How AI Formula Lab works

How are formulations generated?

1. T1 — Loaded ingredient measurements, reviewed data and public nutrition records, with individual provenance.
2. T2 — External MFDS/open-data lookup. Failure, no evidence and success remain distinct; T1/T3 data is retained.
3. T3 — Aggregate partner manufacturing and product-report records; individual customer formulas remain private.
4. T4 — Claude drafts the formula and estimates missing nutrition. Estimates never replace confirmed evidence.

Lookup budget: 10 seconds; assessment: 20 seconds. Timeouts never trigger duplicate generation by another model.

Who confirms nutrition values?

Calculated values require evidence for at least 90% of formula weight and all five core nutrients. Incomplete data and AI estimates remain separately identified; supplier certificates or measurements are needed for missing ingredients.

An admin-supplied label draft applies rounding, labeling units and tolerances. Final labels require prototype analysis and review of the current regulations. Korean food labeling standards

Can a draft go straight to production?

No. Specialists must check food type, additive limits, allergens and formulation, then analyze a prototype. Manufacturing checks do not guarantee production feasibility or advertising legality. Seobifood operates on a general-food basis, not health-functional ingredient validation. Korean Food Code

How accurate are results in practice?

No publishable backtest or prototype comparison metrics are available yet. Accuracy or production success rates are not claimed without verified results and review dates. Market nutrition distributions are comparisons, not evidence-score contributions.

What AI does not do

• Invent approval numbers or report failed external lookups as success.
• Copy customer formulas or substitute raw-food nutrition for powders and concentrates.
• Automatically write public comparison results into ingredient masters or inflate evidence scores.