Five levels, and knowing which one is broken

Alex, 2026-08-21: *"i need multiple levels of thinking. meaning you have to be like a marketing expert not just think about creatives… create angles out of verifiable things that you can scrape or from brain. if not the system is just a low level tool."*

Fair, and the diagnosis is exact. Everything built so far — diversity, payoff, taste — operates on level 1. A system that only reasons about hooks is a copy checker with good manners.

The levels, and what actually moves

The rule that makes this a strategist rather than a checker: you cannot fix at level 1 what is broken at level 3. A better hook on a worse offer loses to a worse hook on a better offer, every time. So the first job on any account is not "write angles" — it is name the binding level, and say so even when the client only asked for creative.

That is also the honest reading of the 50–60% strategy / 30–40% execution split in creative-engine.md. Strategy is levels 3–5. Execution is level 1.

What each level can be built from, without asking permission

The constraint Alex set: *make do with current constraints, from things you can scrape or from brain.* So each level below names its verifiable inputs — a re-checkable citation, per 0043_verifiable_evidence.sql — and separately what should be *suggested* for clearance rather than waited on.

5 — Market

• Scrapeable: the category's shared claims, from competitor sites and their live ads. What everyone says is the thing we must not say — a competitor_claim fact is directly actionable as a *prohibition*, which is the cheapest strategic win available.
• From brain: our own CPL by industry across 33 clients. Nobody else selling creative software has per-industry Singapore CPL bands from real spend.
• Suggest for clearance: any claim about market size or growth. We cannot verify those, so they should never gate work.

4 — Positioning

Most positioning is unusable because it names no alternative. "Premium interior design" positions against nothing.

• Scrapeable: the actual alternative set. For renovation that is *direct contractor*, *ID firm*, *design-and-build*, *DIY project management* — each with published prices and published claims. The alternatives are on the open web.
• From brain: what lost deals chose instead. client_outcomes already holds 55 rows; a loss with a reason is the highest-value positioning fact we own and it costs nothing to read.
• Suggest for clearance: the positioning *statement* itself. Verifiable inputs, human wording.

3 — Offer

The level agencies skip because it belongs to the client. Skipping it is why creative work plateaus.

• Scrapeable: their own published price and terms. Inspire ID's *"From $19,000"* is a fact on a page with a URL and a retrieval date — it needs no approval to be *true*, only to be *repeated*.
• From brain: which packages actually closed, at what value. deals and leads carry deal_value_cents.
• Registry-checkable: CaseTrust, HDB licence, BCA. A credential we can look up is a proof point; a credential a client told us is a claim. That distinction is the whole reason clients.registry_refs now exists.
• Suggest for clearance: any *change* to the offer. Proposing a risk reversal — a fixed-price guarantee, a deposit held in escrow — is strategy work worth thousands, and it is a recommendation, never something the system asserts on a client's behalf.

2 — Angle

Covered by diversity.ts and payoff.ts. One addition from the levels above: an angle must name which level it is arguing on. An angle that argues at level 1 when the account is constrained at level 3 is a well-made irrelevance.

1 — Creative

Where we started. Note that hook quality is 10–20% of performance. It is the last thing to optimise, not the first.

The cross-cutting layer: what a lead is worth

Without this, "good creative" has no definition. A $60 CPL is excellent or fatal depending on close rate and deal value, both of which we hold:

allowable CPL = deal value × close rate × target margin

Inspire ID's own numbers make this concrete — deal_value_cents on won deals, close rate from leads by stage. Any creative recommendation that does not reference this is a preference dressed as a judgment.

This is also the argument that a client cannot dismiss. "We think this hook is stronger" invites debate. "This concept holds CPL at $47 against your $61 allowable, and the alternative does not" ends it.

How this changes the loop

The old loop stalled: nothing could proceed without approval, and 152 facts sat in review.

The new loop, per 0043:

1. Gather what is verifiable — scrape the client's surfaces, read the dossier, check registries. Every fact carries the exact words and a way to fetch them again.
2. Reason across all five levels and name the binding one.
3. Build angles on cited facts. They reach provisional automatically, granted by the database from the citations, and are immediately usable for internal work, proposals and drafts.
4. Publication still needs a human. That is where the risk is, and it is the only place the old gate was actually protecting anything.
5. Clearance becomes a queue we produce, ordered by what is blocking spend — a byproduct of the work rather than a precondition for it.

Provisional cannot be set by hand; the trigger refuses it without a citation, and a citation without a URL or a dossier statement is refused by a check constraint. Loosening one gate meant tightening another, or it would have been a loophole rather than a design.

What to build next, in level order

1. The surface scraper — client site, reviews, registries — writing market_facts and claim_citations together. Nothing else on this list works without it.
2. The alternative set per industry, which is what makes level 4 real.
3. Allowable CPL per client from deals and leads. Small, and it makes every recommendation arguable in the client's own numbers.
4. The binding-level diagnostic: given a client's funnel, say which level is the constraint. This is the output a strategist charges for.