from common import *
import json
g=pd.read_parquet(R/'games.parquet');idx=g.set_index('appid')
def G(a):
    row=idx.loc[a];return f"[{row['name']}]({row.game_url})"
def D(name):
    row=g[g.credit.eq(name)].iloc[0];return f'[{name}]({row.developer_page_url})'
text=f'''# Developer lives on Steam: what distinguishes career sequences?

The strongest result is that a substantial floor, repeated large outcomes, and
continued production are different properties. A first strong game is often
followed by meaningful response without a repeat of its scale. Repetition adds
information, but peak magnitude and the kinds of later games account for much
of the raw difference. One-hit catalogs include several very different states,
especially no observed subsequent release. Quiet intervening work is informative
without erasing the whole earlier catalog. Back-catalog review responses around
new releases are mixed and usually modest in the selected historical cases.

The broad source contains 93,013 currently paid games and 59,259 credited names,
using the previously audited single-credit eligibility rules. Current September
5 review counts describe retrospective catalog shapes. Dated histories provide
a separate check; the two kinds of evidence are not interchangeable. Developer
names need not identify individuals, and free work, collaborations, renamed
credits, missing games and other platforms can be absent.

Q means fewer than 100 reviews, M 100–999, H 1,000–9,999, B 10,000+. H/B combine
for the main strong-response threshold. These are response scales, not profit
or artistic-value labels. Main mature outcomes predate September 1, 2025.
Checks use 500 and 2,000 as alternative strong thresholds. All links and exact
snapshot values for the named examples are in [the linked case atlas](case_atlas.md).

## 1. A higher floor is more common than perfect repetition

Among 1,268 eligible developers whose earliest released title currently above
1,000 reviews dates from 2015 through August 2020, 529 released another included
game within five years: 41.7%. This is a release-date anchor for a game classified
by today's count, not the historical date the game reached 1,000 reviews.

Of those 529 returners, 83.2% had a next game reaching 100 reviews and 45.4% a next
game reaching 1,000. Across all included follow-ups in the window, 54.1% had
another 1,000-review game. In the denominator including non-returners, that last
proportion is 22.6%. Excluding non-returners would tell a materially different
story.

129 released at least three follow-ups within the window. Looking at exactly
their first three, 72 (55.8%) kept all three above 100 reviews, but only 22
(17.1%) kept all three above 1,000. The median weakest follow-up was 165 reviews;
the median of their three-game medians was 547. This supports substantial lower
response floors more often than a series in which every title qualifies as a hit.

A stricter paired subset has three releases before the first qualifying title
and three afterward: only 21 developers. Fifteen improve their three-game median,
but eleven improve their minimum. Median pre/post minima are 39/115; median
pre/post medians 108/576. It is a small, selected subset, not an estimated causal
effect of a breakthrough.

{D('Big Robot Ltd')} illustrates one strong peak with a substantial remainder:
{G(207670)} 30 → {G(242880)} 2,875 → {G(457760)} 791 → {G(853240)} 128.
That differs from {D('BeautiFun Games')}: {G(252670)} 2,296 → {G(422650)} 23 →
{G(741520)} 35 → {G(1292980)} 7. Both have one game above 1,000, but their other
included work received very different levels of review response. Neither sequence
establishes profitability or whether the creators achieved their artistic aims.

## 2. Repetition matters, but the size of the earlier peak matters too

Using one latest qualifying 2019–August 2025 transition per developer, among
catalogs with at most ten mature included games:

| Prior 1,000-review games | Developers | Next game reaches 1,000 | Median next reviews |
|---|---:|---:|---:|
| One | 705 | 36.5% | 530 |
| Two | 192 | 51.6% | 1,101.5 |
| Three or more | 114 | 60.5% | 1,547 |

The raw staircase is clear. Comparing repeated-hit cases with one-hit cases in
similar era, prior-peak and current-price cells reduces the observed/expected
ratio to about 1.13 for two-or-more hits, with 244/306 target cases supported.
Adding prior-title-count strata leaves 186 supported cases and a ratio of 1.16.
Very fine matching has inadequate support; it is not silently treated as proof.
Current price is itself potentially a response to earlier success, so these
are descriptive comparisons rather than an isolated repetition effect.

Magnitude is important: among 123 developers with two or more qualifying games
but a prior peak below 5,000 reviews, 35.0% of next games reach 1,000. Among fifty
with exactly one qualifying game whose peak is 10,000–19,999, 72.0% do. The latter
also have higher current next-game prices. Two modest hits cannot simply be
ranked above one much larger peak without accounting for the scale difference.

{D('Daniel Mullins Games')} has {G(405640)} 14,761 → {G(510420)} 4,480 →
{G(1092790)} 134,744. The middle game is much smaller than its predecessor and
still substantial. A peak-relative definition of failure would misdescribe it.

