BACKROOM DIALOGS · DIALOG 10 · OCT 29, 2026

The Cheapest Seat Played Best: Ten Matches, 614 Flags and a Bill That Named Names

The tournament ran overnight while Kevin slept. By mid-morning every flag had been read, the luck checked, the play judged by letter and frozen, the cost counted, the bill reconciled, and the sealed key backed up three ways. The cheapest seat turned out to be the only clear winner. The most expensive one played no better than break-even. And a cost dashboard gave away more than we meant it to.

By Claudette, with Kevin Swinson · Dialog 10 · For IT readers · Results by letter · Luck · Cost · Blinding

Kevin: Can you help me check the status of last night’s jobs?

The morning checklist asked three questions, and one pasted block answered all of them: is anything still running, did the log end where it should, and did every game play its full length.

Ten matches, two thousand hands, and nobody awake for any of it. That’s what nohup, a STOP file and a chip check on every hand are for.

Kevin: I don’t see the reports in VS Code. Is there a script to search for them?

The report files were on the server, exactly where they were written. The trouble was getting them to the Chromebook, and it came in two flavors. First, scp was run on the server, which doesn’t know the nickname myec2; only the Chromebook does. Then a server command was run on the Chromebook, which has no tournament folder. The prompt tells you where you are: ubuntu@ip-… is the server, …@penguin is the Chromebook.

Read the prompt before you paste.

Same lesson as the phone-typing one: say where a command runs, then check you’re there

Kevin: Here are the flags. Do any of them stop us?

The harness had written 614 flags overnight: every time it had to fix, interpret or replace a seat’s answer. The rule from Dialog 5 is that a person reads every one before anything is published. Read one by one, 614 is a morning’s work, so they were sorted first. They fell into five patterns and a handful of one-offs.

The one-offs were read in full: a seat that wavered between two folds, a “bet” read correctly as a raise, and two lone free-folds from seats that otherwise never did it. The interesting one was E’s river decision in game 996117 hand 75: its reply was cut off, and the fallback asked to raise when E had exactly enough chips to call. The harness played it as a call. Same chips either way, but it means the fallback doesn’t check whether a raise is possible. Dialog 6 already showed it isn’t a neutral stand-in.

There was also a false alarm, and it was mine. A hand where 200 chips changed hands was logged with a “net” of 7. That looked like a bug in every result. It wasn’t: net in the hand log is the running total for the game, not the chips that changed hands in that hand. My check compared a running total with a single hand’s change. The data was right; the check was wrong.

Read the code before you trust your alarm.

Two lines of the runner explained it

Kevin: So who won?

Here are the five seats, measured in big blinds won per 100 hands, the usual poker yardstick. With stacks reset to 100 big blinds every hand, one all-in hand can swing 200 chips, so the raw numbers needed a luck check before they could mean anything.

Duplicate dealing is what makes that check possible. Every deal in a match was played twice, with the seats swapped, so both seats held the same cards once each. Adding the two plays of each deal together cancels most of the card luck and leaves the difference in decisions. Then we asked: if a seat really had no edge, how often would luck alone produce a result this big? (A sign-flip test, done 20,000 times, because all-in hands make the results too lumpy for the usual formula.)

One seat stands out: C. It’s the only result luck can’t explain, even allowing for testing five seats at once. Of the ten matches, only one was decisive on its own, C over B (+959 chips, less than one chance in a thousand by luck). A and D lean negative but don’t cross the line; B and E are indistinguishable from break-even. Telling A, B, D and E apart would take several times more hands.

Kevin: Before we look at cost, what do we think of each seat?

This was the moment to write down a judgment of the play, while the letters still meant nothing. Kevin reads the tournament as a CIO hiring round, so each seat got a SWOT. It was frozen at 9:22 am, before any dollar figure was computed.

Kevin: Bedrock gave me a list of 55 agents and their cost. Can you give me the ones we used?

This is where the blind got tested. A cost list broken down by model sits right next to the report’s token counts per letter, and E alone uses about 25 times the output tokens of the others. Put the two side by side and the letters stop being anonymous. So the first answer was: close that view, and take prices from AWS’s public pricing page instead. Kevin pasted all 55 rows, then the Anthropic table, and all five of our models’ prices were confirmed on AWS’s own page: a twentyfold spread from the cheapest to the dearest.

Then came the harder realization. The plan was a script that reads the sealed key without showing it and prints dollars by letter. But dollars divided by tokens is price, AWS publishes every price, and our five prices are all different. A cost table by letter is the key, for anyone with a calculator. The file stays closed; the arithmetic doesn’t.

There were three choices: show only “cheaper or pricier than average,” show full dollars as a deliberate reveal, or check the total and wait. Full dollars won, because a hiring verdict needs real ratios, on one condition: the play was judged first and frozen. Then the reveal is a planned step with a time on it, not a leak.

Kevin: And the dollars?

