inventory

Phase 1 Inventory — Stats Curriculum

What’s in the homelab DBs, what’s actually usable, and which datasets map to which curriculum modules. Generated 2026-05-06 from claude_reader access on stygies (MariaDB) + phlegethon (PostgreSQL).

TL;DR

Three strong spine datasets:

1. Bandwidth (secv_scrape) — 10.5 years of continuous ISP usage data. Right-skewed, multiple regime shifts, an existing forecasting project to build on.
2. Spotify (your_spotify) — 10 years of listening history, 67K plays, 2 users (paired comparison from late-2023 onward), rich track/artist/genre dimensions. Direct AI-eval analogue (recommendation systems, taste drift).
3. Knoebels queue times (gp0.knoebels) — 690K rows, 62 rides, May 2024 → Oct 2025 of amusement-park ride wait times. Right-skewed, paired-ride comparisons, day/season seasonality, year-over-year drift. Likely the most blog-friendly dataset in the homelab.

Supporting datasets fill in:

• Blocky DNS logs — 5.9M rows / 1 month — count/rate analysis, multi-class outcomes, drift between weeks
• Gatus endpoint results — binary success outcomes across 64 endpoints — natural fit for logistic regression and proportions
• HA states + statistics — 509 sensors, ~5 months of long-term hourly data — physical paired comparisons (HVAC, water, power)
• Wanikani review stats — Bernoulli outcomes (correct/incorrect) by item — small but clean
• Firefly transactions — 4 years of categorical+continuous financial data — robust regression candidate

Datasets with too little data for stats work: SMART monitoring (18 rows), fitness sessions (7 runs), KOA scraper (19 rows), wifi_scans (1 week), HA states (30-day recorder retention).

MCP fix applied 2026-05-06: sql-mcp/server.py now enumerates all non-system schemas in list_tables, and describe_table / table_sample accept schema.table (or a bare name with cross-schema disambiguation). claude_reader was granted on every previously-missed schema. Discovery was re-run after the fix — the three datasets above were among the missed schemas.

Tier 1 — Spine datasets

secv_scrape.scrape_data and scrape_data_intervals (stygies)

ISP usage scraper. The headline dataset.

Companion tables: bandwidth_cap_history, bandwidth_forecast, monthly_totals, rolling_24hr_summary. The cap history is gold for logistic regression (was-this-month-capped is a Bernoulli outcome with real consequences).

Maps to Modules 1, 2, 3, 4, 6, 7 (drift detection capstone uses this).

gp0.knoebels.LZ_attractions_io_queuetimes (phlegethon)

Amusement park ride queue times (Knoebels, PA), scraped from a public attractions API.

Why this is a curriculum gem:

• Right-skewed queue_time distribution (most rides have short waits; a few rides dominate)
• Strong daily/weekly/seasonal seasonality (weekends, summer peak, weather effects if joinable)
• Per-ride paired comparisons (62 rides → many natural pairs to test)
• Year-over-year drift on the same ride is testable (May–Oct 2024 vs May–Oct 2025)
• Clear AI-eval narrative: “predict queue time” is a regression problem with calibration consequences

Maps to Modules 1, 2, 4, 6, 7. Strong blog headliner candidate alongside bandwidth and Spotify.

your_spotify.mart.fact_plays + dims (phlegethon)

10 years of Spotify listening history, recently migrated from MongoDB to Postgres via a staging → mart pipeline. Probably the most blog-friendly dataset in the homelab.

Natural change points: 2nd user joining late 2023 (paired comparison from there onward), the 2020 COVID jump in plays, year-over-year taste drift.

Maps to all 7 modules. Particularly strong for:

• Module 1 — paired t-test on weekly play counts between you and Lindsey post-2023
• Module 4 — regression of daily plays on day-of-week + season + user
• Module 5 — predict P(skip) from track features (if skip data exists in raw — needs verification) or P(explicit-track-played) from owner + hour
• Module 6 — Mann-Whitney comparing genre-listening durations across users
• Module 7 — drift in genre/artist distribution year-over-year (publishable: “ten years of my Spotify, did my taste actually change?”)

blocky.log_entries (phlegethon)

DNS query log from the Blocky resolver in LXC 104.

Maps to Modules 1, 2, 6, 7 — count rate comparisons across clients/days; drift in client behavior week-over-week; permutation tests on client exchangeability.

gatus_db.endpoint_results and endpoint_uptimes (phlegethon)

Uptime monitor results across 64 internal/external endpoints.

Maps to Module 5 (logistic regression on success vs hour/endpoint), Module 6 (non-parametric comparison of duration distributions), Module 7 (drift in latency).

homeassistant.statistics + statistics_meta (stygies)

Long-term hourly sensor aggregates. Note: short-term states table is recorder-retention-bound at 30 days, so we work mostly from statistics.

Caveat on the Minnie example from the handoff: Minnie was born Nov 2025 and joined the household Jan 2026 — that is within the statistics window (~3 weeks of pre, ~4 months of post). The harder issue is that her presence likely doesn’t move most sensor signal: Wes’s expectation is that water usage is the only realistic candidate, and even that’s not guaranteed to show. Better within-window comparisons: heating-season vs shoulder-season power, weekday vs weekend, before/after a specific automation change with a known date.

Maps to Modules 1, 4 (regress room temp on outdoor temp + HVAC state — needs an outdoor source).

Tier 2 — Supporting datasets

wanikani (stygies)

Excellent Bernoulli/binomial structure (correct vs incorrect on review attempts). Small, but clean.

Maps to Module 5 (logistic regression: predict review correctness from item features) and Module 1 (proportion tests across SRS stages).

firefly (stygies) — financial transactions

Maps to Module 4 (regression of monthly spend on category × time), Module 6 (non-parametric — finance is rarely normal). Privacy consideration: blog drafts on this would need to anonymize categories or use ratios, not absolute amounts.

koa_scraper.weekend_scrapes and crowding_estimates (stygies)

Currently only 19 rows — too small. Worth re-checking before Module 7 if scraping has been re-enabled.

campsite.wifi_scans (stygies)

133 BSSIDs over 1 week, 18 open networks. Good shape (categorical security level, signal strength, frequency band) but volume is low. Could feed a side example in Module 1 (proportion of open networks vs secured) but not a primary case.

Tier 3 — Skip (insufficient data)

Module → Dataset Map

Decisions (locked 2026-05-06)

Prerequisites before Phase 2

• Knoebels scraper: Resolved 2026-05-06. Containerized + redeployed to cyrion under /home/nox/docker/kuh-no-bowls/ with a systemd timer (*-*-* 09..23:00/5:00). Missed the first two weekends of 2026 but Memorial Day weekend onward will be captured.