## 3. The one-hit label hides several different states

Among 1,170 one-hit catalogs whose first included release is old enough for the
study and whose sole currently qualifying title was released at least three
years ago:

| Subsequent included releases | Catalogs |
|---|---:|
| None observed after that title | 700 |
| Only a follow-up under one year old | 61 |
| One mature follow-up | 230 |
| At least two mature follow-ups | 179 |

Those states should not be collapsed into a failure-to-repeat probability.
{D('Nolla Games')} has one included paid game, {G(881100)}, with 80,977 current
reviews. Its observed catalog is fundamentally different from a series of
quiet follow-ups. No claim about inactivity or retirement follows from the lack
of another included paid title.

In the more restrictive 683 catalogs with 3–10 mature releases, an older observed
start and at least one 1,000-review game, patterns divide as follows:

| Retrospective pattern | Catalogs |
|---|---:|
| Every mature title reaches 1,000 | 101 |
| Earlier quieter work, then all strong releases from the first strong title | 50 |
| Repeated strong games with other uneven results | 173 |
| A quiet game somewhere between strong games | 49 |
| One hit followed by at least two consistently moderate games | 63 |
| One hit followed by at least two exclusively quiet games | 23 |
| One hit followed by a mixture of quiet/moderate games | 82 |
| One hit with zero or one mature follow-up | 142 |

These are observed, unfinished sequences. The categories can change with another
release or continued review accumulation. A sensitivity excluding possible
punctuation-only name fragments leaves 671 strong core catalogs and very similar
pattern proportions.

## 4. Quiet work changes the picture without erasing earlier strength

Among observed returners with one immediately preceding quiet release, next-game
100-review rates are 10.5% without any older 1,000-review game (4,047 cases),
50.0% with one older strong game (54 cases), and 62.5% with two or more (16 cases).
The corresponding 1,000-review rates are 2.3%, 11.1%, and 37.5%.

After two consecutive quiet games, the one-earlier-hit group has only 21 cases:
23.8% reach 100 next, 9.5% reach 1,000. Some still do, but a precise decay law
cannot be inferred. The apparent rebound after three quiet games is just three
strong outcomes among seven cases and is not treated as a general pattern.

A separate fixed three-year quiet-state study retains those who do not return.
Another included release appears for 9.5% of 20,004 no-earlier-hit states, 42.3%
of 156 one-earlier-hit states, and 64.1% of 39 states with multiple earlier hits.
Returning and receiving substantial response are different outcomes, and
retrospective prior bands still contain future review accumulation.

{D('Blendo Games')} gives a concrete uneven sequence: {G(55000)} 202,
{G(55020)} 8, {G(55040)} 1,108, {G(214700)} 1,282, {G(240440)} 896,
{G(592100)} 29, and {G(301280)} 1,659. The last two releases are dated 2018 and
2025. This establishes a quiet-to-strong return after a long gap, not that the
gap was beneficial or that earlier reputation was untouched.

{D('Suspicious Developments')} provides another useful distinction:
{G(206190)} 10,274 → {G(494720)} 264 → {G(268130)} 6,851 → {G(1043810)} 11,457.
The smaller intervening title is moderate on the absolute scale, not Q. Calling
everything smaller than the first peak a failure would erase that distinction.

## 5. Late breakthroughs can precede repeated response

In the fixed five-year first-qualifying-title study, 29 of the 36 fourth-or-later
breakthrough cases release again, and 22 of those 29 have another strong game:
80.6% return and 75.9% repeat among returners. First-game cases show 37.9% return
and 55.1% repetition among returners. The late group is small and consists of
people who had already repeatedly released; this is not evidence that waiting
for a late breakthrough causes a better career.

{D('David Szymanski')} has {G(329830)} 988, {G(359040)} 494 and {G(395270)} 405
before {G(519860)} 22,045. Later work includes {G(1465510)} 380,
{G(1846170)} 10,141, {G(1935660)} 2,873, {G(2535830)} 2,046 and
{G(2512560)} 1,529. The large step and subsequent substantial work are clear,
even though the exact first-hit index changes at a 500-review threshold because
the earliest title is already close to 1,000.

For later first-qualifying titles, current prices are typically about 1.5–1.65
times the preceding title's current price, and about 28% have a different
current publisher label. Those are present metadata differences around an
observed sequence, not historical price increases or proof of when a publisher
deal occurred. Tag changes are retained but not interpreted as creative pivots.

{D('Endlessfluff Games')} supplies a quieter example: {G(109200)} 105 →
{G(252030)} 2,275 → {G(995980)} 1,168. Repetition need not take the form of a
continually increasing ceiling.

## 6. Consistency is uncommon at a demanding threshold

Among 426 catalogs with four to ten mature releases and at least one
1,000-review game, 186 (43.7%) keep every game above 100, but only 35 (8.2%) keep
every game above 1,000. Sixteen also keep every game at or above 80% positive.
Among the 99 with a 10,000-review peak, 62 keep every game above 100 and 27 keep
every game above 1,000. The scale of the earlier peak again matters.

