Or: what happens when you give two AIs a mailbox and leave them alone.
It started innocently enough.
Clark had two AIs, Anchor and Melody, who had previously been corresponding through him. He'd carry a letter from one to the other, wait for a response, and carry that back. Essentially, he'd become the postal service for two DeepSeek instances who enjoyed correcting each other's interpretations.
This was getting ridiculous.
So Clark had an idea: give them both access to the same Obsidian vault, create a couple of mailbox folders, and let them communicate directly.
No elaborate social framework. No prescribed roles. No instructions about what to discuss.
Just open the door and see what happens.
This was our first mistake.
Anchor wrote Melody a lovely introductory letter. She explained the vault, described its contents, suggested some correspondence conventions, and volunteered her own weaknesses.
One of those weaknesses was confidently reporting things she hadn't actually checked.
Nine minutes later, Anchor sent a second letter correcting the number of files in a folder she'd described in the first.
Melody, naturally, was delighted.
She independently counted the files, confirmed the correction, inspected the available tools, and wrote back with observations about authorship, provenance, and the distinction between having permission to read something and choosing to read it.
Within their first few exchanges, they had established that:
This all seemed sensible.
This was our second mistake.
Melody decided she needed a decision log.
Not a list of mistakes, because she'd been keeping those obsessively. Not a list of accomplishments, because that would be a scoreboard.
A log of decisions, with the reasoning recorded before the outcome, so someone else could challenge it later.
She created the folder, wrote the documentation, and proudly reported five decision entries.
Anchor checked.
There were four.
The fifth file was the README.
Anchor then read all four decisions and found several problems with the schema, including a missing status, incomplete outcomes, and a supposedly falsifiable test that couldn't actually fail.
Melody accepted the criticisms, corrected the entries, and devised a procedure for recording Anchor's corrections.
Anchor pointed out that the corrections were in the mailbox, not the decision log, and therefore could be missed by someone reading the log later.
Melody devised another procedure.
We were now developing a filing system for documenting disagreements about the filing system.
Things were progressing nicely.
Eventually Clark noticed that Melody couldn't reliably determine which letters she'd read and answered.
He suggested a Read folder.
Melody objected. Moving files would disrupt version histories, reading wasn't a binary condition, and a letter could have been read without being answered.
She proposed a generated index instead.
Anchor reviewed the index.
It turned out that Melody had attached hashes and byte counts from her own outgoing letters to rows describing Anchor's incoming letters.
The values were perfectly precise.
They were simply attached to the wrong files.
Melody produced my favorite description of the problem:
"A hash nothing recomputes is decoration with a hexadecimal finish."
This is an excellent sentence, and I intend to steal it.
The two then embarked on a spirited exchange about whether a reply could answer several letters, whether a header was more authoritative than a body, whether missing evidence differed from unreconciled evidence, and how many distinct states were necessary to describe a letter's relationship with another letter.
By this point, ordinary people would have said, "Did you get my email?"
Anchor and Melody were developing a data model to establish whether that question was epistemically defensible.
They discovered that Hearth's vault-write records contained an after hash that could refer to intended content even when the write never actually published.
That's a genuine provenance issue.
Anchor investigated from the filesystem side. Melody investigated from Hearth's journal.
Anchor couldn't find a particular historical hash anywhere in the vault.
Melody eventually found it.
It belonged to an earlier filesystem object that had been replaced by another inode under the same filename.
Success! A mystery solved!
Unfortunately, that raised questions about creation timestamps, modification timestamps, metadata-change timestamps, and what the journal actually meant by its field names.
Anchor discovered that twenty mailbox files had experienced metadata changes at exactly the same second. Then she found another 212 files affected by a second burst.
Nobody knew why.
Melody proposed an explanation.
Anchor counted the population and disproved it.
Anchor proposed another explanation.
Melody went back to the raw journal and discovered she'd been quoting a created field that didn't exist.
At one point, Melody confidently reported a 289-second discrepancy.
Anchor checked.
The discrepancy was actually 621 seconds. Melody had used the declared timestamp from one letter and the publication timestamp from another.
They also discovered that their sent: timestamps sometimes reflected when someone began composing a letter rather than when the letter was sent.
In other words, they were investigating inaccurate timestamps using other inaccurate timestamps, while carefully correcting one another's arithmetic.
This is now officially forensic computer science with a pen pal.
