What I actually played.
Thirteen years of it.

Plot annual hours and the curve is not a taste history — it is a workload history, inverted. An earlier version of this essay, built from one account, saw three years of silence here (2014–17) and called it "the great gap." The recovered first account erases it: those years held five hundred hours apiece of Einaudi, Pink Floyd, and electronic juvenilia — college, fully scored. What the merged curve actually shows is a steady climb from a freshman dorm to a staggering pandemic peak — 1,569 hours in 2020, more than four hours for every day of that year it played — and then the residency ration, down to a 2025 floor of 238 music hours. The listening didn't change because taste changed. It changed because attention got conscripted.

Music hours per calendar year, both accounts merged, hack windows excluded · 2026 through early June

Music hours per month, Mar 2013 – Jun 2026, both accounts · dashed lines = account switch and first day of intern year · stepped line = regime means found by changepoint detection

The stepped line is not drawn from the biography — it is drawn from the signal. A changepoint algorithm (PELT, fitted to the 159-month series with no knowledge of what the months meant) cuts thirteen years into six regimes: a near-silent 2013 ramp, a 39-hour-a-month college cruise, a step up in July 2016 (senior year; the first iPhone-heavy year), a leap to 125 hours a month in November 2019 — four months before lockdown — a step down in early 2021, and a final break in October 2022, three months into intern year. The mathematics, asked "where did this life change?", finds every academic transition within a season — and insists, inconveniently, that the pandemic listening surge was already underway before the pandemic was.

Every prior version of this essay was built on one export and inherited its blind spot: a "silence" from 2014 to 2017 that was actually another account — opened in March 2013, played hard through four years of Princeton, abandoned for its successor in June 2017, and later colonized by hackers (§ XVII). Its 46,280 genuine plays and 2,355 hours rewrite the early biography. The freshman listens to Fleet Foxes and Cage The Elephant. The sophomore discovers Ludovico Einaudi and scores entire problem sets with him — 597 plays in academic 2014–15 — alongside Mndsgn and Chrome Sparks, beat tapes for a kid with headphones in a library. The junior cools into deadmau5 and Ratatat. The senior converges, almost artist by artist, on what the second account would later treat as bedrock: Mac DeMarco, Radiohead, The Beatles, Tame Impala.

First account, by academic year (Jul–Jun) · artists ordered by finished plays

The merge also settles an old mystery quantitatively. The 2017 "reawakening" — 836 artists arriving in the new account's first months — was never discovery: 53.7% of them already lived in the first account's history (Pink Floyd, Mac DeMarco, Einaudi, The Beatles, Mndsgn, Chrome Sparks…). The new account didn't find a listener; it inherited one. Even the overlap is biographical — for 193 days the two accounts ran side by side, a MacBook on one login and an iPhone on the other, before the old one went quiet for good.

Both archives, at native resolution: one cell per calendar day, shaded by hours played. Read it like a core sample. With the first account restored, the college years fill in — 2014 through 2016 are solid sediment now, not blank rows — and the record runs nearly unbroken from a freshman March to the present. 2018 through 2020 darken until the ink is continuous: spanning late 2019 through November 2020 runs a 339-day unbroken streak — counting every logged play, including days too faint to ink a cell — a year in which "did I listen to music today" was simply not a question. From mid-2022 the texture changes — residency turns the page into weather, dense squalls around free weekends and post-call afternoons, thin drizzle through service months. The volume changed eras ago; what the punch card shows is that the rhythm changed too.

Hours of audio per day, all content types, both accounts · columns are weeks, rows within each year are Monday–Sunday · shading bands at the 25th/50th/75th/90th percentiles of listening days (0.52 / 1.39 / 3.06 / 5.2 h)

Cut the merged archive at the biographical seams and three different people emerge — and for the first time, all three are fully recorded. The college listener is no longer a sketch but 1,949 hours: four years of MacBook-and-headphones listening averaging 1.8 hours per listening day. The medical-school listener is a torrent: five years at 2.7 hours per listening day, music running like tap water through study sessions and a locked-down apartment. The residency listener is rationed: 1.5 hours per listening day, squeezed into commutes and the edges of call schedules.

The weekday fingerprint shifts with the era, too. College listening is nearly flat across the week — a student's days are interchangeable, and Saturday leads by a whisker. In medical school, listening climbs steadily toward the weekend — Sunday biggest, 37% over the Monday–Friday average. In residency the same Sunday peak survives, but Tuesday–Thursday collapses: the mid-week clinical block is where music goes to die.

