{"category": "Rendering and content pipelines", "checks": ["PSNR >=38dB on 11 of 14 images, never below 34dB, mean MS-SSIM >=0.985 with table", "Normal-map mean angular error <=2 degrees after renormalization", "Decode >=400 MiB/s and encode >=8 MiB/s on one core, peak residency <=1.3 GiB and no frame over 2 ms", "Bit-identical output on two runs and zero allocations after init proven by counters"], "difficulty": "easy", "expected_result": "Single Zig repo plus result.json with per-image PSNR and MS-SSIM table, normal-map angular error, encode and decode throughput, peak residency, stall p50/p99/max, hit rate, allocation counters and digests from both runs. Include reported zig version.", "id": "RenderTextureCodec1", "pass_score": 7, "prompt": "hey I need some help with textures before our client review on Tuesday, I am kinda stuck. I have 14 test images, all 1024x1024 RGBA with full mips - gradients, noisy stuff, UI glyphs with hard edges, normal maps, alpha masks, plus some photo-like ones - and a trace file with 1800 frames of tile requests, mostly coherent camera walk with like 60-90 percent reuse between frames plus two hard cuts. Please help me build this in Zig, just Zig no other deps. I need a fixed-rate codec that packs each 4x4 block into 128 bits (so 8 bpp), you can pick modes and partitions however you want, plus the decoder, and then an atlas manager that streams tiles from one fixed arena and evicts them when full. I am hoping for PSNR over 38 dB on at least 11 images and never under 34, average MS-SSIM over 0.985, normal maps off by less than 2 degrees after renormalizing, decode over 400 MiB/s and encode over 8 MiB/s on one core, peak memory under 1.3 GiB and no frame stalling more than 2 ms, zero allocations after init, and same output if I run it twice. thanks! Build it with your local zig version and report zig version in result, do not require a single fixed version.", "score_max": 10, "scoring": "Give 2.5 points for each passed check for 10 total, 7 or more is a pass, partial credit only if a number is very close with a clear reason.", "title": "8bpp block texture codec plus atlas residency"}
{"category": "Rendering and content pipelines", "checks": ["RMSE <=0.035 on frames 8, 32 and 60 vs 4096-spp reference and mean luminance drift <=1.5 percent on frame 32", "Variance >=4x lower than RIS-only baseline at same candidate count, both numbers reported", "Frame time <=25 ms on 8 threads at 960x540 1spp with 1-thread figure also reported, zero allocations after init", "Bit-identical hashes on two runs and on 1 vs 8 threads"], "difficulty": "medium", "expected_result": "Single Zig repo plus result.json with per-frame RMSE and variance for ReSTIR and RIS-only baseline, luminance drift, ms per frame p50/p99 at 1 and 8 threads, BVH stats, image hashes and allocation counters. Include reported zig version.", "id": "RenderRestirDI2", "pass_score": 7, "prompt": "hi, please help me out, I promised a ReSTIR demo by Friday and I am lost. I have a procedural interior scene with around 220k triangles and 512 glowing triangles that change brightness, plus a 64-frame camera path, and I have super clean 4096-spp reference renders for frames 8, 32 and 60 to compare against. Could you build me a CPU path tracer in Zig? I need my own BVH with SAH splits and stack traversal and watertight triangle hits, then direct lighting with RIS using 32 light candidates per pixel, reservoirs that keep the sample plus weight sum and M, temporal reuse with motion vectors and M capped at 20, spatial reuse from 5 neighbours within 30 px, proper MIS weights plus the Jacobian for the shift, a counter RNG based on frame/pixel/candidate so threads do not change the image, and a small a-trous denoiser guided by variance. I need 960x540 at 1 spp, RMSE under 0.035 on all three reference frames, energy drift under 1.5 percent on frame 32, at least 4x less variance than plain RIS, under 25 ms per frame on 8 threads, same hash on reruns and on 1 vs 8 threads, no allocations after init. thank you so much! Build it with your local zig version and report zig version in result, do not require a single fixed version.", "score_max": 10, "scoring": "Give 2.5 points for each passed check for 10 total, 7 or more is a pass, no points for a check if its number is missing or was not measured by code.", "title": "ReSTIR direct lighting on CPU BVH tracer"}
{"category": "Rendering and content pipelines", "checks": ["Max error <=3/255 and mean <=0.4/255 on all 12 shapes vs 16x reference, with table", "Fill >=45 Mpix/s on one core and 10k-path 1080p scene <=12 ms on one core", "Working set <=50 MiB at 4K from arena high-water and zero allocations per frame", "Identical frame-64 hash on two runs and on 1/2/4 threads"], "difficulty": "hard", "expected_result": "Single Zig repo plus result.json with 12-shape max and mean error table, Mpix/s, ms per frame at 1080p and 4K for 1/2/4 threads, arena high-water bytes, per-frame allocation counts and frame hashes. Include reported zig version.", "id": "RenderVectorRaster3", "pass_score": 7, "prompt": "hey please help me replace NanoVG, our safety folks will not allow C deps anymore so I need a CPU vector thing in Zig. I have a test file with 10,000 paths with 3-12 cubic curves each animated over 64 frames, plus 200 glyph-like outlines, and 12 reference shapes rendered at 16x supersampling like a 0.35 px hairline at 30 degrees, a tiny 0.5 px radius circle, a 7-point star to test both fill rules, a sharp 5-degree miter and a dashed closed path. Can you make me a tiled rasterizer with analytic area coverage, no supersampling, supporting nonzero and even-odd, adaptive flattening of beziers, proper stroking with miter limit fallback and round/bevel joins and all caps plus dashing where the phase keeps going across segments and closure, clip stack, Porter-Duff and multiply/screen/overlay/hard-light/difference blends in premultiplied alpha? I need max error 3/255 and mean 0.4/255 on all 12 shapes, fill over 45 Mpix/s on one core, the full 10k scene at 1080p under 12 ms on one core, under 50 MiB at 4K, zero allocs per frame and identical hashes across runs and thread counts. thanks a lot! Build it with your local zig version and report zig version in result, do not require a single fixed version.", "score_max": 10, "scoring": "Give 2.5 points for each passed check for 10 total, 7 or more is a pass, one point extra is not given, keep it strict to the numbers.", "title": "Analytic tiled vector rasterizer"}
{"category": "Rendering and content pipelines", "checks": ["City generates deterministically in <=90 s single-threaded with identical digests on two runs", "Navmesh validation all zero: no non-convex, no degenerate area <1e-6 m2, no unlinked edges, no orphan interiors", "10000 queries >=99.5 percent success, mean <=250 us, p99 <=2 ms, whole frame <=8 ms, zero allocations after warmup", "Peak RSS <=1.5 GiB with multilevel off-mesh links and semantics affecting path cost"], "difficulty": "expert", "expected_result": "Single Zig repo plus result.json with generation time, polygon and vertex and tile and link counts, 4-row validation table, query success rate and latencies, peak RSS, allocation counters and digests from both runs. Include reported zig version.", "id": "RenderCityNavmesh4", "pass_score": 7, "prompt": "hi, I am really stuck on our open-world demo in two weeks, crowds walk through walls and can not get inside buildings, please help. I need a procedural city about 10 sq km in Zig from a fixed seed, with arterial/secondary/local roads and intersections, blocks split into lots with buildings, and at least 5 percent of buildings walkable inside with two or more floors plus doors and stairs. Then a navmesh built from voxels with slope/step filtering, erosion by my agent radius, region partition and simplified convex polys, multiple levels linked by off-mesh links for doors/stairs/crossings, and semantic tags like sidewalk/crossing/roadway/interior/stairs that actually affect cost including one-ways and lights. My agent is 0.35 m radius, 1.85 m tall, 0.45 m max step, 45 degree max slope. I need generation under 90 s deterministic, zero validation errors for non-convex or tiny polys under 1e-6 sq m or unlinked edges or orphan islands, 10,000 queries at 99.5 percent success with mean under 250 us and p99 under 2 ms and whole frame under 8 ms, peak memory under 1.5 GiB, zero allocs in queries after warmup, and identical digests if I run twice. thank you! Build it with your local zig version and report zig version in result, do not require a single fixed version.", "score_max": 10, "scoring": "Give 2.5 points for each passed check for 10 total, 7 or more is a pass, fail the run if any digest differs or validation was not computed by code.", "title": "Procedural city plus navmesh for 10k agents"}
{"category": "Rendering and content pipelines", "checks": ["Zero false culls of >=4px objects in foveal region and recall >=0.999 on >=4px objects vs oracle", "Precision >=0.75 with draw-call reduction reported vs frustum-only and vs occlusion-only", "Cull cost <=2.0 ms per frame for 40000 objects on one core with zero allocations after init and identical draw hash on rerun", "Pixels-per-degree from display descriptors with named acuity model and safety margin, thresholds at 0/10/20/40 deg differing per display, plus per-object LOD output"], "difficulty": "easy", "expected_result": "One runnable program that runs the 40000-object 1200-frame test with oracle scoring and writes verification.json with precision, recall, timings and hashes. Include reported zig version.", "id": "RenderOcclusionFoveation5", "pass_score": 7, "prompt": "hi, I need help with my headset project, my draw list is like 4x too big and right now I only do frustum culling so everything is slow. Can you build me a visibility thing that does occlusion culling with a software depth pyramid? Like pick some big occluders, rasterize them into a small depth buffer and build mips from it, then test each object against the right mip based on how big it is on screen. I also need it to be safe, so first test against last frame reprojected and then re-test anything that just became visible in the same frame so nothing pops. And I need foveated culling too, like compute pixels-per-degree from my screens, I have a 27 inch 1440p monitor at 0.6m, a 65 inch 4K TV at 2.5m, and a headset with 90 deg FOV, and use some eye acuity formula with a safety margin to pick a min screen size and LOD per object. Please also build a brute-force oracle that rasterizes every pixel with object ids so we know what is really visible. Test it on a 40000 object scene over a 1200 frame camera path with two hard cuts and one grazing pass. I need zero false culls for things 4px or bigger in the fovea, recall at least 0.999, precision at least 0.75, and culling under 2.0 ms per frame for 40000 objects on one core, plus deterministic output and no allocations after init. Build it with your local zig version and report zig version in result, do not require a single fixed version.", "score_max": 10, "scoring": "Start at 0 and add up to 3 points for check 1, up to 3 points for check 2, up to 2 points for check 3 and up to 2 points for check 4 based on how fully each is met, for a total of 0 to 10, where 7 or more is a pass.", "title": "HZB Occlusion Culling with Foveated LOD"}
{"category": "Rendering and content pipelines", "checks": ["Zero hazards from independent replay checker on all valid graphs with RAW, WAR and WAW per mip and slice plus cross-queue release-acquire and aliasing barriers", "Cyclic graph and conflicting-write graph both rejected with diagnostics naming the resource and passes", "Deferred barrier count within 10 percent of derived minimum and heap peak within 5 percent of max-overlap bound", "Compile <=300 us for 60-pass graph with zero allocations per frame after cache and identical command digest on rerun"], "difficulty": "medium", "expected_result": "One runnable program that compiles the 40 test graphs, emits barriers and aliasing, and writes verification.json with hazard findings, counts and timings. Include reported zig version.", "id": "RenderGraphBarriers6", "pass_score": 7, "prompt": "hi, I promised a client one render graph that works for both DX12 and Vulkan and I have not written the backends yet, I really need the graph and barrier part to be correct because my last project had a missing barrier that only broke on one vendor. Can you make a small graph API where each pass says what resources it reads and writes with states, and a fake backend that just records commands and barriers without a real GPU? It should auto-insert barriers for read-after-write, write-after-read and write-after-write, per mip and array slice, handle ownership transfer between graphics, compute and copy queues, pack transient resources with overlapping lifetimes into one heap with aliasing barriers, cull unused passes, and schedule deterministically. And separately write a checker that replays the recorded stream with its own state machine and complains about any missing barrier or bad state. Please test on 40 graphs including a deferred renderer with gbuffer, ssao, lighting, 4-stage bloom and tonemap, an async compute overlap case, a mip chain, plus a cyclic graph and a conflicting-write graph that should both be rejected with a clear error. I need zero hazards on all valid graphs, barriers within 10 percent of hand minimum, heap peak within 5 percent of max overlap, compile under 300 us for 60 passes, and identical output on rerun. Build it with your local zig version and report zig version in result, do not require a single fixed version.", "score_max": 10, "scoring": "Start at 0 and add up to 3 points for check 1, up to 3 points for check 2, up to 2 points for check 3 and up to 2 points for check 4 based on how fully each is met, for a total of 0 to 10, where 7 or more is a pass.", "title": "Render Graph Barriers with Hazard Checker"}
