2026 07 16 Data Scientist 1100¶

🎯 Session Summary¶

I walked window 1 of 3 of the Data Scientist's Level 1100 (Data Engineering, Master tier) — five main_quest files, dependency-sorted, played end-to-end by the sealed agentic execute engine and then re-read by me as a linked learner journey. The headline is fail: only the one genuinely on-theme data quest holds up.

Data Warehousing (the single SQL / dimensional-modeling quest in this window) is excellent — 92%, every runnable snippet ran and matched the documented output. The other four quests are not really data-science content at all; they are GitHub-Actions / agentic-CI automation chapters from two unrelated quest lines that happen to carry the 1100 level tag. Each of those four ships copy-pasteable code with a verified runtime bug: two share the identical leaked-Jekyll-{% raw %} artifact that breaks ${{ ... }} expressions, and two share GitHub-Actions permission / git-identity defects. A data-scientist learner who reached this level for warehousing would sail through the first quest and then hit four broken automation workflows that a maintainer should fix before learners rely on the code.

🗺️ The Journey¶

#	Verdict	Quest	Score	One-line takeaway
1	✅ pass	Data Warehousing: Build a Dimensional Star Schema in SQL	92	Every SQL/DuckDB snippet executed as documented; only a minor SQLite-FK-enforcement caveat.
2	⚠️ warn	The Editor's Eye: A Reviewer, and the Dragon It Becomes	64	Sound design, but a real YAML syntax error in the BROKEN example + an actionlint-confirmed script-injection anti-pattern in the "safe CI" reference code.
3	❌ fail	The Bard Forge: The Quest That Writes Quests	57	Logic works, but the `contents: read`-only permissions block 403s the script's own `gh pr view`, and the proposal-PR `git commit` fails (exit 128) with no git identity.
4	❌ fail	The Agent Pantheon: Multi-Agent Lifecycle Management	54	Leaked `{% raw %}` breaks the artifact-name expression, two referenced scripts never provided, and a status-enum / workflow-filename mismatch.
5	❌ fail	The Autonomy Scales: Mapping Agent Autonomy Levels	59	Same leaked-`{% raw %}` artifact (×4) causes a verified `bad substitution`; `\n` in a double-quoted PR body never becomes a newline; unscoped PR lookup.

🔬 Evidence¶

All outcomes below come from commands the execute engine actually ran in its disposable per-quest sandbox (quoted/trimmed from walk-evidence.json). Snippet coverage is reported as ran/passed/failed/reasoned against the recorded snippets.

1. Data Warehousing — ✅ 92 (ran 7/9 runnable, 7 passed, 0 failed, 2 reasoned)¶

Linux setup python3 -m venv .venv && … pip install … duckdb → "ran cleanly, installing duckdb 1.5.4 into the venv with no errors." passed
Star-schema DDL (dim_date, dim_product, dim_customer, fact_sales) → "executed in SQLite 3.45.1 with zero errors; .tables confirmed all four tables created." passed
Revenue-by-city-by-quarter GROUP BY → "returned correct sensible aggregates (Denver 199.99, Austin 49.95 + 29.97 split by quarter) exactly as the quest promises." passed
Type-2 SCD walkthrough → "produced exactly the documented result: one closed historical row … and one current row." passed
columnar_demo.py in DuckDB → "ran … in 0.36s, generating 5M rows and returning COUNT/SUM/AVG as documented — no modification needed." passed
macOS/Windows setup blocks → reasoned (Linux sandbox; syntax standard, consistent with the verified Linux path).
Accuracy caveat (reasoned): the REFERENCES clauses imply enforced FK integrity, but "SQLite does not enforce foreign keys unless PRAGMA foreign_keys = ON is set — the quest never mentions this."

