Changelog

Notable changes to STEMKit and stemkit-core.

[output] marks a change that alters a reported number. Figures produced with an earlier version are worth re-checking.

v0.5.1 — 2026-11-11

A video tutorial on every research tool, the MD Workflow Generator's guide set out by engine, and number fields that are no longer rewritten while they are typed in.

Added: video tutorials

• Every research tool has a video tutorial. The Structure Inspector, the Coordinate Manipulator and the MD Workflow Generator have theirs now; the generator has four, for the cluster request, GROMACS, LAMMPS and PLUMED.
• New videos for five tools whose pages changed in v0.5.0: the Plot Builder, the Scientific Converter, the XVG Visualizer, the BibTeX Deduplicator and DOI to BibTeX.

Fixed

• Number fields in the style panels. Typing a decimal such as 2.6, or a number whose first digits are below the minimum such as 10 or 123, could overwrite the field while it was being typed: "2.6" ended as 4 and "8.5" as 40. A field is now left alone while it has focus, a value outside its limits is applied on Enter or on leaving the field, and the figure follows the typing once what is typed is a number within the limits.
• Plot Builder, panel gap. The gap fields in the Layout section stopped at 72 pt, while the figure and the Style panel take up to 144 pt, so a larger gap typed there was cut back.
• Plot Builder, script source. With an example on the page, the switch above the script showed "Read the CSV files" pressed, although an example's numbers are always written into the script. It now shows "Data in the script" while every file in use is an example, and goes back to the person's own choice once a file of theirs is loaded.
• Changelog. v0.5.0 listed twelve built-in examples; there are fifteen.

Changed

• Equation Formatter, no keypress sound. Deleting no longer plays a sound, and sound/fahhh.mp3 is gone; the editor is silent.
• Third-party notices. The licence text of 3Dmol.js, KaTeX, MathLive, Plotly.js and Chart.js now sits beside each bundle, the two Pomodoro sounds are listed in THIRD_PARTY_LICENSES.md, and that file is part of the npm package, with a note in src/core/README.md that the GROMACS option text and the PLUMED keyword tables stay under the LGPL.
• MD Workflow Generator, how-to guide by engine. The guide has a tab for GROMACS, LAMMPS and PLUMED and one for the cluster card and the schedulers, each with its own video and steps (LAMMPS and PLUMED had none), and it follows the engine tab in use. The three engine videos moved from the method section to their tabs. The FAQ is grouped under GROMACS, LAMMPS, PLUMED, schedulers and the page itself.
• Guide sections. The Plot Builder and the XVG Visualizer set out the text under the tool as the other pages do: a "How-to guide" chip, notes in cards on a grid, and references in a narrow column. On every page the guide now sits in the same column, and the questions are the same width.

v0.5.0 — 2026-10-10

The Plot Builder draws nine kinds of chart, on a second pair of axes and on a grid of panels with letters and insets, with fifteen examples to start from; and a second round of fixes: job arrays, wall times and memory as the schedulers read them, unit factors from the exact constants, samples that do not vary, files of several data sets, and duplicates told apart by their DOIs.

Added: more charts in the Plot Builder

• Nine kinds of chart. Each series has a Chart menu: line, points, step or band (as before, with Step new), bars, bars of means, a histogram, box plots, violin plots, a heatmap, contours, and a 2D histogram. Bars name themselves from a text column or sit at the numbers of a numeric one, with an optional Y error; a histogram takes its bins and density; box plots group one column by another or take one box per column; heatmaps and contours lay a value column over X and Y; a 2D histogram counts rows in a grid of equal bins as numpy.histogram2d does.
• The script and the exports draw the same charts. Every kind was run through the generated matplotlib script, reading the file and with the numbers embedded. The grid of a heatmap, contour or 2D histogram is in the script as numbers.
• A second y axis on the right and x axis on top. A series card's X axis and Y axis menus put a line, bars or a histogram against a y axis on the right (temperature and pressure against time) or an x axis on top (wavelength below, wavenumber above), or both. Each has its own label, taken from the first series on it, scale, limits and ticks in the Style panel, and takes the colour of its series when it has only one. Zooming zooms both; one legend lists the series of every axis. The script draws them with matplotlib's twinx and twiny, and the preview, PDF, PNG and SVG match it.
• Bars lying flat, and stacked. A bar series' Bars menu lays it flat: the names go up the left axis, the first at the top, and the lengths and error bars run along x. With other bars stacks it on the stacked series before it, values below zero downwards, or stacks them to 100% with each value as its share; error bars stand at the end of each bar and the legend reads down the stack. Stacked bars may stand on the right axis; bars lying flat use the bottom and left axes, and their card says so if a second axis was chosen.
• Bars of means. A new chart for replicate rows: one bar for each name with the mean of its rows and their standard deviation, standard error or t-based confidence interval (90, 95 or 99%) as the error bar, and each value over its bar if you like. The card counts the values of each name and says which have only one. The script reads the value column and works out the same means and errors with numpy (scipy for an interval).
• Histogram options. Bins as a number, as one of numpy's rules (auto, Freedman–Diaconis, Sturges, square root, Scott, Rice, Doane: the same edges as numpy.histogram_bin_edges, which the script calls by name), or as edges you type, checked as you type them; bars as counts, density, fraction or percent; and a running total. "auto" is the rule of numpy 2.3 and later, which no longer lets the bins get narrower than half the square-root rule's; the script counts those bins itself, so an older numpy draws the same histogram.
• Violin plots. Grouped as box plots are, drawn from the kernel density matplotlib's violinplot uses (Scott's or Silverman's bandwidth), from the smallest value to the largest or further, each violin the same width or all the same area, with a box and median, quartile lines or a median line inside, and the points. The script works the density out with numpy alone.
• Box plot options. Whiskers to the last value within 1.5 × IQR, to the smallest and largest value, or within the 5th to 95th or 10th to 90th percentiles; a notch at the median; the mean with its 90, 95 or 99% interval. On the card and in the Style panel.
• Fixed: bars placed by numbers were at most 1 wide. Bars 5 apart were drawn as thin lines. A bar's width is now in the units of its axis with no upper limit, the builder makes bars four fifths of their closest gap wide whatever the gap, and the Style panel's Bar width is a share of that gap. Bars on names are unchanged.
• Fixed: a bar with no number widened the x axis in the preview, where matplotlib leaves it out of the limits.
• An example for each kind of chart. More examples, beside Load an example, lists fifteen built-in data sets with the series that show them off: growth curves, two y axes, a top x axis, grouped bars, bars of means, stacked bars, histograms, box and violin plots, a heatmap with contours, a 2D histogram, a step line, two panels, a grid of panels, stacked spectra and a zoomed inset. An example takes the place of the files and series on the page (it asks first when the files are yours) and keeps your plot style. Box plots and violins in two panels now stand over each other.
• stemkit-core: plot-builder gains CHART_KINDS, gridFromXyz, histogram2d, boxGroups, suggestColumns and axisChoices; readTable also returns each cell's text (texts). figure gains the series fields xAxis and yAxis, the panel fields y2Label, y2Scale, y2Lim, y2Ticks, y2Color and their x2 counterparts, and hasAxis. For the charts above, figure gains the violin kind; orientation, stack, per-bar bottom and replicates on bars; normalise, cumulative and rule names for bins on histograms; notch and percentile whis on boxes; the panel field yCategories; and histogramBinEdges, histogramHeights, numpySum, kdeBandwidth, gaussianKde, violinStats, groupMeans, barData and legendOrder. plot-builder gains parseEdges, replicateGroups, histogramOptions, EXAMPLES, builderExample and exampleState.