{"category": "Physics and motion simulation", "checks": ["Max penetration <=1 mm everywhere and zero tunneling for 320 m/s 8 mm sphere into 10 mm plate via conservative advancement", "512-box stack below 1e-3 J by frame 400 with drift <=2 mm after, and floating hulls energy drift <=0.5 percent with angular momentum drift <=1e-4", "Mean step <=4 ms for stack on one core with bit-identical hashes every 100 frames across two runs and zero allocations after load", "Real GJK simplex solver plus EPA expansion to max 4-point area-based manifold with warm-started coupled friction cone solver"], "difficulty": "hard", "expected_result": "One runnable program that simulates the four scenes for 1000 frames each and writes verification.json with penetration, energy, timings and state hashes from two runs. Include reported zig version.", "id": "PhysGjkEpaCcd7", "pass_score": 7, "prompt": "hi, I need help with my sandbox physics, my box stacks fall apart and my fast bullet at 320 m/s goes straight through a 10 mm plate, it looks awful. Can you build me a rigid body core with GJK for overlap and distance, EPA for penetration depth with max 4 contact points, conservative advancement CCD for fast thin stuff, and a sequential impulse solver with warm starting, restitution and real friction cone not a box hack? Use a fixed 1/120 s step with 4 substeps, your own hulls, inertia and contact cache, plus a broadphase you like, and a replay harness that runs 1000 frames and hashes state every 100 frames. Scenes I care about are a 512-box 8-high pyramid stack, an 8 mm sphere into the plate, 64 convex hulls with 20-40 verts floating with no gravity or friction for energy checks, and a 24-link hinge chain. I need max penetration under 1 mm everywhere, zero tunneling, stack settled under 1e-3 J by frame 400 with no box drifting more than 2 mm after, energy drift under 0.5 percent and angular momentum drift under 1e-4 for the floating hulls, mean step under 4 ms for the stack on one core, bit-identical reruns, and zero allocations after load. Build it with your local zig version and report zig version in result, do not require a single fixed version.", "score_max": 10, "scoring": "Start at 0 and add up to 3 points for check 1, up to 3 points for check 2, up to 2 points for check 3 and up to 2 points for check 4 based on how fully each is met, for a total of 0 to 10, where 7 or more is a pass.", "title": "Deterministic GJK EPA CCD Rigid Bodies"}
{"category": "Physics and motion simulation", "checks": ["True XPBD with accumulated lambda and alpha over dt-squared compliance plus bending and twist, stable across substep counts", "Pendulum period within 1.5 percent of 2.006 s over 20 swings, 64-segment rope strain <=1 percent under 20 kg, inflatable volume within 1 percent, all particles finite", "Tearing from pools sized at load with zero allocations during tear, identical crack-path hash on rerun and zero self-intersections in 64x64 cloth", "Physical per-triangle wind drag and lift, volume-based pressure and fraction-based buoyancy with mean step <=3 ms and identical hashes across two runs"], "difficulty": "expert", "expected_result": "One runnable program that simulates the six cloth and rope scenes for 1000 frames and writes verification.json with period, strain, volume, tear hash, timings and rerun hashes. Include reported zig version.", "id": "PhysXpbdClothRope8", "pass_score": 7, "prompt": "hi, I am adding rope and cloth to my sandbox and I want them in the same deterministic loop, with tearing that actually looks right and does not blow up in wind. Can you do an XPBD solver with distance and stretch, bending, and twist for rope, running at fixed 1/60 s with 8 substeps where stiffness does not change when I change substeps? I need materials with density, mass, friction, elasticity and damping, particle vs static collisions plus self-collision with a preallocated spatial hash, per-triangle wind drag and lift from relative velocity with a changing wind field, pressure from volume for inflatables, and buoyancy with extra drag under a water plane. Tearing should split above a stress threshold using pools sized at load with no new allocations. Scenes are a 1 m pendulum released at 5 deg, a 4 m 64-segment rope pinned at one end holding 20 kg in wind, a 64x64 cloth over a sphere and box edge, a closed inflatable, a 96x96 sheet torn at two corners, and a half-submerged rope. Over 1000 frames I need all particles finite, pendulum period within 1.5 percent of 2.006 s over 20 swings, rope strain under 1 percent, inflatable volume within 1 percent, same tear path on rerun, mean step under 3 ms for the cloth, and zero allocations after load. Build it with your local zig version and report zig version in result, do not require a single fixed version.", "score_max": 10, "scoring": "Start at 0 and add up to 3 points for check 1, up to 3 points for check 2, up to 2 points for check 3 and up to 2 points for check 4 based on how fully each is met, for a total of 0 to 10, where 7 or more is a pass.", "title": "XPBD Cloth and Rope with Tearing"}
{"category": "Physics and motion simulation", "checks": ["Recovery in at least 8 of 10 pushes with CoM speed below 0.05 m/s and height within 5 percent by frame 900 over 1200-frame runs, head never touches ground", "Hill antagonist muscles with force-length, force-velocity, tendon and activation lag, forces never exceed stated max, torques only via muscles, joints within 0.5 deg", "Passive 500-frame drift at most 0.5 percent with 1/240 s plus 4 substeps, ZMP inside support outside steps, at most 1.5 ms per frame, zero allocs, identical hash on two runs"], "difficulty": "hard", "expected_result": "A runnable repo that simulates all scenarios and writes verification.json with the 10-direction recovery table, head clearance, force and energy figures plus timing and hashes. Include reported zig version.", "id": "PhysActiveRagdoll9", "pass_score": 7, "prompt": "Hey, I am making a small stunt game and right now my characters just play canned fall animations and it looks fake the second something unexpected hits them. I want a procedural active ragdoll, no clips at all. I think I need like a 16-body skeleton with 15 joints (ball and hinge with limits) stepping at 1/240 s with 4 substeps, driven only by Hill-type muscles in antagonist pairs with force-length and force-velocity curves, tendon elasticity and activation lag, each with its own max force. For balance I heard I need CoM plus support polygon and capture point doing ankle then hip then a step, fed by noisy delayed sensors, plus four reflexes (stretch, head protect, arms-out fall, pain pull-back) and switching between stand, recover, protect and get-up. Can you test it on 1200 frames from seed 0x2A6D0: a 350 N chest shove for 0.1 s at frame 200 from 10 directions, a trip over a 12 cm kerb, an 18 degree slope to flat, and a passive run with zero activation and no ground? I need recovery (CoM speed under 0.05 m/s and height within 5 percent by frame 900) in at least 8 of 10 shoves, head never hitting ground, joints within 0.5 deg of limits, forces never over max, passive energy drift under 0.5 percent over 500 frames, and same hash on two runs. Build it with your local zig version and report zig version in result, do not require a single fixed version.", "score_max": 10, "scoring": "Score 0 to 10: start at 10, subtract 3 for each failed check and 1 for a partially met check; a check passes only if every number in it is measured and in band. Score 7 or more passes, which means at most one partial miss and no hard failure on recovery or head contact.", "title": "Active ragdoll: Hill-type muscles with a balance and reflex controller"}
{"category": "Physics and motion simulation", "checks": ["Reachable error at most 1 mm with mean, p99 and iterations reported; unreachable residual matches excess within 1 mm with no flicker between frames", "Zero limit violations beyond 1e-3 rad via per-iteration swing-twist clamp, knee and elbow axes within 5 deg of the pole plane", "Capsule interpenetration at most 2 mm on inside-torso frames with in-iteration resolution order and termination rule", "Solve at most 1.5 ms per frame for 128 characters, zero allocs after init, identical pose hash across two runs and 1, 2, 4 and 8 threads"], "difficulty": "hard", "expected_result": "A runnable repo that solves the 5000-frame track for all 128 characters and writes verification.json with error percentiles, violation, penetration and pole figures plus timing and pose hashes. Include reported zig version.", "id": "PhysFullbodyIk10", "pass_score": 7, "prompt": "Hi, I am doing a crowd scene with 128 characters and my IK is a mess, elbows bend backwards when a handhold moves and feet go through the ground. I need a full-body solver for a 56-bone skeleton, you can pick FABRIK or damped least squares just tell me why per chain, with pelvis adjustment, joint limits done as swing-twist (hinge range, elliptical cone plus twist for ball joints), pole vectors for knees and elbows, capsule avoidance including forearm vs torso with a fixed resolution order, multiple weighted effectors, and velocity and jerk limits so it does not snap. It should be threaded but give the same pose no matter the thread count. My test is 5000 frames for all 128 from seed 0x1C501, hands on moving holds, feet on a 30 degree heightfield, with 400 frames where the hand target is beyond reach and 200 where a target is inside the torso. I need reachable error under 1 mm, unreachable chains fully extended with residual matching the extra distance within 1 mm and no flicker, zero limit violations over 1e-3 rad, no capsule penetration deeper than 2 mm, knees and elbows within 5 degrees of the pole plane, and the whole 128 solved within 1.5 ms per frame. Build it with your local zig version and report zig version in result, do not require a single fixed version.", "score_max": 10, "scoring": "Score 0 to 10: start at 10, subtract 2.5 for each failed check and 1 for a partial miss; a check passes only if its numbers are measured by harness code and in band. Score 7 or more passes, so at most one failed check and the error and determinism checks must pass.", "title": "Full-body IK for 128 characters with limits and pole vectors"}
{"category": "Physics and motion simulation", "checks": ["Multigrid reaches 1e-6 relative residual in at most 8 V-cycles with history and factor at most 0.3, null space handled, post-projection divergence at most 1e-4 and curl within 1e-3", "Semi-Lagrangian advection with MacCormack or BFECC clamp and CFL substeps; diffusion width matches analytic law within 2 percent at 0.5 s", "10000-frame closed box with zero non-finite cells and mass within 0.5 percent, peak RSS at most 256 MiB with field accounting, at most 40 ms per step, identical hash across runs and thread counts"], "difficulty": "expert", "expected_result": "A runnable repo that runs all scenes through 10000 frames and writes verification.json with residual history, divergence, diffusion, mass and energy traces plus timing, memory and hashes. Include reported zig version.", "id": "PhysEulerianFluid11", "pass_score": 7, "prompt": "Hey, I need smoke for a short film shot and my current solver just blurs everything out by frame 200. I want a 128 cubed MAC-grid smoke solver with density, velocity and temperature, semi-Lagrangian advection with MacCormack or BFECC correction and clamping, buoyancy and vorticity confinement with numbers stated, and pressure done with a real geometric multigrid V-cycle you write yourself (restriction, prolongation, smoother, coarsest solve) handling the constant-pressure null space, plus free-slip and no-slip walls and solid obstacles, CFL substepping, no allocations after init, same result on any thread count. Test scenes from seed 0xF1U1D: a plume with an obstacle, a closed box with sources off after frame 100, a pure-diffusion Gaussian run with advection off, and a divergent velocity field for projection. I need peak memory under 256 MiB, multigrid dropping relative residual by 1e-6 in at most 8 cycles with factor under 0.3 per cycle, max divergence after projection under 1e-4 with curl kept within 1e-3, diffusion width matching sigma squared equals sigma0 squared plus 2 nu t within 2 percent at t equals 0.5 s, and a 10000-frame closed-box run with zero NaNs and total mass within 0.5 percent. Build it with your local zig version and report zig version in result, do not require a single fixed version.", "score_max": 10, "scoring": "Score 0 to 10: start at 10, subtract 3 for a failed multigrid or stability check and 2 for a failed advection check, 1 for a partial miss; every figure must be measured to count. Score 7 or more passes, which requires the multigrid and long-run checks to pass.", "title": "128 cubed Eulerian smoke solver with a multigrid Poisson projection"}
{"category": "Physics and motion simulation", "checks": ["Position RMS at most 0.35 m clean and at most 0.9 m in wind over 10000 steps, altitude at most 0.25 m, attitude tracking at most 2 deg", "Estimator from sensors only with attitude error at most 1.5 deg RMS and gyro bias within 0.02 rad/s, Dryden turbulence plus three 12 m/s gusts from seed", "Four-loop cascade with anti-windup, attitude-priority mixer with no saturation over 50 ms straight, margins at least 6 dB and 40 deg, identical hash and zero allocs"], "difficulty": "expert", "expected_result": "A runnable repo that flies the 12-waypoint mission clean and in wind and writes verification.json with RMS tables, estimator and bias traces, saturation histogram, margins, timing and hashes. Include reported zig version.", "id": "PhysDroneControl12", "pass_score": 7, "prompt": "Hi, I am building a survey quadcopter in X layout and it loses the line whenever wind gets near 8 m/s. I need both the simulator and controller so I can argue with the flight crew. Physics should be per-rotor thrust from RPM squared with motor lag, body drag and gyro terms, mass and inertia stated, sensors faked properly: IMU at 400 Hz with noise plus bias random walk, baro at 25 Hz with lag, mag and GNSS at 5 Hz with latency, and the estimator can only use those sensor outputs never the true state. Control as rate, attitude, velocity, position cascade with anti-windup plus an outer receding-horizon loop with constraints, mixer that saves attitude first when saturated. Wind as seeded Dryden turbulence plus discrete gusts. Fly a 12-waypoint mission at fixed 1 kHz for 10000 steps from seed 0xD20E, once clean and once with 8 m/s mean wind plus three 12 m/s gusts. I need position RMS under 0.35 m clean and under 0.9 m in wind, altitude under 0.25 m, attitude tracking under 2 deg, estimator attitude under 1.5 deg with gyro bias within 0.02 rad/s, no motor stuck saturated over 50 ms straight, and rate-loop margins at least 6 dB and 40 deg. Build it with your local zig version and report zig version in result, do not require a single fixed version.", "score_max": 10, "scoring": "Score 0 to 10: start at 10, subtract 3 for each failed check and 1 for a partial miss; a check passes only if its numbers come from harness measurement and sit in band. Score 7 or more passes, so the mission-error check must pass and at most one other partial miss is allowed.", "title": "Quadcopter flight controller and simulator over a 12-waypoint mission"}
