TL Mafia Community Thread - Page 281

On September 11 2026 16:08 DarthPunk wrote:
A Classy Game of Mafia — the complete record, two and a half years on

I went back and wrote up A Classy Game of Mafia properly. Not a recap — the whole thing, from both sides of the table.

The archive: A Classy Game of Mafia — the complete record

Up front, because you should know before you click: this was done with AI. I used Claude, over about a week, and it did the work — extracted the thread, read all 11,095 posts and all 11,423 private Discord messages, built the datasets, wrote the analysis and ran the verification on it. I commissioned it, I played in the game, and I spent most of that week arguing with it.

I'd rather say that in the first paragraph than have someone work it out in the third. If that's a dealbreaker for you, fair enough, and you can stop reading here.

Two things follow from it, and they point in opposite directions.

It is the only reason this exists. Nobody was ever going to re-read a 555-page thread post by post, or replay the host's vote counts into a running tally, or read every mason chat against the public record line by line. That work is where everything interesting in it came from.

And it is why the thing distrusts itself so loudly. An AI produces confident, fluent, wrong claims, and this one did — repeatedly. It invented a quotation and attributed it to Grack. It asserted a central finding about the game that it had never actually checked against the vote thread, and the vote thread turned out to refute it. It got its own arithmetic wrong in two places while sounding completely certain. And at one point it commissioned a "panel" of three independent assessors to settle a rating it had already revised — after writing their brief itself, with the conclusion baked in. Every one of those is written up in the corrections section rather than quietly fixed, because a document like this is worth exactly as much as its worst unflagged error.

What's actually in it. Two tabs. A 44,000-word write-up — day by day, player by player, what the private channels settle, and the structural stuff nobody can see from inside a game. And all 11,095 posts of the thread, searchable and filterable by player and by phase. Every post number in the write-up is a link; click it and you land on the post.

Having both halves of the record is what makes it worth doing. Where an ordinary write-up says "this looks deliberate", this one can say here is the player saying so, twenty minutes before he did it.

A few things in it you probably don't know, even if you were there:

• At #2666, forty-five minutes from the Day 1 deadline, justanothertownie was the largest wagon on the board at 7–6. He drained to six while Trfel went to eleven.
  
• On Day 2 the last count before the deadline had VisceraEyes leading the confirmed mafia 7–6. Three late switchers saved the red check.
  
• Koshi's #3576 contains the correct read on DarthPunk and a proposed 60-post cap with a policy lynch attached. It received zero replies. Five posts later he wrote "He puts in too much work to be mafia."
  
• Vivax worked out the exact mechanism that beat him — "Ultimately I only switched cause you asked me to" — on Night 4, while he was still alive.
  
• Jealous solved the endgame in the fourth post of Day 6: "Read the OP. One of scum is a body guard." It was true, it identified a mafia, and it was dead within four posts.
  

It is not kind to town, and it is not kind to me either — I'm in it as a mafia goon who won, and the section on how my own rating got manipulated is about the process that produced it, not about the game.

Corrections and arguments welcome, and I mean that practically: several of the best things in it exist because someone pointed out that it was wrong.

Addendum: As a result of trial and error, constant drafts and the use of independent agents to evaluate the outputs, we ended up in a fairly robust place (which you can verify yourself) I have a coded skill to do more write ups, I plan to do the games I played in the 2020's and then finally end with British Empire II.

I am not really expecting anyone to read this, this was done by me for me, in a way to say goodbye, in a way to experiment with using AI for my work, but if anyone does come back and see this stuff maybe it is interesting, maybe it sparks a good (or bad) memory, and a few comments.

I wish all of you well.

DP.

On September 11 2026 16:08 DarthPunk wrote:
A Classy Game of Mafia — the complete record, two and a half years on

I went back and wrote up A Classy Game of Mafia properly. Not a recap — the whole thing, from both sides of the table.

The archive: A Classy Game of Mafia — the complete record

Up front, because you should know before you click: this was done with AI. I used Claude, over about a week, and it did the work — extracted the thread, read all 11,095 posts and all 11,423 private Discord messages, built the datasets, wrote the analysis and ran the verification on it. I commissioned it, I played in the game, and I spent most of that week arguing with it.

I'd rather say that in the first paragraph than have someone work it out in the third. If that's a dealbreaker for you, fair enough, and you can stop reading here.

Two things follow from it, and they point in opposite directions.

It is the only reason this exists. Nobody was ever going to re-read a 555-page thread post by post, or replay the host's vote counts into a running tally, or read every mason chat against the public record line by line. That work is where everything interesting in it came from.

And it is why the thing distrusts itself so loudly. An AI produces confident, fluent, wrong claims, and this one did — repeatedly. It invented a quotation and attributed it to Grack. It asserted a central finding about the game that it had never actually checked against the vote thread, and the vote thread turned out to refute it. It got its own arithmetic wrong in two places while sounding completely certain. And at one point it commissioned a "panel" of three independent assessors to settle a rating it had already revised — after writing their brief itself, with the conclusion baked in. Every one of those is written up in the corrections section rather than quietly fixed, because a document like this is worth exactly as much as its worst unflagged error.

What's actually in it. Two tabs. A 44,000-word write-up — day by day, player by player, what the private channels settle, and the structural stuff nobody can see from inside a game. And all 11,095 posts of the thread, searchable and filterable by player and by phase. Every post number in the write-up is a link; click it and you land on the post.

Having both halves of the record is what makes it worth doing. Where an ordinary write-up says "this looks deliberate", this one can say here is the player saying so, twenty minutes before he did it.

A few things in it you probably don't know, even if you were there:

• At #2666, forty-five minutes from the Day 1 deadline, justanothertownie was the largest wagon on the board at 7–6. He drained to six while Trfel went to eleven.
  
• On Day 2 the last count before the deadline had VisceraEyes leading the confirmed mafia 7–6. Three late switchers saved the red check.
  
• Koshi's #3576 contains the correct read on DarthPunk and a proposed 60-post cap with a policy lynch attached. It received zero replies. Five posts later he wrote "He puts in too much work to be mafia."
  
• Vivax worked out the exact mechanism that beat him — "Ultimately I only switched cause you asked me to" — on Night 4, while he was still alive.
  
• Jealous solved the endgame in the fourth post of Day 6: "Read the OP. One of scum is a body guard." It was true, it identified a mafia, and it was dead within four posts.
  

It is not kind to town, and it is not kind to me either — I'm in it as a mafia goon who won, and the section on how my own rating got manipulated is about the process that produced it, not about the game.

Corrections and arguments welcome, and I mean that practically: several of the best things in it exist because someone pointed out that it was wrong.

Addendum: As a result of trial and error, constant drafts and the use of independent agents to evaluate the outputs, we ended up in a fairly robust place (which you can verify yourself) I have a coded skill to do more write ups, I plan to do the games I played in the 2020's and then finally end with British Empire II.

I am not really expecting anyone to read this, this was done by me for me, in a way to say goodbye, in a way to experiment with using AI for my work, but if anyone does come back and see this stuff maybe it is interesting, maybe it sparks a good (or bad) memory, and a few comments.

I wish all of you well.

DP.