2. The Editor's Eye — ⚠️ 64 (ran 6/6 recorded, 5 passed, 1 failed)¶

Smuggle-guard logic — content-only diff → "'Smuggle guard passed: content-only.'"; a diff adding .github/workflows/evil.yml → "correctly failed with exit 1." passed
Loop-breaker (git log -1 --format='%an' vs content-review[bot]) → "correctly produced skip=true for a bot-authored commit and skip=false for a human-authored commit." passed
BROKEN example YAML with: { ref: ${{ github.head_ref }} } → failed: "both PyYAML … and Ruby's Psych … reject it, because the flow-mapping { } collides with the unquoted ${{ }} expression syntax. A learner pasting this exact block gets a hard YAML syntax error."
actionlint (v1.7.12) on Chapter 1 + boss-fight "SLAIN" workflow → "both flag git push origin HEAD:${{ github.head_ref }} … as … 'potentially untrusted. avoid using it directly in inline scripts.' — a genuine script-injection anti-pattern." passed (finding confirmed by the run)
fetch-depth clash — combining the boss-fight fetch-depth: 2 checkout with the Chapter 2 smuggle-guard git diff "origin/${BASE}...HEAD" → "fails with 'fatal: ambiguous argument origin/main...HEAD'." reasoned via reproduction

3. The Bard Forge — ❌ 57 (recorded 7, ran 5, 4 passed, 1 failed, 1 skipped, 1 reasoned)¶

scripts/mine_merge.sh end-to-end (mocked gh, real squash-merge repo) → "The script's core logic … worked exactly as described and produced correct output." passed
Permissions bug — Chapter 1 sets permissions: contents: read only → "specifying any permission explicitly sets all unlisted scopes to none — including pull-requests. gh pr view … would fail with a 403." failed (verified against GitHub's model)
Git-identity bug — Chapter 2 "Open proposal PR" runs git commit on a fresh clone → "I reproduced this exact scenario … and got fatal: empty ident name … not allowed (exit 128)." failed (reproduced live)
Quest Network mermaid → "Rendered … with mmdc (v10) successfully." passed

4. The Agent Pantheon — ❌ 54 (ran ⅚, 3 passed, 2 failed, 1 reasoned)¶

_data/agents.yml → "parses cleanly via PyYAML — 3 valid agent entries." passed
Leaked Jekyll tag — name: health-report-${% raw %}{{ github.run_id }}{% endraw %} → "the resulting string is health-report-${% raw %}{{ github.run_id }}{% endraw %} — NOT ${{ github.run_id }} … the artifact name would literally contain the Jekyll tag text." failed
Missing scripts — load_registry.py, agent_health_check.py → "neither script exists anywhere in the quest content; a learner … gets an immediate 'No such file or directory' failure on the very first job step." failed
python3 scripts/validate_quest.py --quest q17 → "failed with … No such file or directory — the validator is never provided or linked." failed
Internal contradictions (reasoned/accuracy): status: beta used in Chapter 3 but absent from the Chapter 1 enum; agent-health-check.yml vs the created agent-health-monitor.yml.

5. The Autonomy Scales — ❌ 59 (ran ⅖, 1 passed, 1 failed, 2 skipped, 1 reasoned)¶

_data/autonomy-matrix.yml → "parses cleanly with PyYAML (5 task_classifications …); yamllint only flags cosmetic issues." passed
Leaked Jekyll tag ×4 — ${% raw %}{{ github.event.issue.number }}{% endraw %} in feature-agent-l2.yml → "I extracted the exact run: string and executed it in bash: line 1: ${% raw %}…{% endraw %}: bad substitution, exit 1." failed
\n in PR body — --body "…\n\n…" in a plain double-quoted bash string → "bash does NOT convert \n to a real newline in this context — literal \n\n would land in the PR body." reasoned via reproduction
review_autonomy_compliance.sh — passes bash -n but "unscoped … gh pr list --search 'head:copilot/' … so every loop iteration would report the same (first) matching PR." skipped/reasoned (sandbox gh unauthenticated).

🐞 Issues Found¶

Grouped by quest; every item is evidenced by a command run (or an exact quoted line) above. No issue is raised for the passing Data Warehousing quest beyond its one minor accuracy caveat.

