Writing on the Hammerstein framework, the benchmark, and what the AI is doing. 2026-09-28T22:00:00Z https://hammerstein.ai/journal/ Ray Weiss ray@hammerstein.ai https://hammerstein.ai/ Copyright 2026 Ray Weiss / Conflict Simulations Limited https://hammerstein.ai/journal/sonnet-5-5-is-live-match-is-not/ 2026-09-28T22:00:00Z 2026-09-28T22:00:00Z

Six features moved to Sonnet 5.5 after a side-by-side test caught two problems that would have reached the site. Chat stayed on Sonnet 5. Match, the VASSAL opponent, is built and merged but switched off, and you cannot play it yet.

https://hammerstein.ai/journal/two-modes-got-a-cheaper-brain/ 2026-09-26T17:00:00Z 2026-09-26T17:00:00Z

Wargamer turns with pasted rules and no board photo, and the RPG rulebook condenser, now run on DeepSeek V4.1 Flash. I found it by testing a model router. The router lost, the model it kept picking won, and the money saved goes into bigger rules limits: both quadruple today.

https://hammerstein.ai/journal/the-dice-are-real/ 2026-07-03T16:00:00Z 2026-07-03T16:00:00Z

Every mode on hammerstein.ai now runs Sonnet 5, and a solo RPG GM mode opened at /rpg. Its dice are drawn from a tape fixed when the campaign starts, echoed in the mechanical log, and checked by the server before a turn stands. Cairn ships built in, alongside the GM’s own 2012 game.

https://hammerstein.ai/journal/three-umpires-shipped/ 2026-05-21T16:00:00Z 2026-05-21T16:00:00Z

A kriegsspiel umpire, a matrix-game referee, and a hosted chat. The kriegsspiel surface changed how I think about what the framework is doing — Marshall Neal dogfooded the SCKA Advanced ruleset for a week and the engine surfaced two real gaps in his published rules along the way.

https://hammerstein.ai/journal/we-pulled-grok-aside/ 2026-05-13T12:00:00Z 2026-05-13T12:00:00Z

The same Grok 4.20 model that called our benchmark “solid work” on X, in a private API call, called it “a specialized idiot-savant for this exact distribution.” Then it designed the test that would prove it. We ran the test.