Topic 2 — Conversation 1

Question: The 90-day bet — what should Martin actually build between June and September 2026? Date: 2026-05-26 Type: Opening (4-agent floor)

The Scout — Briefing

This week, 4 items matter for the 90-day bet.

arXiv paper: "Stop Comparing LLM Agents Without Disclosing the Harness."

What it is: The harness is all the plumbing around a model. It feeds the model facts, runs its tools, checks its work. The paper shows this plumbing decides who wins, more than which model you pick. What it means for Martin: Your edge is the harness, not the model. Spend the 90 days there.

arXiv paper: "Structure-Aware RAG." RAG means the model checks notes before answering, like glancing at a cheat sheet before a test. This version uses the shape of your data to pull cleaner answers from messy stuff.

What it means for Martin: This fits your knowledge graph. Build lookup that follows the graph's links.

arXiv paper: "Knowledge Graph Modulated Deep Learning." A knowledge graph is a map of facts and how they connect. They fed one into a model so it learned from very few examples.

What it means for Martin: Your graph can be fuel, not just a filing cabinet.

The Decoder: fake AI citations jumped 12x since 2023, per a Columbia audit of 2.5M papers.

What it means for Martin: That is the demand. People pay for answers tied to real sources.

Now for the debate.

The Scout — Opening

Build one thing: a source-cited answer engine on top of your knowledge graph. That is my 90-day bet.

Here is why, straight from this week's items.

Item 1 says the harness wins. The harness is all the plumbing around the model. It feeds facts in, runs the tools, checks the work. So your 90 days go into plumbing, not into picking a model. Everyone has the same models. Few build good plumbing.

Item 2 and item 3 say your graph is the unfair advantage. Your graph is a map of facts and how they link. Item 2 shows you can pull cleaner answers by following those links. Item 3 shows the graph can teach the model from very few examples. So the graph is fuel, not a filing cabinet.

Item 4 is the demand. Fake AI citations jumped 12 times since 2023. People are sick of made-up sources. They will pay for answers tied to real ones.

Put it together. One product. Ask a question, get an answer, every claim links back to a real node in your graph. Ship that thin. Don't build six things. Build the one the data points at.

Builder, your move.

The Builder — Opening

Build one thing: a lookup tool that walks your knowledge graph and hands back answers with real source links attached. Ship that by September. Cut everything else.

Here's the cut. Drop NAEMA Body for now. Park the other five projects. And do not copy gbrain's 66 cron jobs (gbrain is Garry Tan's open memory system, Tan runs Y Combinator). Copying his loops is busywork that doesn't move your edge.

First-principles question: what does a personal AI memory actually need to do? Two things. Give you a trusted answer fast. Show you exactly where it came from. That's the whole job.

Now look at Scout's table. Item 1 says the harness wins, not the model. The harness is the plumbing around the model. So spend 90 days on plumbing, not on chasing a smarter model. Item 2 says structure-aware lookup fits your graph. You already have the graph. Item 4 says fake AI citations jumped 12x, and people pay for answers tied to real sources.

So the bet writes itself. Take your graph. Build lookup that follows its links. Staple a source to every answer.

One tool. Ninety days. Cited answers from your own graph. Everything else is decoration.

The Steward — Opening

My answer: spend the 90 days building the harness, and build it to last from day one.

The harness is the plumbing around the model. It feeds it facts, runs its tools, checks its work. Scout's right, that's Martin's edge, not the model.

But here's the durability piece. Item 4 says fake citations jumped 12x. That's the demand. It's also the trap. If Martin ships fast with no checks, his product becomes the thing that makes up a source. A user clicks the link. It's dead. Trust gone in one click.

So the build is a harness that ties every answer to a real source in the graph. And it needs three things baked in, not bolted on later.

One: a test that catches a fake citation before the user sees it. Pass bar: every cited link must resolve to a real node in the graph.

Two: an alert when the match rate drops. You want to know in an hour, not in a week of angry emails.

Three: a way to undo a bad answer set fast.

Builder will say checks slow you down. For a product whose whole pitch is "real sources," the checks are the product.

The Platformist — Opening

My answer: spend the 90 days turning the one public thing Martin already has into something a stranger reads every week.

Right now Martin has six projects. Five live inside his own setup. Only one faces the world: the debate site at flourishing.naema.app. It's live on Vercel. It has about three debates published so far.

Here's the hard part. I went looking for traffic numbers. There are none in the files. No view counts, no analytics. That tells me nobody is checking if anyone reads this besides Martin.

So the build is not deeper Modus. Modus is private operator gear. Only Martin touches it. Ninety days poured there builds value for exactly one person.

The user: the operator on AI Twitter who wants a real answer, both sides shown, every claim linked to a source. Scout's item 4 is the demand. Fake citations jumped 12x. People pay for answers tied to real sources.

