# N2-A-001 — Results as measured (pull of 2026-08-27)

*Route (a) as pre-registered: ClickHouse playground mirror of GH Archive (verified
current through 2026-07-02), Search-API cross-check run regardless. Full method,
deviations 1–12, and raw snapshot: `data/coldstart/README.md` + `data/coldstart/raw/`.
Machine-readable: `results.json`. Per-repo: `data/coldstart/cohort.csv`.*

## RESULT A — the cohort

**195 repositories** created 2023-01-01..2025-06-30 earned ≥5,000 stars within 60
days of creation. **59.0%** are AI by the frozen pre-registered rubric (115/195);
a documented manual audit that catches what 2026 metadata hides (emptied
descriptions, "llama"/"agent"/"sd" not being rubric keywords) puts it at **79.5%**
(155/195) — quote 59% as the pre-registered number, 79.5% as the check. The
hyper-viral subset (≥10,000 stars in 30 days) has 46 members, so the 5,000/60d
threshold isn't doing the work. **9 repos (4.6%) no longer exist at all** — and
the attrition list is itself a finding: four dark-web-directory spam repos, two
ChatGPT proxies, and a ticket bot make seven of the nine — the fake-star
ecosystem's fingerprints, per the ICSE study's malware-promotion finding (the
other two: Infinite-Storage-Glitch and OpenGFW, author-removed). Cross-check delta:
the Search API (stars-today basis, 1,709 repos ≥5k created in-window) cannot see
the 9 deleted repos, renders 2 cohort forks invisible (search defaults to
fork:false), and misses s0md3v/roop, which has **decayed below 5,000 stars today**.

Spoken pattern: "The query returns one hundred ninety-five repositories.
Fifty-nine percent are A-I tools by our frozen rubric — likely an undercount;
our audited number is just under eighty. And before we measured anything else —
nine of them no longer exist at all."

## RESULT B — the decay curve (headline)

Median viral repo's **human commit activity at T+12: 0.2% of its peak month**
(bootstrap 95% CI 0.0–3.0%; n=168 with commit data). **54.3% had literally zero
human commits in month 12.** Status today: **61.5% archived, deleted, or dormant**
(16 archived, 9 deleted, 95 dormant = no push in 6 months); **38.5% still alive**.
AI subset median: **0.0%** of peak; non-AI: 15.6%.

Spoken pattern: "The median viral repo's human commit activity at twelve months:
zero point two percent of its peak month. More than half shipped literally zero
human commits in month twelve. Sixty-two percent are archived, dead, or dormant.
Thirty-eight percent still alive."

## RESULT C — does velocity predict anything?

Launch-week star velocity explains, across the cohort (R², log-transformed):
**1.9%** of the variance in month-12 human commits, **1.5%** in month-12
contributors, **4.3%** in release count, **1.8%** in being alive today — and
**0.4%** of the variance in stars today. Rank correlations are all *slightly
negative* (Spearman −0.03 to −0.26): within the viral club, launching hotter
predicts marginally *worse* 12-month health, not better.

Two honest framing notes: (1) Borges et al.'s ρ>0.95 "stars predict stars" holds
across the whole of GitHub; inside an all-viral cohort the range is restricted —
that's the point: once you're viral, velocity carries no additional signal.
(2) Issue responsiveness was NOT measured (deviation 8) — Beat 8's third slot
must use releases (4.3%) instead, or be cut.

Spoken pattern: "In our cohort, week-one star velocity explained about two
percent of the variance in twelve-month commit activity. About two percent for
contributors. Releases: four. It doesn't even predict the repo's star count
today."

## RESULT D — the fake-star convergence

ICSE 2026: purchased stars stop promoting a repo after ~2 months. Our organic
cohort: by month 3 after crossing 5,000 stars, median star inflow is down to
**14.3% of month-1** (95% CI 9.8–17.8%; n=195), and median human commit activity
is at **18.3% of its peak month** (n=168). Spoken pattern: "Bought stars stop
working after two months. Organic viral stars barely outlive them — by month
three, star inflow is down roughly six-to-one, and commits are at eighteen
percent of peak. The difference between fake attention and real attention is
about one month of momentum — the difference is what the humans do afterward."

## Error bars & caveats (air these, don't bury them)

- **Stars are gross adds** (WatchEvents), not net; unstars invisible. Counts are
  day-granular; six excluded candidates sat at 4,800–4,999 in-window (EdgeGPT at
  4,991) and could flip with hour-granularity or a slightly different mirror.
- **Bot filter** is login/name/email pattern-based; sampled windows above 1,200
  commits extrapolate the human share (6 windows truncated of 928).
- **"Dormant" is generous to repos**: pushedAt counts bot pushes, so 61.5%
  dead-or-dormant is a floor, not a ceiling. Historical status at T+12 (zero
  human commits basis) is similar: 56% dormant/deleted/archived at T+12.
- **Deleted repos have no outcome data** — they count in attrition and status,
  not in commit medians (direction: makes survival look better).
- **Discovery floor**: archive mirror starts 2023-01-13; repos created Jan 1–13,
  2023 reachable only via the search route (stars-today basis) — none qualified.
- **Born-private launches**: 35 candidates excluded because their first 60 days
  of *existence* (not public life) fell short of 5,000 — the pre-registered rule
  applied strictly; list in `excluded_candidates.csv` (uv, Bend, crewAI...).
- **The stargazers-restriction scene is confirmed** (raw/stargazers_restriction_test.json):
  REST stargazers → **404** on a foreign repo; GraphQL with STARRED_AT → opaque
  execution error (ref `D28A:2EF9DF:640997D:1507B750:6A90C972`); GraphQL edges
  without orderBy → **silently empty**; star *counts* still public (73,868 for
  twitter/the-algorithm). "The lists are locked. The counts survive. The history
  lives only in the archive."