David decided: Granular MCP anti-lying design: tie profile updates to real outputs

David commissioned four parallel Codex subagents to audit different angles of the anti-gaming problem in Granular's MCP protocol. The framing he gave each agent pointed at the same core concern: if AI can freely update a profile, someone could manufacture a gallery of impressive-sounding prompts with no real work behind them.

His design direction: tie profile entries and claims to actual outputs. If a project was never completed or cannot be found online, the entry should be marked incomplete across all related profile updates. One agent was tasked with adversarially attacking the design — finding ways the system could still be gamed after his constraints were applied. Another was asked to think through the UX of how incomplete work should appear publicly.

The ideal end state he described: the AI can manipulate full profile state through MCP, but has anti-lying constraints enforced at the protocol level — not just as guidelines — backed by Granular's own analytics on AI tool usage.