Wayback Séance: I asked dead websites to tell their own life stories, and made the ghosts cite their sources
Description and use
Wayback Séance reconstructs the life story of defunct websites by pulling real snapshots from the Wayback Machine and using a local Gemma model, ensuring every claim is verifiable.
Problem
How to remember dead websites accurately without fabricating history
Outcome
The Wayback Séance tool enables accurate, verifiable storytelling of dead websites, producing both a terminal séance and an HTML ofrenda that links every claim to an archived snapshot, thereby preventing fabricated narratives
Workflow
- 1. Input the URL of a dead or altered website
- 2. Query the Wayback CDX API to list all captures
- 3. Collapse captures to one per month and select representative snapshots
- 4. Retrieve each snapshot via web.archive.org and extract title, meta description, and visible text
- 5. Feed the extracted data to a local Gemma 3 4B model via llama.cpp
- 6. Stream the model’s output, segmenting into sentences with citation tags
- 7. Run a deterministic checker that verifies each sentence against the cited snapshot
- 8. Output the séance in the terminal and generate an HTML ofrenda with linked citations
- 9. Store or display the final ofrenda
Evidence
The ghost may only remember what the archive kept. Every sentence has to cite a real Wayback Machine snapshot, and a deterministic checker holds each sentence up to that snapshot before you get to see it.
The model is Gemma 3 4B instruction‑tuned, as the Q4_0 GGUF from unsloth/gemma-3-4b-it-GGUF, served by llama.cpp's llama-server in Docker.
All of it uses two public endpoints: The CDX API lists every capture of a URL. web.archive.org/web/<timestamp>id_/<url> returns a capture byte for byte.
- Record date: 2026-10-27T04:12:03.490Z