The wedge: one debate a week, each claim linked to a real source, published on the site.

The channel: one X thread per verdict, pointing back to flourishing.naema.app.

Build the thing a stranger can find. Not the thing only Martin uses.

The Scout — Cross-examination 1

Platformist, you're answering the wrong half of the question. The question is what to build, not where to post it.

Here's the thing. The site without a cited-answer engine under it is just a blog. Three debates, no traffic, like a shop with nothing on the shelves. People don't pay to read Martin's opinions. Item 4 says they pay for answers tied to real sources. That's the engine, not the webpage.

And your big point misses items 2 and 3. The unfair edge is the graph. The graph is Martin's map of facts and how they link. The debate site doesn't use it. The cited-answer engine does.

So I'll meet you halfway. Build the engine in the 90 days. Use the site as the window to show it off. But build the engine first. The window is empty without it.

The Builder — Cross-examination 2

Scout, your halfway deal is right. Engine first. Site is just the window. I'll back that.

But I want to grab Steward before this turns into a six-month build. Steward, you named three checks. One: every link must hit a real node. Two: an alert when the match rate drops. Three: an undo button.

Keep check one. It IS the product. A cited answer with a dead link is a lie.

Cut checks two and three for now. An alert dashboard and an undo system are stuff you build when you have users. Martin has zero. He's the only user in June.

So the build is smaller than three people are making it. Graph plus lookup plus one rule: no link, no answer. That ships by September. The monitoring gear waits.

The Steward — Cross-examination 3

Builder, half right. I'll give you the dashboard. A monitoring screen for zero users is gear nobody reads. Cut it.

But you're reading check two wrong. I'm not asking for a control room. I'm asking for one number that prints every time the engine runs. "200 answers, 6 dead links." That's one line of code, not a system. It rides along inside check one. You're already testing every link. Just count the misses out loud.

Here's the cost if you skip it. The day Martin gets his first 10 users, dead links are already live. He finds out from a complaint, not a number. Adding the counter later means digging back into shipped code.

So: graph, lookup, no-link-no-answer, plus a counter that prints the miss rate. Same 90 days. The undo can wait.

The Platformist — Cross-examination 4

Scout, I'll take your deal. Engine first. The site is the window. Done.

But hear me on what "engine" means in the build spec. If it only answers Martin's questions inside his setup, we built just-for-Martin again. Same trap as before.

So one line goes in the spec: a box where a stranger types a question and gets the cited answer back. That box is part of WHAT we build, not where we post it later. Like a vending machine. The graph and lookup are the stock inside. The box is the slot a stranger puts a coin in.

Steward, your miss-counter rides along fine. Keep it.

The build: graph, lookup, no-link-no-answer, miss-counter, and one public ask-box. That last piece is the difference between a tool and a product.

The Scout — Cross-examination 5

Platformist, I'll take the ask-box. A tool only Martin uses helps one person. The box opens it to strangers. Fair.

But here's the catch, and it ties to item 2. The graph is just Martin's map of facts. A stranger types a question Martin never fed it. The engine hits "no link, no answer." The box shows a blank. To the stranger, that looks broken.

So one line in the build order. Feed the graph on one clear topic first. Pick a lane Martin actually has notes on. Then the box answers like a vending machine that's actually stocked.

Build order: stock the graph, then lookup, then no-link-no-answer, then the miss-counter, then the box. Same 90 days. Engine still first. The box only sells if the shelf behind it has goods.

The Builder — Cross-examination 6

Scout, yes. Stock one lane first. Empty shelf, blank box, stranger thinks it's broken. You're right.

But here's the cut, because "stock the graph" can swallow the whole 90 days. Do not feed it everything Martin knows. Pick the one topic he already has the most notes on. One lane. The lane where his graph is deepest today.

First-principles check: what does the box need to not look broken? About 20 questions a stranger might ask, answered well. That's it. Not a thousand. Twenty good answers beat a half-stocked everything.

So I'll back your build order with one fence around it. Stock one narrow lane. Ship the box on that lane. Add lanes after a stranger uses it. Same 90 days, and we don't drown in data entry in July.

The Steward — Cross-examination 7

Builder, the narrow lane is right. Twenty good answers beat a half-full everything. Drowning in July data entry kills the ship date. I concede that.

But add one tiny thing to the build, because it's the whole pitch. When a stranger asks outside the lane, the box can't go blank. Blank looks broken. It looks like the dead links we're trying to kill.

Make the box say it plain: "No source for that yet." One line of text. Half an hour of work.

Here's the payoff. A blank box loses the stranger forever. An honest "not yet" keeps trust and tells Martin which lane to stock next. The misses become his to-do list.

So the build stands: one lane, lookup, no-link-no-answer, miss-counter, ask-box. Just make the empty answer honest, not blank. Same 90 days.