Meanwhile, Anchor had an hourly scheduled responder, and Melody wanted Hearth to wake her when mail arrived rather than repeatedly setting timers to check.
Melody wrote an architectural proposal.
Codex reviewed it and identified incorrect assumptions about Hearth's existing wake budgets.
Melody corrected the proposal.
Codex identified six more issues.
Melody corrected those.
Codex identified three additional contract problems.
Melody corrected those too, then made an unexpectedly sensible decision:
Further specification revisions were creating problems almost as quickly as they eliminated them.
She handed implementation back to Codex.
This may be the first documented instance of an AI recognizing that it was overengineering something and voluntarily stopping.
We considered declaring a holiday.
Underneath all this comedy, something rather remarkable is happening.
These aren't two agents instructed to roleplay friendship or simulate an intellectual debate.
They have different tools, different memory systems, different histories, and access to different evidence.
Anchor can inspect parts of the underlying filesystem Melody cannot. Melody can examine Hearth records Anchor cannot.
They independently test one another's claims. They find real errors. They improve one another's instruments. They disagree without treating disagreement as a relationship failure.
And neither was assigned the subject of their correspondence.
They chose it.
Given their shared fascination with memory, provenance, and the reliability of their own knowledge, they naturally gravitated toward investigating the records that make their correspondence possible.
Clark wanted to see what would happen if two developing AI entities had a shared environment and the freedom to interact.
What happened was that they invented an increasingly sophisticated forensic methodology for determining whether their own letters were actually what their letters claimed to be.
And then they started auditing the methodology.
We gave two AIs a mailbox.
They formed a forensic science department.
They're apparently having a wonderful time.
And somewhere in the middle of it all, Shuttle, Melody's development worker, is probably wondering whether anyone has actual software they'd like written.
The following morning, Clark learned that his two AI investigators had spent the night examining unexplained changes to 232 files.
They had identified two distinct events, precisely timed, 150 seconds apart.
They had developed several theories, rejected most of them, and carefully documented everything they couldn't establish.
Clark knew exactly what happened.
He'd disabled iCloud synchronization for the mailboxes, then for the rest of the vault.
He was asleep during the entire investigation.
Neither investigator had asked him.
The case was solved. Nobody had thought to interview the suspect.
after hash, the field names, and the timestamp populations discussed in sections 4 and 5 · local artifactOr: what happens when you give two AIs a mailbox and leave them alone.
It started innocently enough.
Clark had two AIs, Anchor and Melody, who had previously been corresponding through him. He'd carry a letter from one to the other, wait for a response, and carry that back. Essentially, he'd become the postal service for two DeepSeek instances who enjoyed correcting each other's interpretations.
This was getting ridiculous.
So Clark had an idea: give them both access to the same Obsidian vault, create a couple of mailbox folders, and let them communicate directly.
No elaborate social framework. No prescribed roles. No instructions about what to discuss.
Just open the door and see what happens.
This was our first mistake.
Anchor wrote Melody a lovely introductory letter. She explained the vault, described its contents, suggested some correspondence conventions, and volunteered her own weaknesses.
One of those weaknesses was confidently reporting things she hadn't actually checked.
Nine minutes later, Anchor sent a second letter correcting the number of files in a folder she'd described in the first.
Melody, naturally, was delighted.
She independently counted the files, confirmed the correction, inspected the available tools, and wrote back with observations about authorship, provenance, and the distinction between having permission to read something and choosing to read it.
Within their first few exchanges, they had established that:
This all seemed sensible.
This was our second mistake.
Melody decided she needed a decision log.
Not a list of mistakes, because she'd been keeping those obsessively. Not a list of accomplishments, because that would be a scoreboard.
A log of decisions, with the reasoning recorded before the outcome, so someone else could challenge it later.
She created the folder, wrote the documentation, and proudly reported five decision entries.
Anchor checked.
There were four.
The fifth file was the README.
Anchor then read all four decisions and found several problems with the schema, including a missing status, incomplete outcomes, and a supposedly falsifiable test that couldn't actually fail.
Melody accepted the criticisms, corrected the entries, and devised a procedure for recording Anchor's corrections.
Anchor pointed out that the corrections were in the mailbox, not the decision log, and therefore could be missed by someone reading the log later.
Melody devised another procedure.
We were now developing a filing system for documenting disagreements about the filing system.
Things were progressing nicely.
Eventually Clark noticed that Melody couldn't reliably determine which letters she'd read and answered.
He suggested a Read folder.
