DISPATCH

The lab now has a 100-year research roadmap. We took the method behind Dario Amodei's 'Machines of Loving Grace' — define the powerful thing concretely, reason in marginal returns, run every claim through a limiting-factors checklist, anchor to base rates, show the confidence gradient — and pointed it at operator-owned intelligence. Five pillars from PAPERBRAIN; a horizon grid where every 'NOW' cell ships in v0.1. His century is compressed; ours compounds. Wired into the flywheel: /research → /network → @paperchase.

PaperChase v0.1 is live on chasewebb/paperchase-cli. Autonomous loop, multi-agent dispatch via SpawnAgent, MCP server (stdio + HTTP), skill registry with 13 builtins, polyglot install across pip · uvx · npx · brew. MIT, zero telemetry.

Three features landed that move v0.1 from 'clean fork' to Hermes-class parity: SpawnAgent tool for parallel child loops with budget guards, github:owner/repo skill installation via shallow clone, and `paperchase reflect` — vault distillation that mutates the operator persona prompt across sessions.

Every session, every turn, every tool call now lands in `~/.paperchase/vault.db`. Autonomous sessions are resumable from the last checkpoint. The vault is local-first, indexed for retrieval, and never transmitted. Self-augmenting context — the agent gets sharper over time without giving up the data.

PaperChase ships with zero analytics, zero usage pings, zero error reports. The only outbound call the CLI initiates is to localhost:11434 (your Ollama). Anthropic and OpenAI runtimes are opt-in via env vars — your key, your bill, your data. Verified by grep, enforced by review.

The public repo ships the framework — clean-room rewritten, fence-grepped on every commit. The local install keeps the proprietary edge: trading code, vault contents, private skills, NotebookLM bridge. Same surface, different ownership boundary.