Data Scientist · L1100 · 2026-07-20

Slice data-scientist/1100 · Level 1100 (Data Engineering) · Master tier · Engine verdict ❌ fail (avg 64.0%) · Walked 2026-07-20

🔗 Perfection run · 🏠 Perfection dashboard · 📄 Raw report · 🕘 Change history

🎯 Session Summary

I walked a 5-quest window (window 2 of 3) of the Data Scientist → Level 1100 slice as a learner would, consuming the sealed execute-engine evidence and reading every quest source in plan order. The window is not one learning path — it splices two unrelated tracks that happen to share the 1100 level code: the GH-600 “Agentic Codex” arc (Warden’s Pact → Grand Capstone) and the Data Engineering Mastery arc (Apache Spark → Stream Processing → Data Quality). A data-scientist reading top-to-bottom would whiplash between GitHub governance and PySpark.

Headline verdict: FAIL for the slice. Three of five quests fail with high-severity, copy-paste-breaking bugs verified by real command execution: the Warden’s Pact ships three GitHub Actions workflows with literal {​% raw %​}/{​% endraw %​} tags inside ${​{ }​} expressions (every run: step throws bad substitution), Stream Processing pins a Kafka image (bitnami/kafka:3.7) that no longer exists on Docker Hub so no learner can start Chapter 2, and Data Quality never provides the orders.csv that all four of its snippets read (plus a Great Expectations example broken against the 1.x API pip now installs). The two passing quests are genuinely strong: Apache Spark (95%) ran all 7 executed snippets cleanly, and the Grand Capstone (85%) validator behaved exactly as documented in both failing and passing states. The most actionable maintainer takeaway: the correct fix for the Warden’s Pact templating bug already exists one quest away — the Capstone escapes the identical Liquid mechanism correctly.

🗺️ The Journey

🔬 Evidence

All outcomes below are from the sealed execute-engine pass (walk-evidence.json, mock:false, mode:execute) — commands actually run in a disposable sandbox. I did not re-run the engine (its child processes cannot authenticate from an agent Bash tool); I reasoned over the chain against the quest sources.

1. The Warden’s Pact — 34% ❌ · ran 4 extracted commands, 0 passed / 4 failed, 3 reasoned (1 of 7 snippets literally runnable)

• The only literally-runnable command, python3 scripts/validate_quest.py --quest q19, failed: can't open file '.../scripts/validate_quest.py': [Errno 2] No such file or directory — the script is never provided or linked.
• The engine extracted each run: step from the three workflow YAML blocks and executed them under bash -eo pipefail (Actions’ real default). Every step containing ${​% raw %​}{​{ ... }​}{​% endraw %​} failed with bad substitution, exit 1 — reproduced individually for Ch.2 (gh pr view "${​% raw %​}{​{ github.event.pull_request.number }​}{​% endraw %​}"), Ch.3 (echo "Deploying plan: ${​% raw %​}{​{ needs.prepare.outputs.deploy_plan }​}{​% endraw %​}"), and Ch.4 (the audit-log heredoc). Confirmed against source lines 155, 238, and 275–280.
• CODEOWNERS ordering bug confirmed by reading source (lines 121–133): the catch-all * @team-dev sits at the bottom, and GitHub CODEOWNERS is last-match-wins, so it silently overrides the @team-security rules on /src/auth/ and /src/crypto/ — the inverse of the guardrail’s stated purpose.

2. The Grand Capstone — 85% ✅ · ran 5, 5 passed / 0 failed, 4 reasoned (1 of 9 snippets runnable)

• validate-capstone.sh was created from its heredoc, chmod +x‘d, and run verbatim: against an empty repo it printed 14 ❌ lines and exited 1, exactly as the prose predicts (“one that fails a few times first was written honestly”). Re-run after synthesizing all 14 artifacts, it passed — the machine-verifiable contract works in both directions.
• jq, git, gh all present; the Seal-4 jq -e '.criteria | length >= 3' filter worked against a real JSON file. Both YAML blocks parsed cleanly (yamllint: cosmetic line-length only). The two JSON snippets only fail strict json.loads() because of a leading // path JSONC comment — cosmetic for illustrative config.

3. Apache Spark Mastery — 95% ✅ · ran 7, 7 passed / 0 failed, 1 skipped, 4 reasoned (10 of 12 runnable)

• Linux setup ran end-to-end; pip install pyspark pulled pyspark 4.2.0 and printed the version.
• Ch.1 spark_intro.py: printSchema() + df.show() produced the exact 3-row table shown. Ch.2 lazy snippet: result.show() → checkout | 105.0 and explain(mode="formatted") contained two Exchange nodes, exactly the shuffle marker the quest tells learners to hunt for. Ch.3 groupByKey/reduceByKey both returned identical [('b', 6), ('a', 4)]; coalesce(1) produced one output file and repartition(4,…) produced several. Ch.4 SQL explain("cost") and the cache/unpersist cycle both ran clean.
• Only gap: prereq says “JDK 11/17” (line 127), but pip’s pyspark 4.x requires Java 17+ — stale, not blocking.