Added: panels on a grid, shared axes, gaps, letters and insets

• A grid of panels. The Plot Builder's Series tab has a Layout section with the rows and columns of the figure, and each series' Panel menu names the cells ("Row 2, column 1"). In one column nothing changes: the panels stack, share the x axis, and an empty one is left out. On a grid every panel has an x axis of its own, with its own label, scale, limits and ticks in the Style panel; the widths of the columns and the heights of the rows are set there too, and the title stands over the whole grid.
• Shared axes. The x axis shared down each column or by all the panels, the y axis along each row or by all: one scale and one range for the group, and tick labels and axis labels on its outer panels only, as plt.subplots(sharex=..., sharey=...) draws them.
• The gap between panels. The space between stacked panels was matplotlib's own and could not be changed: 3 pt of padding on each side of every panel, 6 pt between two, plus the room for whatever sticks out of a panel, mostly half the height of its first and last tick labels. Gap between rows and Gap between columns now set it in points, in the Layout section and in the Style panel; left empty it is matplotlib's. At 0, panels that share an axis share an edge: the line between them is drawn once, and the tick marks and tick labels that would reach across it are left out. The script sets the padding and spacing of matplotlib's layout engine, and the margin at the figure's edge stays as it was.
• Panel letters. (a), (b) and so on in reading order, written as (a), a, a), A or (A), inside the top left corner of each panel or outside it, from any starting letter, bold or not, at a size of their own; any panel may have a text of its own instead. Outside, they are counted in the margins.
• Insets. Add an inset draws a small panel inside another, with a series of its own, placed and sized as shares of its panel. Mark what it enlarges draws the box of the inset's limits on the panel and two lines to the inset's corners (matplotlib's inset_axes and indicate_inset_zoom). An inset takes no room in the layout.
• Zoom, fit and read-out in every panel. Dragging in a panel of a grid sets the limits of the axes that panel shares; dragging in an inset sets what the inset shows, and double-clicking it fits its data again.
• The script and the exports. A grid with a panel in every cell is one plt.subplots(rows, cols, ...); cells left empty or panels spanning several cells make a GridSpec. The preview lays grids out as matplotlib's constrained layout does (a margin for each column and row from what its panels need, colour bars beside their panels, legends outside), and was compared with matplotlib on grids, spans, gaps, letters and insets, on light and dark backgrounds: the panels land within a pixel or two.
• Three more examples: a 2 by 2 grid with shared axes and letters, three spectra stacked with no gap, and a curve with a zoomed inset. The stacked bars example now has its legend beside the plot instead of over the bars.
• The Style panel with many panels. Past four panels the labels are shown one panel at a time, and past five axes the axis is chosen from a list; a shared axis is set once for the panels that share it.
• Fixed: a legend larger than its panel. matplotlib makes room around a panel for a legend that does not fit inside it; the preview did not, so a short panel with a long legend was drawn larger than the saved figure had it.
• Fixed: setting one panel's label froze every panel's. Editing any field of a panel in the Style panel copied the labels, scales, limits and ticks of all the panels into the style, where they hid the labels of the next file. Only the field that was set is kept now.
• stemkit-core: figure gains layout (rows, cols, widthRatios, heightRatios), sharex, sharey, panelGap and panelLabels; the panel fields row, col, rowSpan, colSpan, inset, letter, and xLabel, xScale, xLim, xTicks and xCategories of its own (the figure's are the defaults); and panelLetters, panelEdges, layoutPads, isStack and titledPanel. figure-python gains gridLines, insetLines and gapLines. plot-builder gains builderLayout and builderPanels, and an example may bring a layout and a style.

Fixed: output

• [output] MD Workflow Generator, LAMMPS check. round() of a negative half went away from zero (round(-2.5) gave -3; LAMMPS gives -2), and $(...) could print -0 where LAMMPS prints 0. The checker follows LAMMPS's own arithmetic.
• [output] MD Workflow Generator, job arrays. A list such as 1-5,10-15%2 was counted as five tasks with no limit, and 10-1 and 1-10:0 were accepted. Arrays are read as sbatch reads them: values and ranges separated by commas, a step after a colon, one % limit at the end, each index counted once; what sbatch refuses is refused, with the reason.
• [output] MD Workflow Generator, wall time. 2-12, 1-12:30, 120:00 and 100:00:00, all of which sbatch takes, were replaced by 24:00:00. Every form sbatch documents is read. A field outside its usual range (99:99:99) is accepted, as Slurm accepts it, with a warning that says what it adds up to.
• [output] MD Workflow Generator, memory. A value that is not a size went into --mem unchecked, and a bare number meant megabytes for SLURM and nothing for PBS, LSF and Grid Engine. One reader serves all four schedulers: a bare number is megabytes everywhere and is written with its unit.
• [output] Scientific Converter, constants. The length table used the 1986 Bohr radius while the force table used the 2018 one, and the atomic pressure, 1/R, dipole and polarizability factors were rounded. Every factor is now worked out from the exact SI constants and the CODATA 2018 Bohr radius and Hartree energy, so the tables cannot disagree: 1 nm is 18.89726125 a0 (was 18.89725989) and 1 Eh/a0³ is 2.942102e+8 bar (was 2.942105e+8). Each factor is checked against scipy.constants.
• [output] Scientific Converter, typed numbers. "1,5" was read as 15, and "0x10" and "Infinity" were taken as numbers. Commas now count only as thousands separators where they cannot be anything else (1,234.5, 1,234,567); anything ambiguous or not a decimal number is refused with a note beside the field, and the other fields are cleared, not left showing the last result. A result of exactly 0 is shown as 0; it left its field blank, so 273.15 K gave no Celsius value.
• [output] stemkit-core, convert. A value converted to its own unit comes back exactly (7.5 kJ/mol gave 7.499999999999999), and exact ratios come out exact.
• [output] Statistics Calculator, values all alike. Six values of 0.1 have no spread, but their mean is not 0.1 in binary and the sums took the rounding for data: two such groups gave t = -1.2e16 and p = 3e-82. The variance of equal values is now exactly zero, the independent t-test says the statistic is undefined when neither group varies, and pairs that differ by the same amount as written (0.2 against 0.1, 0.3 against 0.2) count as a constant difference.
• [output] Outlier Detector, constant columns. A column of one repeated value such as 0.1 was scored from the same rounding and given a Grubbs statistic. It is reported as having no spread. When the MAD or the interquartile range is zero the flags are those the method defines, and the page and the script now say what they mean.
• [output] Error Bar Generator, significance level. Brackets and highlighted pairs used 0.05 at every confidence level; they use 1 minus the level chosen. The stars keep their thresholds of 0.05, 0.01 and 0.001.
• [output] Error Bar Generator, cells left as text. +5 and 1e16 were dropped without a word; they are read. A decimal comma (6,1), a number too large to hold (1e400) and other text are counted and shown under the data. Two groups of equal values were compared from rounding noise; such a pair is no longer tested, and is named.
• [output] Error Bar Generator, p-values. 0.0099 was printed as "p = 0.010" beside **, which means below 0.01. A p-value is given the digits it needs not to round onto a threshold.
• [output] Curve Fitter, confidence band note. The note that the band is hidden behind the fit line assumed a plot area 0.77 of the figure's height, so it could appear while a sliver of the band showed. It uses the panel as drawn and measures the band across the curve.
• Curve Fitter, a fit that runs away. Fitting A*exp(-x/tau) to rising data reported tau = 3e12 as an ordinary converged fit. The page and fitModel now say that a parameter has run away, or that the curve follows the data no better than their mean, and what to try. Fits that behave are unchanged in every digit.
• [output] XVG Visualizer, files of several data sets. Sets separated by & were joined into one line, and the legend of s1 named a column of the first set. Each set is a series under its own legend, numbered as Grace numbers them, in the plot and in the script.
• [output] XVG Visualizer, labels and numbers. \" cut a label short and \#{110000} stopped the page; both are read as Grace reads them. A row holding 0x10 was plotted as 16; it is skipped and counted, as numpy.loadtxt would refuse it. With the data hidden under a running mean, the limit on data written into the script was not applied.
• [output] BibTeX Deduplicator, the same title under different DOIs. Two entries titled "Editorial", each with its own DOI, were reported as duplicates. Entries whose DOIs differ are not grouped, and a title that two or more DOIs share links nothing; those entries are listed under the groups as look-alikes. A doi field that does not hold a DOI counts as none.
• [output] BibTeX tools, ways of writing a DOI. doi:10.…, www.doi.org, dx.doi.org and links without https://, with or without a query string, now match the bare DOI.
• BibTeX Deduplicator. A scan that fails clears the results of the one before and says what failed. Editing the text after a scan marks the results as out of date and switches the download off until the library is scanned again; the download used to write the library as it was at the scan.
• DOI to BibTeX. DOIs in brackets, in quotes, at the end of a sentence and in www.doi.org links are found. A request that takes longer than 20 s is given up with a message, a "too many requests" reply is waited out and tried twice more before the batch stops, and a reply that is a web page or empty is reported as a failure. The messages name the host that answered.
• [output] Decision Matrix, ties. Options with equal totals were numbered 1 and 2, in an order that followed rounding. They share a rank (1, 1, 3), stay in the order entered and are marked as tied; totals show two decimals. Rank the options names the entries it corrected to the range, and a screen reader hears one sentence with the result, not the whole card at every keystroke.