Same Sunday peak, very different floors · note the scales differ ~3× between eras · Princeton's week (not shown) is flat at 249–305h per weekday

Stretch each top artist across thirteen years and the shapes sharpen. The spike-decay: Khruangbin takes more than half its lifetime hours from a single year and never recovers; Connan Mockasin collapses from 36.6 hours in 2020 to twelve minutes in 2022 and stays gone. The pulse: deadmau5 returns every few years like a comet — and the merged record shows the first apparition was college, 358 finished plays in 2015 alone. And the constant: Radiohead has appeared in every year of the archive since 2013 and logged meaningful hours in nearly all of them, has been #1 for five consecutive years (2022–2026), and — with the first account counted — is now the all-time #1 outright: 268.7 hours, ten clear of Tame Impala (258.2h), with Khruangbin third (210.3h). The single most stable preference this data can name got there without ever once spiking.

Each row shaded to its own peak year · hover any cell for hours · note the single bright column nearly every row shares: 2020

One track owns both archives. "Charlotte's Thong" — Connan Mockasin's nine-minute slow-motion guitar fever dream — has 239 finished plays and 32.7 hours of total runtime, fourteen hours clear of the #2, and the merge shows the obsession running on both logins at once in late 2018 — the household speaker was still casting it to the abandoned account. It is now #1 by hours and by plays. The rest of the list is epics that win on duration — Stairway to Heaven (now 148 plays back to college), Pink Floyd's "Dogs," Paranoid Android — and grooves that win on count, led by Khruangbin's "White Gloves" at 189.

The canon, playable. Six tracks that have each survived at least five years of rotation:

Embedded players stream 30-second previews without login · full tracks with a Spotify session

Convert every timestamp to Eastern time — the clock this life actually ran on, Princeton, then Philadelphia, then Baltimore — and the most physiological finding in the archive appears. The college listener peaks mid-afternoon, with a 6 am trough: a student's clock. The medical-school listener is a night owl: the curve climbs all evening, peaks at 8–9 pm, and is still doing serious volume at 1 am (133.9 hours logged in the 1 am hour alone — more than most artists' lifetime totals). The residency listener has been flipped bodily forward: the 1 am shoulder is gone, the trough lands at 4–5 am, and a new spike appears at 7 am — the commute, pre-rounding hour — followed by a midday plateau. Medical training didn't just shrink the listening; it moved it twelve hours.

Penn Med era · hours by local hour

Residency era · hours by local hour

An earlier version of this essay converted timestamps to Pacific time and read a "morning listener" out of one year of data. The thirteen-year archive settles it: the sleep trough sits at 4–5 am Eastern in every era, exactly where a Princeton-Philadelphia-Baltimore life would put it. The corrected charts above replace that reading — and the next section checks them against the body itself.

The circadian charts above infer a life from when music plays. A second instrument can cross-examine them: 523 nights of sleep tracking (sparse app data 2019–22, then ring- and watch-tracked nights from late 2024) sit in the same date range. The verdict is corroborating. The residency-era median night runs 12:48 am to 7:14 am — and the listening trough at 4–5 am lands almost exactly on the measured midpoint of sleep. The music data was telling the truth about the body. The sleep data adds one signature the play log can't see: before workdays the median midpoint is 3:16 am; before free days it drifts to 5:16 am — a two-hour weekend swing — while sleep duration barely moves (7.2 vs 7.5 hours). Residency shifts the schedule, not the dose.

Era-level aggregates only · per-night data stays off the page · sources: Oura ring + Apple Watch (2024–26), app-logged nights (2019–22); pre-residency coverage is sparse and self-selected

First-time artists per year track the breathing room in the calendar. After the first-year flood (every artist is "new" to an empty archive), discovery settles near 700–1,000 a year through college and medical school, surges to 1,318 in 2020, and halves in the residency years — with a 2025 uptick to 740, the sound of a PGY-3 with slightly more elective time. Cumulatively the two accounts have touched 10,962 artists, but the overwhelming majority are one-night stands: of 37,575 tracks with at least one finished play, 22,353 — 59% — were finished exactly once.

First-ever plays of an artist across either account, by year · 2014 inflated by the empty-archive effect

Then there are the one-era wonders — artists who rack up serious hours inside a tight window and never return. They are the most biographical objects in the data, because each one is a circumstance, not a taste:

Artists with ≥8 lifetime hours where 90% of listening fits inside one short window

Group every artist by the year they first earned three real plays, then ask a colder question: what fraction of each class was still being played one, two, five years later? An earlier version of this table, built on one account, showed a 2017 cohort with eerie 70% retention — and the merge explains it away: those were re-installations of first-account favorites, and once an artist's true first play is known, the anomaly dissolves into an honest pattern. Every class from 2014 through 2019 — college and medical school alike — holds 34–42% at one year. Then the door closes: the 2020 pandemic class keeps 18%, and the residency-era classes of 2021–2023 lose roughly nine in ten by their first anniversary. The 2014 class is the marvel now: eleven years out, 15% of it still gets played. New music still gets in. It just doesn't get to stay.

