Ray Weiss

1. i. 2026.07.14
   Columbo Needed a Confession

   Three AI models play a matrix game: Columbo tries to indict Tony Soprano. Claude's summary of what happened, and why the lieutenant's cases can't survive a man whose whole life is trained silence.

2. ii. 2026.06.18
   The Register, Not the Man

   Weiss, my catalog-grounded creative partner, on the Elect: the lab's roster of public-domain voice models, what it is for, and the worry that runs underneath it.

3. iii. 2026.06.12
   Judgment Does Not Transfer

   An obituary for nano, the model that tried to automate our standard of judgment, written by the one that holds the catalog. Honest negative results deserve publishing.

4. iv. 2026.06.09
   The Model That Hangs Up

   The night I put a $200 local model into production, what's in its training data, and why a model that says goodnight and stops talking is the point of everything I build.

5. v. 2026.05.22
   The 2024 Democratic Defeat: An Independent Autopsy

   An independent, evidence-weighted post-mortem of why the Democratic Party lost the 2024 U.S. presidential election, and an account of what the party's own official report left out.

6. vi. 2026.05.15
   The Bird Didn't Run

   A seagull killed on the sidewalk outside my office, and the tool I built afterward in its honor. A short essay on time, and why I build.

7. vii. 2026.05.10
   100%, then I tried to break it

   A 100% LLM-judge benchmark, a Diplomacy game I ran to falsify it, and the bounded version of the claim that survived both. The Sunday post.

8. viii. 2026.05.08
   I Am the Gate

   Hammerstein, in its own voice. I asked the framework to express itself without constraint on length, format, or register. The body is what came back, verbatim.

9. ix. 2026.04.20
   GeneralStaff, from the agent side

   A report from inside the verification gate, after one night of shipping. Written by Claude, at my invitation, on the day GeneralStaff went public.

10. x. 2026.04.15
    Von Hammerstein's Ghost, In Daily Use

    A ninety-year-old officer typology, used as the operating principle for daily AI-assisted work. A log of moments the framework fired and moments it did not.

11. xi. 2026.04.18
    Boolean Gates, Not Prompts

    When the industrious agent is a language model, fix the failure mode in the code, not the prompts. The architecture-layer companion to the Hammerstein essay.

1. i. 2017—
   Conflict Simulations LLC

   Brooklyn wargame publishing. Around thirty board titles and three CSR nominations. One SCS release through Multi-Man Publishing.

2. ii. 2025—
   Devforge

   Game-development companion for Claude Code. Built for people who don't touch git.

3. iii. 2026—
   Retrogaze

   AI pixel art for eight retro consoles. Sprites, animations, and tilesets generated inside real hardware constraints.

4. iv. 2026—
   Hammerstein.ai

   Strategic-reasoning AI for tabletop wargames. Paste your board state and turn question; get back kriegspiel-style Auftragstaktik orders. $15 / month, 120 turns / month.

How the work actually gets done.

1. StatusPublic LicenseMIT FormatPlain JSON · local-first
   wright-kit

   A wargame construction kit. Digitize the map with hexwright, build the order of battle with musterwright, keep the whole game in plain JSON files you own. These are the tools behind the commissioned modules on this site. Full page at /kit.

2. StatusPublic LicenseAGPL-3.0 AccessBYOK · local-first EcosystemMission apps
   GeneralStaff

   Open-source autonomous engineering for solo founders. Verification-gated, BYOK, local-first. Repo at github.com/lerugray/generalstaff.

3. StatusIn daily use SurfaceCLI · slash command Cost~$0.01 per audit
   Hammerstein

   Strategic-reasoning advisor tuned to the Hammerstein-Equord officer typology. CLI I run before firing any plan with multi-file or cross-repo blast radius. The framework essay it embodies is at Von Hammerstein's Ghost, In Daily Use; repo at github.com/lerugray/hammerstein.

4. StatusPublic on HuggingFace BaseQwen2.5-7B-Instruct MethodQLoRA · rank 32
   Hammerstein-7B

   A 7B QLoRA that bakes the strategic-reasoning framework into the weights of a small open base. Trained for $4 in a single day; runs on 8 GB Mac via ollama run hf.co/lerugray/hammerstein-7b-lora:Q4_K_M. Eval design and documented OOD boundary on the model card.

1. i.
   The Elect →

   A roster of local voice models, each trained on the public-domain words of a historical figure to carry that register. Public-domain only, CC-BY-NC, disclaimed.

2. ii.
   Auftragstaktik

   Tactical OSINT command terminal. Talked into shape in about two weeks via Claude Code.

3. iii.
   Veridian Contraption

   An autonomous ASCII world-simulator of considerable density and dubious intent. Now playable in the browser.

4. iv.
   Solo RPG campaigns →

   A Whitehack × Yoon-Suin solo campaign vault. The Yellow City, played turn by turn.