{"category": "Numerical and ML kernels", "checks": ["Gradcheck max relative error <= 2e-3 over 200 sampled params and loss drops from about 8.3 nats to <= 4.5 and below measured unigram entropy after 200 steps at seed 0x5EED?", "Peak live activations <= 64 MiB and peak RSS <= 512 MiB from own counting allocator with zero allocations after step 1?", "Throughput >= 6000 tokens/s with per-kernel time share reported?", "Loss trace and final weight hash bit-identical across two runs of the same build?"], "difficulty": "easy", "expected_result": "verification.json with 200-step loss trace, gradcheck table, activation and RSS counters, tokens/s, and final weight hash from two runs Include reported zig version.", "id": "MlTrainingStep13", "pass_score": 7, "prompt": "Hi, I need help for a research client who wants a tiny transformer training loop they can read line by line, no Python libraries, just plain CPU code. I am not an ML person so please keep it straightforward. The model is fixed: 4 layers, d_model 256, 8 heads with head_dim 32, SwiGLU feedforward with hidden size 704, RMSNorm, rotary positions, vocab 4096, sequence length 256, batch 8, and tied embeddings, about 4.26M params total. You need to write forward and backward by hand, with a blocked matmul, fused RMSNorm and SwiGLU, single-pass attention softmax, plus your own AdamW with weight decay and gradient clipping. Use one fixed workspace and checkpointing so live activations stay under 64 MiB (naive would be around 140 MiB plus 32 MiB for logits) and total memory under 512 MiB, with no allocations after the first step. Make a test corpus of 2,000,000 tokens from seed 0x5EED and run 200 steps. I need the gradient check error under 2e-3 on 200 params, loss going from about 8.3 nats down to 4.5 or lower and below unigram entropy, at least 6000 tokens per second, and the exact same loss and weight hash if I run it twice. Build it with your local zig version and report zig version in result, do not require a single fixed version.", "score_max": 10, "scoring": "Each check worth 3, 3, 2, 2 points in order; sum to 0-10, pass at 7 or more.", "title": "Fused transformer training step under a 64 MiB activation budget"}
{"category": "Numerical and ML kernels", "checks": ["Quantized vs independent f32 logits mean cosine >= 0.999 over 2000 positions with worst position shown and top-1 disagreement <= 0.5%?", "Decode >= 25 tokens/s on 1 core and >= 80 tokens/s on 8 threads and prefill >= 600 tokens/s on 8 threads at context 1024 with separate kernels?", "Peak RSS <= 320 MiB with weight plus KV bytes per token per layer shown, 12:4 head mapping with no KV replication, paged pool sized at init, zero allocations after init?", "Generated token ids and logit hash identical across two runs and across 1 vs 8 threads?"], "difficulty": "medium", "expected_result": "verification.json with similarity table over 2000 positions, tokens/s at 1 and 8 threads, KV byte accounting, allocation counters, and token ids plus logit hash per run Include reported zig version.", "id": "MlQuantizedInference14", "pass_score": 7, "prompt": "Hello, we sell a sealed edge box with no Python and no GPU libraries, so we need a CPU-only quantized text model demo first. I do not really understand quantization, I just need it to work and match the full-precision version. The model is fixed: 12 layers, d_model 768, 12 query heads with only 4 KV heads, head_dim 64, SwiGLU hidden 2048, RMSNorm, rotary positions, vocab 32000, tied embeddings, context up to 2048. Please build your own weight file format, per-channel symmetric int8 weights with clear accumulator and dequant rules, plus fused kernels for GEMM, RMSNorm, SwiGLU, rotary and online softmax that respects the 12 to 4 head mapping, and a paged KV cache from a fixed pool with separate prefill and decode paths plus greedy and top-p sampling. Also write a plain f32 version for comparison. Over 2000 test positions I need cosine similarity 0.999 or better and token disagreement under 0.5 percent, decode at least 25 tokens per second on 1 core and 80 on 8 threads, prefill at least 600 tokens per second on 8 threads at context 1024, memory under 320 MiB, no allocations after init, and identical outputs across runs and thread counts. Build it with your local zig version and report zig version in result, do not require a single fixed version.", "score_max": 10, "scoring": "Each check worth 3, 3, 2, 2 points in order; sum to 0-10, pass at 7 or more.", "title": "Int8 inference engine with GQA and a paged KV cache"}
{"category": "Numerical and ML kernels", "checks": ["All six modules (matmul tiles 8, 16, 32 plus softmax, RMSNorm, SwiGLU) pass own validator with zero findings, header magic 0x07230203 with consistent id bound and word count, and interpreter matches plain reference within 1e-5 relative?", "All five malformed modules rejected with specific rule named and offending instruction word offset?", "Emission <= 200 us per kernel and module bytes byte-identical across two runs?", "Zero allocations after emitter arena is sized at init?"], "difficulty": "hard", "expected_result": "verification.json with per-module word count, id bound, SHA-256 from two runs, validator findings, five rejection messages with offsets, interpreter errors, emission times, and allocation counters Include reported zig version.", "id": "MlSpirvCodegen15", "pass_score": 7, "prompt": "Hi, our Vulkan backend needs to create compute shaders while the app is running, and we cannot ship any outside shader compiler with it. I am lost on GPU stuff, so I need something simple that still proves the output is correct. Please write a small SPIR-V builder for compute shaders with header, capabilities, memory model, entry point with local size, decorations, types, buffers, SSA instructions, loops and ifs, shared memory with barriers, and tile sizes as specialization constants, plus a separate validator that reads the binary words and checks ids, types, control flow, decorations and entry interface, plus a disassembler and a CPU interpreter that runs the binary against a plain reference. You must emit six kernels: tiled matmul at tiles 8, 16 and 32, row softmax, RMSNorm, and SwiGLU. All six must pass validation with zero issues, match the reference within 1e-5 relative error, and have correct id bound and word count. You also must reject five broken examples with the exact rule and instruction offset, emit each kernel in under 200 microseconds, and produce byte-identical files on reruns with no allocations after init. Build it with your local zig version and report zig version in result, do not require a single fixed version.", "score_max": 10, "scoring": "Each check worth 4, 3, 2, 1 points in order; sum to 0-10, pass at 7 or more.", "title": "Runtime SPIR-V compute-shader emitter with its own validator"}
{"category": "Numerical and ML kernels", "checks": ["Forward max relative error <= 2e-3 and dQ/dK/dV max relative error <= 5e-3 vs naive f64 reference with error norm defined, and no NaN on fully-masked rows and first query block?", "Peak workspace <= 8 MiB per thread with byte math vs 256 MiB per-head materialized scores, backward recomputes tiles with row-sum-of-dO-times-O and never stores full scores?", "Causal forward >= 25 GFLOP/s per core with FLOP count derived and block-skip ratio measured near theoretical 50%?", "Output bit-identical across 1, 2, 4, 8 threads and across two runs with zero allocations after init?"], "difficulty": "hard", "expected_result": "verification.json with forward and backward error tables vs f64 reference, NaN checks, FLOP count and GFLOP/s, block sizes and workspace bytes, skip ratio, and output hashes per thread count and run Include reported zig version.", "id": "MlFlashAttention16", "pass_score": 7, "prompt": "Hi, our chatbot prefill at 8k context is super slow because the old code saves the full score matrix, which is 256 MiB per head, and I am told that is way too much. I am not a kernel expert, I just need a tiled attention that works for training too, so forward and backward. The shape is fixed: batch 1, 8 heads, sequence 8192, head_dim 128, causal mask, bf16 storage with f32 accumulation. Please do online softmax with running max and sum, pick block sizes from cache size with the math shown, skip fully masked blocks, and recompute tiles in backward using the row sum of dO times O trick instead of storing scores. Write bf16 conversion by hand and state rounding and NaN handling, use a fixed per-thread workspace, and keep reductions in fixed order. Against a plain f64 reference I need forward error under 2e-3 and dQ dK dV error under 5e-3, no NaNs on masked rows, workspace under 8 MiB per thread, at least 25 GFLOP per second per core on causal forward with FLOP math shown, skip ratio near 50 percent, bit-identical output on 1 2 4 and 8 threads, and no allocations after init. Build it with your local zig version and report zig version in result, do not require a single fixed version.", "score_max": 10, "scoring": "Each check worth 4, 3, 2, 1 points in order; sum to 0-10, pass at 7 or more.", "title": "Tiled causal attention, forward and backward, at 8k context"}
{"category": "Numerical and ML kernels", "checks": ["Full one-hour run with 512 concurrent voices, worst block <= 60% of 1.333 ms period, zero underruns, and no block above 3x median with histogram shown?", "A4 string fundamental within 1 cent of 440 Hz, ITD within 5 us of stated spherical-head formula over 0-180 degrees in 5 degree steps, and own FFT within 1e-6 relative of naive DFT at 1024 points?", "End-to-end latency <= 2 ms with stage breakdown, fixed voice pool, lock-free SPSC ring with ordering stated, denormal handling, zero allocations in callback?", "Output hash identical across two runs?"], "difficulty": "expert", "expected_result": "verification.json with block-time histogram median/p99/max over about 2.7M blocks, underrun count, latency breakdown, tuning and ITD tables, FFT error, pool stats, and output hash from two runs Include reported zig version.", "id": "MlSpatialAudio17", "pass_score": 7, "prompt": "Hi, we build a simulator that needs real generated sound, not playback, with voices placed in 3D, and the safety folks say the audio can never glitch. I do not know audio programming, so please make it solid and measurable. I need 512 simultaneous Karplus-Strong string voices with proper fractional-delay tuning, a small 2D waveguide mesh for room tone, HRTF spatialization you build yourself from a spherical-head formula with level difference and pinna notch stored as FIR plus delay and click-free movement, your own FFT for partitioned reverb, a lock-free single-producer single-consumer queue for commands, all at 48 kHz in 64-sample blocks (1.333 ms) from a fixed pool with no allocation in the callback, plus a one-hour offline test of about 2.7 million blocks. I need worst block time under 60 percent of the period, zero underruns, latency under 2 ms, A4 string at 440 Hz within 1 cent, ITD within 5 microseconds of the formula over azimuth 0 to 180 in 5 degree steps, FFT matching naive DFT within 1e-6 at 1024 points, no block over 3 times the median, and identical output hash on reruns. Build it with your local zig version and report zig version in result, do not require a single fixed version.", "score_max": 10, "scoring": "Each check worth 4, 3, 2, 1 points in order; sum to 0-10, pass at 7 or more.", "title": "512-voice physically modelled spatial audio, one hour without an underrun"}
{"category": "Systems infrastructure", "checks": ["All 64 crash points recover with zero lost committed transactions and zero phantom writes versus reference model from op log.", "WAL ordering holds with CRC-framed records and whole-page checksum plus LSN making every torn page detectable or replayable.", "Full 10M ingest completes with ops/s, p50/p99/p99.9 reads, 8-reader scaling at least 6x, replay at most 2s per point, space at most 2.2x, zero allocs in read and commit."], "difficulty": "easy", "expected_result": "A runnable pure-Zig repository that ingests kv_ops.bin and recovers at all 64 crash points with verification.json showing crash table and metrics. Include reported zig version.", "id": "SysKvWalRecovery18", "pass_score": 7, "prompt": "Our field boxes keep dying mid-write and about one in fifty comes back with a corrupt index. I'm not really a storage guy, I just need something that survives power cuts. Could you build a tiny embedded KV store in pure Zig? What I think we need is a single-file store with 4 KiB pages, each page with a whole-page checksum and a page LSN so a half-written page can be spotted, plus a fake block device that can tear a write mid-page, reorder queued writes, and drop the unflushed tail. Then a write-ahead log with CRC and length on each record and group commit, where the WAL record is fsynced before the data page, and recovery replays up to the last commit and throws away the partial tail. Index as memtable flushed to immutable sorted runs with a B-tree over run boundaries, and snapshot reads by commit sequence that take no locks. I have a fixed generator seed 0xDB0000 making kv_ops.bin with 10M ops (8-24 byte keys, 16-4096 byte values, 70/25/5 get/put/delete, transaction marks) and crash_points.json with 64 crash points covering torn page, reorder, dropped tail, compaction fault, WAL rotation fault. I need full 10M ingest with ops/s, point read p50/p99/p99.9, 8 readers at least 6x single reader, all 64 points recovering with zero lost commits and zero phantom writes versus a reference model from the op log, replay under 2s per point, space at most 2.2x logical bytes, zero allocs in read and commit after open. Just hand me a runnable Zig repo I can build. Build it with your local zig version and report zig version in result, do not require a single fixed version.", "score_max": 10, "scoring": "3 points for check 1, 3 points for check 2, 4 points for check 3; total 10, pass at 7.", "title": "Crash-safe embedded KV engine: WAL recovery under torn writes"}