4. Stream Processing — 56% ❌ · ran 6, 3 passed / 3 failed, 2 skipped, 3 reasoned (7 of 11 runnable)

• The gating infra command failed live: docker compose up -d against the Ch.2 compose file (image: bitnami/kafka:3.7, source line 258) → manifest for bitnami/kafka:3.7 not found: manifest unknown. Verified docker pull bitnami/kafka:3.7 also fails. This blocks every platform path at the first hands-on step.
• After swapping to bitnamilegacy/kafka:3.7, the broker started in KRaft mode and everything downstream worked with zero further edits — so the compose config itself is correct, only the image reference is stale. producer.py printed produced 5 events; the consumer printed partition=X offset=Y value=… per message and correctly showed offset tracking on re-run; the Ch.4 exactly-once producer (idempotence + transactions) ran and the message was committed/consumable.

5. Data Quality Engineering — 50% ❌ · ran 8, 5 passed / 3 failed, 1 skipped, 0 reasoned (9 of 9 runnable)

• No orders.csv is ever provided — all four snippets (profile.py, validate_ge.py, validate_simple.py, contract.py) call pd.read_csv("orders.csv") and fail with FileNotFoundError for a literal learner; the engine had to synthesize a CSV to proceed.
• Ch.3 Great Expectations snippet failed: context.sources.add_pandas(...) → AttributeError: 'EphemeralDataContext' object has no attribute 'sources' on the great_expectations 1.19.0 that pip install resolves to (the .sources fluent API was removed in the GX 1.0 rewrite).
• Ch.4 pandera contract.py failed: without parse_dates=["created_at"], Column(pa.DateTime) raises SchemaErrors … expected datetime64[ns], got str. Adding it (not in the quest) fixes it.
• What passed: validate_simple.py correctly passed a clean CSV and flagged duplicate/out-of-range/malformed-email on a corrupted copy; anomaly.py raised SystemExit with a −220.8σ deviation; once patched, pandera strict=True/lazy=True rejected an extra column and reported all violations at once.

🐞 Issues Found

high · Stream Processing · Ch.2 docker-compose.yml (line 258) · docker compose up -d fails for all platforms with manifest unknown — bitnami/kafka:3.7 is gone from Docker Hub, blocking the entire hands-on quest at step 1 · Fix: repin to a pullable image (apache/kafka:3.9.0, confluentinc/cp-kafka, or bitnamilegacy/kafka:3.7) and re-verify.

high · The Warden’s Pact · Ch.2/3/4 workflow YAML (lines 155, 216, 232, 238, 263, 275–291) · Literal {​% raw %​}/{​% endraw %​} tags left inside ${​{ }​}; every run: step throws bad substitution (exit 1) when executed · Fix: wrap each YAML fence in one {​% raw %​}…{​% endraw %​} pair (as the Capstone’s Challenge 1.2 already does, lines 175/217) so the ${​{ }​} expressions render intact.

high · The Warden’s Pact · Ch.2 CODEOWNERS (lines 121–133) · Catch-all * @team-dev at the bottom wins under last-match-wins, silently defeating the @team-security scope rule — the opposite of the guardrail’s purpose · Fix: move the catch-all to the top of the file.

high · Data Quality · Chapters 2–4 (line 239 et al.) · No orders.csv is ever provided; every snippet fails FileNotFoundError on first run · Fix: ship an inline CSV block or a small generator snippet (clean + corrupted) before Chapter 2.

high · Data Quality · Ch.3 validate_ge.py (lines 270–289) · context.sources.add_pandas(...) throws AttributeError on the GX 1.19 pip installs today · Fix: update to the GX 1.x context.data_sources.add_pandas(...) API, or pin great-expectations<1.0.

medium · Data Quality · Ch.4 contract.py (line 353) · Column(pa.DateTime) fails WRONG_DATATYPE because created_at reads as string · Fix: add parse_dates=["created_at"] to pd.read_csv.

medium · The Warden’s Pact · Quest Validation (line 341) · python3 scripts/validate_quest.py --quest q19 fails — script never provided · Fix: ship the script or replace with a manual self-check checklist.

medium · The Warden’s Pact · Objectives vs. content (lines 89–93) · Objective 5 (“intentionally trigger each guardrail and verify it fires”) and objective 3 (forbidden-action detection) have no delivered exercise — only an AGENTS.md policy list · Fix: add a trigger-and-verify exercise, or reword objective 3 to match the policy-only delivery.

medium · Stream Processing · Ch.3 / Intermediate Challenge (lines 418–426) · The tumbling-window-over-event-time exercise has no reference code anywhere in the quest · Fix: add a minimal windowing example (kafka-python/Faust/PyFlink).

