goodbadgray
investor demo 05 / 23 the concentration engine
concentration estimation · monte-carlo under constraints

an ingredient list is ordered, not quantified.

brands publish the order, not the amounts. the engine estimates every concentration anyway — an llm prior, disciplined by label order, a cutoff marker and mass balance, then pushed through monte-carlo for honest error bars. below: the real analysis of a real sunscreen.

ingredients estimated
Σ posterior means
90%ci on every estimate

one formula, under the lens

select an ingredient · ↑ ↓ work too

the label, in order

INCI position · posterior mean shown as bar

llm prior
%
point estimate, pre-constraints
posterior mean
%
monte-carlo, all rules applied
90% interval
%
what the engine will defend
illustrative replay
the rulebook — every draw must obey all four
rule · label order order is rank
rule · cutoff the <1% boundary
rule · mass balance it must sum to 100
rule · type priors a sunscreen is not a serum draws are bounded by typical use ranges for sunscreen_spf · lotion formulations, per ingredient role. range numerics are not part of this data harvest — stated, not shown
the bridge into scoring
risk scale factor
multiplies this ingredient's six risk scores at this concentration
benefit scale factor
multiplies its eight benefit scores at this concentration
risk −5‥+5 · benefit 0‥+5 · concentration-adjusted

the whole formula, reconstructed

45 real estimates · log scale · hover any column
posterior mean 90% interval llm prior cutoff
above the cutoff, means fall monotonically with position — the order constraint at work