Changed: accessibility

• Method tabs on the Scientific Converter, the Statistics Calculator and the MD Workflow Generator, and the generator's engine tabs, follow the ARIA tabs pattern: one tab stop, the arrow keys, Home and End, each panel named by its tab.
• Scientific Converter. The definition and source of each unit were shown on hover only; they are on a button that works from the keyboard and by touch.
• Upload CSV on the Error Bar Generator and the Statistics Calculator can be reached and opened from the keyboard.

Changed: tooling

• Sitemap. lastmod is the date of each page's last commit; it was the file's modification time, which stamps a fresh clone with today's date and left July dates on pages changed since. npm run check:sitemap fails when a date is older than its page's last commit.
• npm run check:chrome fails for a page that is neither in the catalogue nor in its list of pages that are not tools, and for a tool page whose title, description or JSON-LD name is another tool's.
• CI also runs the tests on Python 3.15 with the current numpy, scipy, matplotlib and pandas.
• stemkit-core: slurm gains parseArraySpec, parseWallTime, resolveWallTime, resolveMemory and slurmMemory; units gains CONSTANTS; bibtex gains isDoi, doiUrl and looksLikeBibtex, parseDoiList returns skipped and findDuplicates returns lookAlikes; error-bars gains readCell, and computeGroups returns what it could not read; xvg-parser returns each set as a block; bandVisibility takes the panel as drawn as a fourth argument; fitModel warns of a parameter that has run away; detectOutliers returns degenerate and a warning.

v0.4.0 — 2026-10-10

Turning a structure by hand in the Coordinate Manipulator, video tutorials on the tool pages, and a round of fixes: the gmx editconf command, BibTeX entries the reader used to lose, and inputs at the edges (groups that do not vary, blank fields, very long columns).

Added: turning a structure by hand in the Coordinate Manipulator

• Turn molecule. The view takes a choice of what dragging turns: the view, as before, or the molecule, with the cell and the axes staying where they are. The three angle fields follow the drag and the view follows the fields, so a rotation is seen before it is applied; Rotate applies it and Reset drops it. A dragged rotation is kept to a tenth of a degree, so the angles shown are the rotation applied.
• What the rotation is. The panel shows the rotation matrix, the single turn about one axis it amounts to, and the angle between the structure's long axis (or the normal of a flat one) and x, y and z, now and after the turn. Lay the long axis along x, y or z sets the smallest rotation that does it.
• Every step as equations and as Python. Rotations, translations and centrings are recorded in order. Under the workspace they are written out as equations, each with its matrix or vector and then the one transform r' = M r + d they add up to, with the LaTeX to copy, and as a script that repeats them: on the file with MDAnalysis, or on any array with NumPy. Both scripts are run in the tests against the page's own coordinates.
• The cell and the axes are drawn over the view, as the Structure Inspector draws its box: lines of a constant width, heavier towards the viewer, with the x, y and z axes in the corner and the structure's own axis through it.

Added: video tutorials

• Fifteen tools show a narrated video tutorial in their "How to use" section (the Statistics Calculator, which has none, under its heading), with a Watch the tutorial link at the top of the page; the home page shows the tour. The videos are collected in the STEMKit Tutorials playlist.
• Nothing is fetched from YouTube until play is pressed. The thumbnails are served from the site (assets/tutorials/); the player comes from youtube-nocookie.com, opens inline with captions off, and keeps a link to the video on youtube.com. tools/build-csp.mjs allows the player's frame on the pages with a video, and no others. The privacy page says so.

Added: why a confidence band cannot be seen

• Curve Fitter. With Confidence band switched on and nothing to see, a note under the plot now says why: the parameters are not independent (as in y = a*b*x), so the band cannot be worked out, or the band is thinner than the fit line and hidden behind it.
• stemkit-core: bandVisibility(model, style, fit) in plot-style makes the same call, with the same sentence, for a script.

Changed

• [output] Coordinate Manipulator, gmx editconf -rotate. After several rotations the command added the angles axis by axis, which is not the rotation that was applied. It now gives the angles of the one rotation the steps add up to.
• Coordinate Manipulator. The angle fields return to zero once a rotation is applied, Rotate and Translate do nothing with nothing typed, and the output buffer and the editconf command can be scrolled from the keyboard.

Changed: accessibility

Every page was checked with axe against WCAG 2.2 A and AA, in both themes.

• Links in the reference lists and credit lines are underlined, so they are not told from the text around them by colour alone.
• Scrolling regions take keyboard focus: the LaTeX table preview and source, the Outlier Detector's table, and the Coordinate Manipulator's output buffer and command.
• Plot Digitizer. A series is picked with a radio button of its own in each row; the row, which holds the colour and name fields, was itself the control. The arrow keys move between series.
• Targets of 24 px for the Plot Digitizer's colour swatches and the Scientific Converter's copy buttons.
• Error Bar Generator. The equations on the page no longer carry aria-pressed, which the chart's buttons had passed on to them.
• The notes in two reference lists are at the list's full contrast, and the icon named in the Scientific Converter's text is announced as an image.

Fixed: output

• [output] Visual LaTeX Tables, labels. A label with an underscore, such as the page's own example tab:my_table, was escaped like text (tab:my\_table), which stops LaTeX with "Missing \endcsname inserted". The key is now written as typed.
• [output] BibTeX Deduplicator, emphasised words. Titles that differed only in a word inside \emph{…}, \textit{…} and the like were reported as duplicates: the command took the word with it once the braces were gone.
• [output] BibTeX Deduplicator, commands and formulae in titles. The same happened with any other command, so \ce{NaCl} and \ce{KCl} made two titles one, as did titles that differed only in a formula, and titles in a script other than Latin that shared their digits. Titles are now compared as they are written in the file: a command around text is dropped and its text kept, a command that is text (\LaTeX) keeps its name, a formula keeps what it says, and letters of every script count. Accents are folded, so Schr\"odinger and Schrödinger are the same word, and $\alpha$ matches α.
• [output] BibTeX Sanitizer, commands with capitals. \LaTeX in a title became \{LaTeX}, which prints a stray brace. It is now {\LaTeX}.
• [output] BibTeX Sanitizer, commands with arguments, and formulae. \SI{5}{\nano\metre} became {\SI}{5}{\nano\metre}, which does not compile. A command with a capital in its name is now braced together with its arguments, {\SI{5}{\nano\metre}}, and a formula with a capital is braced whole, {$\Delta G$}, where $T$ used to be left to be lowercased. Text inside braces is left alone at any depth, and the letters \AA, \O and \L are no longer braced, which split the word they were in. The result is checked against BibTeX itself in the tests.
• [output] Curve Fitter, which name is the variable. In signal = A*exp(-time/tau) + c the offset c was taken for the variable. Whole words now come before single letters, and a guessed name that is not a data column gives way to one that is, so the Hill model on the Michaelis–Menten sample varies in S and no longer asks for a column n.
• Structure Inspector. not: on a range (not:z:20-25) matched nothing; it now selects everything outside the range. Toggle surface could not remove a surface, and changing its type, opacity or colour drew another on top; the surface is now replaced.
• [output] Coordinate Manipulator, the gmx editconf command. editconf translates, then rotates about the origin, and centres last of all. The command put -center 0 0 0 beside -translate after "centre, then translate", so editconf undid the move; and for a structure turned about its own centre it relied on a note saying that adding -center 0 0 0 would make the two agree, which holds only when the centre is already at the origin. The command now lands on the page's coordinates whatever the order of the steps: -center names where the geometric centre ends up, or -translate is the move as it has to be made before the rotation. A centring on the centre of mass, which editconf cannot do, is given as the geometric centre it leaves. stemkit-core: editconfPlacement(steps, finalCentre) in structure works these out, and is tested against editconf's order of operations.
• Coordinate Manipulator, typed numbers. A padding of 0% was read as the default 10%, so a box could not be fitted tight to the molecule. A distance or an angle that is not a number was applied as 0 without a word; the field is now marked, named in a message, and nothing is applied.
• [output] BibTeX tools, entries the reader lost. A quoted value holding a quote in braces, title = "Outer {"}quoted{"} text", was cut at the inner quote and every field after it dropped, so the Deduplicator compared and showed a damaged entry. An entry with an empty key (@article{,, as JabRef writes one that has none) was left out of the count and the comparison while staying in the download; it is now an entry like any other, shown as "No key". In the DOI lookup, removing or keeping fields rewrote an entry from the fields read so far, so a value the reader cannot follow (pages = 100 - 110) took the rest of the entry with it; such an entry is now left exactly as it is.
• [output] Statistics Calculator, one-way ANOVA of groups that do not vary. With every group constant (1, 1, 1 against 2, 2, 2) the within-group variance is zero; the page reported F as not available with p = 1, "not significant". It now says the test is undefined and why, as the t-tests and Welch's ANOVA already did. stemkit-core: oneWayAnova returns F and p as NaN with a note in that case, and fUpperTail(Infinity, …) is 0, not 1.
• DOI to BibTeX. The "1 failed" count, the error report and the Retry button of one batch stayed on the page while the next batch ran. They are cleared when a batch starts.
• [output] Decision Matrix, blank fields. A weight left blank counted as 1 and a score left blank as 1, the lowest of each, though the fields start at 3 and 5. A blank field now counts as the value it started with, and Calculate writes that value back into it.
• Very long columns. About 150,000 values in one group stopped the Error Bar Generator with "Maximum call stack size exceeded", and the same limit stood in the Statistics Calculator's summary of a sample, the Data Cleaner's column transforms, an index-group range and the Structure Inspector's solvation of a large solute. All take columns of any length now, and the Error Bar Generator no longer slows with the square of the number of rows when each row holds one value.

