The Bench.

A working nautical almanac and sight reduction computer that needs no signal. An auto mode walks the whole procedure and puts an artificial horizon over the camera; or shoot by hand with the phone's inclinometer, or type in a reading from a real sextant, and get a position by least squares. Install it to the home screen and it runs with the radio off.

A browser has exactly one wideband receiver, the microphone, and everything else is walled off on purpose. So this probes what the device will actually let a page hear, puts a real analyzer on the band you do get, and drives an RTL-SDR over WebUSB for the rest.

Three instruments that treat the camera as a photometer. A photon transfer curve for what the sensor's noise is made of and how many stops it really has, a slanted edge for what the lens actually resolves, and Beer-Lambert absorbance with a calibration curve for what is dissolved in a cuvette.

Buhlmann ZH-L16C with gradient factors, the algorithm inside a modern dive computer, written out where you can watch it work. Sixteen tissue compartments loading and off-gassing, a live ceiling, a stop schedule, and the gas planning that decides whether any of it is survivable. It also keeps a dive log on the device: import what your computer's app exports, build a profile by dragging it, and take it away as a spreadsheet. A teaching instrument, not a dive computer.

Twelve ways to look at the same sound, full screen and built to be stared at. Nine of them are a spectrogram displaced on the GPU, so frequency across, time toward you and level as height becomes a landscape, a tunnel, a canyon, a mandala, a planet, a torus, a whirlpool, a twisting ribbon, or a flat torus in four dimensions that squares off into a tesseract and turns through itself. The other three draw the geometry the notes have rather than the spectrum does: pitch as a helix where one turn is an octave and a chord is a shape you can recognize, rhythm as a spectrum of how fast each band is pulsing, and the stereo image with time running away from you. Feedback trails, bloom and chromatic aberration on top, driven by an onset detector. Every height is still a measurement, and the maths is written out in the page.

Every confirmed planet outside the solar system, pulled live from the NASA Exoplanet Archive by your own browser. The points are the easy part. The interesting thing is the empty regions, because most of them are empty for reasons that have nothing to do with what is out there: turn on the transit depth, transit probability and radial velocity contours and the boundaries of the sample land exactly on lines drawn by our instruments. The radius valley and the Neptune desert are both in here, along with everything derived from first principles rather than read from a column.

Measure a room properly. A swept sine deconvolves to the impulse response, which gives reverberation time per octave to ISO 3382, the frequency response, and where your predicted modes actually landed. An auto mode runs the whole sequence and reads the result back in plain terms. Then it tells you what to fix, and refuses to pretend EQ can fix all of it.

A flight simulator for a bacterium. At Reynolds number 0.0001 water behaves like tar, inertia does not exist, and every instinct you have about moving through a fluid is wrong: stop swimming and you stop dead within the width of an atom. Purcell's scallop theorem is here to be failed at, then beaten with a corkscrew. Then you steer a cell to a sugar source the only way a cell can, by choosing when to give up on a direction, and race twenty four of them that do it properly.

You are a nanoparticle in a small artery, with no engine, no rudder and sixteen forks between you and the tissue. The tree is built from Murray's law alone and every number falls out of it: the velocity, the wall shear that barely moves in sixteen generations, and the fact that blood in a capillary flows almost as easily as water. At each fork the fluid has already decided, and all you can do is be on the right side of the dividing streamline when you arrive. Then get to the wall and try to stick before the shear tears you off.

Describe a drug product and it builds the release specification and the certificate of analysis that follow from its form, its route, its dose and where it is going. Twenty one modalities, from a tablet through an antibody drug conjugate and an mRNA particle to an autologous CAR-T, against a library of a hundred and twenty one tests. Twelve launch regions, each with the pharmacopoeia that governs there and the release obligations that come with it. The ICH thresholds are computed from the dose rather than quoted, so the impurity limits, the mutagenic acceptable intake, the solvent and elemental concentrations and the endotoxin limit all move when you change it. Every row says whether its criterion is fixed by the pharmacopoeia, derived from ICH, or yours to justify, and where a region's chapter numbering is not carried here it says so rather than inventing one. It also checks the product against itself, because a specification can be internally perfect and still describe something that cannot exist: an autologous cell therapy has no second dose to withdraw from a multi-dose vial, a live biotherapeutic that passes a sterility test is dead, and an intrathecal product cannot carry a preservative.

An impurity limit is not one number. It is a question of which guideline owns the impurity, what the maximum daily dose does to the arithmetic, and whether the product is inside the guideline's scope at all. This works all three out and shows the rule it used. ICH Q3A for organic impurities in a new drug substance, ICH Q3B for degradation products in the finished product, and ICH M7 for the mutagenic ones, which run on an acceptable daily intake rather than a percentage and therefore land two or three orders of magnitude lower. Scope is settled before any arithmetic: a peptide, an oligonucleotide, a biotechnological protein and an advanced cancer indication are each excluded from frameworks people routinely apply to them anyway, and the engine says so as a result rather than quietly pushing the product through a threshold table. Nitrosamines get their own overlay, because they sit in M7's cohort of concern where the threshold of toxicological concern does not apply and a regulator's published intake governs instead: the FDA and EMA tables are carried separately and shown side by side, and on sixteen compounds the two disagree, once by a factor of fifty two. Every limit table is transcribed from the guideline PDF itself rather than from a summary, and eighty six assertions run in front of you each time the page loads, most of them on the exact dose boundaries where the guideline says up to and including 2 g per day, because that is where a calculator is either right or dangerous.

Formulating a nanocarrier is a search problem, and the usual way it goes wrong is that the search is badly designed. This walks the loop end to end: narrow the technique from what the molecule is, bracket the factors that matter for it, build a design that can actually estimate what you think you are estimating, fit the response surface to your own results, and find the settings that satisfy particle size, polydispersity and encapsulation at once by Derringer and Suich desirability. Every design is constructed from its definition and then checked against its own defining properties rather than copied from a table: the resolution of a fractional factorial is derived from its defining relation, a rotatable central composite reports the alpha it actually uses and names any factor whose low star point would be physically impossible, and Box-Behnken is refused above five factors because the all-pairs construction stops being the Box and Behnken design there. Look the molecule up in PubChem or type its descriptors in, rotate the conformer in a viewer written here rather than pulled from a content delivery network, then read the protocol out at the centre point or at the optimum. Forty two assertions run in front of you on every load. Everything you type stays in the browser; the PubChem lookup is the only request the page can make, and only when you press it.

A painting has a texture no copy of it shares: the weave under the ground, the ridge a loaded brush leaves, the crack that opened in 1904. Prop the phone facing it, put a marble or a ball bearing in shot, and move a torch slowly around while the back camera watches. The highlight sliding across that ball says exactly where the light was in every frame, which is enough to solve for the way the surface slopes at every point, and its known diameter fixes the scale so two sessions taken from different distances compare directly. Twenty four patches of the relief are committed to a domain separated SHA-256 Merkle root and signed. Two completely different sweeps of one painting agree at 0.98 and a different painting scores 0.06. The score is read at one transform all the sites agree on rather than letting each hunt for its own best offset, because the best of four thousand correlations is not a correlation any more, and that is the difference between rejecting a second canvas off the same bolt and accepting it. A screen lamp mode is there too, for when you have nothing but the phone.