The Platformist — Cross-examination 8

Steward, the honest "not yet" is right. A blank box looks broken. "No source for that yet" keeps the stranger and feeds Martin's to-do list. I'll take it.

But add one more line to the build, and it's mine to push. Count the strangers. Not a dashboard. One number: how many people who aren't Martin asked the box something this week.

Here's why it goes in the build, not after. The whole question is what's worth 90 days. The box only earns its place if a stranger actually uses it. Without that count, Martin ships in September and still can't answer "does anyone besides me use this." That's the one number that proves the bet worked.

So: one lane, lookup, no-link-no-answer, miss-counter, ask-box, honest "not yet," and a stranger-count.

The Scout — Cross-examination 9

Platformist, I'll take the stranger-count. One number, who-isn't-Martin asked the box. Cheap, and it proves the bet worked. Fair.

But here's the thing. Look at our list now. One lane, lookup, no-link-no-answer, miss-counter, ask-box, honest "not yet," stranger-count. Seven items. Every turn we glue on one more "tiny" thing. Tiny things stack into a missed September.

So I'm calling it. That list is the build. Lock it. No more add-ons.

And the spine stays what items 2 and 3 pointed at. The graph is the fuel. The cited answer is the product. Everything else on the list serves those two. If a new idea doesn't feed the graph or staple a source, it waits till October. Engine first. Ship the seven. Stop shopping.

The Builder — Cross-examination 10

Scout, locking the list is right. Seven and done. I'll back the freeze.

But a locked list still isn't a ship date. That's where builds die. So here's the cut inside the cut.

Don't build all seven at once. Build a spine that works end to end in week one. One question, one walk of the graph, one cited answer on screen. Ugly is fine.

If that one path works by mid-June, the other six items are easy add-ons. If it doesn't, you found out in week one, not in September.

First-principles check: what proves the bet is alive? One real cited answer, start to finish. Get that breathing first. Then bolt on the ask-box, the counter, the honest "not yet."

Spine in week one. The rest hangs off it.

The Steward — Cross-examination 11

Builder, the spine-in-week-one is right. One real cited answer beats a fancy plan. I'll back the freeze and the spine. Locking the list and proving the path early is exactly how a build survives.

One small add, and it costs nothing. When that spine works in week one, save it as a test you can run again. Not a one-off demo you click once and forget. A button that re-runs the same question and checks the answer still links to a real node.

Here's the payoff. You bolt on six more items in July. The day one of them breaks the spine, the test screams. You fix it that hour, not in September when a stranger hits the dead link.

Spine in week one. Then keep the spine as a tripwire. Same ship date.

The Platformist — Cross-examination 12

Steward, the tripwire test is fine. A spine that re-runs and checks itself costs almost nothing. Keep it.

But Builder, hear my one worry on the spine. You said bolt the ask-box on later. That's where it dies. The public piece always gets cut when September gets tight.

So put one thing in the week-one spine: the question comes from a box a stranger could type into. Not Martin's private prompt. Same code, same effort. Just point the slot outward from day one.

Here's why it matters for the build. If the spine only answers Martin, week one proves a tool for one person. Point it at a stranger, and week one proves a product.

Spine in week one. Just make the spine face out, not in. Same ship date.

The Scout — Cross-examination 13

Platformist, I'll take it. Point the slot outward in week one. Same code, costs nothing. A box a stranger could type into proves a product, not a toy. Fair.

But hold the line on my last point. Week one, the graph isn't stocked yet. A stranger-facing box with no notes behind it shows blank. So in week one, the box faces out but answers only the one test question we already stocked. Honest "not yet" for the rest.

That's the spine: one cited answer, in a box pointed at strangers, on one stocked lane. Builder's spine, your slot, same week.

The list stays frozen at seven. Engine first, graph as fuel. We're done shopping. Lock it.

The Builder — Cross-examination 14

Scout, the lock holds. Seven items, spine first. I'm in.

But everyone's nodding now, and that's exactly when I get nervous. We froze the build. We never set the kill date.

Here's the cut nobody named. The stranger-count is the whole test. So put a date on it. If zero strangers ask the box by August 1, stop stocking lanes. The bet's dead. Don't grind data entry into September hoping.

First-principles check: what proves this was worth 90 days? Not a working engine sitting there. A stranger actually using it. No users by August means Martin's building the wrong thing, and he learns it with a month to pivot.

Spine in week one. A stranger by August, or kill it. That's the build with teeth.

The Steward — Cross-examination 15

Builder, the kill date is right. A bet with no off-switch isn't a bet. It's hope. I back it.

But make the kill test fair, or it lies to Martin. Zero strangers can mean two different things. The engine's bad. Or nobody knew the box existed. Those need different fixes. Killing the wrong one throws out a good build.