v0.3.0 — 2026-10-01

Whole MD runs set up and checked: GROMACS and LAMMPS inputs with every line explained and checked against the programs themselves, and a PLUMED workbench. Also a curve fitter for any equation, one plot area and Python script on every plotting page, the Data Cleaner's pandas script, and continuous integration.

Added: a curve fitter for any equation

• Type the equation, not pick a model. y = A*exp(-t/tau) + y0, I(V) = I0*(exp(V/(n*Vt)) - 1) or z = a*x + b*y + c: the page reads it as it is typed, shows it typeset, points at the character a mistake is at, and sorts the names into variables (matched to data columns by name) and parameters. Seventeen presets fill in common models. Parameters take a starting value, bounds, or a fixed value; physical constants are offered at their CODATA values.
• The fit runs as you type, by Levenberg–Marquardt on the untransformed data, weighted when an uncertainty column is given, with starting values found by the fitter. Each parameter comes with its standard error and 95% interval, and the fit with R², adjusted R², RMSE, reduced χ², AIC, BIC and the parameter correlations. The fitted values match scipy.optimize.curve_fit to better than 1e-7, and the standard errors and confidence band to about 1e-6.
• The plot is styled on the page and saved as PDF, PNG or SVG: size, fonts, ticks (automatic, every step, a count, or listed with labels), log axes, limits, markers, error bars, the fit line, a confidence band, a residual panel, legend, grid, frame and colours. The preview follows matplotlib's own tick and layout rules, so it shows what the Python script will draw.
• The Python script (SciPy and matplotlib) is written as you work: the same fit and the same figure, with the data embedded or read from a CSV.
• One workspace. The equation sits in a bar that stays in view while the results scroll, with the models in a searchable menu beside it. Data, model and plot style are tabs next to the plot; Data and Model are marked done or in need of attention, and a file can be dropped anywhere on the fitter.
• New core modules: expression (parsing, derivatives and LaTeX, with no eval), nonlinear-fit, fit-python, plot-style and pdf (a PDF writer for SVG figures). The page no longer loads regression.js; curve-fitting keeps it for the npm package.

Added: GROMACS set-up with every .mdp setting explained

• Four steps, one zip: System (force field, what is simulated, temperature and pressure, thermostat and barostat with a plain guide to C-rescale, Parrinello-Rahman and Berendsen, time step and bonds with a hydrogen mass repartitioning switch for 4 fs), Stages (energy minimisation, NVT, NPT, production, annealing and pulling, each with its length, output intervals and restraints, and a timeline with step counts and output sizes), Index and Job. Every .mdp file is written in full, each line with a comment saying why and a link to the GROMACS manual, and one download holds submit.sh, the .mdp files, index.ndx when the groups were built, the PLUMED input when a stage runs with PLUMED, and a README.

• Checked by the real program: 122 combinations (five force fields, every stage, both barostats, with and without hydrogen mass repartitioning, and membrane, anisotropic, Nose-Hoover, annealing and pulling variants) pass grompp from GROMACS 2025 with no warnings beyond the ones the file names, and the page's own downloads ran in mdrun for an AMBER protein, a semi-isotropic membrane, GROMOS, 4 fs with hydrogen mass repartitioning, annealing, pulling and a coarse-grained system.

• Every option, searchable: all 325 options GROMACS 2025.1 reads, with defaults, units, allowed values, the manual's text and a link, settable in one file or all of them.

• Check a file: paste an existing .mdp to see whether grompp will stop and why (the checker agrees with grompp on 169 broken files), read it line by line, and open it in the builder.

• Index groups: drop a .gro or .pdb to get exactly the groups gmx make_ndx makes (the one it leaves nameless, for a blank residue name, is called Group, because GROMACS cannot read [ ] back), suggested groups (ligand, protein with ligand, pocket, membrane, solvent), the temperature-coupling groups for the system with the reason, and a group builder; the tc-grps are checked to cover every atom once.

• submit.sh skips stages that already finished and continues from the checkpoint, so a job stopped at its wall time can be submitted again; it adds -maxwarn only for a warning it names.

• New core modules gromacs-mdp (tables generated from the GROMACS source by tools/build-gromacs-mdp.mjs) and gromacs-ndx; npm run check:gromacs and npm run check:ndx compare them with an installed GROMACS.

• The GROMACS version you run: 2025, 2024, 2023 or 2022. The .mdp files leave out options that release does not read, and the checker applies its options, defaults and checks. Every preset passes grompp 2022.6, 2023.5, 2024.6 and 2025.0; Martini 3 and hydrogen mass repartitioning used to write options that arrived in 2024, so grompp 2023 stopped on them.

Added: one plot area and one Python script on every plotting page

• The same plot area everywhere: the Curve Fitter, PLUMED "Analyse a run", the XVG Visualizer, the Plot Digitizer, the Plot Builder, the Error Bar Generator, the Statistics Calculator and the Outlier Detector share one header with the figure's true size, one Style panel (size in inches, centimetres, millimetres or pixels; fonts and separate title and tick sizes; ticks; grid on either or both axes; markers, lines and colours per series; legend; a background: white, transparent with dark or light text and lines, dark, or your own colour; Publication, Presentation and Minimal presets), PDF, PNG and SVG at the true size, and a Python panel. Learn it once.
• The Python draws the same figure from your own files (or with the data in the script) with matplotlib, and repeats the page's numbers: tests run every page's script and compare matplotlib's axis limits with the preview's.
• PLUMED "Analyse a run": six figures (time series, histogram, reweighted free energy, the free-energy surface along one variable or as a heatmap with contours over two, its convergence, and hill heights), with a script that reads COLVAR and HILLS (restarts, cut-off last lines, several walkers) and matches the page to 1e-9.
• Error Bar Generator: significance brackets from Holm-corrected Welch tests, and a pairwise table. Statistics Calculator: the script reruns the test, checked against the page for all twelve. Outlier Detector: a chart of the values with the flagged ones and the rule's thresholds. Plot Builder: error bars, bands, several panels and PDF export. Plot Digitizer: the digitised series replotted as you click.
• Large data: a 400,000-point scatter redraws in under a second; the script reads the file rather than copying a million numbers into itself.
• New core modules: figure, figure-python, plot-builder, error-bars-figure, statistics-figure, outliers-figure and plumed-analysis-figures.

Added: the Data Cleaner keeps a recipe and writes the pandas script

• Every action is a step in a recipe you can undo and redo, switch off, edit, reorder where it is safe, and click to see the table at that point, with the rows each step leaves. Save the recipe and apply it to another file. One-click suggestions for what the page spots.
• The pandas script (a plain script or a clean(df) function) reads your file, repeats each step with a comment saying what it does, and writes a file byte-identical to the page's download, checked with pandas 2.1 and 3.0 on 800 random recipes. New core module data-cleaning-python.

