feat: add quiz.json for seven orphan lessons (07/15-16, 08/19, 10/15-17, 10/34)

[BlueSkyGTM]

Summary

Adds canonical six-question quiz.json files (1 pre / 3 check / 2 post) for seven lessons that already ship docs/en.md and code/ but had no on-site assessment loop.

One commit per lesson (AGENTS.md convention).

Lessons

• 07/15-attention-variants
• 07/16-speculative-decoding
• 08/19-visual-autoregressive-var
• 10/15-speculative-decoding-eagle3
• 10/16-differential-attention-v2
• 10/17-native-sparse-attention
• 10/34-gradient-checkpointing

Questions are grounded in each lesson's docs/en.md and Build demos — not generic trivia.

Not included

• Capstone quiz fixes, audit script changes, agent rules, quiz-factory, or README catalog row updates (happy to open a separate PR for README/ROADMAP wiring if useful).

Test plan

• python3 scripts/audit_lessons.py passes on CI
• Spot-check one quiz against docs/en.md (e.g. 07/15)

[coderabbitai]

📝 Walkthrough

Walkthrough

This PR adds seven quiz JSON files across multiple lesson modules in the curriculum, each containing structured multiple-choice assessments with six questions organized into pre-, check-, and post-stage categories to support learner self-evaluation and concept reinforcement.

Changes

Transformer and Optimization Technique Quizzes

Layer / File(s)	Summary
Transformer architecture and attention variant quizzes phases/07-transformers-deep-dive/15-attention-variants/quiz.json, phases/07-transformers-deep-dive/16-speculative-decoding/quiz.json, phases/08-generative-ai/19-visual-autoregressive-var/quiz.json	Three quizzes added covering sliding-window attention compute and KV cache effects, interleaved global attention, speculative decoding acceptance mechanics and KV rollback, and visual autoregressive model generation order, latent representations, scale-wise attention, and ImageNet scaling properties.
LLM training and optimization technique quizzes phases/10-llms-from-scratch/15-speculative-decoding-eagle3/quiz.json, phases/10-llms-from-scratch/16-differential-attention-v2/quiz.json, phases/10-llms-from-scratch/17-native-sparse-attention/quiz.json, phases/10-llms-from-scratch/34-gradient-checkpointing/quiz.json	Four quizzes added covering Leviathan rejection sampling with EAGLE-3 draft selection, differential attention operations and learned scalar modulation, native sparse attention trade-offs and three-path architecture, and gradient checkpointing memory-compute trade-offs with recursive vs. uniform strategies.

Estimated code review effort

🎯 1 (Trivial) | ⏱️ ~4 minutes

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly describes the main change: adding quiz files for seven specific lessons with their identifiers, which matches the changeset content.
Description check	✅ Passed	The description is directly related to the changeset, providing context about what quiz.json files were added, which lessons were updated, and explaining the scope of work.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[coderabbitai]

🧹 Nitpick comments (1)

phases/10-llms-from-scratch/15-speculative-decoding-eagle3/quiz.json (1)

1-78: ⚡ Quick win

All correct answers are in position 0 across all four quiz files.

Every question in this file (and all three other quiz files in this PR) has "correct": 0. This pattern creates predictable quizzes where learners can achieve perfect scores by always selecting the first option, reducing educational value.

Consider varying the position of correct answers across questions to prevent gaming and ensure learners are genuinely engaging with the content rather than pattern-matching.

As per coding guidelines: Quiz files must follow the quiz.json schema with exactly 6 questions: 1 pre + 3 check + 2 post. correct is zero-indexed. (The schema is met; this comment addresses quiz quality.)

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@phases/10-llms-from-scratch/15-speculative-decoding-eagle3/quiz.json` around
lines 1 - 78, The quiz file 15-speculative-decoding-eagle3 has every question's
"correct" index set to 0, making answers predictable; update the "correct"
values across the question objects so the correct option index varies (not
always zero) while preserving the schema (6 questions: 1 pre, 3 check, 2 post)
and keeping the explanations intact—edit the "correct" fields in each question
object in the JSON to appropriate zero-based indices matching the true correct
option for that question.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Nitpick comments:
In `@phases/10-llms-from-scratch/15-speculative-decoding-eagle3/quiz.json`:
- Around line 1-78: The quiz file 15-speculative-decoding-eagle3 has every
question's "correct" index set to 0, making answers predictable; update the
"correct" values across the question objects so the correct option index varies
(not always zero) while preserving the schema (6 questions: 1 pre, 3 check, 2
post) and keeping the explanations intact—edit the "correct" fields in each
question object in the JSON to appropriate zero-based indices matching the true
correct option for that question.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: 8788b571-9e64-422d-b062-9b8107358d62

📥 Commits

Reviewing files that changed from the base of the PR and between 86bbb61 and 3b05e98.

📒 Files selected for processing (7)

• phases/07-transformers-deep-dive/15-attention-variants/quiz.json
• phases/07-transformers-deep-dive/16-speculative-decoding/quiz.json
• phases/08-generative-ai/19-visual-autoregressive-var/quiz.json
• phases/10-llms-from-scratch/15-speculative-decoding-eagle3/quiz.json
• phases/10-llms-from-scratch/16-differential-attention-v2/quiz.json
• phases/10-llms-from-scratch/17-native-sparse-attention/quiz.json
• phases/10-llms-from-scratch/34-gradient-checkpointing/quiz.json

[BlueSkyGTM]

Vercel check (not a PR defect)

The failing Vercel status is \Authorization required to deploy\ — the upstream project needs a maintainer with Vercel team access to authorize deployment from this fork PR. Same as #237.

This PR only adds \quiz.json\ files (no \site/, no lesson \code/). CodeRabbit passed. \�udit_lessons.py\ validates locally.

Related: catalog wiring for these lessons is in #237 (README/ROADMAP only). Independent PRs; merge in either order.

No graphify rebuild required for quiz-only changes.