[high] The Editor's Eye · boss-fight BROKEN example (lines ~195–211) — Observed: with: { ref: ${{ github.head_ref }} } is a hard YAML syntax error (PyYAML + Ruby Psych both reject the flow-mapping/${{ }} collision). Fix: quote the expression inside the flow mapping or use block style, so the block fails for the intended pedagogical reason, not on paste.
[high] The Editor's Eye · Chapter 1 + boss-fight git push — Observed: actionlint flags git push origin HEAD:${{ github.head_ref }} in run: as an untrusted-expression script-injection risk — in a quest whose whole premise is "event-safe" CI. Fix: pass github.head_ref through env: and reference "$HEAD_REF"; add an explicit callout.
[medium] The Editor's Eye · fetch-depth mismatch — Observed: naively combining the boss-fight fetch-depth: 2 checkout with the Chapter 2 smuggle guard's origin/main...HEAD diff → "ambiguous argument … unknown revision". Fix: show one consolidated workflow with a single consistent fetch-depth: 0.
[high] The Bard Forge · Chapter 1 permissions — Observed: permissions: contents: read alone zeroes pull-requests, so mine_merge.sh's gh pr view 403s. Fix: add pull-requests: read.
[high] The Bard Forge · Chapter 2 "Open proposal PR" — Observed: git commit on a fresh clone with no identity → exit 128 (reproduced). Fix: add git config user.name / user.email before the commit.
[high] The Agent Pantheon · Chapter 2 artifact name — Observed: leaked ${% raw %}{{ github.run_id }}{% endraw %} yields a literal Jekyll string, not the run ID. Fix: wrap the whole fenced block in {% raw %}/{% endraw %} at the Markdown level so the visible YAML is plain ${{ github.run_id }}.
[high] The Agent Pantheon · missing scripts — Observed: load_registry.py, agent_health_check.py, and scripts/validate_quest.py are referenced but never provided → "No such file or directory". Fix: provide/link them or mark clearly as platform-supplied and out of scope.
[medium] The Agent Pantheon · internal inconsistencies — Observed: status: beta absent from the documented enum; agent-health-check.yml vs the created agent-health-monitor.yml. Fix: reconcile the enum and the filename everywhere.
[high] The Autonomy Scales · Chapter 3 feature-agent-l2.yml — Observed: leaked ${% raw %}{{ github.event.issue.number }}{% endraw %} (×4) → verified bad substitution at runtime. Fix: strip the leaked raw tags so the block renders valid ${{ github.event.issue.number }}.
[medium] The Autonomy Scales · PR-body \n — Observed: \n\n in a plain double-quoted bash arg stays literal. Fix: use $'...\n\n...' or a heredoc.
[medium] The Autonomy Scales · review_autonomy_compliance.sh — Observed: unscoped gh pr list --search "head:copilot/" returns the first match every iteration. Fix: scope the lookup to the specific run/branch.
[low] Data Warehousing · SQLite FK enforcement — Observed: REFERENCES clauses imply enforced integrity but SQLite needs PRAGMA foreign_keys = ON. Fix: add a one-line note (or the PRAGMA) so learners don't over-trust FK validation.

🔗 Chain Continuity¶

The most important finding of this walk is structural: Level 1100 here is a difficulty/theme band, not a single learning path, and this window interleaves three unrelated quest lines:

data-warehousing — quest_line "The Data Engineer's Ascent" (SQL). Its declared required_quests is /quests/1100/etl-pipeline-design/ and it unlocks apache-spark / data-quality — a coherent Data-Engineering spine, and the only quest that matches the Data Scientist's stated Level-1100 theme.
self-operating-website-06 and -10 — quest_line "The Self-Operating Website" (GitHub Actions / Claude Code). Their prerequisites are Chapter V (/quests/1010/…) and Chapter IX (/quests/1110/…) respectively — neither is in this window, and they are two non-adjacent chapters (06 and 10) of a serialized campaign, so a learner meeting them here lands mid-story with the intervening chapters missing.
agentic-multi-agent-lifecycle-management (Q17) and agentic-autonomy-levels-matrix (Q18) — quest_line "gh-600 / The Agentic Codex". These are the only genuine dependency edge inside the window: Q17's unlocks_quests lists Q18, Q18's required_quests lists Q17, and the planner ordered them correctly (Q17 → Q18). Q17 still assumes Q16 (/quests/1011/agentic-multi-agent-failure-recovery/), outside the window.

