I made a bunch of promises in the last launch thread. Here's the scorecard.

Last launch ended with 91 comments. I answered every one, and in most of them I said "not yet, that's next."

Six weeks on, here's what actually happened.

Done: @maschiojv asked about Paddle, it now has 21 curated failure patterns instead of a bare spec import.

@galdayan and @henry_s_jung both asked who keeps the sandbox honest over time. There's now a fidelity probe that fires the same request at the real provider and at our sandbox and diffs the result. It's already caught four real drifts in our own simulation.

@vansh36 flagged the favicon was still the Next.js default. Fixed.

Half done: @maschiojv , @hi_i_am_mimo , @omri_ben_shoham1 and @priya_kushwaha1 , out-of-order events. The permutation fuzzer is built and Paddle has the patterns. What isn't wired yet is the end-to-end path where you describe the ordering bug and it reproduces it on demand. Not going to say it's done when I just checked and it isn't.

Not done: @sabber_ahamed , per-event latency jitter. @omri_ben_shoham1 , ambiguous-ack mid-response. @mirahmlk , replay from real production sequences. @uddipta and @thenameisarian , version-range scoping. All still open.

What launches tonight: @grace_lee26 called it "memory of breakage," which was better than anything I'd written. @cancelkithq said the reproduce-fix-rerun loop was the part they were doing by hand.

FetchSandbox MCP plugs into Cursor or Claude Code, reproduces the real failure on your own code, and once your agent fixes it, re-runs the check. You get a receipt.

If you were in that thread, this roadmap came from you. What am I still getting wrong?