Deep Reference Library
The projects are the primary path. These chapters are the deep library. Open one when the current build creates a concrete question, then return to code.
Rust language by project pressure
Section titled “Rust language by project pressure”| When the project forces this question | Reference |
|---|---|
| Who owns the website, media, or context buffer? | Ownership for AI payloads |
| How can a parser return views without copying? | Borrowing, slices, and lifetimes |
| Which states belong in structs versus enums? | Structs, enums, and patterns |
| How should absence and failure cross boundaries? | Types and errors |
| How do model and tool adapters stay replaceable? | Traits and workspaces |
| When should runtime provider selection use trait objects? | Trait objects and runtime providers |
| How do fallible pipelines stay readable? | Iterators and fallible transformations |
| How do shared buffers cross tasks safely? | Smart pointers, interior mutability, Send, and Sync |
| What must an FFI or native model boundary guarantee? | FFI and unsafe boundaries |
Production Rust by product pressure
Section titled “Production Rust by product pressure”| Need | Reference |
|---|---|
| Async model calls and cancellation | Async and backpressure |
| Bounded workers and queues | Channels, semaphores, and backpressure |
| A typed command surface | Clap CLI architecture |
| HTTP request boundaries | Axum router, state, and errors |
| Database schema and atomic state | SQLx migrations and transactions |
| Restart-safe execution | Durable jobs, leases, and idempotency |
| Upload and event streaming | Streaming uploads and SSE |
| Identity and data isolation | Authentication, authorization, tenancy, and quotas |
| Logs, metrics, traces, and redaction | Structured observability |
| Tests beyond the happy path | Unit, integration, golden, property, fuzz, and chaos tests |
| Deployment and recovery | Containers, health, shutdown, rollout, and rollback |
Model and multimodal foundations
Section titled “Model and multimodal foundations”- What a model artifact contains
- Tokens, embeddings, tensors, and transformer mechanics
- Sampling, seeds, and reproducibility
- Prefill, decode, KV caches, and context cost
- Capability, knowledge, instructions, and calibration
- Remote, embedded, or service inference
- Quantization, memory, batching, and device
- Evidence envelopes
- Provenance and locators
- Images
- Audio
- Video
- Timeline alignment
- Multimodal retrieval
Harness, agents, and evaluation
Section titled “Harness, agents, and evaluation”- Behavior specifications and completion
- Task, environment, policy, budget, and trace
- Structured tools
- Typed tools, effects, capabilities, and receipts
- Agent loops, progress, stopping, and repair
- Context manifests and budget allocation
- Retrieval contracts and snapshots
- Layered validation
- Retry, fallback, and idempotency
- Causal traces
- Strict and forked replay
- Datasets and graders
- Executable product requirements
- Trajectories and experiments
Small models, tuning, security, and operations
Section titled “Small models, tuning, security, and operations”- Decompose tasks
- Verify and route
- Spend test-time compute
- Fine-tune, retrieve, or redesign
- LoRA, QLoRA, and DPO
- Threat-model the agent
- Observe quality, latency, and cost
Advanced packaged labs
Section titled “Advanced packaged labs”After the product day introduces a boundary, use the Connected Mini-Projects to attack it more deeply. The packaged behavior, context, memory, tool, and replay labs are advanced hardening exercises—not prerequisites for creating your first useful artifact.