Six-week calibrated post-core / BYO-artifact overlay

Hard prerequisite gate

Do not begin W1 until both conditions are true:

1. You completed the S01–S12 core route (and the hard path if that is your chosen artifact), or you already have equivalent harness experience. If you are new to harnesses, do S01–S12 first.
2. You can name an inspectable agent/eval artifact and reopen its green, banked baseline. The hard-path café host can satisfy this artifact gate; otherwise bring your own.

Record the gate before starting with the study/PROGRESS.template.md record:

mkdir -p study/evidence
cp study/PROGRESS.template.md study/PROGRESS.md

Then update study/PROGRESS.md during each week.

Learning-purpose emphasis

This route pairs one defensible model-layer build with an existing agent/eval artifact. Choose which learning emphasis leads; complete and bank both evidence sets.

Choose the learning emphasis before ordering the evidence narrative. The same evidence set supports both emphases; only the leading evidence changes.

Evidence over claims

1. Bank before claiming. "Completed" means the named receipt exists, can be reopened, and identifies the run or commit that produced it. A watched lecture or course-completion badge is not implementation evidence.
2. Separate observed from inferred. Record commands, configuration, environment, raw measurements, failures, and the interpretation separately.
3. Keep work learner-owned. Bank your code, tests, measurements, and restatements; do not copy assignment text, external notebooks, course answers, or solution code.
4. Use honest tense. Until the A1 loss curve and sample generations are banked, say "building CS336 A1," not "built a Transformer from scratch."
5. Prefer one inspectable artifact to five vague bullets. Each week below has one required receipt. Stretch work never substitutes for it.

Scope boundaries before the calendar

• Stanford CS336, Spring 2026 supplies the single deep build: Assignment 1 (A1) only. A2 (systems) is deferred. A3 is scaling and A5 is alignment; they are named for orientation, not scheduled as six-week implementation deliverables.
• Stanford's honor-code boundary applies. Agents may help answer high-level and low-level questions, but may not solve an assignment. All assignment implementation, debugging decisions, tests, and explanations remain learner-owned.
• The DeepLearning.AI RLHF course is a managed Google Cloud pipeline experience. It is not a from-scratch reward-model (RM) or PPO implementation; describe the receipt accordingly.
• The DeepLearning.AI vLLM course is the optimize-deploy-benchmark block. Its receipt must preserve the optimization configuration, deployment command/configuration, benchmark workload, environment, raw result, and at least one before/after or baseline comparison number.

The six-week calendar

The hour split is a cap, not a suggestion to fill unused time with new scope.

W1 — Artifact baseline + managed post-training pass

• Base: Run the local harness verification gates and record their exact results. Complete the RLHF short course as a managed Google Cloud pipeline concept pass; do not describe it as hand-building an RM or PPO loop.
• Stretch: Draft a one-page artifact narrative that distinguishes mechanisms, measurements, and claims.
• Defer: New lesson authoring, new harness features, or a second post-training course.
• Banked evidence artifact: study/evidence/w1-baseline-and-rlhf.md, containing dated local gate commands/exit results plus a learner-written account of what the managed pipeline demonstrated and did not demonstrate.

W2 — Optimize, deploy, benchmark inference

• Base: Complete the vLLM optimize-deploy-benchmark sequence. Preserve enough detail for another engineer to distinguish optimization effects from workload or hardware changes.
• Stretch: Add the strongest benchmark comparison to the harness artifact narrative, with its workload and environment beside the number.
• Defer: CS336 A1. Starting it early is not permission to omit the inference receipt.
• Banked evidence artifact: study/evidence/w2-vllm-receipt.md, containing the optimization configuration, deployed model/server configuration, exact benchmark workload, environment, raw output location, and one labeled baseline-versus-result number. Do not include copied course solution code.

W3 — CS336 A1: concepts + tokenizer

• Base: Use the Spring 2026 CS336 materials needed for A1, then implement and test the A1 tokenizer yourself. Ask agents questions; do not ask them to solve the assignment.
• Stretch: Sketch the Transformer module interfaces in your own words before implementing them.
• Defer: A2 systems work, A3 scaling implementation, A5 alignment implementation, and any unrelated model project.
• Banked evidence artifact: study/evidence/w3-a1-tokenizer.md, containing the learner-owned commit SHA, test command/results, one failure investigated, and a restatement of the tokenizer invariants.

W4 — CS336 A1: Transformer + training loop

• Base: Implement the A1 Transformer and training loop. Debug with the smallest useful run before spending on a larger run. The code and debugging decisions stay learner-owned.
• Stretch: Capture an early training trace if the base implementation is already tested.
• Defer: A2, A3, A5, benchmark polishing, and harness feature work.
• Banked evidence artifact: study/evidence/w4-a1-training-smoke.md, containing the learner-owned commit SHA, architecture/configuration, test results, first successful training-smoke log, and one explained defect.

W5 — Finish and bank A1

• Base: Finish A1 and bank the result that makes the build claim truthful: a labeled loss curve plus sample generations tied to a commit and run configuration. Keep harness work to maintenance only.
• Stretch: Write a short comparison between the W1 managed RLHF experience and the A5 alignment concepts, without attempting A5 or copying course material.
• Defer: Every second coding assignment, especially A2. A3 scaling and A5 alignment remain conceptual orientation, not completed deliverables.
• Banked evidence artifact: study/evidence/w5-a1-result.md, containing the commit SHA, run configuration, loss-curve path, sample-generation path, test results, and limits on what the run proves.

W6 — Ship the harness artifact + evidence narrative

• Base: Re-run the harness gates, bank the release commit, and write the evidence narrative. Explain what the harness receipts demonstrate, what the A1 evidence demonstrates, and what neither evidence set demonstrates.
• Stretch: Take the optional free Anthropic API course for platform fluency. It does not replace the release receipt.
• Defer: The CCA exam. There is no CCA exam objective in this six-week route.
• Banked evidence artifact: study/evidence/w6-release.md, containing the release commit SHA, exact verification commands/results, artifact location, supported claims by evidence source, and an explicit list of unsupported claims.

Slip policy

Protect the two needle-movers: the learner-owned CS336 A1 build and the shipped harness artifact. If a week slips:

1. Cut stretch work first.
2. Carry only the unfinished base item that blocks A1 or the artifact receipt.
3. Keep A2, A3 implementation, A5 implementation, the optional Anthropic API course, and the CCA exam deferred.
4. Do not schedule a make-up binge. Re-estimate the remaining weeks using their original hour caps; if the receipts no longer fit, extend the calendar.

Honest longer alternative

If full CS336 A1 + A2 is non-negotiable, use roughly 10–12 weeks at 8–10 hours per week. Do not compress both assignments into this route. The six-week version buys A1 depth, bounded inference/post-training breadth, and a shipped harness artifact; it does not buy the full CS336 core.