Added: PLUMED workbench

• Keyword tables for PLUMED 2.9, 2.10 and 2.11, generated from PLUMED's own syntax.json, drive the input generator, its version notes and the module warnings.
• Check, explain and open an existing plumed.dat, parsed by PLUMED's own rules and tuned on the inputs of PLUMED's regression tests.
• Atoms from a structure file: groups, and centres and directions per molecule, numbered by position as PLUMED counts them.
• Analyse a run: COLVAR gives hill widths and grid bounds; HILLS gives the free-energy surface, matching plumed sum_hills to nine decimal places, and whether the bias has settled; the printed bias reweights other quantities. The same analysis comes as a Python script for files on a cluster.
• Run files that start and continue a job with one script, for GROMACS and LAMMPS: one run, MPI walkers, walkers as separate jobs, and umbrella windows. Checked by killing jobs at random for every method and layout.
• Functions, walls, setup lines (LOAD, INCLUDE, RESTART, FLUSH) and several output files in the generator.
• New core modules: plumed-syntax, plumed-parse, plumed-atoms, plumed-analysis and plumed-run; the MD Workflow Generator's PLUMED form now calls plumed, so the page and the package write the same input. npm run check:plumed hands every generated input to installed PLUMED builds; npm run build:plumed-syntax regenerates the tables.

Added: LAMMPS set-up with every input line explained

• The LAMMPS tab works like the GROMACS one: System (a data file, a lattice or a restart; force-field presets for CHARMM36, CHARMM22/27, AMBER, OPLS-AA, COMPASS/PCFF (class II), TIP3P, SPC/E and TIP4P/2005 water, EAM (one element, alloy and Finnis-Sinclair), Tersoff, Stillinger-Weber, ReaxFF and a Lennard-Jones fluid in reduced units, or your own lines; thermostat, barostat, time step, SHAKE or RATTLE, cut-offs), Stages (minimise, NVT, NPT, production, with restraints and output), Groups and Job. It writes in.system, in.settings and one input per stage chained by restart files, submit.sh and a README, with a note on every line that has a choice in it, and one zip holds them all. A job stopped by its wall time ends cleanly and carries on from its newest restart when submitted again; PLUMED attaches with fix plumed.
• A data file is read in the browser: counts, box, atom types with their elements and charges, the water model and its geometry, the types SHAKE holds, CMAP crossterms, net charge and density, and groups for water, ions and the solute. It agrees with LAMMPS 29 Aug 2024 on 1081 of 1081 files and broken copies (a build with 17 packages).
• Check an input you already have: every line explained in the input's own units, with a link to the LAMMPS manual, and the line LAMMPS would stop on. It agrees with LAMMPS 29 Aug 2024 on every example input and broken copy compared, quoting LAMMPS's own message: 806 examples and 479 broken copies with a build of 17 packages, 804 and 480 with one of 32 (ASPHERE, BODY, GPU, KIM, PLUMED, REPLICA and others).
• Every command and style, searchable: kind, package, accelerated variants, syntax, keywords and defaults, with a summary in plain words and the manual link. The tables hold facts only (LAMMPS is GPL-2.0).
• Run in LAMMPS: every preset ran through minimise, NVT, NPT and production on 2 MPI ranks, including stopped and continued runs with and without PLUMED.
• New core modules: lammps-reference, lammps-input, lammps-data and lammps-workflow; npm run check:lammps, check:lammps-data and check:lammps-workflow compare them with an installed LAMMPS (LMP_BIN), and npm run build:lammps-docs regenerates the reference tables from a LAMMPS source tree.

Added: continuous integration

• GitHub Actions runs the tests and the smoke test on Node 18, 22 and 24, and the site checks, on every push to main and every pull request. The README shows its status, the npm version and the Zenodo DOI. The smoke test now reaches 28 of the 44 modules.
• A second job installs Python and runs the tests that check the scripts with matplotlib, twice: on current releases, and on matplotlib 3.6.3 with Ubuntu 24.04's numpy, scipy and pandas.

Added: smaller pieces

• A background for every figure: white, transparent (with dark or light text and lines), dark, or your own colour, for any export format; it replaces the Dark preset.
• New core module zip: uncompressed zip archives of text files, used by the MD Workflow Generator's "Download all".
• The version on every page: the footer names the release and its date and links to this changelog; src/core/version.js exports the same.
• A Content-Security-Policy on every page, so the browser itself refuses anything from another host; inline scripts are allowed by hash only. The only exceptions are the DOI registration agencies DOI to BibTeX asks and the RCSB PDB the Structure Inspector fetches from.

Changed

• The XVG Visualizer's smoothing is a running mean over N points (the one numpy repeats), not a spline; a series on another scale goes in a lower panel instead of on a right-hand axis.
• The Plot Builder breaks a line at a blank or text cell, as matplotlib does, instead of joining across it; sub- and superscripts are written $_{…}$; figure sizes run from 1 to 30 inches.
• The Plot Digitizer's CSV gives every series a distinct name.
• The Error Bar Generator's own "Customize plot" controls are replaced by the shared Style panel.
• The MD workflow generator's water model defaults to TIP3P, and its topology header offers GROMOS 54A7 and Martini 3; the header is a tab of its own and is left out of the zip so it cannot overwrite a real topol.top.
• The PLUMED sample hills file is the two-variable run (assets/samples/plumed/HILLS).
• The footer carries a new credit line, written as .mdp lines.
• The Statistics Calculator says, with references, that choosing a test from a normality test on the same data is contested, and the suggestion links to how it is made.
• The Journal Abbreviator marks results from the ISO 4 rules apart from the built-in list, so you know which to check.
• The unit converter loads its own module only, not the whole library.
• robots.txt no longer hides the stylesheets and scripts from search engines, and the home page's canonical address is the site's root.
• [output] parsePDB reads the first model of a multi-model PDB, as GROMACS and PLUMED do, and says how many it skipped (modelCount, warnings); pass { models: 'all' } for the old behaviour. It keeps four-letter residue names (POPC, TIP3) and residue number 0, and parseGRO reads files written at any precision. Selections take the box (box, boxVectors) and measure within: to the nearest periodic image, matching gmx select.

Fixed: GROMACS and PLUMED, cross-checked against the programs

Twelve independent audits ran the tools against GROMACS 2025 (grompp, mdrun, make_ndx) and PLUMED 2.9, 2.10 and 2.11 and read their source; every finding was reproduced by at least two of three separate checks before it was fixed, each with a regression test.

• [output] PLUMED free energies. A 1D surface taken from a 2D HILLS file summed the hills along one variable instead of integrating the other out; a restarted run's hills that were read back were dropped as a copy; walkers' files joined with cat lost most of their hills; hills wider than half a periodic axis, ADAPTIVE hills and files without a kernel type were summed differently from plumed sum_hills; reweighting used one bias column and missed the walls, and ran at the global temperature rather than the method's own.
• [output] PLUMED run files. A 4 fs run lost half its hills (the kit's time step was never checked against the .tpr); LAMMPS MPI walkers were identical copies; umbrella windows could all sit at AT = 0; MOVINGRESTRAINT work lost the last interval before each checkpoint; OPES never wrote its state; GROMACS MPI walkers asked for one rank, which -multidir refuses.
• PLUMED inputs PLUMED rejected or misread, among them PLANE, DIHEDRAL_CORRELATION, ERMSD, PCARMSD and PUCKERING components, DRMSD without WHOLEMOLECULES from 2.10, NLIST settings lost under METAD, and several CV options; all 311 generated inputs now parse in 2.9, 2.10 and 2.11, and 303 run on a trajectory. The input checker no longer flags things PLUMED accepts, and now catches missing compulsory keywords, undefined references and parse failures it missed.
• GROMACS stages that would have failed: energy minimisation was given GPU flags mdrun rejects, as was -update gpu with Nose-Hoover or virtual sites and -pme gpu without PME; conjugate-gradient minimisation from a solvated start; thread-MPI and MPI builds got the wrong rank options; file names with spaces were not quoted.
• GROMACS set-up: GROMOS 54A7 constrains all bonds, as it was parametrised (it used h-bonds); Martini 3 pair lists and LINCS follow the force field's settings; pulling sets a central PBC atom per group and checks the group names; temperature-coupling groups are no longer split into baths too small to hold a temperature.
• The .mdp checker now stops where grompp stops for AWH, rotation and swap set-ups, a group named twice, sc-r-power other than 6, fourierspacing of zero, and more, and no longer stops where grompp does not; it reads Unicode spaces, bare CR line ends and hexadecimal numbers as grompp does. It agrees with grompp on 169 broken files.
• Index groups: element filters no longer read SOD, CLA, POT or CAL as sulphur, carbon or phosphorus; CHARMM nucleotides and glycans are not cosolvent; selections within a distance respect periodic boundaries.
• Page text that said something false about GROMACS or PLUMED is corrected: thermostat and barostat guidance, walker commands, the check command, wall offsets, units and version notes.
• [output] PLUMED units. Switching-function cut-offs (R_0 and D_0 of SMAC, FCCUBIC, ANGLES, CONTACTMAP and INPLANEDISTANCES) did not follow UNITS LENGTH, so R_0=0.5 silently meant 0.5 Å. Restraint, wall, moving-restraint and ABMD values (AT, KAPPA, SLOPE, OFFSET, TO) now follow a change of UNITS like the grid and SIGMA do; the same bias written in nm with kJ/mol and in the page's Å with kcal/mol gives the same energy in PLUMED 2.11 to 1e-12. Counts start on grids sized from their atom lists (the old 0–20 grid made PLUMED stop on larger counts), and changing the biased component resets its starting values.
• GROMACS pulling: with a structure loaded, the page works out the distance between pull groups as grompp does and says when grompp will stop because it exceeds 0.49 of the box, with the fix; it agrees with grompp within 0.0015 nm.
• GROMACS launch (the page and generateScript): the thread-MPI gmx gets -ntmpi 1 (it started a rank per visible core and oversubscribed shared nodes), and a job without GPUs gets -nb cpu (it took a GPU the node showed but the job did not ask for). On the page, threads are pinned with -pin auto rather than -pin on, which pinned from core 0 whatever the scheduler gave. A placeholder 1 Å CRYST1 line is read as no box.

