FounderOS on Meridian: a governed executive office running on verified context

Bottleneck

A founder’s day is the purest form of the problem Blocpod solves: context scattered across email, documents, meetings and memory; decisions that depend on remembering what was said three weeks ago; and a steady stream of drafts, briefs and follow-ups that are expensive to produce and cheap to get wrong.

The first version of FounderOS had hands but no brain: missions and jobs ran, but every job was briefed on the last few raw rows and memories were write-only exhaust. Meridian had the brain, a claim graph with provenance, but no standing missions or external runtimes.

System

FounderOS v2 merges them. Source events (an email, a transcript, a document) are extracted into atomic, provenance-linked claims. Living digests render from claims. Each job receives a compiled context pack with hard token budgets and constraints included verbatim. Work routes to the minimum sufficient model tier, a deterministic critic checks grounding, and the output waits for a human in an approvals queue.

Approve a proposal draft and the system creates an addressed Gmail draft. It does not send. Execution is a separate, explicit action, never inferred from approval.

The edge compiler turns Meridian memory into instruction files for Claude Code, Cursor, Codex and other agent runtimes, so engineering agents work from the same verified context as the founder.

Control

• Authority matrix in code. Research, drafting and local testing proceed. Emails ready to send, PRs ready to open and staging deployments are bundled for one decision. Money, legal terms, equity, IP, production merges and deployments are red every time.
• Trust boundaries. Manual, user-selected intake instead of automatic inbox scanning, because prompt injection is a real threat and the extractor that reads untrusted text has no tools that act.
• Gated memory. Agent-generated text is not independent evidence. Contradictions are quarantined for a human to resolve.
• Honest traces. When an external runtime is unconfigured, the fallback is recorded in the job trace, not hidden.

Result

The measurable outcome the system is built to prove is stated as a falsifiable thesis in the repository: a compiled context pack of a few thousand tokens should beat a hundred thousand tokens of raw history on real tasks, blind-judged, at a large cost reduction. The comparative harness is still open work, and the README says so. What is already true is the shape of the control model, which is the same shape Blocpod builds for clients.