The weekend estimate is correct.

The orchestration is a weekend. Determinism, sourcing and a graded track record are not.

Request access The same page, twice

Orchestration

A weekend

Agents that read a page and return structured JSON.

What breaks

Month two

Verdicts move when the page did not.

Track record

Never

It accumulates at the speed of real campaigns and cannot be backdated.

The honest part

You probably can build it.

Five role prompts, a loop, a scoring pass, a report template. It will look like our output by Sunday night.

A weekend gets you

Agents that read a landing page.
Personas holding different briefs.
Structured JSON coming back.
A report that reads well.

A weekend does not get you

A seal store, so an untouched finding replays instead of being recomputed.
A verifier, proving each quote exists in the page it cites.
Author classification, so a rival’s page is never read as a buyer’s words.
A graded record, sealed before outcomes exist.

Side by side

What you own on Monday, and what you owe by month two.

	Build it yourself	WhyUser
	WEEKEND ONE	WEEKEND ONE
Agents that read a page	A weekend	Included
Personas holding different briefs	A weekend	Included
A report that reads well	A weekend	Included
	MONTH TWO	MONTH TWO
Same page in, same verdict out	Hard. Temperature zero is not determinism	The verdict is computed in Python. The model only writes the narrative
Behavioural mix	You will guess. An even split gives 20 to 30% simulated CTR, which is fiction	Weights derived from published channel benchmarks and the 95:5 rule
Committee handoffs	Rarely built. Most in-house versions stop at per-role output	Seven chains scored healthy, at risk or broken
Rare dealbreakers	Averaged away unless you count runs rather than summarise them	Flagged even when only one run hits them
	YEAR ONE	YEAR ONE
Every model update	Re-baseline and re-validate. Yours to own, forever	Ours
A track record	Starts at zero today. Cannot be backdated	473 claims sealed before outcomes existed, graded by customers
Who is accountable when it is wrong	The person who built it, in front of the CMO	Us. That is what you are paying for

Three things that break

None of them show up in a demo.

Drift

Your verdict moves when your page did not.

The same prompt can still return different text, so you cannot tell your edit from the noise.

Even splits

Simulated click rates come back at 20 to 30%.

An even mindset split over-represents the actively shopping buyer more than tenfold. Real cold LinkedIn is under 1%.

No handoff graph

Every role approves and the deal still dies.

Catching this means scoring the edges, not the roles. Most in-house builds stop at per-role output.

Common questions

Answers, briefly.

Can I build this with Claude Code and agents?

Probably. The orchestration is a weekend. The gap is everything after it.

What is the real cost?

Maintaining simulation infrastructure through every model update, for exactly one customer.

When should I build it?

When simulation is your product, or you have a research function that owns methodology.

Can I copy the calibration?

No. You can copy a sprint. You cannot copy a quarter of graded outcomes.

Keep reading Claude vs WhyUser → How to evaluate a method → Determinism & replay → All comparisons →

Before you spend

Fail in the test.
Win in the market.

The orchestration is a weekend. Determinism, sourcing and a graded record are not. Try the finished one first.

Request access, run it on your next campaign →The same page, twice