The same process produces every product page, section by section, with every factual claim traced to a source and adversarially checked before it reaches you. This is the workflow, the content formula, and how each section’s variations are chosen.
See the live Turmeric (D2) page →This is the direct fix for the June QA failures. The engine physically blocks any draft that breaks this rule — content cannot leave the system until it’s clean.
Each SKU runs the same path. Every step below carries an honest status: ✓ running = a command you can run live today · ◐ pilot = built and proven on the 5 pilot SKUs, being generalised · ○ next = designed, build in progress.
Pull this SKU’s facts from the JV legacy DB. bun run db:build (ingest + doses + links).
Attach verbatim GB/EU authorised wording. 41 claims in the library; linked per-SKU.
PubMed RCT/meta-analysis by PMID. bun run db:fetch-evidence — 69 packs on file.
Compute satisfaction stats from the SKU’s own reviews. bun run review-stats <SKU>.
bun run assemble builds a section from the fact stores (scientific V7 live; other sections being added).
bun run verify <draft> — checks every claim resolves to a held source. Runs today.
The gate emits a refuter-packet.json of claims to challenge. Auto-agent loop is the next build; today it’s a reviewed packet.
You approve; export to Matrixify/Shopify. Proven on the pilot batch.
assemble generalises it section-by-section.bun run verify — deterministic checks live today; adversarial refuter packet emitted, auto-loop next. Returns three verdicts:A claim can’t be traced to a real source. Do not send. This is real and runs — it caught 2 unverifiable citations in Turmeric and forced the fix.
Deterministically clean, but the refuter packet must still be worked. Today that pass is reviewed by hand; the auto-agent is the next build.
Clean — every claim resolves to a held source. Turmeric is here today, verified.
No section is free-written. Each one is assembled from a fixed formula: a benefit/positioning line the customer reads in <1 second, a proof element drawn from the strongest source we hold, and a citation that substantiates it.
The proof slot auto-selects per SKU: vitamins/minerals use their authorised EU claim (compliant by default); botanicals like turmeric use a study count or formulation spec. That single rule keeps all 75 SKUs consistent and compliant.
Every section can be expressed at different levels of boldness. Bolder = more persuasive but more compliance sign-off; safer = lower risk, scales faster. The layout never changes — only the wording.
| Section | Variations | What changes | Risk spectrum |
|---|---|---|---|
| Above the fold title · pills · ticks |
A1 → A2 → A3 | Positioning intensity: outcome promise → category positioning → pure formulation fact | A1 promise A3 fact |
| Scientific studies 3 stat cards |
V1 → V7 (7 options) | What the cards measure: invented % (banned) → real cited figures → mechanism → review-derived % in Umar’s exact format | V1 invented V7 review % |
| Results timeline Month 1–4+ |
T1 → T2 → T3 | Voice: real customer quotes → usage guidance → benefit-narrative | T1 quotes T2 usage |
| How we compare Umar’s fixed rows |
C1 → C2 → C3 | Row values only (rows are fixed, incl. Cost per Day): efficacy claims → provable facts → binary product facts | C1 efficacy C2 facts |
| FAQ 7 questions |
Fixed framework | Rows 1–3 product-specific · rows 4–7 brand-wide (template-fixed wording) | structurally safe |
DETERMINISTIC_PASS, and the 91/74/41% stat cards computed by bun run review-stats JV-TURMERIC500 over 2,195 verified reviews. Status today: facts ledger, evidence ingest, review stats, and the verify gate run live; section assembly is proven on 5 pilot SKUs and being generalised to the full catalogue; the automated refuter loop is the next build. We commit timelines for the remaining SKUs once the SKU list is confirmed.