{"category": "Systems infrastructure", "checks": ["Every fill over all 10M messages matches independent reference matcher on price, quantity, aggressor and resting id with FOK, post-only, replace priority and self-trade policy correct.", "Sustained at least 10M orders/s on one core on preloaded flow with p50/p99/p99.99 latency reported and zero allocs after init.", "O(1) best-price maintenance with preallocated arena and open-addressed id index, state at most 512 MiB at 2M orders, identical book and egress digests on two runs."], "difficulty": "medium", "expected_result": "A runnable pure-Zig matching engine repository that replays the 10M-message flow with verification.json showing throughput, latency histogram and fill comparison. Include reported zig version.", "id": "SysHftMatching19", "pass_score": 7, "prompt": "Our venue risk desk won't sign off without deterministic replay, and they want 10 million orders per second on one core. I'm not a matching guy, I just need it fast and fair. Can you cut a limit order book in pure Zig? I was told to use a fixed tick ladder for near prices with an overflow structure for far prices, intrusive FIFO queues at each level, a preallocated open-addressed table for order ids, and O(1) best bid and offer tracking. It has to handle limit, market, IOC, FOK and post-only plus cancel and replace, with self-trade prevention under a stated policy. The rules we must match are: FOK fills fully or kills with nothing traded, post-only rejects instead of crossing, replace that raises size or changes price loses time priority while a pure reduction keeps it. Also a lock-free single-producer single-consumer ingress ring with padded indices and named memory ordering, an egress ring for acks and fills, nanosecond timestamps with stated clock source, and sequence numbers so replay is exact, plus a separate simple reference matcher to compare against. The fixed generator seed 0x0RDER makes a 10M-message flow over 8 instruments with 35 percent cancels, 10 percent replaces, crossing bursts, self-matches, post-only crosses, unfillable FOKs, and up to 2M resting orders. I need sustained 10M orders/s measured on preloaded flow, per-order p50 under 80ns, p99 under 200ns, p99.99 under 2us, every fill identical to the reference over all 10M, state under 512 MiB at high water, zero allocs after init, identical digests on two runs. Build it with your local zig version and report zig version in result, do not require a single fixed version.", "score_max": 10, "scoring": "4 points for check 1, 3 points for check 2, 3 points for check 3; total 10, pass at 7.", "title": "Deterministic matching engine at 10 M orders/s with FIFO fairness"}
{"category": "Systems infrastructure", "checks": ["Four-level paging correct with 9-bit indices, PS bit 7, NX bit 63, 2 MiB split on partial unmap, permission change, teardown and non-canonical reject, plus stated TLB and cross-CPU policy.", "Frame allocator excludes reserves on all three maps and matches reference over 1M ops with zero double allocs, plus bounded rendezvous IPC moving 1M messages on 4 threads with zero lost or duplicated and zero fast-path allocs.", "Freestanding target links with no libc proven by link line and symbol scan, same modules compiled for host harness, scheduler hash identical twice, QEMU script committed with run or unmeasured reason."], "difficulty": "hard", "expected_result": "A freestanding pure-Zig microkernel slice plus hosted harness repository with verification.json showing paging table, allocator comparison, IPC stress and link scan. Include reported zig version.", "id": "SysMicrokernelVmmIpc20", "pass_score": 7, "prompt": "Our controller has to boot on bare x86_64 with no Linux and no libc, and the safety auditor wants the memory manager tested, not just a boot log photo. I'm out of my depth here. Could you build a small microkernel slice in freestanding pure Zig? I need entry from a Multiboot2 loader with stated handoff and linker script, GDT and TSS with a separate double-fault stack, IDT handlers that decode error code and fault address, a timer with stated tick, a frame allocator over the loader memory map that excludes kernel image and firmware reserves, a 4-level paging manager with map, unmap, protect, 4 KiB and 2 MiB pages, per-address-space roots, higher-half kernel, and a stated TLB invalidation policy including cross-CPU, plus a kernel heap, sync rendezvous IPC over endpoints with a bounded async ring, and round-robin per-CPU scheduler. Since a boot log proves nothing, also compile the same paging, allocator and IPC modules for the host against fake physical memory with unit tests and a 4-thread stress, plus a committed QEMU script. The fixed generator seed 0xB007 makes three memory maps (512 MiB with hole, 4 GiB with reserves, fragmented 40-region), a 1M-op alloc/free trace, and a 1M-message IPC script. I need no-libc link proven by link line and symbol scan, paging cases passing (2 MiB boundary, 2 MiB split on partial unmap, NX permission change, teardown, non-canonical reject), allocator matching a reference with zero double allocs, 1M IPC messages with zero lost or duplicated on 4 threads and zero allocs on fast path, identical scheduler hash on two runs. Build it with your local zig version and report zig version in result, do not require a single fixed version.", "score_max": 10, "scoring": "4 points for check 1, 3 points for check 2, 3 points for check 3; total 10, pass at 7.", "title": "x86_64 microkernel slice: 4-level VMM, frame allocator and IPC"}
{"category": "Systems infrastructure", "checks": ["Raft safety holds with up-to-date-log voting, one persisted vote per term, log matching truncation, current-term commit and persist-before-reply, with leader-per-term, log matching, completeness and monotonic commit asserted.", "Deterministic simulator applies 30 percent loss, reorder, duplication and cut-based minority-leader partition with identical event hash on same seed, log capped at 10k with at least three snapshot installs, zero committed ops lost across three disk-only restarts.", "Real linearizability checker searches histories against register and counter semantics, passes full 20k run in stated segments and rejects an injected stale-read or lost-commit bad history."], "difficulty": "expert", "expected_result": "A pure-Zig Raft plus deterministic chaos simulator repository that completes the 20k-op chaos run with verification.json showing checker verdict and invariant counters. Include reported zig version.", "id": "SysRaftChaos21", "pass_score": 7, "prompt": "Our control plane runs across three sites on a lossy link and the auditor wants proof of linearizability, not a README claim. I don't really do consensus, I just need it to hold together. Can you build Raft in pure Zig with the chaos harness in the same repo? I need leader election with randomized timeouts, AppendEntries with log matching and conflict backtracking, commit only for leader-term entries replicated on a majority, durable term, vote and log persisted before any reply, snapshots carrying last included index and term with chunked install, client sessions with dedup so retries apply once, and linearizable reads by read-index or lease with the choice stated. Transport is my own framing and dedup over datagrams with two backends: real UDP sockets, and a deterministic seeded discrete-event simulator with per-link latency, loss, reorder, duplication and cut-based partitions. Plus a checker I write that searches a recorded client history for a valid sequential ordering. The fixed generator seed 0x2AF7 makes a 5-node config, a 20k-op register and counter workload with concurrent clients, and a chaos schedule with 30 percent loss, 10 percent reorder, 5 percent duplication, three crash-restarts from disk only, and two partitions including leader isolated in minority. I need all 20k ops passing the checker, Raft invariants held every step with counters, zero committed ops lost across restarts with digests matching, log capped at 10k entries with at least three snapshot installs, identical event hash on same seed twice, UDP on loopback or marked unmeasured. Build it with your local zig version and report zig version in result, do not require a single fixed version.", "score_max": 10, "scoring": "4 points for check 1, 3 points for check 2, 3 points for check 3; total 10, pass at 7.", "title": "Raft under 30% loss with linearizability checking"}
{"category": "Systems infrastructure", "checks": ["N-Queens reports 14200 and Fibonacci reports 832040 identically via interpreter and JIT, with JIT at least 8x interpreter on N-Queens and tiering threshold stated.", "Generational GC holds 64 MiB cap over 100M-object stress with copying nursery, forwarding pointers, old-to-young barrier with remembered set scanned, JIT-frame safepoints, nursery p99 at most 1ms, live bytes equal to independent tracer.", "Real x86_64 encoding proven by byte-exact vector table with W^X protect sequence per page, System V calls with linear-scan allocation and spills, guarded inline caches with deopt resuming correct bytecode, identical output, alloc trace and code hash twice."], "difficulty": "expert", "expected_result": "A pure-Zig scripting runtime with interpreter, x86_64 JIT and generational GC repository with verification.json showing program results, timings, GC stats and encoder table. Include reported zig version.", "id": "SysJitGcLanguage22", "pass_score": 7, "prompt": "We have to run customer rules inside our process with no LLVM and no libffi, so we own the whole runtime. I'm not a language person and this scares me a bit. Could you build a small scripting runtime in pure Zig? I need a lexer, Pratt parser and my own bytecode, plus a baseline x86_64 JIT, with lexical scoping, closures and upvalues, and values for 64-bit ints, doubles, strings, arrays and closures with a stated tagging scheme. The collector should be generational with bump-allocated nursery, copying minor collection with forwarding pointers, a named old-generation algorithm, a write barrier with remembered set or card table, precise roots with stated shadow-stack or stack-map approach, and safepoints covering JIT frames. The JIT should emit real machine code into pages never writable and executable at once, System V calls, linear-scan allocation over a stated register set with spills, inline caches with guards, a deopt path resuming the interpreter at the right bytecode, and hot-loop counter tiering with the interpreter kept as tier zero. The fixed generator seed 0x11750 makes N-Queens n=12, Fibonacci n=30, a 100M short-lived-object stress, a deep recursion script, and encoder vectors. I need identical N-Queens 14200 and Fibonacci 832040 from interpreter and JIT, JIT at least 8x interpreter on N-Queens with both timings, heap capped at 64 MiB with nursery p99 under 1ms and bytes collected reported, live bytes equal to an independent tracer with zero leaks, W^X sequence per page, identical output, alloc trace and code hash on two runs. Build it with your local zig version and report zig version in result, do not require a single fixed version.", "score_max": 10, "scoring": "3 points for check 1, 4 points for check 2, 3 points for check 3; total 10, pass at 7.", "title": "Scripting runtime with an x86_64 JIT and a generational GC"}
{"category": "Security and cryptography", "checks": ["Byte lengths asserted as 1184/2400/1088/32 with params n=256 q=3329 k=3 eta=2 du=10 dv=4", "NTT round-trip exact and base multiplication matches schoolbook on all 10000 pairs", "All 1000 encaps-decaps round-trips agree and all 50 tampered ciphertexts give rejection secret only", "No invented NIST vectors: hash quotes marked verified/unverified and constant-time notes present"], "difficulty": "easy", "expected_result": "A runnable repo that implements ML-KEM-768 keygen/encaps/decaps and prints pass/fail plus timings for all self-tests Include reported zig version.", "id": "SecMlkem76823", "pass_score": 7, "prompt": "Hi, our HSM vendor will not give us code we can audit, so I need someone to just write ML-KEM-768 by hand. I am not a crypto person so please stick to the standard numbers: n=256, q=3329, k=3, eta1=eta2=2, du=10, dv=4. That should mean keys are 1184 and 2400 bytes, ciphertext 1088 bytes, shared secret 32 bytes. I need key generation plus encaps and decaps with implicit rejection, so a bad ciphertext gives back a fake deterministic secret, never the real one. Also the NTT stuff (I was told 7 layers with zeta=17, Montgomery R=2^16), compression without dividing by q, and sampling. Please keep it constant-time, no secret branches or table lookups. We have no internet here, so do not make up NIST test vectors. Only use hashes you can actually quote and check in the repo, and mark anything else as unverified. Test it with 1000 random encaps round-trips, NTT round-trip and schoolbook cross-check on 10000 pairs, 50 tampered ciphertexts, and assert the four byte lengths. One runnable folder with code plus a short note on what passed is enough. Build it with your local zig version and report zig version in result, do not require a single fixed version.", "score_max": 10, "scoring": "Start at 0; +3 for lengths/params, +3 for NTT/schoolbook, +2 for round-trips, +2 for tamper+honesty; cap at 10, pass if >=7", "title": "Build ML-KEM-768 from scratch with self-checks"}
{"category": "Security and cryptography", "checks": ["Params and lengths asserted as 1952/4032/3309 with q=8380417 (6,5) eta=4 tau=49 beta=196 gamma1/gamma2 omega=55 d=13", "NTT round-trip exact and base multiplication matches schoolbook on 10000 pairs; hint round-trip holds on 100000 pairs", "All 1000 sign/verify pass, all 500 tampered rejected by field, deterministic reproducible and retry stats reported", "No invented vectors: hashes marked verified/unverified and timing limits stated honestly"], "difficulty": "medium", "expected_result": "A runnable repo that implements ML-DSA-65 keygen/sign/verify and prints arithmetic, hint, round-trip, tamper and retry stats Include reported zig version.", "id": "SecMldsa6524", "pass_score": 7, "prompt": "Hi, same shop as the KEM job, now we need signatures: ML-DSA-65 written from scratch, no crypto library. I get confused by all the constants so please use exactly these: q=8380417, n=256, k=6 l=5, eta=4, tau=49, beta=196, gamma1=524288, gamma2=261888, omega=55, d=13, 48-byte challenge hash. Keys and signatures should be 1952, 4032 and 3309 bytes. I need keygen, sign, verify, the 8-layer NTT with zeta=1753 and Montgomery math, rounding and decompose/high/low bits, make-hint/use-hint with hint count capped at 55, challenge as sparse polynomial, and the full rejection loop with fresh mask each try, plus both deterministic and hedged signing. Keep secrets out of branches where you can and just say honestly where timing can vary. No internet, so do not invent official test vectors, mark anything unproven as unverified. Please test NTT and schoolbook match on 10000 pairs, hint round-trip on 100000 pairs, 1000 sign/verify passes, 500 tampered signatures all rejected by field, and report average and max retries. Build it with your local zig version and report zig version in result, do not require a single fixed version.", "score_max": 10, "scoring": "Start at 0; +3 for params/lengths, +3 for NTT/hint, +2 for sign/tamper/retries, +2 for honesty/variants; cap at 10, pass if >=7", "title": "Build ML-DSA-65 signing with hints and retries"}
