One sentence, seven compiled plans, and seven places it stops.

1. Reversible

   01

   Say who you want to reach

   A sentence in plain English — "London-based technology decision makers, £15,000 a month, four weeks, awareness" — becomes a canonical persona object: semantics, geography, firmographics, demographics, intent topics, exclusions. Every extracted element comes back as an editable chip with its own confidence.

   You getAn editable persona, with per-element confidence.
   where it stops

   Nothing is inferred silently. An age bound the brief did not state stays absent rather than defaulting to the full range, because a default is a claim you did not make.

2. Reversible

   02

   See what it becomes on each channel

   The persona compiles into seven native targeting specifications — the segment ids, attribute combinations, show allowlists, venue packs, venue types, keyword seeds and interest terms each channel actually accepts — each with a fidelity score, a grade and a written rationale.

   You getSeven compiled plans, scored 40 to 85, with a lawful-basis manifest each.
   where it stops

   The language model extracts semantics and never emits a vendor identifier. Deterministic registry tables do that, by matching your terms against catalogue keywords, which is why a hallucinated segment cannot reach a buy.

3. Reversible

   03

   Get a plan that adds up

   Budget, objective and flight produce one funded line per provider, with the split, forecast reach and frequency, CPM, fidelity badge and per-line reasoning. Or skip the generator and pick the billboards, shows, channels and venues by hand — manual planning is a first-class path, not a fallback.

   You getA costed plan, line by line, with the data CPM uplifts visible before approval.
   where it stops

   Deterministic code does all the arithmetic, grounded in adapter forecasts. The model explains the plan; it does not compute it. A model that is wrong about a budget is wrong in a way that looks right.

4. Reversible

   04

   Approve it, and check the creative

   Generate or select same-channel creative for every line. The pre-check predicts Clearcast, CAA and publisher flags before you commission anything, and format limits are enforced as schemas so invalid copy fails at the boundary.

   You getPrechecked creative bound to each line, with an immutable fingerprint.
   where it stops

   Approval records evidence. It cannot enable spend. Those are two separately granted actions on purpose, because the most common way a platform spends money by accident is treating them as one.

5. The funding gatethe only point where money leaves

   05

   Fund it, and launch

   The wallet funds the campaign and the launch runs. Sandbox rails run immediately. Live search and social rails stage every campaign, ad group and ad PAUSED or DISABLED, persist the provider ids in an exact staged manifest, and begin no spend.

   You getA live campaign, or a staged one waiting on a human decision.
   where it stops

   A launch the wallet cannot fund does not launch: the check fails closed by default and only a sandbox preview may opt into warning instead. Enabling spend on a staged live line needs a fresh, separately granted resume that first re-proves the upstream hierarchy still matches the manifest exactly.

6. After launch

   06

   Watch it deliver

   Delivery pacing, booking timelines for the assisted rails, cross-channel frequency, and the journey explorer that shows a postcode sector exposed to a billboard and then to a CTV ad. Every outbound side effect writes a provider attempt before the call and settles it after.

   You getOne report across seven channels, with delivery labelled by provenance.
   where it stops

   An ambiguous provider outcome is never retried. It moves the line to reconciliation-required and hands an operator the vendor, the operation and the idempotency token to go and check.

7. After launch

   07

   Ask whether it caused anything

   One button runs a matched-market geo-lift test against held-out UK conurbations, with difference-in-differences readout. Optimiser proposals arrive explained and wait for approval.

   You getA powered causal result, or an honest refusal — and today, in sandbox, always the refusal or a clearly labelled modelled demonstration.
   where it stops

   The power analysis runs first, at α = 0.05 and 80% power. An underpowered design returns the sample days it would need and no lift figure at all, because a number from an underpowered test is noise that looks like evidence.

Two of those refusals are worth reading twice. At step 02 the language model never emits a vendor identifier, which is the mechanism that stops a hallucinated segment reaching a buy. At step 04 approval and spend are separately granted actions, because the most common way a platform spends money by accident is treating them as one thing.