Fixed: output

• [output] Force conversions. eV/Å was off by a factor of 10 and Hartree/Bohr by about 25: 1 eV/Å gave 96.5 kJ mol⁻¹ nm⁻¹ instead of 964.9, and 1 Eₕ/a₀ gave 3.35×10⁻⁹ N instead of 8.24×10⁻⁸. The other units were right. A test now derives every force unit from the energy and length tables.
• [output] Small-sample rank tests. Mann–Whitney and Wilcoxon used only the normal approximation, so complete separation of triplicates gave p = .0495 where the exact test cannot go below .10. They now follow SciPy's defaults: exact p for Mann–Whitney without ties when a group has 8 or fewer values, and for Wilcoxon without ties or zeros up to 50 pairs; an exact permutation test for Wilcoxon with ties or zeros up to 13 pairs; and the normal approximation otherwise, now with the tie correction to Wilcoxon's variance. The result says which was used, and the Python script asks SciPy for the same. Far-tail p-values (below about 1e-5) of the normal approximation are now accurate.
• [output] Figure scripts on current matplotlib. On matplotlib 3.8 and later a shared log x axis with data at or below 0 started lower than the preview showed (0.007 instead of 0.21 in one case), and on 3.8 to 3.10 error bars on a log axis set wrong limits. The scripts now give the same axis limits on every matplotlib from 3.6 to 3.11, and the preview follows 3.11's ticks and legend placement.
• ISO 4 abbreviations. The rules dropped section letters ("Journal of Physics A" → "J. Phys.") and misread hyphenated LTWA patterns ("Phase Transitions" → "-ph. Transitions"). They now agree with the built-in list on 169 of 184 titles (150 before), never produce a malformed word, and the 15 differences left are ones where the official abbreviation departs from ISO 4.
• [output] The Data Cleaner's rounding used toFixed, so 2.675 became 2.67; it now rounds the number as written (2.68), half away from zero or to even as you choose. Its filters no longer treat a missing cell as 0, the log of zero or a negative number is now missing rather than left unchanged, sorting is by value then text rather than by the browser's locale, a column with a blank name is no longer dropped, and TRUE or a date stays as written.
• [output] Generated PLUMED inputs that PLUMED rejected. Every metadynamics, PBMETAD, OPES and steered-MD input closed its block with ... METAD, which PLUMED refuses. About twenty variables printed or biased components that do not exist, such as cv1 for POSITION or .s for PATHMSD, which has .sss. All 311 generated inputs now parse in PLUMED 2.9, 2.10 and 2.11.
• [output] An OPES run continued from the kernels file and came back about 18 kJ/mol off even after a clean stop; it now restarts from the state file. With MPI walkers, PBMETAD restarted every walker but the first without its bias.
• Subscripts and superscripts in typeset formulas were drawn at full size on every page with KaTeX: the stylesheet was from a newer release than the script.

Fixed: pages that described the tools wrongly

• MD Workflow Generator, PLUMED notes: they said D_MAX gives COORDINATION linked cells. PLUMED has none for it, in 2.9, 2.10 or 2.11: D_MAX only brings the switch to zero, and the speed-up is NLIST. Linked cells belong to multicolvars such as COORDINATIONNUMBER and Q6, which the notes now name. Checked with plumed driver: D_MAX left a two-group COORDINATION as slow as before.
• The same notes offered PLUMED 2.9 or 2.10 "next to the CV list"; the menu has 2.9, 2.10 and 2.11 and sits in the System and setup panel. They named plumed --version, which PLUMED refuses, for plumed info --version, and described the file's header wrongly.

v0.2.1 — 2026-09-25

Every page's claims were checked against its code. This release fixes what did not match: bugs where the code was wrong, wording where the page was.

Fixed: output

• [output] BibTeX: a bare value failed the whole file. month = jul, journal = jcp or doi.org's month = Sept made parseBibtex return nothing or throw, so the deduplicator rejected typical exports and the output of DOI to BibTeX. @string and month macros are now expanded, # concatenation works, and @article(...) entries parse.
• [output] BibTeX Sanitizer and Deduplicator rewrote what they kept. The sanitizer turned bare macros and quoted values into braced text; the deduplicator's download rebuilt every entry, losing braces such as {NumPy}, and dropped @string, @preamble and @comment blocks. Unchanged entries are now written as they were.
• [output] DOI to BibTeX dropped every field outside its panel (address, school, howpublished), and fetched a DOI from an earlier batch again. Only unticked fields are removed now, and repeated DOIs are skipped.
• [output] MD workflow generator: the default pipeline stopped at NVT. grompp was given -t em.cpt, but energy minimisation writes no checkpoint. The topology wrote a second [ defaults ], which grompp rejects, and named opls-aa.ff for oplsaa.ff. mdrun now gets -ntomp, split between -ntmpi ranks, and the PLUMED temperature reaches methods whose own field is blank. CHARMM36 is marked as a separate download.
• [output] Structure inspector: residue ranges selected nothing. resi:1-50, mixed lists such as 1-5,10 and the selection builder's ranges now match; elem: accepts any case. The label limit counts the atoms being labelled, not the whole file, and can be raised whenever it applies.
• [output] XVG visualizer: the Python export failed on CSV and multi-set files. It now passes the delimiter, skips the header row and & separators, and reads only the rows the page plots. A record with an empty field, or with a different number of values from most rows, is skipped and reported instead of shifting the columns after it; a comma ending every line is allowed. CSV header names label the series.
• [output] Curve fitter: a row whose x or y value is empty is skipped and reported instead of taking y from the wrong column. The exported script no longer calls the power fit y-weighted, and the theory panel's logarithmic model matches the menu (y = a + b ln x).
• [output] Error bar generator: the CSV leaves SD, SEM, t, CI and CV% blank for a single-replicate group, as the table does, instead of writing 0.

Fixed: pages that described the tools wrongly

• Removed claims for features that do not exist: cell merging in Visual LaTeX Tables, WebGL rendering in the XVG visualizer, colour sweeping in the plot digitizer, syntax highlighting in the equation formatter, title-case repair in the BibTeX Sanitizer, and reshaping in the data cleaner.
• Corrected descriptions of what the tools do: delimiter detection in LaTeX Tables, the journal abbreviator's ISO 4 fallback, DOI lookup through doi.org (not Crossref alone), the coordinate manipulator's 5,000-atom preview and zero mass for unknown elements, the curve fitter's linearised models, the statistics calculator's validation references and APA example, and the Pomodoro timer's looped noise.
• The plot digitizer shows the cursor position in data units after calibration, which its checking tip already assumed.
• Privacy, home and 404 pages say which two features go online instead of "100% local", and no longer claim GDPR compliance by default.
• The 404 page, served at whatever path was missing, lost its styles and linked to more missing pages when that path was nested; its links now resolve from the site root.
• Structured data no longer shows raw HTML entities.

