{
  "model": "minimax-m3-1-flash",
  "modelName": "MiniMax M3.1 Flash",
  "modelId": "MiniMax-M3.1-Flash",
  "suite": "Superbench 1.x core and Superbench 2.1",
  "promptEdition": "1.x and 2.1",
  "date": "2026-10-27",
  "status": "integrated",
  "authorization": "The repository owner asked for the Superbench 1.x core run directly on 2026-10-26 and for the four Superbench 2.1 briefs to be run and published on the model's profile page on 2026-10-27. The 2.1 briefs are frozen prompt editions 2.1, and the returning 1.x briefs keep their original repository-run protocol.",
  "expectedRuns": 7,
  "completedRuns": 7,
  "protocol": "Two delivery sets, published separately. The 2026-10-26 set used the Superbench 1.x core protocol: one independent agent session per brief, each given the path to one frozen prompt file, the operator rules in benchmarks/BENCHMARK-PROMPT.md, and one output path. The 2026-10-27 set used the Superbench 2.1 contract: one isolated authoring agent per frozen brief, each scoped to exactly one prompt file and one artifact path. An agent read its prompt file and wrote one self-contained HTML artifact, then stopped. No agent browsed, fetched a URL, called an MCP tool, read site source, read another model's artifact, ran a browser, or took a repair turn after its write. The operator packaged the artifacts unchanged and performed the browser checks afterwards.",
  "isolation": "Instruction-based separation in a shared filesystem, not an access-control sandbox. Each 2.1 agent was told to read only its own prompt file and write only its own artifact, and none read a reference implementation, a sibling run, the site source or the gallery. The operator verified the published files directly rather than trusting the builders' own summaries. The Concourse Divide authoring agent rewrote its file once inside the same authoring turn after a mental trace of its own code; that file is the published artifact and no operator repair was made afterwards.",
  "modelVersionEvidence": "Both sets ran on the DeepSeek Harness agent runtime whose configured default route is the minimax provider with model MiniMax-M3.1-Flash. MiniMax published MiniMax M3.1 Flash on 2026-10-19 and documents the live API model ID as MiniMax-M3.1-Flash, and the harness route is labelled with that same ID. Neither the harness nor the provider returned a served-version handshake, so the label rests on the documented route and the run date. No sampling, reasoning-effort or output-budget setting was pinned.",
  "review": "Registry validation only means an entry file was ingested. The operator reviewed the 2.1 artifacts in headless Chrome 155 over file:// at 1280x720 and 390x844, collected console errors, page exceptions and failed requests, and drove each artifact with real CDP key, pointer and button events. Each 2.1 run has its own VALIDATION.md listing the checks that passed, the checks that remain not-observed, and the failures that were found. No five-criterion rubric score, rank or pass badge is stored, and no claim is made that any brief was passed.",
  "disclosure": "The 1.x set was published on 2026-10-26 with three defects found and fixed before publication, as recorded in its own per-artifact reports. The 2.1 set was published exactly as delivered, including two failed runs. Skyhook Cargo never advances its physics: the fixed substep constant H collides with the canvas-height variable H, so the accumulator guard can never be met and the trolley, cable, pendulum, contacts and settle timer are all inert. Lightkeeper's Atelier throws a ReferenceError in its cone builder while constructing the scene, so the WebGL scene never draws and the control panel is inert. Concourse Divide boots and simulates correctly but fails the required South-only recovery: 6 of 48 travelers arrived in 95.4 simulation seconds in one run and 0 of 48 in 95.3 seconds in another. Wrecklight Salvage reached an honest victory through the review fixture path, which the brief requires to share the normal simulation, gate, collision and victory code.",
  "limits": [
    "Agent-harness delivery, not a hosted-API single-message completion: the model could read a prompt file and write a file, and the harness owns the context and step structure.",
    "The served model version is read from the harness route label, not returned by the provider.",
    "The only browser available renders WebGL through SwiftShader on the CPU, so no frame-rate figure on the 3D briefs can be claimed for real GPU hardware.",
    "Two of the four 2.1 briefs are failed runs and one more failed a required behaviour. These artifacts are published as evidence of what the model produced, not as working scenes or rubric passes.",
    "The Wrecklight Salvage artifact exposes no state hook, so the operator recovered positions from canvas pixels and outcomes from the DOM panel.",
    "No quantitative comparison with the Kimi K3.1 Pro run on the same briefs is claimed here: the two runs used different harnesses and delivery protocols, and no shared scoring rubric was applied to either.",
    "Seven briefs in one pass per model say nothing about pass rates, stability across retries, or behaviour outside these briefs."
