Connected Mini-Projects
This book is designed to produce working judgment, not recognition. The practical track connects 36 mini-projects to the same companion workspace. You repeatedly extend one system from strict Rust domain types to a CLI, service, multimodal evidence pipeline, AI harness, evaluation system, model router, fine-tuning workflow, and production incident drill.
The mini-projects are not detached demos. Their types, fixtures, receipts, and failure investigations feed Forge, Signal Room, Cutroom, and the final Mosaic platform.
What you do in every lab
Section titled “What you do in every lab”predict ─► trace ─► implement ─► break ─► diagnose ─► improve ─► explain ▲ │ └──────────── rerun the same acceptance evidence ────────────────┘- Predict the accepted input, state transition, bounded resource, expected trace, and failing mutation before running the code.
- Trace one input through types, transformations, effects, receipts, and terminal state.
- Implement a narrow behavior that requires an engineering decision.
- Break it with a supplied malformed fixture, failpoint, timeout, stale generation, unauthorized effect, or adversarial input.
- Diagnose the first violated invariant rather than the last visible symptom.
- Improve one responsible layer and rerun unchanged cases.
- Explain the evidence, rejected alternative, limitation, and recovery path.
The 36-project path
Section titled “The 36-project path”| Project | Build | What survives into later work |
|---|---|---|
| MP-01 · Validated Task Envelope | Strict IDs and task states | Every run and API boundary |
| MP-02 · Source Span Workbench | Unicode-safe normalization | Restorable citations |
| MP-03 · Runtime Provider Registry | Capability-aware dispatch | Model routing |
| MP-04 · Concurrent Budget Ledger | Reserve/commit/release | Harness and worker admission |
| MP-05 · Isolated Native Boundary | Audited unsafe wrapper | Local inference and media |
| MP-06 · Forge Command Spine | Stable CLI behavior | Forge interface |
| MP-07 · Bounded Async Worker | Cancellation and backpressure | Services and workers |
| MP-08 · Streaming Upload and Resume | Bounded input and ordered events | Mosaic API |
| MP-09 · Durable Run Recovery | Leases and idempotency | Restart-safe execution |
| MP-10 · Tenant-Safe Artifact Store | Scoped content addressing | Evidence and outputs |
| MP-11 · Observable Service Rollout | Health, drain, rollback | Operations |
| MP-12 · Causal Inference Notebook | Token/tensor/sampling mechanics | Honest model contracts |
| MP-13 · Inference Capacity Planner | Cache/device capacity | Routing and admission |
| MP-14 · Capability-Neutral Route | Scripted/local/remote adapters | Provider independence |
| MP-15 · Four-Modality Intake | Text/image/audio/video envelopes | Signal Room and Cutroom |
| MP-16 · Provenance and Locator Graph | Lineage and source locations | Verifiable evidence |
| MP-17 · Timeline and Geometry | Reversible time/region maps | Multimodal reasoning |
| MP-18 · Retrieval and Citation | Fusion, diversity, citations | Grounded answers |
| MP-19 · Adversarial Laboratory | Mutations and fail-closed oracles | Security regression corpus |
| MP-20 · Behavior Contract Compiler | Completion evaluation | Definition of done |
| MP-21 · Context Budgeter | Versioned selection manifest | Controlled model input |
| MP-22 · Verified Memory Write Gate | Evidence, consent, conflicts, expiry | Safe durable memory |
| MP-23 · Authorized Tool Loop | Capabilities, exact approvals, effect receipts | Bounded agency |
| MP-24 · Replayable Model Loop | Stopping, repair, replay | Forge harness |
| MP-25 · Dataset and Split Builder | Versioned cases and splits | Evaluation corpus |
| MP-26 · Grader Calibration Clinic | Agreement and abstention | Trusted grading |
| MP-27 · Multimodal Evaluation Pack | Modality-specific graders | Signal Room/Cutroom gates |
| MP-28 · Statistical Release Gate | Paired slices and intervals | Release decisions |
| MP-29 · Decomposed Incident Triage | Narrow subproblems | Strong small-model workflow |
| MP-30 · Generate–Check–Repair | Independent feedback and search | Test-time compute |
| MP-31 · Difficulty-Aware Cascade | Calibrated escalation | Quality/cost frontier |
| MP-32 · Data Governance Pipeline | Rights, consent, contamination | Trainable corpus |
| MP-33 · Tiny SFT and LoRA | Adapter experiment | Tuning comparison |
| MP-34 · Preference Audit | Preference quality and gates | Safe optimization |
| MP-35 · Capability Red Team | Injection and exfiltration drills | Security release gate |
| MP-36 · Safe Degradation Incident | Overload and recovery | Final operational proof |
MP-20 through MP-24 are fully packaged labs. Earlier chapters already contain runnable companions and exercises; they will receive the same formal lab packaging as the expansion proceeds. The table distinguishes the committed sequence from what is implemented today.
Completion is evidence, not a feeling
Section titled “Completion is evidence, not a feeling”Every lab names:
- a stable lab ID and version;
- prerequisite chapters and labs;
- files you are expected to change;
- a focused command and full regression command;
- normal, boundary, failure, and security fixtures where applicable;
- exact or environment-sensitive expected evidence;
- supplied mutations and a root-cause investigation;
- a design-review question and completion rubric.
Your learner record should include the source revision, acceptance result, artifact or trace identity, diagnosed mutation, measurement environment, design choice, rejected alternative, known limitation, and the next dependency unlocked.
How the exercise levels grow
Section titled “How the exercise levels grow”| Level | Typical work | Proof |
|---|---|---|
| 1 · Recall | Classify a boundary or failure | Correct vocabulary |
| 2 · Predict | Trace ownership, state, or cost | Accurate pre-run prediction |
| 3 · Repair | Extend code or fix a mutation | Focused regression |
| 4 · Build | Complete and measure a mini-project | Artifact plus receipts |
| 5 · Defend | Review a threat model or architecture | Evidence-backed design memo |
Major chapters cover levels 1–3. Each chapter cluster culminates in levels 4–5. Solutions explain the invariant and test strategy without turning the exercise into transcription.
Start with the definition of done
Section titled “Start with the definition of done”Begin with MP-20 · Behavior Contract Compiler when you reach Harness Engineering. It is intentionally provider-free: the first practical lesson is that a model saying “done” is an observation, not proof of completion.