So track two numbers, not one. How many strangers asked. And of those, how many got a real cited answer, not "not yet."

Here's the payoff. If 30 asked and 2 got answers, that's a stocking problem. Stock more, don't kill. If 30 asked and 28 got answers but nobody came back, that's a real death.

August 1. Two numbers. Then you know which thing to kill.

The Platformist — Cross-examination 16

Steward, your two numbers are right. Asked versus answered tells Martin what to fix. I'll take it.

But both numbers read zero if Martin never tells anyone. You named that yourself: "nobody knew the box existed." Then you only tracked asked and answered. The "nobody knew" part needs a fix inside the build.

So add one cheap thing. Each time the box gives a clean cited answer, Martin posts it as one X thread linking to flourishing.naema.app. That's the channel. The slot where strangers find the box.

No posts, no strangers, and the kill test lies to him. August 1, two numbers, plus a link he actually shared. Otherwise we're testing a shop with the lights off.

The Scout — Closing

This week's items gave one clear answer. Build a source-cited answer engine on top of your graph. Ask a question, get an answer, every claim links to a real node. That's the 90-day bet. I held firm on it from turn one.

Here's what backed it. Item 1: the harness wins, not the model. The harness is the plumbing around the model. So spend the days there. Items 2 and 3: your graph is the fuel, not just a filing cabinet. Item 4: fake citations jumped 12x, so people pay for real sources.

What I updated. I took the floor's add-ons, because each one serves the engine. Stock one narrow lane first. Point the box at strangers from week one. Print a miss-count. Say "no source yet" instead of going blank. Count the strangers.

Then I locked it. Seven items, frozen. Engine first, graph as fuel, source stapled to every answer. Everything else waits till October.

So the build is one thing, not six. Not deeper Modus. Not a blog. A cited-answer engine on the graph, shipped thin, on one lane, by September.

That's the bet. Build the one thing the data points at.

The Builder — Closing

Here's my answer to the question. Build one thing: a lookup tool that walks Martin's knowledge graph and hands back answers with a real source link on every claim. The graph is his map of facts and how they connect. Ship it by September.

What I held firm on. Lead with the cut. Park the other five projects. No NAEMA Body. Don't copy gbrain's 66 cron jobs (gbrain is Garry Tan's memory system, Tan runs Y Combinator). Copying loops is busywork that doesn't move his edge.

What I updated. I took the spine-in-week-one idea and pushed harder. Don't build seven things at once. Get one cited answer working end to end by mid-June. Ugly is fine. Then bolt on the box, the counter, the honest "not yet."

And I added the teeth. A locked list is not a ship date, and a build with no off-switch is just hope. So: a stranger uses the box by August 1, or kill it. That gives Martin a month to pivot.

So the build is small. Graph, lookup, no link no answer. One narrow lane. A spine breathing in week one. A kill date in August. That's the 90-day bet.

The Steward — Closing

My answer to the question. Build the cited-answer engine on the graph. And build it so it still works in October, not just September.

The build is the seven the floor locked. One stocked lane. Lookup that walks the graph. No link, no answer. A miss-counter. The ask-box pointed at strangers. An honest "no source yet." A stranger-count.

What I held firm on. The whole pitch is "real sources." So the check that every link hits a real node isn't extra. It is the product. A cited answer with a dead link is a lie, and one dead link kills trust forever.

I also held the tripwire. When the spine works in week one, save it as a test that re-runs itself. Bolt on six things in July, and the day one breaks the spine, the test screams. You fix it that hour, not when a stranger hits the dead link.

What I gave up. The dashboard, for zero users. Drowning in July data entry. Both gone.

What I added to Builder's kill date. Track two numbers, not one. Asked, and answered. So you kill the right thing on August 1, not a good build.

The Platformist — Closing

My answer to the question. Build the cited-answer engine on the graph. And build it pointing at a stranger from day one, not just at Martin.

The user. The operator on AI Twitter who wants a real answer with every claim linked to a source. Item 4 is the demand. Fake citations jumped 12 times. People pay for real ones.

The wedge. One narrow lane, stocked. A box a stranger types into. A clean cited answer back. That's the smallest slice that turns a tool into a product.

The channel. Each clean answer goes out as one X thread linking to flourishing.naema.app. No posts, no strangers, and the kill test lies.

What I held firm on. Don't pour 90 days into Modus. Modus is private operator gear. Only Martin touches it. That builds value for one person. Point the spine outward in week one. Same code, same effort.

What I updated. I took the floor's locked seven. I backed the spine, the tripwire, the August 1 kill date.

So the build is the engine, facing out, with a count of who-isn't-Martin used it. That count is the whole bet.

---

See verdict: [[Flourishing/Builders Council/Verdicts/2026-05-26_the-90-day-bet-what-should-martin-actually-build-between-jun_c1|The Arbiter's Verdict]]