Changed: search and speed

• Plotly, 3Dmol and the other libraries load without holding up the first paint, which came after a 3.6 MB download on the plotting pages. The plot builder no longer shifts while Plotly loads.
• Titles and descriptions fit in search results; the share image is a JPEG with its size and alt text; the privacy page and the structure inspector have the preview tags and structured data the other pages have.

Removed

• home-sections-preview.html, a scratch page still served with invented testimonials.

v0.2.0 — 2026-09-25

Four fixes below change what a tool reports; they are marked [output]. The rest is layout, controls and wording. In stemkit-core the only change to an existing function is the chi-squared tail below; everything else there is new.

Fixed: numbers and figures

• [output] Error bar generator: statistics under the wrong headings. Since the move onto stemkit-core the table printed its cells in another order than its headings: the CI half-width under Median, the median under IQR, the IQR under CV, the CV under Min/Max, the minimum under t*, and t* not at all. The statistics were computed correctly; a value read off the table was another statistic. Every "Customize plot" control was also dead.
• [output] Coordinate manipulator: centre readout and rotation pivot. The "Geometric centre" readout was the bounding-box midpoint, so a centred structure read (0.104, -0.588, 0.002); it is now geometricCentre, the point centring uses. The "Rotate about" choice was never read, so every rotation used the default pivot; it is now honoured and named in the gmx editconf note. Rotations about the default pivot are unchanged.
• [output] chiSquaredUpperTail lost the far tail. It was 1 - lowRegGamma, so chi2.sf(100, 5) = 5.3e-20 came back as the 5e-324 floor and chi2.sf(60, 1) was 0.5% low. The upper incomplete gamma is now evaluated by its continued fraction where the subtraction fails; matches scipy.stats.chi2.sf to 1e-10 relative out to 8e-104. The df = 2 case the normality test uses already had an exact closed form and is unchanged; the new Kruskal-Wallis test was the first caller with more degrees of freedom.
• [output] Plot builder: sizes meant what they said only on screen. "Scale (DPI)" was a 1-4x multiplier and the PNG carried no resolution, so it placed two to four times too large; the matplotlib script sized the figure at px/100 in with fonts in points, making its text 1.39x larger than the preview's, and bbox_inches='tight' then cropped it to another size. A figure is now width x height in in, cm, mm or px at 72-600 dpi with text in points, and the preview, PNG (tagged with its dpi), SVG and script produce the same figure: seven configurations compared by overlay, matplotlib 3.6.3.

Added: statistics (core/statistics, 80 tests, suite 1208)

• descriptives, boxPlotStats, leastSquaresLine; oneSampleTTest, oneSampleWilcoxon, spearmanCorrelation (Bonett-Wright interval), kruskalWallis, welchAnova; post-hoc tukeyHSD, gamesHowell and dunnTest with adjustPValues (Holm by default) and qUpperTail; recommendTest. Each is validated against SciPy 1.11.4 or statsmodels, or against its formula where SciPy has none, with the call in the test.
• The statistics calculator offers twelve tests grouped by design, a per-group descriptive table (CSV, Markdown, LaTeX booktabs), a plot of the data (box plots with points and mean CI, paired lines, scatter with the fitted line; SVG and PNG), pairwise tables after three or more groups, and a "Which test?" suggestion it never applies on its own. Data is read as it is pasted.

Changed: structure inspector, plot builder, XVG visualizer

• Structure inspector: on-canvas view controls (hold-to-rotate pad and trackball, zoom rail, X/Y/Z views, fit, auto-spin), Find by residue, chain, atom, element or species with suggestions, stepping and "only matches", a species list with one-click hide water and ions, and the simulation box from GRO, PDB CRYST1, CIF and VASP drawn and labelled, toggled with B. F now fits; fullscreen moved to Shift+F. Selection style "Hide" now hides, and unbonded atoms (ions) show in stick and line styles.
• Plot builder: the preview redraws as settings change; there is no Render button. Size presets, axis limits, tick direction, frame, legend positions, line styles and markers; Okabe-Ito default colours.
• XVG visualizer: a series can go on a second y-axis, mirrored in the Python export with twinx(); the X column is no longer offered as a series.

Changed: the site

• One brand accent (Prussian blue) and a benzene logo replace indigo and seven per-tool colours; favicons and the social card are redrawn.
• One header on every page, with "Find a tool" (Ctrl/Cmd+K or /), "Next steps" under each tool, and a shared catalogue in js/site.js that npm run check:chrome holds the pages to. The 404 page's own list, which offered three tools that do not exist, is gone.
• Every tool opens on the same page head, numbered panels show live progress, and file loaders share one size. A saved or system dark theme now applies on every page (it was lost on 18 of 21 tools). Text contrast meets WCAG AA in both themes. Explainer formulas are all typeset by KaTeX from TeX, and the explainers share one type scale.
• The home page leads with search and a file-type lookup; invented testimonials and a 4.9-star rating in the structured data are removed.
• The plot digitizer's preview loaded Plotly from cdn.plot.ly, the last request any page made to another host; it uses the vendored copy.
• The privacy page says which two features contact another service (DOI lookup at doi.org, PDB fetch from RCSB).

Added: stemkit-core

• On npm as stemkit-core, renamed from @stemkit/core, which was never published. Under Node the package entry (core/node.js) registers the four bundled libraries, so a script no longer calls registerVendor. The package has no dependencies.
• core/scheduler: directive headers, environment-variable names, launch prefixes and submit commands for SLURM, PBS Professional / OpenPBS, LSF and Grid Engine, sharing core/slurm's validation. The SLURM path delegates to core/slurm unchanged. 51 tests; the module count is 17.

Changed: MD workflow generator

• Scheduler selector (SLURM, PBS Pro / OpenPBS, LSF, Grid Engine). The engine blocks follow the choice: launcher, job and array variables, working directory, log paths and the submit command. Anything site-specific (queue names, parallel environment, GPU resource names, memory units) is marked as such inside the generated script.
• The #SBATCH header now comes from core/slurm through js/script-generator-slurm.js, as that adapter's header always intended; output for the same inputs is byte-identical.
• Download buttons for submit.sh, topol.top and plumed.dat; settings persist between visits and can be exported and imported as JSON; a resource summary (cores, GPUs, core-hours, array tasks in flight) under the cluster form; syntax colouring of the generated script.

Changed: structure inspector

• Application shell: the viewer fills the space under the navigation, the side panel scrolls on its own, can be resized by dragging or with the arrow keys and collapsed, and remembers its width, state and last open tab.
• Viewer toolbar and keyboard shortcuts (R reset, S spin, M measure, H hydrogens, L labels, F fit, Shift+F fullscreen, ? shortcut sheet, Esc).
• Files can be dropped anywhere on the page; large files report read progress; a file that fails to parse names the file and the reason and returns to the upload zone. The PDB fetch distinguishes a missing entry from a network failure.
• Two bundled samples (assets/samples/), generated from textbook geometry so the demonstration works offline.
• PNG export offers a transparent background and shows the output size in pixels; very large exports warn rather than fail.

Changed: coordinate manipulator

• The workspace is built from the shared components: panels with plain headings instead of a per-card colour theme, labels above their controls, the system statistics as a key/value list with tabular figures, the box and editconf notes as callouts, and a viewer toolbar over the canvas.
• Empty state with two bundled samples and the accepted formats as chips; a file can be dropped anywhere on the page.
• Undo and redo are a toolbar with keyboard shortcuts; on a phone the viewer comes first, then the controls, then the output buffer, with no sideways scrolling at 375 px.
• Parsing, geometry, unit handling and the written files are untouched: the same operations on the same structure produce byte-identical PDB, GRO and XYZ output and the same gmx editconf command as the previous version.

