📐 Mission Plan: 03_ml_workflow (Deep Analysis)

1. Chapter Context

Chapter Title: ML Workflow: Orchestrating the Lifecycle.
Core Invariant: The Iteration Tax (Velocity compounds into Quality).
The Struggle: Balancing the speed of experimentation against the depth of each individual experiment.
Target Duration: 45 Minutes.

Track	Persona	Fixed North Star Mission	The "Iteration Tax"
Cloud Titan	LLM Architect	Maximize Llama-3-70B serving on a single H100.	Data Egress. $90k/PB moving between regions.
Edge Guardian	AV Systems Lead	Deterministic 10ms safety loop on NVIDIA Orin.	Real-world Testing. Safety certs take 2 weeks per change.
Mobile Nomad	AR Glasses Dev	60FPS AR translation on Meta Ray-Bans.	App Store Review. 1-week cycle for every FW update.
Tiny Pioneer	Hearable Lead	Neural isolation in <10ms under 1mW.	Data Collection. Manual ear-sim tests take 1 month.

Objective: Map the time-on-task for each of the 6 lifecycle stages and identify the "Bottleneck Stage."
The "Lock" (Prediction): "Which stage currently accounts for >60% of your total development time?"
The Workbench:
- Sliders: Sensitivity of 6 Stages (Data Prep, Labeling, Training, Eval, Deploy, Monitoring).
- Instruments: LifecycleWaterfall (Time per stage), CostPropagationGauge (The 2^{N-1} multiplier).
- The 5-Move Rule: Students must simulate at least 5 different "Optimization Strategies" (e.g. automating labeling vs buying GPUs) to find the highest leverage point.
Reflect: "You automated 'Training' but the project iteration time only dropped 5%. Identify the 'Hidden Tax' in your lifecycle using the waterfall data."

Objective: Compare a "Heavy/SOTA" model vs. a "Light/Iterative" model over a 6-month window.
The "Lock" (Prediction): "If the SOTA model starts with 5% higher accuracy but iterates 10x slower, which model will win in Month 6?"
The Workbench:
- Sliders: Model Complexity (Large/Slow -> Small/Fast), Engineering Budget Allocation ($), Iteration Count.
- Instruments: compound_accuracy_plot (Time vs Accuracy growth), IterationTaxMeter.
- The 15-Iteration Rule: Students must find the "Knee of the Iteration Curve"—the point where further reducing model size results in "Speed without Signal."
Reflect: "Your CFO wants to buy the larger model because it has better 'Day 1' metrics. Use the 6-month projection to prove why iteration velocity is a higher-value feature for this mission."

Objective: Negotiate a "De-scoped" mission that is actually achievable within the project deadline.
The "Lock" (Prediction): "Given our current iteration tax, what is the maximum achievable accuracy before the 6-month deadline?"
The Workbench: All lifecycle variables unlocked.
The "Stakeholder" Challenge: The Product Manager demands "99% or bust." The student must use the Constraint Propagation math to prove that 99% requires more iterations than the calendar allows. Propose an achievable target.
Reflect (The Ledger): Define your final "Engineering Pace." How many iterations per week are you committing to? Justify the risk of the "Verification Gap" in your plan.

Primary: IterationProjectionPlot (Showing the compound effect of fast cycles).
Secondary: LifecycleWaterfall (Showing where the weeks go).
Transparency: Toggle for Cost(N) = 2^{N-1} propagation formula.