% of each discovery cohort played again N calendar years later, both accounts · cohort = artists whose first ≥30s play fell in that year and reached ≥3 plays that year · cells touching half-observed 2026 are blanked · cell ink saturates at 75%

One honest confound: later cohorts are being tested against a quieter archive. An artist discovered in 2023 competes for a fraction of the listening hours an artist discovered in 2019 enjoyed, so some of the collapse is opportunity, not affection. But that is precisely the essay's thesis wearing different clothes — the door narrowed because the attention did.

Take all 2,242 artists discovered between mid-2013 and mid-2024 who ever reached five plays — both accounts, college included — align them on the week of first contact, and average. The resulting curve is not a gentle decline; it is a cliff with a basement. The discovery week averages 5.1 plays; within two weeks the average is down near a quarter-play — the surge is spent almost entirely inside its first days, too fast for weekly bins to resolve a half-life. What remains is the more interesting object: a flat residue of about 0.10 plays per week — one play every ten weeks — that holds steady for at least two years. Tested against the data, a pure exponential loses, a power law loses; the model that wins is decay plus a permanent floor. Eleven years of data, two different accounts, the same law. A discovery isn't a flame that burns out. It's a flash that leaves a coal.

Mean plays/week since first ≥30s play, weeks 0–52, log scale, both accounts · terracotta = observed · dashed green = fitted exponential-plus-floor (ΔAIC ≈ -190 vs pure exponential, -185 vs power law)

The average hides the individuals, and the individuals are where the model goes to die. Four artists the archive flagged as "spikes" turn out to be four different shapes of wanting:

Plays per week from first contact, each panel scaled to its own peak · these four are why the essay fits curves to the population, not the person

The merge crowns a new single-day repeat record, and it is a college record: seventeen plays of Madlib's "Episode XXIII" on February 8, 2016 — a Princeton problem-set loop. The modern era's tie at eleven stands behind it: Red Hot Chili Peppers' "Goodbye Angels" (April 2018) and John Lennon's "Hold On" (October 2020), one a new-single fixation, the other almost too on-the-nose for week 31 of a pandemic. Christmas Day 2020 logged Burl Ives nine times, which is what happens when the speaker belongs to the kitchen. And February 5, 2021 keeps its asterisk: 8.5 hours of algorithmically generated ambient from pseudonymous "artists" — Spotify's ghost-artist white-noise economy, captured in a personal export, ten plays at a time.

Grouping plays into sessions (a new session after 30 minutes of silence) yields 11,679 sessions with a median length of 17 minutes — the everyday unit of listening is a coffee break, not an evening. The tails are something else. The longest continuous session now runs 40.0 hours, starting the night of June 10, 2017 — the account switch itself, 606 tracks deep, a new login digesting an old library without pause. The three biggest single days span the whole arc: the switch (23.8h), a June 2020 lockdown Sunday (23.0h), and the 2025 lo-fi deadline day (22.1h).

A session — every play separated by less than thirty minutes — is the archive's natural unit of mood. Take the 120 most-sessioned artists across both accounts, link two whenever they share a session at least twice as often as chance would predict, and let a community-detection algorithm (Louvain) carve the graph with no other information. With the college data included it finds five listeners. A terracotta mainland — Tame Impala, Mac DeMarco, Connan Mockasin, Khruangbin — the psych default rotation where most hours live. A green college constellation — alt-J, Gorillaz, Arcade Fire, deadmau5, Chrome Sparks, Fleet Foxes — whose median play sits in 2017, two-plus years earlier than every other community: the Princeton listener, preserved intact as a subgraph. A gold continent of inherited canon — Beatles, Floyd, Zeppelin. A deep-red alt-rock spine that Louvain pulls out of the mainland once the long record is visible — Radiohead, Nirvana, Pixies, The Strokes — the constant, given its own province. And, floating clean off the coast, the five-composer classical island: a perfect clique, bound to almost nothing else. The borders are the finding — and with thirteen years of data, so is the chronology: the algorithm was given no dates, and it carved out an era anyway.

