"This is where it ends", hidden in a 1996 HTML comment nobody was meant to read, is the most haunting line in any Ghost in the Machine entry. The grounding rules are what make it work: nothing outside SOURCE and RECORDS, quotes copied exactly. Since the rules already demand exact quotes, a cheap post-check that every quoted span appears verbatim in the snapshot would turn "the ghost should be honest" into "the ghost is provably honest" without another model call. Lovely way to bring the old web back.
Thank you! That comment really felt like finding a message in a bottle from 1996, so I'm thrilled that beat landed.
Also, the distinction between "the ghost should be honest" and "provably honest" is fantastic. A lightweight deterministic post-check is dead simple to wire up and completely closes the hallucination loophole without another model call. Adding that check into the pipeline tonight. Really appreciate the sharp feedback!
The font ligature bug is such a good reminder that evidence can change at the display layer too. I'd keep a tiny rendering test with literal HTML comments, ampersands and non-ASCII text, then compare what gets copied with the stored snapshot. Since you're decoding old charsets as well, I'd show the capture URL and charset beside the remains so I can trace a surprising character back to the archived bytes.
You hit the nail on the head. A renderer can hallucinate just as easily as a model, just in a different layer of the stack.
Adding the raw capture URL, timestamp, and original encoding badge to the Remains dossier makes total sense—it turns the tab into a true audit log back to the Internet Archive's CDX records. Thanks for the sharp eye!
The CDX link is a nice addition. I’d keep the archived bytes and the decoded text side by side for that ligature example. It gives the reader a way to locate where the difference first appeared.
Electronics engineer. 25 years making different systems work together under real conditions: devices, transactions, backend, data, security and operations. Now, edge AI on real hardware.
Location
Ciudad de México
Education
B.Eng. Electronics and Communications, Universidad Iberoamericana
Kozmo's ghost answering "that never made it into my HTML" is the behavior I wish more pipelines had: saying what isn't on record instead of filling the gap.
On the verbatim check you're adding tonight: it catches invented quotes, but there's a failure it lets through. I had a local Gemma write a morning lab report from a JSON of the night's incidents, and the worst error wasn't an invented sentence. It moved a real one: a fact with a real number, from an incident six hours earlier, attached to the last one. Every word existed in the source; it was pinned to the wrong thing.
Your séance has the same shape with its two blocks. A coroner line shown in marigold as if the site had written it would pass a "this text appears in the prompt" check. Checking each quote against the block it claims to come from, SOURCE for marigold and RECORDS for [n], closes that one too.
Have you caught a ghost quoting the coroner as its own words?
That Gemma incident-report failure is such a sharp diagnosis of the flat-context problem. When every token is technically "real," temporal and entity misattribution is ten times more dangerous than outright hallucination because simple checks wave it right through.
To answer your question: yes, absolutely. In earlier runs, I caught ghosts lifting cold autopsy language from the bankruptcy post-mortems in RECORDS and rendering it in marigold as if it had been headline copy on their 1999 homepage. The moment a ghost casually narrates its own SEC filings in the first person, the séance breaks—it stops being a trapped, bounded historical artifact and turns into an omniscient search engine.
Partitioning the assertion—marigold strictly verified against SOURCE, bracketed citations strictly verified against RECORDS—is the exact architectural fix. It enforces an existential boundary between what the site knew while alive and what only the coroner could see. Brilliant catch.
Electronics engineer. 25 years making different systems work together under real conditions: devices, transactions, backend, data, security and operations. Now, edge AI on real hardware.
Location
Ciudad de México
Education
B.Eng. Electronics and Communications, Universidad Iberoamericana
Thank you, Kanak, and for answering with your own runs. Autopsy language from RECORDS showing up in marigold is the clearest example of the problem I've seen: every word is real, and it still lies about who said it. That boundary is exactly what my morning report was missing: what was known when the incident happened, and what only came later.
"What was known then vs. what was known later" sums up the entire challenge. Once grounding is treated as an epistemic and temporal boundary rather than just raw token inclusion, pipelines get dramatically more reliable. Really enjoyed swapping notes on this!
Software Developer & Creator focused on building Windows desktop applications, productivity tools, and useful software.
Creator of Task Monitor and FazaFix. Interested in C#, .NET, Windows Development
Yep! They are 31 iconic websites and personal homepages from the 90s and early 2000s that shut down years ago—like Kozmo.com, old GeoCities Area51 homepages, AltaVista, and Pseudo.com.