medium · Stream Processing · Title/framing · “& Flink” in the title but Flink is never used (all snippets are kafka-python) · Fix: add minimal Flink content or drop it from the title.

low · Apache Spark · Prereq (line 127) · “JDK 11/17” is stale — pip’s pyspark 4.x needs Java 17+ · Fix: state “JDK 17+” and consider pinning pyspark==4.2.0.

low · Grand Capstone · Completeness (Challenges 1.3, 2.3, 3.2–3.3, 4.2–4.3, 5.1–5.4, 6.1–6.2) · ~2/3 of sub-challenges are “Refer to Q_” stubs with no inline artifact/filename hint; the Domain Coverage Rubric (/18, /19…) is never tied to the validator’s pass/fail · Fix: add a one-line expected-artifact note per stub and state the rubric↔validator mapping.

🔗 Chain Continuity

• The window is two arcs, not one path. Quests 1–2 belong to quest_series: agentic-ai-mastery (GH-600 Agentic Codex); quests 3–5 belong to quest_series: Data Engineering Mastery. They share only the 1100 level code. For a data-scientist character, the Data Engineering sub-chain is the on-theme track; the two agentic-governance quests read as a different curriculum bolted onto the same level. This is a slicing/windowing artifact worth flagging to a maintainer — the level bundles disjoint learning tracks, so no single reader is the intended audience for all five.
• Data Engineering sub-chain (3→4→5) is well-ordered. Spark’s frontmatter unlocks Stream + Data Quality, and Stream unlocks Data Quality; the plan presents them in that dependency order, and each opens by referencing ETL Pipeline Design as the shared prerequisite. Conceptually the chain is coherent and Spark leaves a learner genuinely ready for the next two. The break is environmental, not pedagogical: Spark (self-contained local PySpark) passes, but Stream (needs a live Kafka broker) and Data Quality (needs a dataset + version-sensitive libs) both fail the moment a learner leaves pure-Python territory. A beginner who finishes Spark cleanly would hit a wall at Stream’s very first docker compose up.
• Agentic sub-chain (1→2): the prerequisite carries a broken artifact forward. The Warden’s Pact unlocks the Capstone, and the Capstone’s Seal 6 / Challenge 6.2 explicitly refers back to Q19 for its three guardrails. A learner who copy-pasted Q19’s workflows would carry three bad substitution-failing files into the capstone — yet the capstone’s validator only checks test -s .github/CODEOWNERS and grep -qi forbidden AGENTS.md, so the capstone would still “pass” despite the upstream workflows being non-functional. The chain hides the Q19 defect rather than catching it.
• The fix already exists inside the slice. The Capstone’s Challenge 1.2 (source lines 175/217) wraps its YAML fence in a single {​% raw %​}…{​% endraw %​} pair and tells the learner “drop them when you copy the YAML” — the correct handling of this site’s Liquid escape. The Warden’s Pact does the same escape wrong (tags interleaved inside ${​{ }​}). A maintainer fixing Q19 can copy the sibling quest’s proven pattern verbatim.

🧠 Reasoning & Method

• Mode: execute. The machine evidence in walk-evidence.json / walk-evidence.md was pre-computed and sealed by the workflow in a disposable sandbox (mock:false). Per the skill, I consumed it as-is and did not re-run, edit, or regenerate it — the engine’s child claude processes can’t authenticate from an agent Bash tool. Every passed/failed above is a command the engine actually ran; items I judged from reading the source (CODEOWNERS ordering, the two-arcs observation, the capstone-hides-Q19 continuity finding) are labelled as source-line reasoning, not execution.
• What I ran vs. reasoned: I ran no quest commands myself; I read all five quest sources in plan order and cross-checked each engine finding against the exact source lines it cites. The linked-journey analysis (§ Chain Continuity) is my own reasoning over those sources plus the sealed per-quest verdicts.
• Coverage & limits: This is window 2 of 3 — quests 11–15 of the 15-quest level 1100. Quests 1–10 (windows 1 & 3, including the Autonomy Levels Matrix that Q19 depends on and ETL Pipeline Design that the Data Engineering arc assumes) were not in this session; the ledger accumulates them across runs. Snippet coverage was partial by design: the Warden’s Pact exposed only 1 of 7 snippets as literally runnable (the rest are illustrative YAML/CODEOWNERS/markdown, reasoned statically), and the Capstone 1 of 9. Full execution coverage was strongest on Spark (10 runnable) and Data Quality (9/9 runnable). No network access beyond what quests required; the Kafka/Docker and pip installs happened inside the sealed engine sandbox, not here.
• Confidence: High on the three high-severity blockers (each backed by a reproduced error string) and on the two passes (Spark and the Capstone validator both executed end-to-end with matching output). Medium on the “two-arcs” slicing observation, which is a curriculum-design judgment rather than an executable fact.