The determinism receipt
"Deterministic" is an easy word to put on a landing page and a hard one to stand behind. This page is the precise version of the claim, written so you can test it rather than take it.
A language model reads your brief. Everything after that is deterministic physics over a hand-curated catalog: same inputs, same formula.
The boundary, drawn exactly
The structured generator
Composing from an archetype, descriptor goals, and a concentration is explicit math: descriptor weights, volatility filtering, accord scaffolding. The same structured brief produces the same formula, to four decimals, every time.
The computed read-back
The numbers on any formula, whether composed here or pasted in: the pyramid, contextual roles, projection, longevity, sillage, cost, the radar, and the density-aware lab math. Run the read-back twice on the same composition and the numbers do not move.
The craft tooling
Substitution ranking, the cost optimizer, reformulation around an unavailable material, build-check against your inventory, dilution and batch math. All of it is computed from catalog data and your formula. No model in the loop.
The plain-language brief box
When you describe a scent in your own words, a language model interprets the words and proposes the palette and the starting proportions. That step is creative on purpose, so the same sentence can compose differently on different runs. Every figure computed about the result, the pyramid, the roles, performance, cost, and the lab math, comes from the deterministic engine over those materials.
Stories and names
The scent story and name suggestions are written by a model. They are prose for a client deck. They never touch a percentage.
The labeled AI assists
A few optional assists use the model and say so in the interface: the AI Insights commentary on a formula, whose one-click suggestion changes percentages only when you choose to apply it, the inventory advisor that drafts formula ideas from what you own, and accord naming. Where an assist proposes materials or percentages, those are the model's proposals. The read-back you see about any result is still computed by the engine.
Check it yourself
All three steps run on a free account.
- Open the generator, pick an archetype and the same descriptor goals, and compose twice. Compare the two formulas line by line. They match.
- Paste the same formula twice and compare the two read-backs. They match, number for number.
- Now write a one-line brief in the plain-language box and run it twice. You may get two different compositions. That is the language model doing its one composing job, and it is the reason this page exists.
What determinism buys you
Reproducibility and auditability. A formula you can regenerate is a formula you can version, hand to a client, and defend line by line. Every figure traces back to the catalog's physical data, not to whatever a model felt like saying that day.
What it does not claim
Determinism does not claim the formula smells finished. The output is a structured, buildable starting point with honest numbers attached. Whether it is any good is decided at the bench, by you. The numbers themselves are computed estimates, and we keep a public account of the methods behind them.
Related reading: methods and provenance · your formulas are yours · how it works