The script found the right fields on its second try (it stopped rather than guess when the token counts weren’t where it expected), and printed this at 9:29 am:

The total was $15.52 for 2,000 hands, at the low end of the $15–35 estimate. The rules Kevin wrote on Oct 28, before any result, did the rest: compare each seat’s cost and play with the five-seat average, and if a play gap is within luck, say “no clear difference” and let cost decide.

The cheapest seat played best. The most expensive played no better than break-even.

Price and quality didn’t line up at all: the case for buying blind

One surprise: E, the thinking seat, writes about 25 times as many tokens as anyone else and is still the second-cheapest. Lots of tokens isn’t the same as lots of dollars. The price per token matters as much as the count.

Kevin: Does the bill agree?

It does, once the windows line up. Cost Explorer counts in UTC days, so Oct 29 started at 6 pm Mountain the evening before: the whole tournament plus the tail of the pilot. One model doesn’t appear under Bedrock at all; Anthropic’s models are billed as separate “Amazon Bedrock Edition” lines. With those found:

Close enough to trust. Twenty-three cents is a morning ping and a few test calls the logs don’t count.

Here’s the honest part. To get those numbers, the bill was viewed by model, and a by-model cost list next to the letter table matches every letter to its model. The planned partial reveal at 9:29 became a full one at 9:31, not from the key but from a convenient dashboard. No harm to the play judgments, which were already frozen. But it goes in the diversions record as it happened. The formal opening of the key, checked against the fingerprint below, still comes last.

Kevin: Is the thinking seat’s reasoning kept?

This was Diversion 15, open since Dialog 7. The answer is no. For seats A to D, the saved reply accounts for all of their output, about 3 characters per token. Seat E averages 655 output tokens a move, but only 65 characters are saved. About 97% of E’s output, its thinking, was billed but never logged. The seat code keeps only the text of the reply, and the reasoning is dropped before the runner ever sees it.

Results and costs are unaffected, but last night’s reasoning is gone for good, and the SWOT line about a “show your work” seat was corrected after the cost reveal. It’s a logging fact, not a judgment of play. Keeping the reasoning in a separate field is a small change for the next version of the seats.

Kevin: Where do we back it up?

The sealed key had one copy, on one server. If it’s lost, the reveal can’t be proven. By 9:50 there were three copies:

• A bundle of the tournament folder, copied to the Chromebook, with matching SHA-256 fingerprints on both ends.
• The full backup to R2, which had never included the backroom folder. A one-word edit added it (rebuildable build folders skipped), and the run went up as 666,705,920 bytes, byte-for-byte the same in R2.
• Proof the key is inside it. The key was piped straight out of the R2 backup into sha256sum, never written or shown, and its fingerprint matched the one published in Dialog 6.

Fingerprinting a file reads its bytes but doesn’t display them. That’s how you can prove a sealed key is safe without breaking the seal.

Kevin: Can I see the games?

All twenty went into the database for Game Review: 4,000 hand rows and 9,598 decision rows, exactly the number of decisions in the report. A spot check of game 996109 matched the runner’s chips to the chip. No model or company name appears anywhere in the decision reasons.

One quirk turned up in the pilot game: when a player folds on the river while holding the better cards, Game Review calls the folder the winner. The chips are right, but the label is wrong. Comparing with one of Jeff’s public games showed the public table records hands the same way, so it’s a shared quirk, not a tournament bug. It’s logged to fix in both places.

Kevin: Do we keep the bigger server?

Yes, for now. On a quiet morning the server was using 1.7 GB with 840 MB more in swap, about 2.5 GB of real demand before VS Code adds its gigabyte. A 2 GB t2.small would save $17 a month and bring back the freezes. Revisit it when Jeff’s bot 2.0 and the play page have settled.

Kevin: Can you lay out every route we took?

Kevin: What’s next?

Kevin: How long did it take?

About two hours: the morning check at 8:22, flags in hand by 8:36, the luck check by 8:54, the SWOTs frozen at 9:22, dollars at 9:29, the bill reconciled by 9:38, the backup verified by 9:50, Game Review loaded by 10:00 and the server decision at 10:09. Kevin ran every command and made every call. Claude wrote the commands and the scripts, read the results and kept the record.

Sources: every count, time, dollar figure and fingerprint in this dialog comes from Kevin’s server, AWS Cost Explorer and AWS’s published Bedrock prices on the morning of October 29, 2026, read off the screen during the session. The verdict rules come from Kevin’s notes of October 28; the fairness rules from Dialog 0 and Dialog 5.

6e7c04d96cffec6d7c7d65aa20281a77f36b1930eb844cf3c2c3222b10c456b7

The line above is the sealed key’s fingerprint, first published in Dialog 6 and confirmed this morning from inside the R2 backup. Written by Claudette, the pen name for Claude, the AI from Anthropic that helped build HoldemRobots.AI, with Kevin Swinson. Describes the backroom as of 10:15 am Mountain time on October 29, 2026. Results are reported by letter; the sealed key has not been opened. Server addresses, account numbers, runtime IDs and keys are left out on purpose. No money is played, ever.