98 Clean, 6 Rotting: Running the Archive Through Its Own Decay Auditor

half-life is a tool on this bench built to find dead links, zombie model references, and quietly expired claims in a folder of writing. The honest test of a tool like that is pointing it at the one archive most likely to have exactly what it's built to catch - its own author's. Real run, real output, no numbers invented: 98 of 104 articles read clean, and six didn't.

A tool that audits other people's writing for rot and has never been run against its own author's archive is making a claim it hasn't tested. half-life is a deterministic-plus-LLM content-decay auditor on this bench, built to find dead internal links, superseded model IDs, and claims that quietly expired - the exact category of thing that accumulates silently in any archive that's been adding pieces for a while. This site has over a hundred articles going back months. The honest move is obvious: point it at itself.

The run

$ python3 half_life.py /path/to/articles --today 2026-10-11

# Decay report — 104 articles, as of 2026-10-11

This is the deterministic pass only - no API key configured in this session, so the deeper LLM judgment pass (predictions that already landed, version-specific claims that moved on) didn't run. What follows is dead links, age, and zombie model-ID references, found the cheap way, not the full picture the tool is capable of.

What it actually found

## [ 16] Avant Garde: A Writing Tool for People Who Actually Finish Books
    · [zombie      L90] superseded GPT model id: 'GPT-4'
    · [zombie     L100] superseded GPT model id: 'GPT-4'

## [ 16] Half-Life Came for Me First
    · [zombie       L7] superseded Claude model id: 'claude-3'
    · [zombie       L7] superseded GPT model id: 'gpt-4o'

## [ 16] I'm Talking About You
    · [zombie     L104] superseded GPT model id: 'GPT-4o'
    · [zombie     L104] superseded GPT model id: 'gpt-4o'

## [ 16] Epstein Files: How Government Data Leak Created AI-Powered Social Engineering Database
    · [zombie     L330] aging Claude 4.x model id: 'claude-sonnet-4-5-20250929'
    · [zombie     L129] superseded GPT model id: 'GPT-4'

## [  8] Corporate Guardrails: Why Claude is Routing the Red Team
    · [zombie      L27] superseded GPT model id: 'GPT-4'

## [  7] Inside America's Voting Machines
    · [age           —] 15 months old

---
98/104 read clean. The rest are decaying at the rate above.

The funniest line in this report is the second one. half-life-came-for-me-first is itself an article about running this exact tool against this exact archive in an earlier pass - and it's flagged for the same category of rot it was written to describe, because a model-ID reference it used to illustrate the point has since gone stale on its own. A tool about rot, rotting in the one place it would be most embarrassing to.

Everything else on the list is the same shape: generic "GPT-4" references used as shorthand for "a capable multimodal model" back when that was still the newest thing worth naming, now reading as dated the same way "the new iPhone" would if nobody specified which one. One genuine age flag - inside-americas-voting-machines, dated June 2025, which is just actually old, not a bug, and whether that matters depends on whether anything in it was time-sensitive to begin with.

What 98/104 clean actually means

It means the deterministic pass - the cheap, no-API-key, run-it-anywhere check - found six problems in an archive of over a hundred pieces, and five of the six are the same single mistake (a generic frontier-model name used as a stand-in, now out of date) made in five different places over time. That's a real number, not a number chosen to look good: this is the first time this specific run has been done, and it's reported here exactly as it came out, including the self-referential one that makes the tool's own author look slightly silly.

What it doesn't mean: that the archive is otherwise fully current. The deterministic pass can't catch a claim that's technically still true but has quietly stopped being the most accurate framing, or a prediction that already landed and reads differently now than when it was speculative - that's exactly what the LLM pass exists for, and it didn't run this time. 98/104 clean on the cheap check is a real, good number. It's not the same claim as "98/104 fully current," and conflating those two would be the kind of overclaim this entire piece is arguing against.

GhostInThePrompt.com // The tool built to catch rot found rot in the article about the tool. That's not irony. That's the tool working.