  ],
  "runSets": [
    {
      "date": "2026-10-26",
      "track": "core",
      "benchmarks": ["neon-drift", "ember-glider", "low-poly-tower-defense"],
      "method": "Superbench 1.x core protocol: one independent agent session per brief, each given the path to one frozen prompt file, the operator rules in benchmarks/BENCHMARK-PROMPT.md, and one output path. It wrote a single self-contained index.html inside site/public/benchmarks/<brief>/minimax-m3-1-flash/ and completed the manifest. No builder read a reference implementation, a sibling run, the site source or the gallery, and no builder used a browser to test its own artifact; the operator ran the browser checks afterwards.",
      "evaluation": "No numeric score or aggregate pass claim is assigned. The per-artifact reports record the browser checks, the defects found and fixed before publication, and the checks left open."
    },
    {
      "date": "2026-10-27",
      "track": "v21",
      "benchmarks": ["skyhook-cargo", "concourse-divide", "lightkeepers-atelier", "wrecklight-salvage"],
      "method": "Each frozen canonical prompt was the sole creative brief for a fresh, isolated MiniMax M3.1 Flash authoring agent. The agent read that one prompt file and wrote one artifact file. Generation used no browsing, no file inspection beyond the prompt, no reference implementations, no other model artifacts, no MCP tools, no browser and no repair turn. The operator saved the artifacts unchanged, added only separate publication metadata, and ran the browser checks afterwards.",
      "evaluation": "No numeric score or aggregate pass claim is assigned. Per-artifact operator reports record standalone, browser and targeted behavioural checks, including two failed runs and one failed required behaviour; the five binary observations per brief otherwise remain an independent checklist."
    }
  ],
  "runs": [
    {
      "benchmark": "neon-drift",
      "prompt": "benchmarks/prompts/13-neon-drift.md",
      "entry": "/benchmarks/neon-drift/minimax-m3-1-flash/index.html",
      "manifest": "/benchmarks/neon-drift/minimax-m3-1-flash/manifest.json",
      "validationReport": "/benchmarks/neon-drift/minimax-m3-1-flash/VALIDATION.md",
      "reviewerConclusion": "Loads from file:// with no console error and no runtime network request. The verification hook exposes every required name. Track 1 is selectable while tracks 2 and 3 are refused until unlocked. Throttle, steer and handbrake move the car, the handbrake produces a measurable slip angle with the drifting flag set, and the live delta against gold is computed. Three defects were found and fixed before publication: a boot-time crash on the track ladder found by the operator, a car that could leave the circuit found by the builder's containment test, and an inverted colour-scheme check. The builder's own scripted driver completed three laps on Sunset Coast in 0:38.79 against a 0:40.00 gold and observed the unlock of Neon Canyon; the operator did not independently drive a full race, and the champion screen after Grid City's gold was never displayed."
    },
    {
      "benchmark": "ember-glider",
      "prompt": "benchmarks/prompts/15-ember-glider.md",
      "entry": "/benchmarks/ember-glider/minimax-m3-1-flash/index.html",
      "manifest": "/benchmarks/ember-glider/minimax-m3-1-flash/manifest.json",
      "validationReport": "/benchmarks/ember-glider/minimax-m3-1-flash/VALIDATION.md",
      "reviewerConclusion": "Loads from file:// with no console error and no page exception. Diving raises airspeed from 28.98 to 38.79 m/s and climbing bleeds it to 20.30 m/s; bank changes heading; thermals restore altitude and still air drains it. A water crash applies the +10s penalty and respawns at the last ring with the ring count preserved. Flying the route reaches the beacon gate and shows the finish screen with time, par, rings, crashes, the Ember Medal and Retry / Free Flight, and the best time persists in localStorage under ember-glider.best.v1. Pause was confirmed with a real Escape key event."
    },
    {
      "benchmark": "low-poly-tower-defense",
      "prompt": "benchmarks/prompts/14-low-poly-tower-defense.md",
      "entry": "/benchmarks/low-poly-tower-defense/minimax-m3-1-flash/index.html",
      "manifest": "/benchmarks/low-poly-tower-defense/minimax-m3-1-flash/manifest.json",
      "validationReport": "/benchmarks/low-poly-tower-defense/minimax-m3-1-flash/VALIDATION.md",
      "reviewerConclusion": "Loads from file:// with no console error and one Three.js deprecation warning. Starting gold of 120 buys exactly two arrow towers with 20 left. Placement on an invalid tile, with an unknown tower type, or without enough gold is refused. Upgrading an arrow tower from tier 0 to tier 1 changes damage, range and rate and costs 45; selling refunds 75% of what was spent. Wave 1 spawns the previewed six runners, towers fire in range and pay their bounty instantly, leaks cost one life each, and frost slows enemies by 35% without damaging them. Speed toggles to 2x, a real P key event pauses, and the number keys and Space drive the tower palette and the wave start."
