It validates that your case file is complete: ten required fields, available hours between 1 and 168, number of plants between 1 and 5,000 — and that the case really belongs to you.

It looks for a real profile for your species and stage. It has an explicit hard rule: if your plant is a tree (apple, orange, lemon, avocado), it blocks immediately — Yolia does not cover trees today. If the profile found is generic, it tells you so, not pretending a precision it does not have.

It requires exactly 25 active rules against your declared environment: pressure, altitude, water and soil pH (hard block outside 4–9), conductivity, wind, radiation, humidity, and extreme temperatures. From this comes which "functional bases" your formula needs and which it must avoid.

A real beam-search optimizer (beam search, width of 96 simultaneous states) builds the module-by-module selection, evaluating coverage, synergy, redundancy — and immediately discards any combination with a hard conflict.

It groups the candidate modules by real catalog family (it never invents a family) and selects up to 3: primary, supporting, and specialized, with up to 4 guaranteed “anchor modules” per family.

It expands the core by adding, one by one, the module with the highest verifiable actual gain, as long as there is loading margin per phase — in three modes (balanced/maximum/dracula, the latter for severe climate stress). If two modules are incompatible, it doesn't eliminate power: it relocates them to different phases.

It models your plant as a real transport system: root absorption, xylem/phloem, reserves, source-sink. It hard-blocks if the saturation risk reaches 92% or more — if you're asking it to process more than it can handle.

The cross-audit of everything above: 13 simultaneous checks, zero vegetative-nitrogen modules in the reproductive phase during flowering/fruiting/ripening. It's the only one at this stage that already writes to the database.

The last gate before modules turn into a formula: unique identity, hours within what is available, at least one ingredient, assigned phase.

It keeps your own crop from showing up as an ingredient in its own formula. If it detects the match, it looks for a real substitute and scales it; if there's no substitute, it removes the ingredient instead of letting it slip through unnoticed.

It translates your botanical profile and the modules into concrete biochemical functions — and literally states, in its own output: "Does not demonstrate causality or field results."

It adds up the validated ingredients at two levels: global total and per-phase total. It does not assume everything is mixed in a single tank — it respects the phase separation from the previous engines.

It turns the reference recipe into the real plan for your batch. If a reference data point is missing, it withholds the amount instead of making it up. It calculates whole batches of recipe, always rounding up — it never promises an unvalidated fraction of fermentation.

It builds the real work sequence per module: steps, hours, pH and temperature window, storage conditions — validating that nothing exceeds your available hours.

It converts each ingredient to grams and adds up the declared NPK load. It detects specific hard conflicts, like a target pH outside the physical 0-14 range.

It combines botany + biochemistry + chemistry into a single hierarchical decision: if any is blocked, everything is blocked. Only with all three domains complete does it reach "preliminary approval" — and it explicitly states which validation stages remain pending (regionalization, advanced chemical safety, dosing, experimental validation).

The most demanding gate: 24 simultaneous checks which must all be true, from the case footprint to saturation and waste staying below 92%. If something fails, it blocks and lists exactly what.

It assembles the final 9-section report you are about to see — but only once the structural audit has already passed.

It writes your formula, its version, and the entire scientific evaluation to the real database — the only point where the case file becomes permanent.

The final gate: confirms valid identifiers, unblocked science, a consistent report, at least 2 phases — and that the notice "prototype not authorized for field use until dosing, regionalization, stability and final safety are complete" is present.

The last link: it logs the approval and marks your formula as structurally approved. If the science was blocked, it rejects the output even if it got this far.