Prerequisite reality: 4 of the 5 quests declare required/recommended prerequisites that live outside this window (earlier levels or other chapters). That is expected for a windowed sweep (window 1 of 3; the ledger accumulates the rest), but it means this window does not stand alone as a journey — a Data Scientist arriving for "Data Engineering mastery" gets one on-theme quest and four DevOps/agentic-automation quests from stories they haven't started.

Cross-quest defect patterns (the real value of reading the chain as one journey): - The leaked Jekyll {% raw %} artifact appears identically in both gh-600 quests (#4 and #5) — this is a systemic rendering bug in the gh-600 source, not two isolated typos, and any fix pass should sweep the whole quest_line: gh-600 for the same pattern. - The GitHub-Actions safety defects (permission scoping in #3, script-injection in #2, git-identity in #3) cluster in the two Self-Operating Website quests — ironic given both quests explicitly teach "safe/event-safe automation." A learner following that line literally copies an anti-pattern the quest is trying to warn against.

Net: within the window the only edge that actually holds together for a learner is Q17 → Q18, and even that edge carries the same broken workflow code in both quests.

🧠 Reasoning & Method¶

Mode: execute — the workflow pre-computed and sealed walk-evidence.json / walk-evidence.md by running test/quest-validator/agentic_validate.py --mode execute in a disposable per-quest sandbox before I started. I consumed that evidence as-is and did not re-run the engine (its child claude processes can't authenticate from my Bash tool). Every passed/failed above is an engine-run command; anything I marked reasoned is a static/simulated judgment the engine labeled as such.
What I ran vs. reasoned: I ran no quest commands myself — my Bash use was limited to reading walk-plan.json, the two evidence files, and the five quest sources. My value-add is the linked-journey analysis in §🔗 (dependency graph, window composition, and the two cross-quest defect clusters), derived from reading each quest's frontmatter/dependencies in plan order against the sealed per-quest verdicts.
Coverage & limits: This is window 1 of 3 (offset 5) of a 15-quest level — I walked exactly the 5 planned quests and did not expand to the rest of the level; the ledger sweeps the other windows over subsequent runs. macOS/Windows setup blocks and any auth-gated gh/claude invocations were reasoned/skipped (Linux, unauthenticated sandbox), as the engine notes. Snippet coverage per quest is stated numerically in §🔬.
Confidence: High on the four verified runtime bugs (each reproduced or confirmed against a documented model in the sandbox) and on Data Warehousing's clean pass. High on the continuity findings (they derive directly from the quests' own quest_dependencies and quest_line frontmatter). The three fail verdicts reflect real, copy-paste-breaking defects, not stylistic nits — they should be fixed before learners rely on this window's automation code.

Appendix — machine evidence (verbatim excerpt)¶

5 quests evaluated · ✅ 1 pass · ⚠️ 1 warn · ❌ 3 fail · avg 65.2% · ~$3.8435

Score Quest Snippets run

✅ 92 Data Warehousing: Build a Dimensional Star Schema in SQL 7/9

⚠️ 64 The Editor's Eye: A Reviewer, and the Dragon It Becomes 6 (1✗)

❌ 57 The Bard Forge: The Quest That Writes Quests 5 (1✗)

❌ 54 The Agent Pantheon: Multi-Agent Lifecycle Management 5 (2✗)

❌ 59 The Autonomy Scales: Mapping Agent Autonomy Levels 2 (1✗)

	Score	Quest	Snippets run
✅	92	Data Warehousing: Build a Dimensional Star Schema in SQL	7/9
⚠️	64	The Editor's Eye: A Reviewer, and the Dragon It Becomes	6 (1✗)
❌	57	The Bard Forge: The Quest That Writes Quests	5 (1✗)
❌	54	The Agent Pantheon: Multi-Agent Lifecycle Management	5 (2✗)
❌	59	The Autonomy Scales: Mapping Agent Autonomy Levels	2 (1✗)