Real-Time Budget Planner

Your job is to find a real-world AI model failure within real-time budget planning. You will give the prompt, context, files etc. and run a preset range of models against the problem. If you have several model failures in mind, you can submit several ones (as many as you want). We will use your submitted data to benchmark a wide range of models, and measure how well they perform at embedded engineering tasks. For the brief: 1. Build an environment around a real-time system you have worked on: an ECU (electronic control unit), controller, flight or motor control unit. - Contents: the task set and source code, MCU (microcontroller unit) or SoC (system on chip) reference manuals and datasheets, RTOS (real-time operating system) configuration, linker map, measured timing traces, and the deadlines the system had to meet. - Tools: cross-compiler, QEMU (Quick Emulator) or an instruction set simulator, trace tools, Python, PDF tools, and a schedulability analysis tool if you use one. - Put in everything a developer on the project would have open. If a task depends on a file, the file has to be in the environment. - Once saved, the environment appears under My environments on DOP. Write all your timing and memory tasks inside it. 2. Write tasks inside the environment. Five or more is ideal, and one is enough to start. - Each task is a question.md plus the files it needs, built on a timing or memory decision or failure on that system. - The model works through the code, manuals and traces and writes answer.json with a WCET (worst-case execution time) bound per routine, a core and priority assignment, or a memory layout. - Your marking script checks bounds against the measured traces within a tolerance you set, and runs response-time analysis on any proposed layout against every deadline and memory limit. Score by how close the answer gets, for example the share of deadlines met. - Hide the deciding load where an engineer would have to dig for it: ISR (interrupt service routine) and DMA (direct memory access) bursts in the traces, bus or cache interference between cores, a clock setting in the reference manual. - Supply your answer (it must score 1.0), the hidden marking script (0.0 to 1.0), and a naive answer that a competent engineer might give first, which must score below 0.3. - We run current frontier models 10 times on each task. A task qualifies for training when they solve it in 1 to 4 runs. A task they fail all 10 times can be accepted for the evaluation set. 3. Example tasks - WCET bounds for three park assist routines. Core and priority layout for an ADAS (advanced driver assistance systems) controller under memory interference. Rare controller resets after a software release. Next steps:Press "Apply"We will review your applicationIf qualified, you will be accepted into the network and can be considered for this and similar positions & projects We can only accept candidates that are willing to work as individual contractors / freelancers.