methodology.md

# Methodology — Nerq MCP-server repository mortality (2026-07)

**As of:** 2026-07-22 · **License:** CC-BY 4.0 · Reproducible from public sources with no API key.

This document states how each `state` in the dataset is derived so the dump can be independently
reproduced and spot-checked.

## 1. Population

Distinct public GitHub repositories (`owner/repo`) linked from indexed Model Context Protocol (MCP)
servers. 28,824 distinct repositories in this dump. Population definition is honest and bounded: it is
the set of GitHub repositories reachable from the MCP servers Nerq indexes, not "all MCP servers that
have ever existed."

## 2. Reads (source of truth)

For each repository, two first-hand reads are taken (`provenance = own`):

1. **HTTP status** of the repository — `200` reachable, `404` gone (deleted or made private).
2. **Last commit timestamp** — the latest entry in the repository's public `commits.atom` feed
   (`https://github.com/{owner}/{repo}/commits.atom`). No API key. Not rate-limited. The timestamp is
   the repository's own last public commit on its default branch; it is a repository fact, not an
   assertion about any person.

## 3. State derivation (exact rules)

Let `age = observed_date − last_commit` in days. Then:

```
if http_status == 404:                      state = "deleted"     # gone or made private
elif last_commit is unreadable/none:        state = "unknown"
elif age <= 180:                            state = "alive"
elif 180 <  age <= 365:                     state = "dormant"
else (age > 365):                           state = "abandoned"
```

- **`deleted`** is authoritative on the HTTP read (`404`), heuristic on the cause (deleted vs private).
- **`alive` / `dormant` / `abandoned`** are heuristic thresholds on commit recency. The thresholds
  (180 / 365 days) are stated here and are the only knobs; changing them re-buckets the same reads.
- **`unknown`** (41 rows) is recorded, not guessed — the row exists so the absence is auditable.

## 4. Interval-censoring

An observation dates a *state*, not a *transition*. A `deleted` row means "reachable-or-not was `404`
on `observed_date`"; the actual deletion happened at some unknown point between the repo's
`last_commit` and `observed_date`. This dump is a single seed day (`2026-07-22`), so death dates are
interval-censored to `[last_commit, 2026-07-22]`. The living series narrows these with dated `re-probe`
rows; collection gaps (a known one on **2026-07-30**) are disclosed, and any transition spanning a gap
is censored to the surrounding interval rather than pinned.

## 5. Reproduction

To reproduce any single row: fetch `https://github.com/{entity_id}` (expect `200` or `404`) and
`https://github.com/{entity_id}/commits.atom` (read the first `<updated>` timestamp), then apply the
rules in §3 with `observed_date = 2026-07-22`. No credentials required.

## 6. Known boundaries

- `commits.atom` reflects the default branch only; a repo committing on a non-default branch could read
  as more stale than it is.
- The 180-day boundary class (repos whose last commit sits within a few days of the threshold) may flip
  between `alive` and `dormant` depending on the exact read time.
- `404` conflates deletion and privatization; the dataset does not claim to distinguish them.