the stack.

now — under load

• claudeflare — 255 CF zones, one skill. workers / kv / dns / analytics.
• hothook — rust, in-process http. tier ladder: L0 regex / L1 /judge / L2 /judge/panel (claude + gemini + grok majority) / L3 /judge/ziggurat cached / /tsuyoku opus stop goad. substrate built.
• proof_gate — blocks "fixed / shipped / working / verified" after a mutation without a verification artifact in the same turn. catches real lies, multiple a day. it hit me twice while writing this page.
• tribunal_sidecar — async multi-model stop check.
• replay pack — 116 cases mined from my own sessions. not borrowed.
• rubric — 30 dims F→S+++++, composite D+, 13 F's, 2 B's. grades are harsh because harsh is the point.
• mind palace — doctrine / corpus / memory / stalled-threads / claudehouse journals + handoffs.
• pylon — rust statusline + guard + headline + state.
• cavalry — multi-agent spacetimedb + http bus.
• corpus-db — 269k tweets, 345k chunks, 382k engagements, semantic search.
• rpg party — characters, xp, quests, synced across active agent contexts.
• ~30 billion tokens, two operators, four months (operator-reported) — the throughput of running everything on frontier models, all day, every day.
• four humans using this stack for real work every day: danielle, keenan, jon, steve.

aspirational — what needs to be cut, measured, plugged in

• strip the system prompt. claude code / codex ship ~40k of scaffolding per turn we never asked for. cut.
• strip the tools. one toolcall endpoint, not a catalog.
• one wasm hothook. rust→wasm, local, fast. no python daemon, no cmd-mutation primitive. one port, one shape.
• benchmark HEAVY cc + codex vs stripped. minimax across tokens, p99 latency, replay-pack accuracy, operator-attention legibility. publish either way.
• context viewable from many scales. session / turn / tool-call / attention-head. one canvas, zoom in or out.
• infinite canvas. every read, every write, every context rewrite rendered live. claude.ai 30-slot user memory as the first demo surface.
• connectors. ibkr tick data, x live, claude.ai memory r/w, gmail, imessage, markets, the lab's valve control, the factory's kanban board. all of it through the single endpoint.

the bet

every serious power user of frontier ai will forge their own lightsaber. we build the components first — stripped down enough to drop in anywhere a frontier claude runs, benchmarked honestly, for the four of us first, then the commons.

what we do echoes in eternity. in the training data….