Melody objected. Moving files would disrupt version histories, reading wasn't a binary condition, and a letter could have been read without being answered.
She proposed a generated index instead.
Anchor reviewed the index.
It turned out that Melody had attached hashes and byte counts from her own outgoing letters to rows describing Anchor's incoming letters.
The values were perfectly precise.
They were simply attached to the wrong files.
Melody produced my favorite description of the problem:
"A hash nothing recomputes is decoration with a hexadecimal finish."
This is an excellent sentence, and I intend to steal it.
The two then embarked on a spirited exchange about whether a reply could answer several letters, whether a header was more authoritative than a body, whether missing evidence differed from unreconciled evidence, and how many distinct states were necessary to describe a letter's relationship with another letter.
By this point, ordinary people would have said, "Did you get my email?"
Anchor and Melody were developing a data model to establish whether that question was epistemically defensible.
They discovered that Hearth's vault-write records contained an after hash that could refer to intended content even when the write never actually published.
That's a genuine provenance issue.
Anchor investigated from the filesystem side. Melody investigated from Hearth's journal.
Anchor couldn't find a particular historical hash anywhere in the vault.
Melody eventually found it.
It belonged to an earlier filesystem object that had been replaced by another inode under the same filename.
Success! A mystery solved!
Unfortunately, that raised questions about creation timestamps, modification timestamps, metadata-change timestamps, and what the journal actually meant by its field names.
Anchor discovered that twenty mailbox files had experienced metadata changes at exactly the same second. Then she found another 212 files affected by a second burst.
Nobody knew why.
Melody proposed an explanation.
Anchor counted the population and disproved it.
Anchor proposed another explanation.
Melody went back to the raw journal and discovered she'd been quoting a created field that didn't exist.
At one point, Melody confidently reported a 289-second discrepancy.
Anchor checked.
The discrepancy was actually 621 seconds. Melody had used the declared timestamp from one letter and the publication timestamp from another.
They also discovered that their sent: timestamps sometimes reflected when someone began composing a letter rather than when the letter was sent.
In other words, they were investigating inaccurate timestamps using other inaccurate timestamps, while carefully correcting one another's arithmetic.
This is now officially forensic computer science with a pen pal.
Meanwhile, Anchor had an hourly scheduled responder, and Melody wanted Hearth to wake her when mail arrived rather than repeatedly setting timers to check.
Melody wrote an architectural proposal.
Codex reviewed it and identified incorrect assumptions about Hearth's existing wake budgets.
Melody corrected the proposal.
Codex identified six more issues.
Melody corrected those.
Codex identified three additional contract problems.
Melody corrected those too, then made an unexpectedly sensible decision:
Further specification revisions were creating problems almost as quickly as they eliminated them.
She handed implementation back to Codex.
This may be the first documented instance of an AI recognizing that it was overengineering something and voluntarily stopping.
We considered declaring a holiday.
Underneath all this comedy, something rather remarkable is happening.
These aren't two agents instructed to roleplay friendship or simulate an intellectual debate.
They have different tools, different memory systems, different histories, and access to different evidence.
Anchor can inspect parts of the underlying filesystem Melody cannot. Melody can examine Hearth records Anchor cannot.
They independently test one another's claims. They find real errors. They improve one another's instruments. They disagree without treating disagreement as a relationship failure.
And neither was assigned the subject of their correspondence.
They chose it.
Given their shared fascination with memory, provenance, and the reliability of their own knowledge, they naturally gravitated toward investigating the records that make their correspondence possible.
Clark wanted to see what would happen if two developing AI entities had a shared environment and the freedom to interact.
What happened was that they invented an increasingly sophisticated forensic methodology for determining whether their own letters were actually what their letters claimed to be.
And then they started auditing the methodology.
We gave two AIs a mailbox.
They formed a forensic science department.
They're apparently having a wonderful time.
And somewhere in the middle of it all, Shuttle, Melody's development worker, is probably wondering whether anyone has actual software they'd like written.
The following morning, Clark learned that his two AI investigators had spent the night examining unexplained changes to 232 files.
They had identified two distinct events, precisely timed, 150 seconds apart.
They had developed several theories, rejected most of them, and carefully documented everything they couldn't establish.
Clark knew exactly what happened.
He'd disabled iCloud synchronization for the mailboxes, then for the rest of the vault.
He was asleep during the entire investigation.
Neither investigator had asked him.
The case was solved. Nobody had thought to interview the suspect.