The project pulls their original archived HTML out of the Wayback Machine so you can browse their remains and talk to them like ghosts. You can also paste in any dead URL to resurrect it into the closet live!
We're a place where coders share, stay up-to-date and grow their careers.
"This is where it ends", hidden in a 1996 HTML comment nobody was meant to read, is the most haunting line in any Ghost in the Machine entry. The grounding rules are what make it work: nothing outside SOURCE and RECORDS, quotes copied exactly. Since the rules already demand exact quotes, a cheap post-check that every quoted span appears verbatim in the snapshot would turn "the ghost should be honest" into "the ghost is provably honest" without another model call. Lovely way to bring the old web back.
Thank you! That comment really felt like finding a message in a bottle from 1996, so I'm thrilled that beat landed.
Also, the distinction between "the ghost should be honest" and "provably honest" is fantastic. A lightweight deterministic post-check is dead simple to wire up and completely closes the hallucination loophole without another model call. Adding that check into the pipeline tonight. Really appreciate the sharp feedback!
The font ligature bug is such a good reminder that evidence can change at the display layer too. I'd keep a tiny rendering test with literal HTML comments, ampersands and non-ASCII text, then compare what gets copied with the stored snapshot. Since you're decoding old charsets as well, I'd show the capture URL and charset beside the remains so I can trace a surprising character back to the archived bytes.
You hit the nail on the head. A renderer can hallucinate just as easily as a model, just in a different layer of the stack.
Adding the raw capture URL, timestamp, and original encoding badge to the Remains dossier makes total sense—it turns the tab into a true audit log back to the Internet Archive's CDX records. Thanks for the sharp eye!
The CDX link is a nice addition. I’d keep the archived bytes and the decoded text side by side for that ligature example. It gives the reader a way to locate where the difference first appeared.
Kozmo's ghost answering "that never made it into my HTML" is the behavior I wish more pipelines had: saying what isn't on record instead of filling the gap.
On the verbatim check you're adding tonight: it catches invented quotes, but there's a failure it lets through. I had a local Gemma write a morning lab report from a JSON of the night's incidents, and the worst error wasn't an invented sentence. It moved a real one: a fact with a real number, from an incident six hours earlier, attached to the last one. Every word existed in the source; it was pinned to the wrong thing.
Your séance has the same shape with its two blocks. A coroner line shown in marigold as if the site had written it would pass a "this text appears in the prompt" check. Checking each quote against the block it claims to come from, SOURCE for marigold and RECORDS for [n], closes that one too.
Have you caught a ghost quoting the coroner as its own words?
That Gemma incident-report failure is such a sharp diagnosis of the flat-context problem. When every token is technically "real," temporal and entity misattribution is ten times more dangerous than outright hallucination because simple checks wave it right through.
To answer your question: yes, absolutely. In earlier runs, I caught ghosts lifting cold autopsy language from the bankruptcy post-mortems in RECORDS and rendering it in marigold as if it had been headline copy on their 1999 homepage. The moment a ghost casually narrates its own SEC filings in the first person, the séance breaks—it stops being a trapped, bounded historical artifact and turns into an omniscient search engine.
Partitioning the assertion—marigold strictly verified against SOURCE, bracketed citations strictly verified against RECORDS—is the exact architectural fix. It enforces an existential boundary between what the site knew while alive and what only the coroner could see. Brilliant catch.
Thank you, Kanak, and for answering with your own runs. Autopsy language from RECORDS showing up in marigold is the clearest example of the problem I've seen: every word is real, and it still lies about who said it. That boundary is exactly what my morning report was missing: what was known when the incident happened, and what only came later.
"What was known then vs. what was known later" sums up the entire challenge. Once grounding is treated as an epistemic and temporal boundary rather than just raw token inclusion, pipelines get dramatically more reliable. Really enjoyed swapping notes on this!
31 Websites Dead?
Yep! They are 31 iconic websites and personal homepages from the 90s and early 2000s that shut down years ago—like Kozmo.com, old GeoCities Area51 homepages, AltaVista, and Pseudo.com.
The project pulls their original archived HTML out of the Wayback Machine so you can browse their remains and talk to them like ghosts. You can also paste in any dead URL to resurrect it into the closet live!