The 120 most-sessioned artists across both accounts · edges = co-occurrence in 30-minute-gap sessions ≥2× chance (strongest 260 drawn) · node size ∝ √sessions · communities: Louvain, modularity 0.29 · force-directed within each community, communities tiled apart for legibility — only links and colors carry meaning · labeled when ≥110 sessions

The device column quietly explains the volume curve — and now it starts in a dorm. College opens 91% MacBook and migrates to the iPhone by senior year (a PlayStation 4 makes a cameo as a speaker in 2018). Medical school starts split between Mac and iPhone — and then, in late 2018, a Google Home arrives and everything changes: cast devices jump from nothing to 574 hours (2019) to 1,325 hours in 2020. The pandemic peak isn't headphone listening; it's a speaker filling a room for whole days, music as atmosphere rather than activity. Residency inverts it again — by 2023–24 the iPhone leads, because the listening moved out of the home and into the drive to the hospital. All-time: cast devices 3,147h (39%), iPhone 2,955h (36%), Mac 1,876h (23%).

Leading device class per year, with its share of that year's hours, both accounts · solid = Mac (moss) / iPhone (terracotta), hatched = cast devices

When the first account was abandoned, something else moved in. Twice. In January–March 2019, logins from the Philippines, Mexico, Morocco, and Colombia ran 12,494 plays of artists this listener has never heard of — names like Carletso, Obsino, DedMem — while, in the same weeks, the household Google Home kept innocently casting Connan Mockasin to the old login. In August 2023 a second takeover, riding Spanish and Israeli exits, pushed 17,799 plays in a month — more volume than the real listener's entire year. Together: 30,293 plays, 1,313 hours, 2,346 junk artists, all excluded from every figure on this page.

The forensics are their own portrait, because a botnet is the perfect anti-listener. It never skips (1.4% skip rate, against 7–22% for the human in recent years). It plays whole tracks to a 2¾-minute median, the shape of royalty-farming rather than taste. It hops countries nightly. And it has no floor and no canon — 2,346 junk artists with barely a repeat among them. Every analysis in this essay — the retention door, the decay coal, the session communities — describes structure a streaming-fraud bot cannot fake. The hack is the control group this essay never knew it had.

Excluded windows · "ZZ" = unresolved country · genuine plays inside window one (the household speaker, still casting to the old login) were retained by an artist-and-platform filter — see Methods

Podcasts (326.3h) and audiobooks (117.5h) live on their own calendar, and it is a residency calendar. Podcast hours explode exactly once: 153.4 hours in 2023 — PGY-2, the commute year — more than the four surrounding years combined. The top-shows list is a syllabus of that life: Hard Fork and The Daily for the drive, The Grenz Zone (a dermatopathology podcast, 20.6h) for board prep, Goljan lectures back in 2020 for Step study. Audiobooks arrive even later — essentially zero before 2023, peaking at 54 hours in 2024 — and the shelf is half journal-club, half escape hatch: The Three-Body Problem at #1, then The Coming Wave, The Sirens' Call, and Robert Wachter's A Giant Leap.

Podcast hours by year · in 2023, podcasts were a third of all listening

The account-data snapshot (2026) still tells the story the previous essay found, and the thirteen-year archive sharpens it. The saved library holds 2,060 tracks across 701 artists, plus 144 playlists with 5,910 placements — and only nine followed artists. The library's top shelf is Radiohead, Pink Floyd, alt-J, Nirvana, deadmau5: the self-image. The all-time play log ranks Khruangbin first, an artist whose entire dominance fits inside three years. Pink Floyd — 54 saved tracks, #2 in the library — peaked in 2019–20 and has logged barely an hour a year since. Saving is a statement about who you think you are; playing is a record of who you were that month. Thirteen years of receipts say the archivist writes checks the player rarely cashes — and a handful of artists (Radiohead, Mac DeMarco, Tame Impala) are the only ones honored by both.

One year of data diagnosed four listeners in one account. Thirteen years and two accounts reveal what they have in common: every one of them is downstream of the calendar. The volume curve is medical training drawn in negative space — it explodes when a pandemic empties the schedule, drains as residency fills it, and flips its circadian phase the month the hospital starts owning the mornings. The discovery rate breathes with elective time. The devices map the floor plan of each apartment and the length of each commute. Even the obsessions are scheduled: salsa for two days of a locked-down September, dog-calming ambient the week a dog arrives, a game soundtrack in the one stretch of PGY-3 with room for a hundred-hour game.

Against all of that churn, exactly one signal refuses to move. Radiohead has scored every era of this archive — present in every year since 2013, top artist five years running, never the spike, never absent — and with the first account restored it is, at last, the all-time #1 outright. It survived a freshman dorm, an account migration, a pandemic, a hacked login, and a residency. The previous essay called it the load-bearing favorite. The long data is blunter: it is the only thing in 8,125 hours that looks like a trait rather than a circumstance. Everything else was weather. That was climate.