# Developer lives on Steam: sequences, floors, repetition and back catalogs Authorized September 7, 2026 after discussion of seven research avenues. Main data are the independent September 5 snapshot. New dated review collection is limited to six selected paid catalogs. All results are observational and about listed Steam work, not personal biographies, retirement, production effort, profits, verified audience transfer, or the counterfactual effect of practice. ## Identity and maturity Reuse the earlier audited name/credit eligibility table, joined to the independently normalized snapshot: 93,013 games, 59,259 credited names. Joint/missing credits, ambiguous names and explicit promotional titles were excluded by that table. Names are the identity keys; creator-page IDs are used only for navigation because pages can represent multiple developers. Name changes, staff changes, homonyms, ports, editions, missing free work and delisting remain limitations. The additional punctuation-only identity audit flags 793 possible name fragments or placeholders; it does not merge them. Removing them in a sensitivity affects 61 core catalogs and twelve core catalogs with a strong title. The principal pattern proportions change little. This is not an exhaustive identity audit. Mature means released before September 1, 2025. Younger titles are retained for observed continuation and linked examples, but excluded from mature response sequences. Core readable shapes have 3–10 mature included games and first observed release from 2010 through August 2020: 2,651 catalogs, of which 683 have a game with at least 1,000 current reviews. The consistency study instead uses 4–10 mature games, first observed from 2010 onward: 2,014 catalogs, 426 with a 1,000-review peak. These denominators must not be interchanged. Q is under 100 current filtered reviews, M 100–999, H 1,000–9,999, B 10,000+. H/B are combined for the principal strong-response threshold. Threshold checks use 500, 1,000 and 2,000. Recommendation percentages remain separate. These levels are not financial-success or artistic-value classifications. ## Questions and populations 1. First strong-release events: the earliest released title currently above the threshold, released January 2015–August 2020, with a five-year next-release window. This anchor is NOT the historical date it crossed the review threshold. All examples start from an observed catalog beginning in 2010 or later. No-return cases remain in the denominator. Floor comparisons using three follow-ups require at least three releases in that window, then use exactly the first three. The paired before/after subset also requires three earlier games and is much smaller. 2. Repetition: latest qualifying transition per developer, 2019–August 2025, gap >=30 days. Main catalogs have at most ten mature titles; all-size results remain separate. Compare prior-hit counts, peak sizes, date/current-price strata and prior-title counts. Narrow cells often lack support. Coarser results and exact matched denominators are retained rather than imputing unsupported comparisons. Current price may itself follow success. 3. One-hit catalogs: include no-follow-up, only-immature-follow-up, one-mature- follow-up and multiple-follow-up states; split by age of the qualifying title. 4. Quiet sequences: current retrospective prior strength and consecutive sub-100 counts, with small cells explicit. A separate latest quiet-state study uses 2015–August 2022 anchors and fixed three-year follow-up, counting return and subsequent response separately. It still does not recover historical bands. 5. Early/late first strong title: compare subsequent presence, response and current metadata around the event. Publisher/price fields are today's observations, not historical contracts or launch prices. Current tag changes do not establish creative pivots or mechanical change. 6. Consistency: all-game minimum, median, strong fraction, peak dependence, current publisher labels and reported franchise coverage. Missing franchise labels are not proof of unrelated new IP. All-strong plus >=80% positive is retained as a separate dual criterion. 7. Back catalogs: dated histories below; no inference of transferred players. ## Historical collection Six deliberately selected contrasting catalogs: Abbey Games, BeautiFun Games, Big Robot Ltd, Blendo Games, Endlessfluff Games, Erik Asmussen. Twenty-six games, including the immature Factory Town 2 for complete current paid coverage. 370 paced API requests collected 32,974 currently returned Steam-purchase review records, without errors/retries. Caps were 450 requests and 40,000 reviews, 2 MB per response, >=2 seconds between starts, stop on errors. All pages are sanitized: review IDs, creation/update times, purchase and EA flags only. No review texts, account/profile IDs or media were collected. The scoped policy is now inactive; the root bulk-catalog policy remained paused. Combine these with the previous bounded 24-game paid history set. The result is 48,419 records across 50 games, covering eleven complete *currently included paid* catalogs plus one partially covered catalog (Eliza only for Zachtronics). Only complete catalogs enter the historical sequence and back-catalog study. Partial coverage is explicit in `history_coverage.csv`. Dates of surviving reviews do not recover deleted or filtered records. Current votes/text are not historical sentiment and are not used. A recorded date is operationally consistent if the first surviving review lies from seven days before to thirty days after it. A late first review can simply mean quiet early activity, so failure of this check is uncertainty, not a proven bad release date. Equal-age sequence summaries additionally require a post-2014 date and a mature 365-day window. Whole-current and equal-age sequences can therefore contain different subsets; app IDs are retained and missing observations are not zeros. Historical prior states count surviving reviews created before the later release. This checks how current outcome labels differ from the information available then; it cannot calibrate those differences across all Steam careers. ## Back-catalog comparisons Eligible later launches have mature 90-day follow-up by September 1, 2026 and a consistent date. Earlier included games must be at least one year old. Compare their arrivals in the 90 days before and after the launch. Isolated events have no other included paid release by that name within 180 days either side. This does not rule out omitted work, discounts, news or other simultaneous events. A seasonal comparison uses the same old-game subset around the equivalent date one year earlier, requiring those games already be a year old then and excluding nearby included releases. The smaller old-game subset can differ from the main event's full back catalog. A separate control check uses games from the prior random 60-game lifetime sample: similar pre-event review counts within factor two, age within three years, different developer and no included release within 180 days. Require 3–5 controls. Only 18 of 49 eligible old-game/event pairs have support, covering twelve events/seven developers; pooled and median results are descriptive and not a reliable causal launch multiplier. Current patterns, first-qualifying-release windows and historical launch-time states are distinct analyses. Nothing turns their rates into an individual probability or a prescription to raise prices, repeat a franchise, wait, or ship. ## Links, examples and reproduction The case atlas contains 18 developer names and 88 games. Developer links use unshared recorded Steam creator pages or developer-filtered Steam catalogs. All eighteen returned HTTP 200 during bounded link checks. Erik Asmussen's recorded page is titled 82apps. Each game links by its observed app ID. Counts in the atlas are snapshot totals; historical tables identify their own windows. Using `../.venv/bin/python`, offline scripts run in this order: `prepare.py`, `catalogs.py`, `transitions.py`, `robustness.py`, `evolution.py`, `history.py`, `backcatalog.py`, `backcatalog_controls.py`, `examples.py`, `charts.py`, `validate.py`. `collect_history.py` and `check_links.py` are network provenance, not required to reproduce the saved analyses. Do not reactivate collection merely to reproduce. Current scripts retain their cohort specifications, source/target hashes, memberships, request provenance, matching peers, and exclusions. Read `findings.md` for interpretation and `case_atlas.md` for all linked examples.