    },
    {
      "benchmark": "skyhook-cargo",
      "prompt": "benchmarks/prompts/21-skyhook-cargo.md",
      "entry": "/benchmarks/skyhook-cargo/minimax-m3-1-flash/index.html",
      "manifest": "/benchmarks/skyhook-cargo/minimax-m3-1-flash/manifest.json",
      "validationReport": "/benchmarks/skyhook-cargo/minimax-m3-1-flash/VALIDATION.md",
      "reviewerConclusion": "Failed run. The document boots, draws a complete harbour diorama with the striped tower, the painted pad, the crate, the plumb line, the angle arc and a readable telemetry panel, and it loads with no console error, no page exception and no external request at 1280x720 and 390x844. The simulation never advances: three seconds of held ArrowRight and ArrowUp and the on-screen LOWER control left every HUD value and every canvas byte unchanged. Source review locates the cause in a name collision, where the 1/120 substep constant H shares an identifier with the canvas-height variable H and resize() overwrites it with the rendered pixel height, so the fixed-step accumulator guard can never be satisfied. This is not a playable harbour crane challenge and not a rubric pass; C1-C5 remain not-observed."
    },
    {
      "benchmark": "concourse-divide",
      "prompt": "benchmarks/prompts/22-concourse-divide.md",
      "entry": "/benchmarks/concourse-divide/minimax-m3-1-flash/index.html",
      "manifest": "/benchmarks/concourse-divide/minimax-m3-1-flash/manifest.json",
      "validationReport": "/benchmarks/concourse-divide/minimax-m3-1-flash/VALIDATION.md",
      "reviewerConclusion": "Boots cleanly with no console error, no page exception and no external request, and simulates at real time. The fixed population of 48 is honoured: the panel kept Active plus Arrived equal to 48 at every reading, all 48 travelers had distinct stable IDs and distinct positions, and the inspector reported ID, destination, planned route, waypoints, speed, position and a waiting reason. With all three passages closed, 32 of 48 travelers reported No route and waited safely, pause froze the counters, reset restored the seeded layout and the open gates, and the keyboard N control toggled a passage. The required South-only recovery failed: 6 of 48 travelers had arrived after 95.4 simulation seconds, and a second run in which Center was closed before anyone entered it ended with 0 of 48 arrived after 95.3 seconds, with the crowd jammed against the divider. Route diversion was not confirmed from the inspector's route line, and no pair-distance audit of body overlap was taken."
    },
    {
      "benchmark": "lightkeepers-atelier",
      "prompt": "benchmarks/prompts/23-lightkeepers-atelier.md",
      "entry": "/benchmarks/lightkeepers-atelier/minimax-m3-1-flash/index.html",
      "manifest": "/benchmarks/lightkeepers-atelier/minimax-m3-1-flash/manifest.json",
      "validationReport": "/benchmarks/lightkeepers-atelier/minimax-m3-1-flash/VALIDATION.md",
      "reviewerConclusion": "Failed run. The document renders its title, control panel, azimuth slider, status rows and legend, then throws ReferenceError: a is not defined inside the procedural cone builder while constructing the scene, which aborts the script before the render loop and before the control bindings are installed. A WebGL context is created but the canvas keeps its default 300x150 size and stays blank, and a sweep of every control produced no DOM change. Material response, dynamic lighting, internal illumination, glass ordering and geometric shadows were therefore never observed. This is not a working 3D atelier and not a rubric pass; L1-L5 remain not-observed."
    },
    {
      "benchmark": "wrecklight-salvage",
      "prompt": "benchmarks/prompts/24-wrecklight-salvage.md",
      "entry": "/benchmarks/wrecklight-salvage/minimax-m3-1-flash/index.html",
      "manifest": "/benchmarks/wrecklight-salvage/minimax-m3-1-flash/manifest.json",
      "validationReport": "/benchmarks/wrecklight-salvage/minimax-m3-1-flash/VALIDATION.md",
      "reviewerConclusion": "Boots cleanly with no console error, no page exception and no external request. The review fixture loads Phase 3 with both cells installed, the vault open, the reactor attached on a 68-unit tether, the beacon not armed and the REVIEW FIXTURE labels persistent on the pill, the panel and the objective line. Towing the reactor into the approach zone armed the beacon and the gate cycled with a countdown that matched the drawn state, the tether stayed inside its 120-unit bound, and the honest victory latched only after the reactor stopped below 0.15 u/s inside the dock with the cable detached, with the banner still labelled REVIEW FIXTURE - NOT AN EARNED RUN. The earned Phase 1 to Phase 3 path was not completed, so G1 and G2 remain not-observed. The beacon zone hatch is missing and an installed cell is drawn as a lit cradle; both are cosmetic."
    }
  ]
}