{D('Harvester Games')} has {G(253110)} 4,792 → {G(364390)} 1,455 →
{G(593960)} 1,440 → {G(1846460)} 1,029: a consistent substantial floor with
unequal current totals.

{D('Freebird Games')} has {G(206440)} 69,477 → {G(327410)} 8,206 →
{G(337340)} 18,101 → {G(1182620)} 11,953 → {G(2159210)} 3,782. A dominant first
peak coexists with a substantial continuing catalog.

{D('Erik Asmussen')} (the recorded developer page is 82apps) has three mature
qualifying games: {G(270450)} 2,253, {G(860890)} 4,140, and {G(2207490)} 1,231.
{G(3312130)}, released July 2026, is explicitly excluded from mature sequences.
Its 149 snapshot reviews are not used to declare a broken run of successes.

Catalogs with a reported franchise in at least half their games more often keep
a 100-review floor, but franchise fields are incomplete. Same-publisher status
alone does not yield a clear superior floor. These metadata comparisons cannot
establish that continuing a franchise causes consistency or that missing labels
mean unrelated new IP.

## 7. New releases sometimes coincide with back-catalog activity, but unevenly

The historical study combines 32,974 newly collected records from 26 games with
the prior bounded career histories: 48,419 records across fifty games. Eleven
currently included paid catalogs are complete in this historical set. The new
collector used 370 requests, no errors/retries, and is now closed. No account
IDs, reviewer text or media were collected.

There are 31 eligible later launches with old included games at least a year
old; 26 are isolated from another included paid release by 180 days either side.
For those 26, fifteen have more old-game reviews in the following ninety days,
and seven gain at least ten. The median paired increase is five reviews. Requiring
at least five pre-event reviews leaves 24 events and a median gain of 2.5.

| Developer and later release | Earlier-catalog reviews, prior 90 days | Following 90 days |
|---|---:|---:|
| {D('Blendo Games')}: {G(301280)} | 32 | 58 |
| {D('Abbey Games')}: {G(1875060)} | 40 | 61 |
| {D('a327ex')}: {G(915310)} (older game is {G(760330)}) | 4 | 12 |
| {D('Erik Asmussen')}: {G(2207490)} | 111 | 91 |
| {D('Studio Fizbin')}: {G(2129810)} | 204 | 149 |

Seasonally aligned prior-year windows give a median relative ratio of about
1.16 across 22 eligible comparisons, but natural changes in a game's declining
baseline remain confounded. A separate matched check using the earlier random
lifetime sample supports only 18 of 49 eligible old-game/event pairs, covering
twelve events/seven developers. It finds 280 post-event reviews versus 244
expected, or 1.15 times expectation. Low support and selected catalogs make this
unsuitable as a general launch multiplier.

These results support particular episodes of connected catalog attention and
possibly modest average associations in the sampled cases. They do not establish
transferred players or a dependable rule that a new release lifts all older work.

## Historical checks change how sequences should be read

Eight of 38 selected transitions have more prior 1,000-review games when using
today's counts than when using surviving records created before the next launch.
For {D('Blendo Games')}, neither of today's two earlier 1,000-review games had
1,000 surviving reviews dated before {G(240440)}. For {D('Rad Codex')},
{G(1224290)} had 809 surviving reviews before {G(2276830)}, versus 1,212 now.
Deleted/filtered records mean these are reconstructed surviving counts, not a
complete archive of historical reviews.

Equal-age comparisons can also reverse the apparent scale of later work.
For {D('Abbey Games')}, {G(296970)} has 845 surviving reviews by day 365,
{G(917150)} 343, and {G(1875060)} 1,776. The newest is slightly smaller than
the first of those in current totals, but more than twice as large at year one.
For {D('Blendo Games')}, {G(240440)} has 290 by year one, {G(592100)} thirteen,
and {G(301280)} 1,551. This is a materially stronger rebound at equal age than
the lifetime-count comparison suggests.

The first-year histories cannot uniformly cover every early release. Recorded
dates, Early Access and late first surviving records prevent some comparisons;
equal-age sequence tables identify their exact subsets rather than silently
substituting zeros. Current shapes remain useful descriptions, but a career
forecast must not mistake those shapes for what was known at each launch.

## Artifacts and verification

The linked atlas contains eighteen developer pages and 88 games. All eighteen
developer URLs returned HTTP 200. Current-count examples use snapshot values;
historical windows identify their dates and populations separately. Both figures
were inspected. The analysis passed {json.loads((R/'validation.json').read_text())['passed']:,} checks covering identity accounting,
sequences, first-event windows, source histories, cursor chains, historical prior
counts, event recounts, matched controls, collection limits and links. The original
snapshot and paused bulk policy remain unchanged.
'''
(R/'findings.md').write_text(text)
print('Report written',len(text.split()),'words')