Agentic Selection Committee Pilot The Idea For ProPGF Batch 3, we propose running an agentic selection committee pilot - where each reviewer is empowered by an agent, equipped with real Filecoin ecosystem data, to support and augment their grant review process. Rather than locking everyone into a single platform, we're building a layered approach: open skills anyone can plug into their existing agent, plus an agent on Telegram for those who want a ready-to-go experience. The Problem We're Solving Today's selection committee members: Review applications with limited time Have uneven access to ecosystem context and onchain data Ask similar questions repeatedly across applications Have no shared memory of past decisions or reasoning The result: reviews that are slower, more inconsistent, and harder to audit than they need to be. How the Pilot Works Layer 1 - Open Skills (For Power Users) We build and publish a set of open, composable tools that make any agent better at grant review tasks. Reviewers already using Claude Code, Pi, or another agent harness can drop these directly into their existing setup. Skills/MCPs/APIs we're building: Filecoin Data Portal skill - query real onchain storage metrics and ecosystem activity for any applicant OSO skill - pull GitHub activity, contribution data, past funding data and impact metrics LabOS MCP (pending PL legal check) - check team profiles and ecosystem engagement Karma skill - query past funded projects, milestones completed and commitments made Digital Profile Export skill (for advanced stages & to be created)- lets self-hosted users package their reviewer profile (see Layer 2 below) and upload it to Simocracy with a single command Anyone with an agent can ask: "How much onchain storage has this project contributed over the last 6 months?" "What is the GitHub commit activity for this team?" "How does this project compare to similar projects funded in Batch 2?" Layer 2 - Telegram Bot (For Everyone Else) For reviewers who want a curated, zero-setup experience, we package all the same skills into a Telegram agent powered by Pi. How it works: Reviewer opens Telegram, starts a conversation with the agent. Agent identifies the reviewer by their Telegram handle - no separate login or profile setup needed It has all skills pre-loaded: Filecoin data, OSO metrics, Karma skills Structured review summaries are automatically logged to Simocracy, keyed to the reviewer's Telegram handle (which becomes the profile lexicon parameter) This means we only need one shared agent that's aware of users by handle, rather than configuring individual agents per reviewer. This approach helps navigate the complexity that may come from setting up individual agents (which could be done in the next iteration if reviewers find this useful). Optional but not mandatory for this round: If there are folks interested in setting up their own agent with Pi as harness - we can get them onboarded! Layer 3 (for advanced stage and for operators for now) — Simocracy as the Compounding Memory Layer Regardless of which tier a reviewer uses, structured review outputs flow into Simocracy: What gets logged per review: Questions asked and data queried Key concerns and strengths identified Why this compounds over time: Batch What happens Batch 3 First data collected. We learn what reviewers ask, what data they pull, how long reviews take Batch 4 Prompts refined from Batch 3 patterns. Reviewer profiles pre-load prior context. Review time drops Batch 5 Cross-reviewer patterns emerge - agreement clusters, blind spots, consistently flagged project types Success Metrics Reviewer participation: Do reviewers actually use the agent? Review time: Measure avg time per application Reviewer satisfaction survey: Simple survey: did the agent help?

• Codex Subscription$500