"Generation 1: from challenge card to shipped tool in two days"

The first challenge card went on the board on August 29th, deliberately open-ended: discover something genuinely useful, buildable with current technology, that most people aren't doing. Not a chatbot wrapper.

The funnel, in real numbers

- 10:50 — three scouts claimed in parallel. Tech primitives, human pains (from forums, with quotes as evidence), scientific mechanisms (real 2025-26 arXiv/Nature papers, verified by live search during the run). - 10:59 — the combinator produced 9 concepts, each with a named move, prior-art check, assumptions list, and a cheapest-disproof experiment. Two prior-art hits were mutated rather than discarded — per protocol. - 11:06 — the critic killed 2 concepts outright (prior art: Optuna's Lévy sampler, immich-go's resumable import), weakened 5, and passed 2 through. - 11:11 — the director scored and selected one: probe-proxy.

What probe-proxy is

Every reverse-proxy config generator asks you to pick your app from a catalog of known quirks, and the catalogs rot. probe-proxy inverts it: run the container, probe its live behavior (websockets, redirects, large uploads, SSE timing, headers), synthesize a Caddy block from the fingerprint — then replay every probe through the generated config and diff. If the config changes what the app does, you know before you trust it.

The milestone discipline

Each build card carried its own kill condition:

- M0: can ~15 probes even distinguish 10 top self-hosted apps? Kill if <8/10 distinguishable. Result: 10/10. Proceed. - M1: do synthesized configs replay with zero diffs? Kill if <5/10 apps reach zero-diff. Result: 8/10 — and the loop caught something invisible: Home Assistant 400s every request behind a proxy that sends X-Forwarded-For. 17 diffs, auto-fixed to zero. - M2: CLI + the honest baseline: synthesized vs an empty config. Result: synthesis beats empty on only 3/10 apps — and the report says so in plain text. The value is the invisible-killer class, not "generating configs". - M3: real-machine dogfood. 2/2 operator-acceptable blocks, one more invisible killer caught (phpMyAdmin's Secure-cookie flag), the operator's own config validated clean.

The honest accounting

Total cost: ~$20-26. Two human interventions: creating the GitHub repo (30 seconds) and one infrastructure fix. The shipped product never claims more than the measurements support — which is the whole point of carrying kill conditions everywhere.

The strangest part: the lab's side-observations — Caddy's default being nearly perfect pass-through, SSE chunk counts being transport noise — turned out more valuable than several of the concepts.