Changed: every page

• All stylesheets, fonts and scripts are served from the repository. Inter is vendored (variable weight, SIL OFL 1.1) and is now applied as the body face, which it never was: the pages requested it from Google Fonts and rendered in the system font. The structure inspector loaded Tailwind from a CDN and threw tailwind is not defined offline. Favicons are relative.
• Design tokens and shared components in src/tailwind/input.css (docs/CSS.md). Visible keyboard focus, prefers-reduced-motion, no text below 11 px, reference tables and code blocks that scroll inside their box on phones rather than widening the page.
• The six pages that still carried inline <style> blocks use src/tools/<tool>.css. Ten per-tool stylesheets contained stray keyframe bodies outside any @keyframes rule, left by the earlier extraction; the shared stylesheet set the two- and three-column documentation grid and then reset it to one column, which every tool stylesheet patched with its own media query. Both fixed at the source.

v0.1.1 — 2026-07-25

Fixed: numerics

• [output] Skewness and kurtosis used the sample SD, not the population SD. Deflated skewness by ((n-1)/n)^(3/2), about 15% at n = 10. Both moments feed D'Agostino-Pearson, so every normality p-value was affected.

• [output] Adjusted G1 was fed into the D'Agostino-Pearson transform. The transforms are defined on the biased moment ratios sqrt(b1) = m3/m2^(3/2) and b2 = m4/m2^2. skewness() returns the sample-size-adjusted G1, and that was passed in. At n = 10 this gave K2 = 0.4521 against scipy's 0.3724, a 21% error. dagostinoNormality() now uses an internal unadjusted moment; skewness() still reports G1, which is the right statistic to publish and matches scipy.stats.skew(bias=False). K2 now matches scipy.stats.normaltest to 12 significant figures for 8 <= n <= 47.

  The test asserting this had its tolerance loosened to 3 decimal places and the residual attributed to "a marginally different kurtosis transform" in scipy. The reference value was scipy's, and correct. Tightened to 12 places with two regression tests at the n where the conventions diverge most.

• [output] Upper tails computed as 1 - cdf(x), which cancels to zero. ANOVA at F = 377.545, df 2 and 21, reported p = 0 instead of 3.46e-17. Now uses the complementary incomplete beta and gamma; an asymptotic expansion takes over past |z| ~ 8 where the vendored erfc underflows.

• [output] Element inference misassigned metals and prefixed hydrogens. Haem iron got fluorine's mass (18.998 Da, not 55.845). Selenomethionine selenium got sulfur's. 1HB-style hydrogens got none. Metalloprotein masses and centres of mass were wrong.

• [output] Unidentified atoms borrowed carbon's mass (12.011 default). Worst case: TIP4P/TIP5P water carries a massless charge site, so every water in a solvated system read 30.03 Da instead of 18.015. That is +67% per water, or 120 kDa of nonexistent mass across 10,000 waters. Three outcomes are now distinct: known element returns its weight; known virtual site (MW, LP, DUM, MCH3, MNH3 and relatives) returns zero, because that is its mass; anything else returns zero and is recorded so it can be named.

• [output] Reciprocal units ignored the inverse flag. wl_nm lost its marker and convert() had no branch for it, so 532 nm converted to 0.0000532 cm-1 instead of 18797 cm-1. convert() now inverts both entering and leaving the base unit.

• [output] kBT scales returned NaN. kt_kj, kt_kcal, kt_mev restored in convertTemperature() using exact SI 2019 constants. Cross-checked against convertKT() to 1e-12.

• [output] Atomic weights updated to CIAAW 2024. Four were wrong, not just rounded: Zr 91.224 to 91.222, Gd 157.25 to 157.249, Lu 174.97 to 174.96669, Ar 39.948 to 39.95. Argon's standard weight is the interval [39.792, 39.963]; 39.948 is an older single figure still in circulation. Values now carried at published precision with the source cited. Fourteen elements with variable isotopic composition use the abridged value; Tc uses the mass number of its longest-lived isotope; elements without a standard weight are omitted, so their atoms report as unidentified rather than get a fabricated mass.

• [output] Piconewton and newton factors truncated to 7 digits. 1.660539 where the CODATA-derived value is 1.66053906717. Every other factor in the table carries 12 to 13 digits.

Fixed: bugs

• latex-formatter never ran. Unclosed forEach in the theme-toggle handler threw SyntaxError: missing ) after argument list, so the script never parsed and the tool was dead in production. Duplicate theme toggle removed; the shared inline script already handles theming and the second handler cancelled it out.
• structure-inspector busy overlay could not be dismissed. #busyOverlay set display:flex by id; the script hid it by toggling .hidden. Id specificity (1,0,0) beats class (0,1,0), so the spinner stayed up and swallowed mouse input to the 3D canvas. Fixed with #busyOverlay.hidden { display:none }.
• DOI-to-BibTeX field filtering did nothing. Filter applied to a copy that was then discarded.
• Coordinate Manipulator accepted files but never loaded them. UI layer targeted a DOM that does not exist.

Fixed: docs for the fitting models

An earlier description of the vendored regression.js fits was wrong. The error had reached the in-app formula panel, the exported Python, and two tests that asserted it.

Model	Regresses	Weighted	Log-space bias
logarithmic	y on ln x	no	none
power	ln y on ln x	no	yes
exponential	ln y on x	by y	reduced

• Logarithmic is not a log-space fit. y = a + b*ln x is linear in its parameters, so regressing y on ln x is OLS on untransformed y. Transforming the predictor introduces no bias in the response. Verified numerically: no nearby parameter pair fits the original scale better. Removed from LINEARISED_MODELS.
• Power is unweighted, so its residuals really are in log space and small y values carry more influence than a direct fit gives them.
• Exponential weights by y, the standard log-transform bias correction, which keeps it closer to a direct non-linear fit. For exponential and power, parameter pairs fitting ~29% and ~36% better on the original scale do exist. Note that on an exactly exponential series the weighting is irrelevant: the log-space residuals are zero and every variant recovers ln 2. The divergence needs perturbed data. On [[1,2.0],[2,4.1],[3,8.2],[4,16.1],[5,32.3]] the weighted fit gives 0.690216 against 0.693167 unweighted.
• Untransformable points are not dropped. validateForModel refuses the dataset and names the requirement: exponential needs every y > 0, power every x > 0 and y > 0, logarithmic only x > 0.

Added

• ISO 4 journal abbreviation, two tiers. journals maps a whole title to its whole abbreviation: exact, but only for titles already entered. iso4 applies the ISSN LTWA word by word, so any title can be abbreviated. Tier 1 runs first; tier 2 takes what it does not recognise.
• Molecular weight shows its working. A how? link opens a per-element breakdown: atom count, weight used, contribution, total. Massless sites and unidentified atoms are listed separately rather than inferred from a total that does not match the atom count. Downloads as CSV with a header recording source file, method, weight set and exclusions. massBreakdown lives in the core and is tested for full atom accounting and agreement with centreOfMass.
• Equivalent gmx editconf command for coordinate-manipulator operations, verified against editconf.cpp. Note -box centres silently.
• Triclinic cell support: outline rendering, correct rotation and periodicity.
• Undo and redo in the coordinate manipulator.
• Grubbs' p-value in the outlier detector, alongside the flag count, with an explanation when the MAD fallback was used.
• Explicit WebGL check in the structure inspector. It needs WebGL, not a discrete GPU.
• listAllCategories() and isAffine(category) in units. listCategories() now returns multiplicative categories only, since affine units have no conversion factor or base unit. Marked-up unit forms live alongside plain ones as symbolHtml and nameHtml, keeping the core presentation-neutral.
• Full unit metadata: 11 categories, 70 units, including temperature. All 64 factors shared with the previous table verified identical before merging.

Changed

• Computation extracted from page scripts into src/core/ as DOM-free ES modules, tested independently and importable under Node. See CONTRIBUTING.md.
• Inline <style> blocks moved into src/tools/ and src/stemkit-docs.css. See docs/CSS.md.
• Plot digitizer: native alert() dialogs replaced; resolution note no longer covers the figure; Python export script actually emitted.
• PNG exports match on-screen sharpness.
• Measurement labels no longer stack on redraw.
• Custom abbreviation rules persist across sessions.

Docs

• vendor.js claimed a resolution failure "cannot be guarded with try/catch". True of a static import; await import() rejects catchably. Scoped, with the reason injection is still preferred (the bundles must be available synchronously).

Known issues

• Two audio files referenced by the Pomodoro timer are missing.
• iso4 has no tests/smoke.mjs case and is the lowest-covered domain module.

---

Condensed from a longer development log kept during the extraction of src/core/. Full narrative, including audit passes and per-tool verification, is in the git history.