MLSysBook Labs

Reusable belts and knobs for track-aware lab content

A local reference for the components used to assemble labs: fixed pedagogical flow, track-specific realization, interactive controls, evidence views, source traces, reflections, decisions, and downloadable reports.

4 canonical tracks: iPhone, Oura Ring, RoboTaxi, Cloud Fleet

5 repeatable belts for opening, prediction, trade-off, synthesis, and report

18 modalities across controls, plots, evidence, and decisions

1 source-of-truth rule: facts and equations come from MLSysIM

System

Every lab combines the same three layers. The layer boundary is what keeps track changes clean instead of scattering constants through notebooks.

Layer 1

Pedagogy Kernel

Concept, prediction, evidence, reflection, decision, and takeaway remain stable across tracks.

Layer 2

Track Realization

Track profile changes narrative, defaults, hardware refs, metrics, thresholds, and stakeholder pressure.

Layer 3

Modality Stack

Selectors, sliders, plots, tables, callouts, ledgers, and report components make the idea visible.

Boundary

Single Source

MLSysIM owns hardware, model, infrastructure facts, and solver equations. Labs own only typed scenario variants.

Canonical Tracks

Students should pick one track once. Later labs read that track from the Design Ledger unless an instructor allows switching.

iPhone Mobile

Hardware source: Hardware.Mobile.iPhone15Pro
Primary metrics: Battery drain, thermal headroom, on-device latency, memory
Typical narrative: Ship a responsive local model without overheating or draining the phone.

Oura Ring TinyML

Hardware source: Hardware.Tiny.OuraRing
Primary metrics: SRAM and flash fit, battery life, OTA payload, sampling cadence
Typical narrative: Fit sensing and inference into a tiny wearable budget.

RoboTaxi Edge

Hardware source: Hardware.Edge.RoboTaxi
Primary metrics: p99 and p999 latency, rare-event recall, power, reliability
Typical narrative: Keep perception and planning inside tight latency and safety guardrails.

Cloud Fleet Fleet

Hardware source: Cloud fleet profile backed initially by Hardware.Cloud.H100
Primary metrics: Throughput, p99 latency, cost/request, utilization, carbon
Typical narrative: Run a service under SLA, cost, capacity, and sustainability constraints.

Belts

Belts are repeatable lab flows. Pick the flow, then plug in the track variant and modality stack.

Flow selector

Opening Belt

Pedagogical job

Core modalities

Output

Knob Demo

The same component can render different narratives and constraints after the track changes. Values below are illustrative placeholders; implementation should pull real facts from MLSysIM.

Inputs

Track

Primary knob 6 bit

Strategy

Stack builder

Calibration pass Guardrail validation Rollback plan

iPhone realization

Pareto Frontier

Phase Diagram

Budget Stack

Timeline / Topology

Modality Catalog

Search or filter by role. Each entry names when to use it, what the student does, and the shape of the output contract.

No modalities match the current filters.

Recipes

These recipes show how belts and knobs combine into a lab part or an entire compression pilot.

Lab unit	Pedagogical move	Recommended stack	Report evidence
Opening	Anchor the chapter and selected track.	Track selector, chapter recap, scenario strip, lab map.	Track ID, hardware ref, scenario ID, learning objectives.
Part A	Make a misconception visible before reveal.	Multiple-choice prediction, strategy selector, constraint budget, source trace.	Prediction, actual outcome, failed or binding constraint.
Part B	Show that the answer is a frontier, not a single knob.	Numeric prediction, slider, Pareto frontier, phase diagram, table fallback.	Selected point, dominated choices, guardrail violations, reflection.
Part C	Build a deployable recipe and name residual risk.	Stack builder, budget stack, before/after comparison, decision card.	Final configuration, primary metric, guardrail metric, residual risk.
Synthesis	Defend an engineering judgment.	Evidence summary, big takeaways, source trace, report export.	Markdown report plus optional JSON snapshot.

Source Documents

This page is a visual companion to the planning docs. The docs remain the planning source.

Modality Catalog Reusable components, controls, plots, evidence views, and report artifacts. Structure Contract Required lab headers, part headers, synthesis sections, and report schema. Source-Of-Truth Policy Ownership boundary for MLSysIM facts, solver equations, and lab variants.