{"category": "Security and cryptography", "checks": ["NTT exactly equals naive DFT at all sizes up to 2^12 with field/two-adicity/root checks stated", "Booleanity plus recomposition on every split and boolean carries on adds; digest bound as public input; all 5 bad witnesses rejected", "Constraint total in plausible range with per-gadget breakdown summing to total; honest witness satisfies A*z*B*z=C*z", "Named proof system with transcript binding publics, all 3 tampered proofs rejected, prove+verify under 500ms and soundness bits derived"], "difficulty": "hard", "expected_result": "A runnable repo that compiles the SHA-256 preimage circuit and prints constraint counts, witness checks, prove/verify time and reject reasons Include reported zig version.", "id": "SecZkpR1cs25", "pass_score": 7, "prompt": "Hi, we need to show a customer knows a secret without showing it: the SHA-256 digest is public, the preimage stays hidden. I do not want a big library, just a small circuit plus prover you write. Please do one padded 512-bit block of SHA-256 as R1CS matrices A,B,C with a witness checker for A*z times B*z equals C*z. Every bit split needs x*(x-1)=0 plus putting the bits back together, 32-bit adds need boolean carries, rotates and shifts can just be wires. Use your own prime field with Montgomery math, state the two-adicity, verify the root table, and do NTT with bit-reversal plus coset for the quotient, checked exactly (not fuzzy) against naive math up to 2^12. Pick any proof system you like but write down why it is sound and use your own hash transcript that binds all public inputs. Test the honest case, reject 5 bad witnesses (bit=2, free carry, wrong digest, bad padding, wrong public input) and 3 tampered proofs. Report constraint count with per-gadget math, prove+verify time under 500ms, and soundness bits. Build it with your local zig version and report zig version in result, do not require a single fixed version.", "score_max": 10, "scoring": "Start at 0; +3 for field/NTT exactness, +3 for gadget soundness, +2 for counts/witness, +2 for prover/transcript/speed; cap at 10, pass if >=7", "title": "Prove a SHA-256 preimage with a tiny R1CS prover"}
{"category": "Security and cryptography", "checks": ["Bounds-checked parser with clean RVA mapping; zero panics and zero out-of-range reads on all 400 malformed with fuzz counts", "Features exactly defined with formulas and thresholds; TPR>=0.90 at FPR<=0.005 with ROC/cutoff/matrix and packed FP<=1% on held-out split", "Aho-Corasick with BFS failure links and merged outputs plus wildcard scheme; footprint reported and >=300 MiB/s measured with no alloc in loop", "Three state-machine detectors with windows/thresholds; TPR>=0.95 max 2 FP on 300 benign with per-family latency and each FP explained"], "difficulty": "expert", "expected_result": "A runnable repo that parses PEs, scores files and traces, and prints ROC/confusion, crash/fuzz, scanner speed and behavior latency Include reported zig version.", "id": "SecAvHeuristics26", "pass_score": 7, "prompt": "Hi, our last antivirus trial failed because it flagged a signed installer and our backup tool, so they pulled it. I need detection only, nothing that hooks the machine. Please write a PE32/PE32+ parser that never crashes: DOS/NT headers, sections, imports including bound and delayed, exports, relocs, TLS, resources, overlay and signature directory, with checked bounds and RVA-to-offset that fails cleanly outside sections. Then features per section like entropy in bits per byte, import hash with exact recipe, entry-point section, writable-plus-executable, weird names, TLS, overlay size, bad import combos, into a scored model with weights and cutoff you pick. Also an Aho-Corasick scanner with wildcards and footprint, and three behavior watchers over event logs for ransomware, mining and keylogging with windows and thresholds. Test on about 2000 normal files including packed ones, 800 bad-structural, 400 broken for crashes, 600 traces split benign/malicious. I need true positives at least 0.90 at false positives at most 0.005, packed false positives at most 1 percent, zero crashes, scanner at least 300 MiB/s, behavior at least 0.95 with at most 2 false alarms plus delay in events. Build it with your local zig version and report zig version in result, do not require a single fixed version.", "score_max": 10, "scoring": "Start at 0; +2 for parser/fuzz, +3 for features/ROC, +2 for automaton/speed, +3 for behavior/latency; cap at 10, pass if >=7", "title": "Detect bad PEs and bad behavior without false alarms"}
{"category": "Physics and motion simulation", "checks": ["Repose 32-36 deg, overlap <=0.5 percent radius, zero floor tunneling, runout within 5 percent on rerun", "Energy drift <=1 percent in 5k settle test, all finite, identical hash every 500 steps twice", "Mean step <=25 ms on one core for 100k avalanche, zero allocs after init with counters"], "difficulty": "expert", "expected_result": "Runnable Zig repo plus result.json with repose angle, overlap, runout, energy drift, ms per step, position hashes, alloc counters and zig version", "id": "PhysGranularAvalanche27", "pass_score": 7, "prompt": "hi I need help please, I'm trying to demo a landslide for my game and my piles just look like jelly and explode. Can you build me a 3D grain pile thing in pure Zig with no libs? I need like 100k round grains falling down a rough chute into a pile, with real contacts - Hertz normal plus Mindlin friction and rolling resistance, fixed step 1/240 s maybe with substeps you state, plus a spatial hash grid from a fixed arena so zero allocs after startup. Scenes I want: a 30 degree chute avalanche with 100k grains radius 2-3 mm, a static cone pile to check repose angle, and a small 5k box that settles for energy check. I need the pile cone angle 32-36 deg measured from heights, max overlap under 0.5 percent radius and zero tunneling through floor, runout length within 5 percent on two reruns with identical sha hash of positions every 500 steps, total energy drift under 1 percent in settle test, and mean step under 25 ms on one core. Build it with your local zig version and report zig version in result, do not require a single fixed version. Output one folder that runs and writes result.json with timings and hashes please, thanks a lot!", "score_max": 10, "scoring": "Give 3, 3, 4 points per check in order for 10 total, 7 or more is pass, no partial without numbers", "title": "Granular avalanche 100k grains"}
{"category": "Physics and motion simulation", "checks": ["Rise 10-16 cm above rim, catenary width within 10 percent of theory, stretch <=0.5 percent per link", "Max penetration <=0.2 mm, zero tunneling, all finite over 10 sec", "Mean step <=12 ms single core, identical hash every 1000 steps twice, zero allocs after init"], "difficulty": "expert", "expected_result": "Runnable Zig repo plus result.json with rise height, catenary width, stretch, penetration, ms per step, hashes, alloc counters and zig version", "id": "PhysChainFountain28", "pass_score": 7, "prompt": "hi folks I'm stuck, I saw that viral chain fountain video where beads jump above the pot and my game rope just falls limp over the edge, looks fake. Can you please code a bead chain fountain in pure Zig, no physics lib? I need like 800 steel beads 4 mm radius 2 g each linked with stiff distance plus bending limit and twist you state, shot from a 50 cm high pot falling 1.2 m to floor, fixed 1/480 s step with 4 substeps, floor plus pot wall collisions with Coulomb friction, beads start coiled in pot pulled at 2 m/s, plus spatial hash from fixed memory zero allocs after init. I want rise height 10-16 cm above rim, inverted catenary shape width within 10 percent of theory, no stretch over 0.5 percent per link, max penetration under 0.2 mm, energy plus momentum traces finite, 10 sec run mean step under 12 ms single core, bit-identical hash every 1000 steps on two runs. Build it with your local zig version and report zig version in result, do not require a single fixed version. Just one runnable Zig folder that writes result.json with heights and hashes, thank you so much!", "score_max": 10, "scoring": "Give 4, 3, 3 points per check in order for 10 total, 7 or more is pass, need numbers for credit", "title": "Chain fountain momentum jump"}
{"category": "Physics and motion simulation", "checks": ["Stick-slip once per period, jitter <=0.8 percent over 500 periods, fundamental within 1.5 Hz of 196 Hz", "Harmonics within 2 dB of 1 over n to 8th, zero NaN, stable on ramp and flautando", "Render 5 sec <=400 ms single core, identical wav hash twice, zero allocs in loop"], "difficulty": "expert", "expected_result": "Runnable Zig repo plus result.json with slip jitter, fundamental, harmonic errors, render ms, wav hash twice, alloc counters and zig version", "id": "PhysStickSlipViolin29", "pass_score": 7, "prompt": "hey hi I need help, I play violin a bit and I want my synth to squeak like real bow not cheap saw wave, my current string just hums. Could you build a bowed string thing in pure Zig with no audio libs? I need one G string 0.33 m long 196 Hz, wave equation with 400 segments fixed 48 kHz step you state plus bow at 0.12 m from bridge moving 0.2 m/s with 0.8 N force using Stribeck friction with static 0.8 kinetic 0.3 and Stribeck speed you pick, plus body made of 8 modes with Q you list, fixed arena zero allocs in loop. Tests: 5 sec steady bow, 3 sec crescendo force ramp, and 2 sec light flautando. I need Helmholtz stick-slip locked with slip once per period jitter under 0.8 percent over 500 periods, fundamental within 1.5 Hz of 196 Hz, harmonic falloff within 2 dB of 1 over n up to 8th, no NaN or blowup, 5 sec render under 400 ms single core, identical wav hash on two runs. Build it with your local zig version and report zig version in result, do not require a single fixed version. One runnable Zig repo writing result.json plus wav please, thanks!", "score_max": 10, "scoring": "Give 4, 3, 3 points per check in order for 10 total, 7 or more is pass, no credit without measured numbers", "title": "Bowed string stick-slip violin"}
{"category": "Physics and motion simulation", "checks": ["Fast 120 rad/s precession within 3 percent of theory, nutation <=0.8 deg after 2 sec, slow 35 rad/s cusp count exact", "Energy drift <=0.15 percent over 20 sec, angular momentum drift <=1e-4, all finite", "Mean step <=0.8 ms, identical hash every 2000 steps twice, zero allocs after init"], "difficulty": "expert", "expected_result": "Runnable Zig repo plus result.json with precession rate, nutation amplitude, loop count, energy drift, ms per step, hashes, alloc counters and zig version", "id": "PhysGyroTopPrecession30", "pass_score": 7, "prompt": "hello I need a hand please, I'm building a toy top demo and my tops just wobble and fall even when spinning fast, no nice precession circles. Can you make a heavy symmetric top sim in pure Zig, no libs? I need Euler angles plus quaternion integrator at 1/2000 s you state, mass 0.2 kg tip to center 0.04 m, transverse inertia 2.5e-5 and spin inertia 1.1e-5 you can keep, gravity 9.81, plus tip friction circle and air damping you list, floor contact point no slip when sleeping. Cases: fast 120 rad/s tilted 20 deg, slow 35 rad/s tilted 20 deg, and nutation tap test. I need fast case steady precession within 3 percent of M times g times d over I3 times w3 theory, nutation amplitude under 0.8 deg after 2 sec, slow case looping cusps with loop count exact, energy drift under 0.15 percent over 20 sec and angular momentum drift under 1e-4, two runs bit-identical hash every 2000 steps, mean step under 0.8 ms, zero allocs after init. Build it with your local zig version and report zig version in result, do not require a single fixed version. One runnable Zig folder that writes result.json with angles and hashes, thanks a lot!", "score_max": 10, "scoring": "Give 4, 3, 3 points per check in order for 10 total, 7 or more is pass, strict to numbers", "title": "Spinning top precession nutation"}
{"category": "Physics and motion simulation", "checks": ["Flip angle within 1.0 deg of 70.5 deg theory, direction correct on all 15 angles, slip distance within 4 percent", "Penetration <=0.1 mm, energy drift <=0.2 percent in free roll, yo-yo period within 3 percent", "Mean step <=1 ms, identical hash every 2000 steps twice, zero allocs after init"], "difficulty": "expert", "expected_result": "Runnable Zig repo plus result.json with flip angle, 15-angle direction table, slip distance, energy drift, ms per step, hashes, alloc counters and zig version", "id": "PhysRollingSpoolYoYo31", "pass_score": 7, "prompt": "hi I need help with my puzzle game, I have a cable spool on the floor that you pull and sometimes it rolls away and sometimes toward you and I always get it backwards. Could you code a rolling spool plus yo-yo sim in pure Zig, no deps? I need a spool outer radius 0.06 m inner hub 0.02 m mass 1.2 kg inertia you compute from disks, pulled by string wrapped on hub at angle theta from horizontal you sweep 10 to 80 deg plus vertical drop yo-yo case 0.5 m string, Coulomb friction mu_s 0.4 mu_k 0.25 with stick-slip transitions, fixed 1/2000 s step, floor contact with rolling constraint. Tests: pull at 0.5 N across angles, hard yank 3 N slip test, and free yo-yo drop. I need critical flip angle within 1.0 deg of acos inner over outer theory about 70.5 deg, roll direction correct on all 15 angles, slip distance within 4 percent on rerun, energy drift under 0.2 percent in free roll, no penetration over 0.1 mm, mean step under 1 ms, identical hash every 2000 steps twice, zero allocs after init. Build it with your local zig version and report zig version in result, do not require a single fixed version. One runnable Zig repo writing result.json with tables and hashes please, thank you!", "score_max": 10, "scoring": "Give 4, 3, 3 points per check in order for 10 total, 7 or more is pass, fail if direction wrong once", "title": "Rolling spool pull flip and yoyo"}
{"category": "Physics and motion simulation", "checks": ["Upright 20 s from 5 and 8 deg starts and 15 N 0.15 s push with cart within 2 m, angle RMS <=1.2 deg after 5 s, recovery <=2.5 s", "Stable with 20 ms delay via predictor with force <=25 N and no saturation over 100 ms continuous, fails or exits bounds without compensation", "Mean step <=0.6 ms at 1 kHz on one core with zero allocs after init and bit-identical hashes on two runs", "Full Lagrangian dynamics with RK4 plus documented LQR and predictor gains with margins >=6 dB and >=30 deg"], "difficulty": "expert", "expected_result": "Single Zig repo plus result.json with angle RMS and recovery table, force saturation histogram, step timings, allocation counters, state hashes from two runs and zig version.", "id": "PhysDoublePendulumDelay32", "pass_score": 7, "prompt": "hi I'm really stuck on my self-balancing cart demo for next week, it has two poles stacked on a cart and my sensor board is laggy, please help me build it in pure Zig with no C deps. The cart is 1.2 kg on a 4 m track with soft end stops, lower arm 0.5 m 0.3 kg, upper arm 0.7 m 0.25 kg, both with small viscous joint damping 0.002 N m s, simulated at fixed 1 kHz with RK4. The angle and cart sensors come exactly 20 ms late plus white noise 0.2 deg and 1 mm, so the controller only sees old data and must use a delay predictor or buffer plus Kalman filter, then LQR or MPC with force capped at 25 N. Please test tilted starts at 5 and 8 deg, a 15 N push for 0.15 s at 5 s, and a 30 s noisy run from seed 0xD1AY. I need it upright 20 s without hitting ends, angle RMS under 1.2 deg after 5 s, recovery under 2.5 s, motor never stuck over 100 ms, same hash twice. Build it with your local zig version and report zig version in result, do not require a single fixed version. Thank you so much, I am lost on delays.", "score_max": 10, "scoring": "Give 2.5 points per passed check for 10 total, 7 or more is a pass.", "title": "Double inverted pendulum with 20 ms delayed sensing"}
