkeepers-atelier/gpt-6-1-terra/manifest.json",
   "validationReport": "/benchmarks/lightkeepers-atelier/gpt-6-1-terra/VALIDATION.md",
   "reviewerConclusion": "Failed run. The WebGL program fails to link (Precisions of uniform 'uShadow' differ between VERTEX and FRAGMENT shaders), so the page shows only 'WebGL could not initialize this scene.' No 3D scene, material, light or shadow is drawn, and no rubric item could be observed."
  },
  {
   "benchmark": "wrecklight-salvage",
   "prompt": "benchmarks/prompts/24-wrecklight-salvage.md",
   "entry": "/benchmarks/wrecklight-salvage/gpt-6-1-terra/index.html",
   "manifest": "/benchmarks/wrecklight-salvage/gpt-6-1-terra/manifest.json",
   "validationReport": "/benchmarks/wrecklight-salvage/gpt-6-1-terra/VALIDATION.md",
   "reviewerConclusion": "Boots offline with no console error, page exception or external request. The operator completed the full earned run with real keyboard events. The vault stayed closed with one cell and opened with two, and the reactor attached, released, reattached and towed out. Phase 3 began, the beacon armed the 4 s open and 3 s closed gate cycle, the reactor went through the open gate, and the reactor was detached inside the dock. SALVAGE SECURED appeared after exactly 2.0 s of settle. Review: Phase 3 loads a labelled fixture with no victory. Closed-gate blocking and the doorway linger test were not run separately."
  },
  {
   "benchmark": "ember-glider",
   "prompt": "benchmarks/prompts/15-ember-glider.md",
   "entry": "/benchmarks/ember-glider/gpt-6-1-terra/index.html",
   "manifest": "/benchmarks/ember-glider/gpt-6-1-terra/manifest.json",
   "validationReport": "/benchmarks/ember-glider/gpt-6-1-terra/VALIDATION.md",
   "reviewerConclusion": "Working game. A scripted pilot using only pitch and bank flew the full route: 14 of 14 rings, 0 crashes, 1:51 against a 3:00 par, Ember Medal earned, and the best time persisted. Thermals, glide decay, stall warning, the water crash with the +10s penalty and respawn, the over-par no-medal finish, pause, Retry, Free Flight, safe landing and reduced motion all behave. Partial: the off-screen indicator handles only left and right. Splash and stall feedback are text only."
  },
  {
   "benchmark": "low-poly-tower-defense",
   "prompt": "benchmarks/prompts/14-low-poly-tower-defense.md",
   "entry": "/benchmarks/low-poly-tower-defense/gpt-6-1-terra/index.html",
   "manifest": "/benchmarks/low-poly-tower-defense/gpt-6-1-terra/manifest.json",
   "validationReport": "/benchmarks/low-poly-tower-defense/gpt-6-1-terra/VALIDATION.md",
   "reviewerConclusion": "Working game. A scripted player using the public hook cleared all 10 waves, boss included, and won with 17 of 20 lives. Placement rejection, bounties, 1- and 3-life leaks, cannon splash, the 36% frost slow, both upgrade tiers, the exact 75% sell refund, 2x speed, pause, keyboard-only play, defeat, victory and Restart without reload all behave. Missing: hover inspection and muzzle flash. Balance is gentle."
  }
 ]
}