Dataset · Cold Start Data
Every repository that went viral, and what happened to it after.
On June 30, 2026, GitHub restricted the API that told you when a repository's stars arrived. The public event archive still records every one of them, so we pulled the cohort ourselves: all 195 repositories created between January 2023 and June 2025 that gained 5,000 or more stars within 60 days of creation, and their outcomes twelve months on.
This is the whole measurement — the frozen pre-registration, the as-run method with every deviation from it, the raw pulls, and the four scripts that produced the numbers. It backs the episode “We Tracked Every Viral AI Repo. Most Stars Predict Nothing.” on Cold Start Data.
What it found
195
repositories in the cohort
Every public repo created between 2023-01-01 and 2025-06-30 that gained 5,000 or more stars within 60 days of creation. Nine of them no longer exist at all.
0.2%
median human commit activity at month 12, as a share of peak
95% CI 0.0–3.0, n=168 with commit data. 54.3% recorded zero human commits in month twelve. The confidence interval is wide and we report it rather than the point estimate alone.
≤4.3%
of the variance in any health measure explained by launch-week stars
R² on log-transformed values, across commits, contributors, releases, stars today, and still-alive. The rank correlations run slightly negative — more launch-week stars, marginally worse outcomes.
Two numbers in the episode carry a deliberate split: the AI share of the cohort is 59.0% under the rubric frozen before the pull, and 79.5% under a documented manual audit that catches what 2026 metadata hides. The pre-registered number is the one we quote; the audit is published beside it so you can see the gap and judge it yourself.
The dataset
Start with the README — it carries the method and the deviations.
| README.md | 9 KB | The as-run method in full, including the complete list of deviations from the pre-registration. |
| PRE-REGISTRATION.md | 1 KB | The sample design, frozen before any data was pulled and reproduced verbatim. |
| cohort.csv | 40 KB | The final cohort, one row per repository — pre-registered columns first, derived columns after. |
| excluded_candidates.csv | 12 KB | Discovery candidates that failed the creation-window rule, each with its reason. |
| ai_label_borderline_calls.json | 22 KB | Every borderline call made by the frozen AI-labeling rubric, with the evidence behind it. |
| ai_label_manual_review.json | 6 KB | The secondary audit layered on the rubric — every override and every judgment call left standing, with reasons. |
| results.json | 3 KB | The four result blocks the episode narrates, machine-readable. |
| results.md | 6 KB | The same numbers written out with the caveats attached to each. |
Raw pulls
Unmodified, as fetched on 2026-08-27. The discovery and outcome scripts re-run against live sources, so exact counts drift with time — this is the frozen snapshot the episode was built on.
| raw/daily_series.tsv | 3.2 MB | Daily star counts per repository from the public event archive — the series everything else is derived from. |
| raw/cohort_broad.tsv | 13 KB | The unfiltered discovery pull, before the creation-window rule was applied. |
| raw/createvent_creations.tsv | 3 KB | Repository creation timestamps as recorded in the event stream. |
| raw/search_crosscheck.tsv | 143 KB | The independent Search API route, run as the pre-registered cross-check. |
| raw/repo_meta.jsonl | 105 KB | REST metadata for every cohort repository, as served on the run date. |
| raw/outcomes.jsonl | 225 KB | GraphQL outcome pulls: commit history, contributors, releases, archival status. |
| raw/entries.jsonl | 150 KB | The assembled per-repository records the analysis reads. |
| raw/stargazers_restriction_test.json | 862 B | The test that documents what the June 2026 API restriction actually blocks. |
| raw/clickhouse_meta.txt | 115 B | The event-archive currency check: span and event count at run time. |
Reproduction
Run in order. The first script needs no credentials at all.
| scripts/01_discovery_clickhouse.py | 5 KB | Discovery against the public event archive. No token required. |
| scripts/02_search_crosscheck.py | 4 KB | The Search API cross-check. Needs a GitHub token. |
| scripts/03_outcomes_graphql.py | 15 KB | Outcome measurement at T+6 and T+12. Needs a GitHub token. |
| scripts/04_analyze.py | 17 KB | Everything downstream: cohort assembly, labeling, correlations, intervals. |
Take it and argue with it
We publish the measurement because a finding nobody can check isn't a finding. Re-run it, extend the window, disagree with the labeling rubric — attribution is appreciated, and corrections more so. If you find an error, we'll fix it visibly, per our editorial standards.
hello@vantedgevault.com