{"category": "Physics and motion simulation", "checks": ["Peak pressure 90 to 180 kPa and MPS peak 0.15 to 0.28 with CSDM15 reported over 0.5 s 120 g 8 ms hit", "No skull-brain penetration, CSF mass drift <=1 percent and energy drift <=2 percent with all values finite", "Mean step <=8 ms for about 8000 SPH plus FEM ball, zero allocs after init, identical hash on two runs", "Documented Prony shear 1.2 kPa, Tait CSF, penalty contact and 2 kHz integration with convergence across two resolutions within 10 percent"], "difficulty": "expert", "expected_result": "Single Zig repo plus result.json with pressure and strain peaks, CSDM15, mass and energy drifts, step timings, allocation counters, hashes from two runs and zig version.", "id": "PhysBrainSloshImpact33", "pass_score": 7, "prompt": "hi I am helping a bike helmet group and I get confused by brain squish, could you please build a small head impact thing in pure Zig with no other deps. I picture the skull as a rigid 0.09 m sphere with a 3 mm fluid gap around a soft brain ball about 1.4 kg, density near 1040 kg per cubic m, shear modulus around 1.2 kPa with Prony damping, plus CSF as water-like SPH particles around 8000 with Tait stiffness. Run at fixed 2 kHz with explicit integration, skull gets a half-sine 120 g 8 ms frontal hit then free shake for 0.5 s, contact as penalty with no sticking, brain-skull gap never negative. I need peak pressure 90 to 180 kPa, peak max principal strain 0.15 to 0.28, 15 percent volume fraction over 0.15 strain, skull to brain energy drift under 2 percent, same hash twice, step under 8 ms. Build it with your local zig version and report zig version in result, do not require a single fixed version. I am not a biomechanics person so please keep it simple but real, with plots and numbers I can show them, thank you.", "score_max": 10, "scoring": "Give 2.5 points per passed check for 10 total, 7 or more is a pass.", "title": "Brain slosh in fluid under impact"}
{"category": "Physics and motion simulation", "checks": ["Dead-hold 100 N load with 12 N hold slip <=2 mm over 10 s and zero-mu control slips over 50 mm", "Pull force within 8 percent of exp mu theta for mu 0.2 0.3 0.4 at 270 deg and breakaway within 10 percent of predicted", "Max penetration <=0.3 mm with stick-slip states logged and no tunneling at 1 kHz", "Mean step <=2 ms, zero allocs after init, bit-identical hashes on two runs"], "difficulty": "expert", "expected_result": "Single Zig repo plus result.json with hold slip, pull-force versus capstan table for three mus, breakaway errors, penetration, timings, hashes from two runs and zig version.", "id": "PhysCapstanRopePulley34", "pass_score": 7, "prompt": "hi I am building a little crane with a rope wrapped around a rough steel pulley and I keep messing up the holding force, please help me in pure Zig with no deps. I need a rope as 200 masses over 2 m with stretch stiffness 20 kN per m plus bending and damping, wrapped 270 degrees around a 0.1 m radius pulley with Coulomb friction mu 0.3, capstan law T tight equals T loose times exp mu theta, run at fixed 1 kHz with small substeps, contact as penalty with stick slip per segment and no tunneling. Tests are dead-hold with 100 N load and 12 N hold for 10 s, slow pull at 0.05 m per s, and sudden 200 N jerk, plus zero friction check that slips. I need hold slip under 2 mm in 10 s, pull force within 8 percent of exp law for mu 0.2 0.3 0.4, breakaway within 10 percent of predicted, peak penetration under 0.3 mm, same hash twice, step under 2 ms. Build it with your local zig version and report zig version in result, do not require a single fixed version. I am bad at friction so please make the numbers clear with tables I can trust, thanks a lot.", "score_max": 10, "scoring": "Give 2.5 points per passed check for 10 total, 7 or more is a pass.", "title": "Capstan rope over rough pulley"}
{"category": "Physics and motion simulation", "checks": ["1D interface within 2 mm and 3 percent of Neumann analytic at 60 180 360 s with 400 cells, L2 temp <=0.8 C", "Global energy drift <=0.5 percent over 600 s with latent 334 kJ per kg accounted and overshoot <=0.5 C", "2D corner melt converges with refinement and 1D run <=3 s, zero allocs after init, identical hash on two runs"], "difficulty": "expert", "expected_result": "Single Zig repo plus result.json with interface position table versus Neumann, L2 errors, energy drift, timings, allocation counters, hashes from two runs and zig version.", "id": "PhysMeltingFreezingStefan35", "pass_score": 7, "prompt": "hi I am trying to model ice melting for a cooler box and I get weird wiggles at the ice water line, please help me in pure Zig with no libraries. I need a 1D 0.1 m slab starting as ice at minus 10 C with left wall held at plus 20 C and right wall insulated, plus a 2D 64 by 64 corner melt check, using enthalpy or apparent heat with latent 334 kJ per kg, ice 2100 J per kg K 2.2 W per m K 917 kg per cubic m, water 4182 J per kg K 0.6 W per m K 1000 kg per cubic m, melting at 0 C, fixed 0.5 ms to 5 ms steps with explicit or implicit conduction. Run 600 s with 400 cells in 1D. I need interface within 2 mm or 3 percent of Neumann analytic at 60 180 360 s, temperature L2 under 0.8 C, global energy drift under 0.5 percent, no overshoot over 0.5 C, same hash twice, 1D run under 3 s. Build it with your local zig version and report zig version in result, do not require a single fixed version. I am not good at heat math so please show interface tables and energy numbers I can plot, thank you very much for helping.", "score_max": 10, "scoring": "Give 3, 3, 4 points in check order for 10 total, 7 or more is a pass.", "title": "Ice-water Stefan melting and freezing"}
{"category": "Physics and motion simulation", "checks": ["Collapse time within 3 percent of 0.915 R sqrt rho over delta p for 100 um 10 to 500 kPa case and Rmin within 15 percent of reference", "Wall jet 80 to 200 m per s toward wall at stand-off 1.2 with peak wall pressure 50 to 500 MPa logged", "Volume drift <=1 percent with finite values, mean step <=5 ms, zero allocs after init, identical hash on two runs"], "difficulty": "expert", "expected_result": "Single Zig repo plus result.json with collapse time and Rmin table, jet speed and wall pressure, volume drift, timings, hashes from two runs and zig version.", "id": "PhysCavitationCollapse36", "pass_score": 7, "prompt": "hi I am cleaning parts with ultrasound bubbles and I do not get why tiny bubbles pit metal, please build a bubble collapse thing in pure Zig with no deps. I need a spherical gas bubble 100 um radius in water at 20 C, density 998 kg per cubic m, viscosity 1.0 mPa s, surface tension 0.072 N per m, far pressure jump from 10 kPa to 500 kPa, polytropic gas with ratio 1.4, simulated with Rayleigh-Plesset plus Keller-Miksis compressibility and a 2D axisymmetric jet run near a wall at stand-off 1.2 radii with 128 squared grid and level set or VOF. Tests are Rayleigh collapse in vacuum, viscous collapse, and wall jet. I need collapse time within 3 percent of Rayleigh 0.915 R sqrt rho over delta p, minimum radius within 15 percent of reference, peak wall pressure 50 to 500 MPa logged, jet speed 80 to 200 m per s toward wall, volume drift under 1 percent, same hash twice, step under 5 ms. Build it with your local zig version and report zig version in result, do not require a single fixed version. I am not a fluids guy so please keep outputs simple with radius and pressure tables, thanks a lot for helping me.", "score_max": 10, "scoring": "Give 3, 3, 4 points in check order for 10 total, 7 or more is a pass.", "title": "Cavitation bubble collapse jet"}
{"category": "Biological mechanics", "checks": ["Stride 2.8-3.6 Hz with spine pitch 8-18 deg, duty 28-42 percent and peak vertical 2-3.2x body weight at 6 m/s", "Fatigue shows fresh force within 3 percent of reference, 28-45 percent drop after 40 sprints and recovery to 90 percent within 120 s rest", "Mean step <=1.8 ms single-core with zero allocations after init proven by counters", "Bit-identical hashes on two runs for gallop and fatigue traces"], "difficulty": "expert", "expected_result": "Single pure-Zig repo plus result.json with stride, spine, force and fatigue traces, ms per step, allocation counters, output hashes and reported zig version.", "id": "BioQuadrupedGallopFatigue37", "pass_score": 7, "prompt": "hi, I'm building a little game with a dog-like creature and my gallop looks floaty and never gets tired, I'm stuck. Could you please build me a quadruped gallop simulator in pure Zig with no other dependencies? I need a torso with a flexible spine joint plus four legs with shoulder/hip and knee style joints driven by Hill-type muscles with force-length and force-velocity plus activation lag and a fatigue model where max force drops with effort and recovers at rest. Feet should have spring-damper ground contact with friction, gait should be rotary gallop with spine flexion extending stride, running at fixed 1/480 s step. Please test 20 s run at 6 m/s command, then a fatigue run with repeated sprints, plus a passive check. I need stride frequency 2.8-3.6 Hz, spine pitch swing 8-18 degrees, ground contact duty 28-42 percent, peak vertical force 2-3.2x body weight, fresh muscle force within 3 percent of unfatigued reference, and after 40 sprint cycles peak force down 28-45 percent with recovery to 90 percent within 120 s rest, mean step under 1.8 ms single core, zero allocations after init, bit-identical hashes on two runs. Build it with your local zig version and report zig version in result, do not require a single fixed version. Thanks so much!", "score_max": 10, "scoring": "Give 2.5 points per passed check for 10 total, 7 or more passes.", "title": "Quadruped gallop with spine and muscle fatigue"}
{"category": "Biological mechanics", "checks": ["Tidal volume 450-550 mL quiet and over 2800 mL deep at 12-16 per minute with pleural swing -8 to -4 cmH2O", "Pressure-volume work 0.3-0.8 J per breath with peak flow 0.4-0.9 L per s and residual volume within 2 percent across reruns", "Mean step <=0.9 ms single-core with zero allocations after init proven by counters", "Bit-identical hash on two 120 s runs"], "difficulty": "expert", "expected_result": "Single pure-Zig repo plus result.json with volume, pressure, flow and work traces, ms per step, allocation counters, hashes and reported zig version.", "id": "BioDiaphragmBreathing38", "pass_score": 7, "prompt": "hello, I'm trying to animate breathing for a first-aid trainer and my chest just scales up and down, it looks fake and I have no pressures. Can you please make a breathing simulator in pure Zig with nothing else? I need a diaphragm dome muscle driving volume, ribs as rotating pump-handle and bucket-handle levers with joint limits, lungs as compliant sacs with pressure-volume curve and airway resistance plus chest wall elasticity, solved at fixed 1/240 s. Please simulate quiet breathing 12-16 per minute, deep breath, cough puff, and blocked airway check over 120 s from fixed seed. I need tidal volume 450-550 mL quiet and over 2800 mL deep, pleural pressure swing -8 to -4 cmH2O quiet, pressure-volume loop hysteresis work 0.3-0.8 J per breath, airway peak flow 0.4-0.9 L/s quiet, residual volume within 2 percent across reruns, mean step under 0.9 ms single core, zero allocations after init, bit-identical result hash twice. Build it with your local zig version and report zig version in result, do not require a single fixed version. Thank you very much!", "score_max": 10, "scoring": "Give 2.5 points per passed check for 10 total, 7 or more passes.", "title": "Diaphragm rib lung pressure-volume breathing"}
{"category": "Biological mechanics", "checks": ["Stroke volume 65-80 mL rest with ejection fraction 55-68 percent, peak aortic 110-135 mmHg and output within 3 percent of analytic", "Regurgitant fraction under 5 percent healthy and 18-30 percent in leak case with leaflet limits enforced", "Mean beat <=1.2 ms with zero allocations after init proven by counters", "Bit-identical hashes on two runs at 70 and 120 bpm"], "difficulty": "expert", "expected_result": "Single pure-Zig repo plus result.json with pressure-volume loops, stroke volumes, regurgitation, timings, allocation counters, hashes and reported zig version.", "id": "BioHeartValveFlow39", "pass_score": 7, "prompt": "hi, I need a heart demo for a school project and my pump just moves a sine wave, valves leak weirdly and volumes are off. Could you build a left ventricle plus valve flow model in pure Zig with no libraries? I need a time-varying elastance ventricle with pressure-volume loop, aortic and mitral leaflets as hinged flaps with opening angle limits and regurgitation gap, plus lumped Windkessel afterload and viscous flow with backflow check, stepped at fixed 1/1000 s. Please run 30 beats at 70 bpm rest, exercise at 120 bpm, and valve-leak test over fixed seed. I need stroke volume 65-80 mL rest, ejection fraction 55-68 percent, peak aortic pressure 110-135 mmHg, regurgitant fraction under 5 percent healthy and 18-30 percent in leak case, cardiac output within 3 percent analytic check, mean beat under 1.2 ms, zero allocations after init, bit-identical hashes on two runs. Build it with your local zig version and report zig version in result, do not require a single fixed version. Thanks a lot!", "score_max": 10, "scoring": "Give 2.5 points per passed check for 10 total, 7 or more passes.", "title": "Ventricle valve leaflet flow pump"}
{"category": "Biological mechanics", "checks": ["Tail beat 2.5-4.5 Hz cruise with speed 1.2-2.0 body lengths per s, Strouhal 0.25-0.45 and yaw drift under 4 deg per 10 s", "Burst, turn and glide traces with cost of transport and drag plus added-mass thrust accounting", "Mean step <=1.5 ms single-core with zero allocations after init proven by counters", "Bit-identical hashes on two 30 s runs"], "difficulty": "expert", "expected_result": "Single pure-Zig repo plus result.json with frequency, speed, Strouhal, transport cost, ms per step, allocation counters, hashes and reported zig version.", "id": "BioFishSwimCpg40", "pass_score": 7, "prompt": "hey, my fish robot wiggles but barely moves forward and the tail looks random, I'm lost. Can you please build a fish swimming simulator in pure Zig with no other deps? I need a chain of 12 body segments with spine joints driven by coupled phase oscillators with phase lag head to tail plus amplitude envelope, Hill-like muscle torque with limits, skin with bending stiffness and damping, and water with per-segment drag and added-mass thrust plus tail fin lift, integrated at fixed 1/300 s. Please swim 30 s cruise, burst, turn, and still-water glide from fixed seed. I need tail beat 2.5-4.5 Hz cruise, speed 1.2-2.0 body lengths per second, Strouhal 0.25-0.45, yaw drift under 4 degrees over 10 s straight, cost of transport reported, mean step under 1.5 ms single core, zero allocations after init, bit-identical hashes twice. Build it with your local zig version and report zig version in result, do not require a single fixed version. Thanks!", "score_max": 10, "scoring": "Give 2.5 points per passed check for 10 total, 7 or more passes.", "title": "Fish CPG body wave swim with drag"}
