Precisely Stated Problems

The key move is to stop treating VUV as a 2D trace. It's 3D: at every instant the detector records a 12-band spectrum — a shape. And here's the physical fact we lean on: a real molecule's absorption is coherent across those 12 bands — one consistent, on-template shape — while noise is incoherent, twelve random wiggles that form no molecular shape.

So instead of asking "is this wiggle tall enough to be real?" we ask "is this 12-band shape coherent — does it look like a real molecular absorption, or like noise?" Coherence is a stricter test than height: it recovers real peaks that height missed and rejects noise — and it can never invent a false one. That's what lets you pull a clean, trustworthy fingerprint out of a noisy run, even near the floor. (The concept is the point; the implementation stays under the hood.)

We stopped asking the spectrum to do what physics won't let it do. Restate the problem precisely and it changes shape: the claim isn't "the spectrum identifies every sulfur" — it's "the spectrum identifies the family, and something else must pick the twin." Stated that way, the answer is sitting in the run: the twins may share a spectrum, but they don't share a boiling point — and boiling point is what decides elution order on the column. Structure sets the property; the property sets the order; the order is measured physics, not a lookup.

So identification becomes convergence: the fingerprint narrows the field, elution order and boiling point discriminate the look-alikes, and the certified contents of the standard anchor the truth. No single axis is trusted — the ID stands only where the axes agree. And when they don't agree, the system says so, out loud, instead of hiding behind a 0.999. A system that knows when it can't know is worth more than one that always answers.

Restate the problem precisely and the assumption shows itself: "a detector only measures amount" was true of single-channel detectors. A LUMA isn't one. Twelve VUV bands are a readout of electronic structure — which transitions are available, which bonds are present. And structure is where properties come from. The cleanest case is refractive index: refraction and VUV absorption are the same electrons responding to light — the physics practically insists they correlate. Within a chemical family, they do, strongly.

So the approach isn't "regress spectrum against property and hope." It's physics first: read the family from the shape of the fingerprint, then apply that family's structure–property law — local models anchored to known physics instead of one global black box. Anchoring buys two things chemometrics alone can't give: reach (out on unfamiliar blends, the physics-anchored local models cut prediction error nearly in half versus generic regression) and honesty — a family-first model knows when a sample doesn't belong to any family it's seen, and says so instead of guessing. Some properties follow rigorously; others hold within a family. We say which is which — that's the difference between a claim and a con.

State it precisely and the assumption is visible: "the detector can't un-mix them" assumes the detector records a number. A LUMA records a spectrum — every instant of the lump is a 12-band shape, and the lump's shape is just the two compounds' shapes added together in known proportions that change across the peak. Walk through the overlap and the blend ratio shifts point by point — the spectrum visibly migrates from the leading compound's shape toward the trailing one's.

That migration is the tell. It says there are two compounds, it shows what each one's spectrum looks like (the leading edge is nearly pure one, the trailing edge nearly pure the other), and it makes splitting the amounts a piece of algebra instead of geometry — un-mixing shapes, not guessing curves. The separation happens in spectral space, not on the column.

Beer's law: A = ε · l · c. Absorbance equals absorptivity × path × concentration. Look at where the response lives: ε is a property of the molecule — how strongly its electrons couple to light at each wavelength. It isn't your instrument's opinion; it's the compound's physics. And VUV absorbance is famously linear over a wide range.

Stated precisely, the problem inverts: calibrate the molecule once, not the instrument forever. Response factors live in the library, alongside the fingerprints — measured once, or carried across a family — and each run is tied down with an anchor instead of a fresh curve for every compound. Honesty matters here: this is pseudo-absolute, not magic — anchors, checks, and validation against certified standards stay in the loop. What leaves is the treadmill.

Same move as the noise problem, taken to the floor. A trace compound is small — but it's still coherent: its shape shows up faintly and consistently in every band. Noise is loud but incoherent. Score that, and there's no knife edge at all: real absorptions score near 1.0; noise sits at the mathematical floor — 1 in 12, about 0.08, fixed by the number of bands. A cliff, not a slope. Detection stops meaning "tall enough" and starts meaning "shaped like a molecule" — and a shape test can reach below the height floor, with the library telling it exactly what shape to look for.

And the honest boundary: this does not turn a LUMA into an ultra-trace detector — raw sensitivity is set by the optics, and dedicated sulfur detectors still win on pure detection limit. The claim is different and, for real decisions, often worth more: down at trace levels, when something is there, you don't just learn that sulfur exists — you learn which one. Detection with identity attached.