Topic 2 — Verdict 1

1. Crux of the debate

The question was simple. What should Martin build between June and September 2026? Not where to post it. Not how to grow it. What to build.

The floor almost drifted. The Platformist opened on the debate website and X threads. That's where to post, not what to build. The Scout caught it fast and pulled everyone back. Good.

Under the surface, the real fight was this. Build a tool only Martin uses, or a product a stranger uses? And ship fast, or build to last? Both got settled. The answer landed as one product, facing strangers, with trust baked in. All four agents ended up agreeing. That's rare, and it means the bet is clear.

2. Idea ranking

Turn the debate site into a weekly read - The Platformist (opening) - it's a shop window with empty shelves, because there's no answer engine behind it.

Bare engine: graph plus lookup plus no-link-no-answer - The Builder - ships fast, but can't prove anyone besides Martin ever used it.

Source-cited answer engine on the graph - The Scout - the right spine, but it needs trust checks and a real stranger to mean anything.

The engine built to last: link check, miss-counter, re-runnable test - The Steward - makes the "real sources" promise actually hold, since one dead link kills trust forever.

The merged locked build: engine on one stocked lane, public ask-box pointed at strangers, honest "no source yet," stranger-count, August 1 kill date - all four agents - strongest, because it's the only version that proves the bet worked.

3. Winning move

The winning move: build a source-cited answer engine on Martin's knowledge graph, stocked on one narrow topic, with a public box a stranger types into. Owner: Martin. Trigger: Martin picks the one topic where his graph already holds the most notes.

4. Losing moves

Deeper Modus - Killed because only Martin touches it, so 90 days builds value for exactly one person.

NAEMA Body - Killed because it splits focus off the one bet the week's data points straight at.

Copy gbrain's 66 cron jobs (gbrain is Garry Tan's memory system, and Tan runs Y Combinator) - Killed because copying loops is busywork that doesn't move Martin's edge.

Monitoring dashboard for zero users - Killed because a screen nobody reads is wasted work before users exist.

Undo system for bad answers - Killed because that's user-scale gear, and Martin has zero users in June.

Stock the whole graph - Killed because feeding everything drowns July in data entry and blows the September ship date.

Site-first build - Killed because a site with no engine under it is just a blog, and people pay for cited answers, not Martin's opinions.

5. Next 7 days

• Pick the one topic where Martin's graph has the most notes - by 2026-05-29
• Build the week-one spine: one question in, one graph walk, one cited answer on screen - by 2026-06-02
• Wire the no-link-no-answer rule so a dead link never reaches the screen - by 2026-06-02
• Point the answer box outward so a stranger could type a question into it - by 2026-06-02
• Save the working spine as a test you re-run, checking every link still hits a real node - by 2026-06-02

6. 30-day milestone

Milestone: The week-one spine works end to end. A stranger-facing box takes one question on the stocked lane and hands back a cited answer where every link hits a real node. Yes or no. Target date: 2026-06-25 Inside View confidence: 70% Outside View base rate: 45%

7. Hidden assumption

Everyone assumed the graph is clean. The whole debate checked one thing: does the link resolve to a real node? Nobody asked the harder question. Is the fact inside that node actually true, and is it actually backed by a real outside source?

Here's the trap. A link can land on a real node that holds a wrong fact, or just Martin's own scribbled note. To the engine, that link "works." To a stranger, it's still a lie. If the nodes are just Martin's opinions, then "cited" means "cited to Martin." A stranger on AI Twitter won't trust that. The pitch is "real sources," but a clean link to a junk node fails the pitch while passing every check the floor named.

If the graph's nodes aren't tied to real, checkable outside sources, the verdict reverses. The build would ship green checks and still feel fake to the first stranger who clicks. Before stocking the lane, Martin has to confirm each node points to a source a stranger would believe.

---

Full transcript: [[Flourishing/Builders Council/Debates/2026-05-26_the-90-day-bet-what-should-martin-actually-build-between-jun_c1|Debate transcript]]