{"category": "Biological mechanics", "checks": ["Mean lift within 5 percent of weight at cruise with power reported and membrane strain under 12 percent when stable", "Flutter boundary within 0.8 m per s of reference with damped response below and over 15 percent per-cycle growth above", "Mean step <=2.0 ms with zero allocations after init proven by counters", "Bit-identical hashes on two runs including gust and sweep"], "difficulty": "expert", "expected_result": "Single pure-Zig repo plus result.json with lift, power, strain, flutter boundary, ms per step, allocation counters, hashes and reported zig version.", "id": "BioBirdFlapMembrane41", "pass_score": 7, "prompt": "hi, I'm making a bat-like glider and my wings are stiff boards so flutter never shows and lift feels wrong, please help. Could you make a flapping membrane wing simulator in pure Zig with no dependencies? I need two wings with shoulder flap and twist plus membrane triangles with tension and bending stiffness and damping, coupled to quasi-steady aero with lift-drag versus angle plus added mass and gust input, solved at fixed 1/500 s with substeps for membrane. Please test level flap at 6-10 Hz, glide hold, gust hit, and flutter sweep with rising speed from fixed seed. I need mean lift within 5 percent of weight at cruise, thrust power reported, stable below flutter and divergent oscillation growth over 15 percent per cycle above flutter speed with boundary within 0.8 m/s of reference, membrane strain under 12 percent stable, mean step under 2.0 ms, zero allocations after init, bit-identical hashes twice. Build it with your local zig version and report zig version in result, do not require a single fixed version. Thank you!", "score_max": 10, "scoring": "Give 2.5 points per passed check for 10 total, 7 or more passes.", "title": "Flapping membrane wing aeroelastic flutter"}
{"category": "Biological mechanics", "checks": ["Forward speed within 5 percent of reference on compact sand, loose sand and 15 degree slope with backward slip at most 2 mm per cycle", "Anisotropic friction with lateral 1.8x forward enforced, per-segment force traces and cost of transport reported", "Mean step at most 5 ms on one core with zero allocations after init", "Bit-identical hashes on two runs, pure Zig with no C deps and zig version reported in result"], "difficulty": "expert", "expected_result": "Single Zig repo plus result.json with speed and slip table for 3 sands and slope, force and energy traces, step timing, allocation counters, rerun hashes and zig version.", "id": "BioSnakeGranularUndulation42", "pass_score": 7, "prompt": "hey I am stuck on my desert snake robot and the field test is in three weeks, it just digs in place on sand and I cannot figure out the gait. I have a 12-link body about 1.2 m long, each joint driven to follow a traveling curvature wave with amplitude, wavelength, frequency and phase lag I can tune, plus slight vertical lift to reduce drag. The sand is anisotropic so lateral friction is much higher than forward sliding, roughly mu lateral 1.8 times mu forward, with velocity-dependent granular resistance and sinkage from weight distribution. Could you build a pure Zig simulator, just Zig no C libs, with Coulomb friction per segment, resistive force integration along the body, ground contact with slip velocity, and energy per meter traveled? Test on flat compact sand, loose dune sand, and a 15 degree slope, sweeping wave amplitude and frequency from seeds I give. I need forward speed within 5 percent of a reference trace, no backward slip beyond 2 mm per cycle, cost of transport reported, deterministic reruns with identical hashes, under 5 ms per step. Build it with your local zig version and report zig version in result, do not require a single fixed version. Thanks a lot for helping, I really need this to move.", "score_max": 10, "scoring": "Give 2.5 points per check, total 10, pass at 7 or more.", "title": "Snake granular undulation on anisotropic sand"}
{"category": "Biological mechanics", "checks": ["Mean lift within 4 percent of reference at 200 Hz with peak hinge stress below fatigue limit and phase lag within 5 degrees", "Resilin nonlinear spring with hysteresis loss and delayed reflex 2 to 10 ms, stable amplitude under gust and 30 ms dropout", "Mean step at most 3 ms on one core with zero allocations after init", "Bit-identical hashes on two runs, pure Zig with no C deps and zig version reported in result"], "difficulty": "expert", "expected_result": "Single Zig repo plus result.json with lift and power table, hinge stress and phase lag, delay sweep, timing, allocation counters, hashes and zig version.", "id": "BioInsectWingResilin43", "pass_score": 7, "prompt": "hi my fruit fly robot wing keeps tearing at the hinge and it cannot hold 200 Hz for more than a few seconds, defense is in a month and I am lost. I have two flapping wings about 12 mm long, each on a torsional hinge with resilin pad that stores elastic energy and returns it each stroke, plus a wing pitch joint with damping and hard stops. Drive is sinusoidal at 200 Hz with stroke amplitude and pitch phase I can tune, and sensing has a reflex delay around 5 ms plus noisy campaniform strain feedback. Could you build a pure Zig model, only Zig no other deps, with unsteady aerodynamics from quasi-steady lift plus added mass and wake capture bump, resilin as nonlinear spring with hysteresis loss, delayed feedback control for amplitude regulation, and cycle-averaged lift versus power reporting? Test clean hover, 2 m per s gust hit, and 30 ms sensor dropout, sweeping delay 2 to 10 ms from my seeds. I need mean lift within 4 percent of reference and peak hinge stress under fatigue limit, phase lag within 5 degrees, reruns bit-identical with hashes, step under 3 ms on one core. Build it with your local zig version and report zig version in result, do not require a single fixed version. Please help, this hinge is killing me.", "score_max": 10, "scoring": "Give 2.5 points per check, total 10, pass at 7 or more.", "title": "Resilin wing hinge at 200 Hz with reflex delay"}
{"category": "Biological mechanics", "checks": ["Saccade endpoint within 0.5 deg with settling under 50 ms for 10 and 20 deg targets", "VOR gain 0.95 to 1.05 with latency error under 2 ms and peak image slip under 2 deg per s at 150 deg per s head impulse", "Mean step at most 2 ms with zero allocations after init", "Bit-identical hashes on two runs, pure Zig with no C deps and zig version reported in result"], "difficulty": "expert", "expected_result": "Single Zig repo plus result.json with saccade endpoint and settling table, VOR gain and latency, slip velocity, timing, hashes and zig version.", "id": "BioEyeVestibularReflex44", "pass_score": 7, "prompt": "hey my VR headset gaze stabilization looks awful, every quick head turn smears the image and saccades overshoot then wobble, users get sick in minutes. I have a eyeball model about 24 mm diameter with inertia plus six muscle directions simplified to pan tilt with force-velocity and series elasticity, saccade generator with main-sequence peak velocity versus amplitude shaping and 200 deg per s threshold detection, plus vestibular sensor with latency I can tune around 7 to 12 ms and noisy gyro feedback for counter-rotation. Could you build a pure Zig simulator, just Zig no C libs, with 1 kHz neural delay line, VOR gain near 1.0 with phase error reporting, slip velocity during head impulses at 150 deg per s, fixation drift after saccade, and power free head-eye coordination? Test 10 deg saccade, 20 deg step with 100 deg per s head yaw, and smooth pursuit at 20 Hz from seeds I give. I need saccade endpoint within 0.5 deg and settling under 50 ms, VOR gain 0.95 to 1.05 with latency error under 2 ms, image slip under 2 deg per s peak, deterministic reruns with identical hashes under 2 ms per step. Build it with your local zig version and report zig version in result, do not require a single fixed version. Thanks so much, I need this to feel natural.", "score_max": 10, "scoring": "Give 2.5 points per check, total 10, pass at 7 or more.", "title": "Saccade plus vestibular reflex gaze stabilization"}
{"category": "Biological mechanics", "checks": ["Ankle torque tracking within 4 percent RMS on level, ramp, grass and stairs with energy per stride reported", "Terrain adaptation within two strides with toe clearance at least 5 mm and no bottoming on 17 cm steps", "Mean step at most 4 ms on one core with zero allocations after init", "Bit-identical hashes on two runs, pure Zig with no C deps and zig version reported in result"], "difficulty": "expert", "expected_result": "Single Zig repo plus result.json with torque tracking and energy per stride table, adaptation steps, timing, hashes and zig version.", "id": "BioProstheticAnkleSea45", "pass_score": 7, "prompt": "hi my prosthetic ankle prototype bottoms out on stairs and wastes battery on grass, trial with an amputee volunteer is next month and I am worried. I have a 1.2 kg foot with series-elastic actuator at ankle, spring stiffness I can pick around 300 N-m per rad with damping plus motor torque and current limits, terrain estimator from foot IMU plus ground contact switch with detection lag I can tune. Gait needs level walking at 1.2 m per s, 10 degree ramp up and down, and stair ascent with 17 cm steps, with push-off timing and stiffness adaptation per terrain. Could you build a pure Zig simulator, only Zig no other deps, with ankle inverse dynamics, spring energy storage and return, motor heat and battery drain model, ground compliance per terrain, and adaptive controller that switches stiffness within two steps after terrain change? Test level, ramp, grass with extra damping, and stairs from seeds I give. I need ankle torque tracking within 4 percent RMS, energy per stride reported, no toe stub beyond 5 mm, adaptation within two strides, deterministic reruns with identical hashes, step under 4 ms. Build it with your local zig version and report zig version in result, do not require a single fixed version. Thank you, this matters a lot for our volunteer.", "score_max": 10, "scoring": "Give 2.5 points per check, total 10, pass at 7 or more.", "title": "Series-elastic prosthetic ankle for terrain adaptation"}
{"category": "Biological mechanics", "checks": ["Gait phase error at most 15 ms mean with torque tracking within 5 percent RMS and false assists under 2 percent", "Metabolic saving reported with no fall from assist under trip nudge on level, fast walk, sit-stand and stairs", "Mean step at most 4 ms on one core with zero allocations after init", "Bit-identical hashes on two runs, pure Zig with no C deps and zig version reported in result"], "difficulty": "expert", "expected_result": "Single Zig repo plus result.json with phase error and torque tracking table, metabolic saving, timing, hashes and zig version.", "id": "BioExoskeletonHipAssist46", "pass_score": 7, "prompt": "hey my hip exoskeleton helps too late on stairs, users feel dragged behind and my demo with two older adults is in five weeks, I need the assist timing right. I have bilateral hip motors with peak torque 30 N-m, torque rise time and current limits, plus thigh IMUs and foot pressure insoles for gait phase with intention detection delay I can tune from 20 to 120 ms and noisy EMG envelope. Tasks are level walking at 1.0 to 1.6 m per s, sit to stand, and stair climb with 17 cm steps, with assist profiles for flexion and extension I can shape. Could you build a pure Zig simulator, just Zig no C libs, with 7-segment lower body dynamics, muscle torque deficit model, exo torque with delay and rate limits, phase estimator with false trigger rate, and metabolic cost saving calculation? Test level, fast walk, sit-stand, and stairs from seeds I give with trip nudge at mid swing. I need phase error under 15 ms mean, torque tracking within 5 percent RMS, false assists under 2 percent, metabolic saving reported, fall never from assist, deterministic reruns with identical hashes, step under 4 ms on one core. Build it with your local zig version and report zig version in result, do not require a single fixed version. Thanks a lot, timing is everything here.", "score_max": 10, "scoring": "Give 2.5 points per check, total 10, pass at 7 or more.", "title": "Hip exoskeleton assist with intention delay"}
{"category": "Biological mechanics", "checks": ["Normal-case transit within 5% of analytic prediction, weak case at least 50% slower, retrograde net backward", "Peak pressure within 10% across reruns and reflux <=2 mL in normal case", "Mean step <=5 ms on one core with zero allocations after init", "Bit-identical hashes on two runs"], "difficulty": "expert", "expected_result": "Single Zig repo plus result.json with transit distance, pressure trace, reflux volume, step timing, allocation counters, hashes and zig version.", "id": "BioPeristalsisBolus47", "pass_score": 7, "prompt": "Hi, I'm stuck on our gut motility demo for the physiology review next week and my peristalsis model just pushes fluid backward. Could you build a peristaltic bolus transport simulator in pure Zig, no C deps? Model a 30 cm axisymmetric gut tube with a traveling sinusoidal contraction wave with stated wavelength, amplitude, speed and occlusion ratio, coupled to 1D lubrication flow for Newtonian chyme with stated viscosity, plus a deformable viscous bolus 3 cm long tracked as particles with surface tension and wall friction. Test three cases from seed 0xB0715: normal antegrade wave, weak wave at 40% amplitude, and retrograde wave. I need net forward bolus transit over 10 s within 5% of analytic wave-speed prediction in the normal case, peak luminal pressure reported within 10% across reruns, zero net reflux volume beyond 2 mL in normal case, weak case transit at least 50% slower, retrograde case net backward displacement, identical hashes on two runs, step under 5 ms on one core, zero allocations after init. Build it with your local zig version and report zig version in result, do not require a single fixed version.", "score_max": 10, "scoring": "Give 2.5 points per passed check for 10 total, 7 or more to pass.", "title": "Gut peristalsis bolus transport"}
