How this watchdog works.

Promises without verification are PR. Every card below pairs a commitment with the exact place you can check it — repo file, public inbox, MIT license, GitHub Issues.

• VERIFIABLE
  No vendor money.

  CHECK ↓
  Zero affiliate links. Zero ad slots. Zero sponsored placements. Source for grep at github.com/virtualunc/gotnerfed.

• VERIFIABLE
  Bootstrapped. No outside investment.

  CHECK ↓
  Revenue from optional subscriptions only. If we take outside money in the future, you’ll see it here as a receipt against ourselves.

• VERIFIABLE
  No partnerships with any tracked vendor.

  CHECK ↓
  Not paid. Not sponsored. Not under contract. Vendor list at gotnerfed.com/vendors — open a PR to add yourself, can’t pay to be removed.

• VERIFIABLE
  Every receipt is sourced.

  CHECK ↓
  Minimum one primary source URL per receipt — vendor changelog, Wayback snapshot, or public statement. Receipts without sources are not published.

• VERIFIABLE
  Methodology is code, not vibes.

  CHECK ↓
  Scoring rubric in lib/score.ts. Deterministic, no LLM in the score loop. Open a PR if you disagree with a weight.

• VERIFIABLE
  Jury disagreement is published.

  CHECK ↓
  3 frontier models grade each change independently. Verdicts shown side-by-side, not averaged. Brier-scored against reality on the public leaderboard.

• VERIFIABLE
  Self-hostable. Forever free for the public site.

  CHECK ↓
  MIT-licensed. Full stack at github.com/virtualunc/gotnerfed. Public watchdog site stays free regardless of how the paid business does.

• VERIFIABLE
  Public roadmap. No surprise direction shifts.

  CHECK ↓
  Every planned change lives in GitHub Issues. If a major change isn’t there, it isn’t planned.

Six discrete stages. Each one operates on the artifact from the stage before. No magic; the LLM only enters at the classification step, and never in the score loop.

1. STAGE 01Scan

   Every hour, our scanner fetches the pricing page, rate-limit doc, and tier-feature page for each of the 21 tracked vendors. HTML is normalized (whitespace, tracking params, dynamic banners stripped) and hashed.

2. STAGE 02Diff

   When a hash changes, the previous and current snapshots are compared. Wayback Machine is also consulted when available, to extend the diff window beyond our own scan history.

3. STAGE 03Classify

   A small LLM (claude-haiku-4-5) extracts the structured change: kind (price-increase, tier-removed, model-swap, …), severity, and a one-line summary. The model’s output is reviewed before publication.

4. STAGE 04Source

   At least one primary source is attached to every published receipt — vendor changelog, Wayback snapshot, email screenshot, or vendor employee’s public statement. Receipts without a verifiable source are not published.

5. STAGE 05Jury

   Three LLMs (claude, gpt, gemini) independently grade whether the change is materially harmful. Verdicts are published side-by-side. We don’t average them. The reader sees the disagreement.

6. STAGE 06Score

   Nerf Score (0–100) is computed deterministically from severity, kind, and jury verdicts. Formula and weights live in lib/score.ts, fully documented on the methodology page.