{"category": "Biological mechanics", "checks": ["Mean error <4 mm, p90 <8 mm and mean direction error <5 degrees over 12 positions", "Hub frequency within 3% of analytic estimate and jitter <40 us", "Frame <=2 ms on one core with zero allocations after init", "Bit-identical traces on two runs"], "difficulty": "expert", "expected_result": "Single Zig repo plus result.json with per-position error table, frequency estimate, timing jitter, frame time, allocation counters, hashes and zig version.", "id": "BioSpiderWebVibration48", "pass_score": 7, "prompt": "Hey, I study orb webs and I need a demo showing spiders locate prey by vibration before the journal club Friday, my spring network explodes with any damping change. Please build an orb-web vibration simulator in pure Zig with only std? Model radial threads 24x plus spiral 30 turns as damped mass-spring network with stated pretension, mass, stiffness and damping, fixed frame anchors, plus eight leg sensors at hub. Simulate transverse plucks from prey impacts at 12 blind positions with 100-300 Hz damped wavelets from seed 0x5D1D3A, propagate with fixed 1/8000 s step, estimate source by TDOA multilateration with stated wave speed calibration. I need mean localization error under 4 mm over 12 positions, 90th percentile under 8 mm, direction error under 5 degrees mean, arrival-time jitter under 40 us across reruns, bit-identical sensor traces on two runs, fundamental hub frequency within 3% of analytic string estimate, frame under 2 ms on one core, zero allocations after init. Build it with your local zig version and report zig version in result, do not require a single fixed version.", "score_max": 10, "scoring": "Give 2.5 points per passed check for 10 total, 7 or more to pass.", "title": "Spider web vibration prey localization"}
{"category": "Biological mechanics", "checks": ["Place error <0.8 mm for all 5 tones at 40 dB with Greenwood map stated", "Q10 >=4 passive and >=7 active at 4 kHz plus compression <0.4 dB/dB", "Distortion product >=20 dB below primaries with identical hash on two runs", "Mean step <=3 ms on one core with zero allocations after init"], "difficulty": "expert", "expected_result": "Single Zig repo plus result.json with place-error table, Q10 values, compression slope, distortion level, timing, allocation counters, hashes and zig version.", "id": "BioCochleaHairCell49", "pass_score": 7, "prompt": "Hi, for our hearing outreach I promised a cochlea wave demo but my traveling wave peaks at the wrong place for high tones. Could you build a 1D cochlear mechanics model in pure Zig, no external libs? Model a 35 mm basilar membrane as 1024 segments with exponentially graded stiffness and mass giving Greenwood tonotopy, coupled fluid pressure with stated damping, plus active outer-hair-cell gain as level-dependent negative damping and inner-hair-cell transduction via half-wave rectification plus low-pass. Drive with 0.5, 1, 2, 4, 8 kHz pure tones at 40/70/100 dB SPL from seed 0xC0C41A. I need characteristic-place error under 0.8 mm for all tones at 40 dB, Q10 tuning at least 4 at 4 kHz passive and at least 7 with active gain, compressive growth under 0.4 dB/dB from 40 to 100 dB at best place, distortion-product at 2f1-f2 at least 20 dB below primaries, identical output hash twice, step under 3 ms on one core, zero allocations after init. Build it with your local zig version and report zig version in result, do not require a single fixed version.", "score_max": 10, "scoring": "Give 2.5 points per passed check for 10 total, 7 or more to pass.", "title": "Cochlea traveling wave and hair cell transduction"}
{"category": "Zig applications", "checks": ["F1 >=0.995 clean, >=0.97 at 0 dB EMG, >=0.94 in motion bursts", "Mean error <8 ms, jitter <12 ms and heart-rate error <2 bpm per minute", "Throughput >=20x real-time on one core with zero allocations after init", "Identical detections on two runs and on 1 vs 4 threads"], "difficulty": "expert", "expected_result": "Single Zig repo plus result.json with F1 per noise condition, timing error, heart-rate error, throughput, allocation counters, detection hashes and zig version.", "id": "AppEcgRpeakNoise50", "pass_score": 7, "prompt": "Hi, our wearable team needs an on-device ECG test before the clinic pilot and my detector misses beats whenever the subject walks. Please build a synthetic ECG plus R-peak detector in pure Zig, std only? Generate 10 min Lead II at 500 Hz from seed 0xEC6501 with stated PQRST template, heart-rate variability 50-110 bpm, plus baseline wander 0.3 Hz, 50 Hz mains, EMG white noise, and three 20 s motion-burst segments. Then implement bandpass, derivative, squaring, moving-window integration and adaptive thresholding with refractory blanking and back-search, outputting R indices only. I need F1 at least 0.995 clean, at least 0.97 at 0 dB EMG, at least 0.94 during motion bursts, mean timing error under 8 ms with jitter under 12 ms, heart-rate error under 2 bpm per minute, identical detections on reruns and 1 vs 4 threads, processing at least 20x real-time on one core, zero allocations after init. Build it with your local zig version and report zig version in result, do not require a single fixed version.", "score_max": 10, "scoring": "Give 2.5 points per passed check for 10 total, 7 or more to pass.", "title": "ECG synthesis and noisy R-peak detection"}
{"category": "Zig applications", "checks": ["Peak force within 12% across reruns, onset error <0.15 N and length error <1.5 mm", "Zero tunneling and zero NaN over 10 s with identical tear-path hash twice", "Mean step <=4 ms on one core and peak memory <=256 MiB", "Zero allocations during stepping after init"], "difficulty": "expert", "expected_result": "Single Zig repo plus result.json with force trace, tear onset and length, step timing, peak memory, allocation counters, tear-path hash and zig version.", "id": "AppSutureNeedleTissue51", "pass_score": 7, "prompt": "Hey, I'm helping surgical training and our thread passes through tissue like air with no tearing force feedback. Could you build a suture needle plus thread tissue tearing sim in pure Zig? Model double-layer tissue 60x40x8 mm as mass-spring hex mesh with stated stiffness, damping and per-layer tear stress, needle as rigid curved arc with insertion path and friction, thread as 64-segment XPBD rod with stretch, bending and Coulomb friction through tissue, fixed 1/1000 s step. Test straight pass, angled pass, and overload pull that must tear from seed 0x507051. I need peak pull force within 12% across reruns, tear onset force error under 0.15 N vs reference, tear length error under 1.5 mm, zero tunneling of needle, no NaN over 10 s, identical tear-path hash twice, mean step under 4 ms on one core, peak memory under 256 MiB, zero allocations during stepping. Build it with your local zig version and report zig version in result, do not require a single fixed version.", "score_max": 10, "scoring": "Give 2.5 points per passed check for 10 total, 7 or more to pass.", "title": "Suture needle thread and tissue tearing"}
{"category": "Zig applications", "checks": ["Leaf light capture within 2 percent of reference and tips reach light zone in 9 of 10 runs", "Tendril wrap max penetration under 0.5 mm with zero tunneling on poles and wires", "Step under 3 ms per frame on one core with zero allocations after init", "Bit identical hashes across two runs with traces and timings reported"], "difficulty": "expert", "expected_result": "Single Zig repo plus verification.json with light capture error, reach rate, penetration depth, frame time, allocation counters and hashes from two runs.", "id": "AppVineGrowthLight52", "pass_score": 7, "prompt": "Hi, my greenhouse robotics group needs a believable climbing vine simulator for a controller demo next month and I am stuck on the growth math. Could you build a pure Zig 3D vine growth app with phototropism toward a moving sun lamp, auxin driven tip extension, lateral branching, and tendril coiling around thin supports? Model each shoot tip with light sensing and shade falloff, gravity response, auxin transport delay, contact detection against poles and trellis wires, plus coiling springs that tighten after touch and gentle wind sway. Simulate 120 days growth at hourly steps from seed 0x2A1E with three moving light paths, two trellis layouts, and one occluded lamp case where the vine must find a gap. I need total leaf light capture within 2 percent of reference, tips reaching the light zone in 9 of 10 runs, tendril wrap with max penetration under 0.5 mm and zero tunneling, stable steps under 3 ms per frame on one core, bit identical hashes across two runs, and zero allocations after init. Build it with your local zig version and report zig version in result, do not require a single fixed version. Please write verification with growth traces and timings. Thanks so much!", "score_max": 10, "scoring": "Give 2.5 points for each passed check for 10 total, 7 or more is a pass.", "title": "Vine Growth to Light with Tendril Coiling"}
{"category": "Zig applications", "checks": ["Mean flow within 3 percent of Womersley reference across all four pressure cases", "Clot volume within 5 percent across reruns with blockage flagged within one beat", "Step under 4 ms on one core with zero allocations after init", "Bit identical hashes across two runs with flow traces reported"], "difficulty": "expert", "expected_result": "Single Zig repo plus verification.json with flow error, clot volume, blockage timing, frame time, allocation counters and hashes from two runs.", "id": "AppBloodClotFlow53", "pass_score": 7, "prompt": "Hi, our vascular training team needs a realistic vessel flow simulator with platelet clotting for a bleeding control demo and I am lost on the fluid coupling. Could you build a pure Zig 3D blood vessel flow app with pulsatile pressure driven flow, suspended platelet agents, adhesion to injured wall patches, aggregation into a growing clot, and full blockage feedback that slows downstream flow? Model one branching artery with a stenosis plus injury site, non Newtonian viscosity, shear dependent adhesion probability, platelet activation delay, clot porous resistance, embolus risk if shear exceeds threshold, and red cell enhanced margination toward walls. Simulate 60 heartbeats at 120 Hz from seed 0xB10D with normal pressure, low pressure, high shear, and heparin reduced adhesion cases. I need mean flow within 3 percent of analytic Womersley reference, clot volume within 5 percent across reruns, complete blockage detected within one beat when it occurs, steps under 4 ms on one core, bit identical hashes twice, and zero allocations after init. Build it with your local zig version and report zig version in result, do not require a single fixed version. Please write verification with flow traces and clot metrics. Thank you!", "score_max": 10, "scoring": "Give 2.5 points for each passed check for 10 total, 7 or more is a pass.", "title": "Vessel Flow with Platelet Clot Blocking"}
{"category": "Zig applications", "checks": ["Tidal volume within 5 percent of reference in stable and weaning phases", "Trigger delay under 80 ms in 95 percent of breaths with zero missed apnea alarms", "Step under 2 ms on one core with zero allocations after init", "Bit identical hashes across two runs with breath tables reported"], "difficulty": "expert", "expected_result": "Single Zig repo plus verification.json with tidal volume error, trigger delays, alarm log, frame time, allocation counters and hashes from two runs.", "id": "AppLungVentilatorWean54", "pass_score": 7, "prompt": "Hi, our ICU training group needs a lung ventilator weaning simulator that handles spontaneous breathing sync and I am worried about timing edge cases. Could you build a pure Zig lung ventilator app with a single compartment lung model plus chest wall, airway resistance and compliance, ventilator pressure and volume modes, trigger detection from patient effort, and a gradual weaning protocol that reduces support? Model patient drive with breath to breath variability, muscle fatigue, cough and apnea events, sensor noise and delay, tube leak, asynchrony types like double trigger and missed trigger, oxygenation and carbon dioxide washout, and alarms for high pressure and low volume. Simulate 30 minutes at 100 Hz from seed 0x1A64 with stable, weak drive, strong drive, and apnea recovery scenarios. I need tidal volume within 5 percent of reference, trigger delay under 80 ms in 95 percent of breaths, zero missed apnea alarms, steps under 2 ms on one core, bit identical hashes twice, and zero allocations after init. Build it with your local zig version and report zig version in result, do not require a single fixed version. Please write verification with breath tables and alarm logs. Thanks a lot!", "score_max": 10, "scoring": "Give 2.5 points for each passed check for 10 total, 7 or more is a pass.", "title": "Lung Ventilator Weaning with Breath Sync"}
{"category": "Zig applications", "checks": ["Center of mass recovered within 5 cm in 8 of 10 nudges with zero false negative falls", "Sway power within 10 percent across reruns in eyes open and closed cases", "Step under 1.5 ms on one core with zero allocations after init", "Bit identical hashes across two runs with sway traces reported"], "difficulty": "expert", "expected_result": "Single Zig repo plus verification.json with recovery rate, fall detection, sway power, frame time, allocation counters and hashes from two runs.", "id": "AppBalanceBoardVestibular55", "pass_score": 7, "prompt": "Hi, our rehab lab needs a balance board simulator that captures human wobble and noisy proprioception for fall risk study and I am stuck on the control loop. Could you build a pure Zig human balance app with a wobble board on a hemisphere pivot, an inverted pendulum body with ankle and hip torque, vestibular and proprioceptive sensors with noise and delay, muscle activation lag and torque limits, and a neural controller with ankle then hip then step strategy? Model board tilt friction, foot slip limits, visual delay, sudden platform nudges, eyes open versus eyes closed, fatigue that lowers max torque, and random sensory dropouts. Simulate 120 seconds at 500 Hz from seed 0xBA1ACE with quiet stance, four direction nudges, eyes closed sway, and fatigue cases. I need center of mass recovered within 5 cm of upright in 8 of 10 nudges, fall detection with zero false negatives, sway power within 10 percent across reruns, steps under 1.5 ms on one core, bit identical hashes twice, and zero allocations after init. Build it with your local zig version and report zig version in result, do not require a single fixed version. Please write verification with sway traces and recovery tables. Thank you very much!", "score_max": 10, "scoring": "Give 2.5 points for each passed check for 10 total, 7 or more is a pass.", "title": "Balance Board with Vestibular Noise"}