Eikos changelog

October 8, 2026 · 1.0.19

Causality with large tables. With 20,000 rows, the cross-validated cutoff window went from tens of seconds to tenths, and moving a control no longer recomputes every section.
Research opens instantly. The power simulation runs when you reach that section and its result is remembered.
Analyze loads a large CSV in half the time: the missing-data tests are computed when you reach their block.
Windows stays on 1.0.18 for now.

October 1, 2026

Windows reaches 1.0.18. The Windows installer was still on 1.0.14; it now ships the same version as the web and macOS.

September 28, 2026 · evening · 1.0.18

The Solver, in English. Steps, templates and answers for almost 700 exercises now read in English.
English statements. Percentiles, means, angle conversion, recurrences, dice and coins give the same answer as in Spanish.
562 solution steps now show a proper title instead of an internal identifier.
Machine-translation slips fixed (mean, null space, ratios).

September 28, 2026 · 1.0.17

New Practice section. Eight question types graded by the engine, which names your mistake.
Lab. “Break the data”, “Inside a formula”, the lineup and the prior before the data.
Analyze. Mediation, Bayes factors, reliability, robust regression, state space and copulas, verified against R and Python.
Interface reviewed on phone, tablet and desktop.

September 26, 2026 · 1.0.16

Try the AI assistant free. Your account includes calls, no key or card needed.
The assistant explains why it cannot answer and how to continue.
English interface and guides thoroughly reviewed.

September 22, 2026 · 1.0.15

Solver. Success probabilities across trials, eigenvalues and curve intersections.
Grapher. R² is no longer computed for a constant outcome, and RMSE uses the correct denominator.
More checks for mathematical reading and calculation.

September 21, 2026 · afternoon

Coming back from payment and the app not noticing. Checkout opens outside Eikos —in the system browser— so the purchase happens where the app cannot see it. And nothing was checked on return: the only moment the subscription was re-read carried a ten-minute pause. Someone who had just paid could open the manuscript and be asked to pay again. It is now checked on returning to the foreground, with no wait if you are coming back from checkout.
A fault on our side took your subscription away. If the database did not answer, the server did not distinguish “has no subscription” from “could not check”, and replied with the first. An infrastructure outage turned into a paywall, and only for people who pay. It now says it could not check, and the app keeps the access it already had.
What you pay for on web or desktop works on your phone. It always did by design —the subscription follows your account, not your device— but no test required it. Now all four platforms do.

September 21, 2026

Three test suites were failing and nobody ran them. They lived in the browser-test folder, which only picks up files ending in .spec.mjs, so neither that runner nor npm test —which only looks in its own folder— ever executed them. That was 348 invisible checks while the count read 210 suites passing. The guard that stops a suite from being left unrun now looks there too.
The variance-ratio interval and Dunnett’s test were answered with another exercise’s numbers. The workshop bank comes first in the routing and two of its templates carry hard-coded figures: one returned 1.526 —which is (1.05/0.85)²— to anyone typing n₁=16, s₁=4.2, n₂=21, s₂=2.8; the other returned 14 for a Dunnett test, because it matched the “14” inside “14.8”. The engine that does compute them existed and never got to run. A template that does not use the statement’s numbers now steps aside, and that is checked by measuring the answer, not by asking the template.
Orthogonal contrasts returned zero. The series reader did not accept brackets, so with “n = [10, 10, 10], means = [15.2, 18.4, 22.1]” the sample sizes were taken as the means: ψ = 0 and ψ = 0, with “the three groups are aligned” printed beside them. They are 6.9 and 0.5. The same fault left Cook’s distance, the hat matrix, leverage and four others asking for data the statement already carried.
Stokes on a curved surface now draws. Only the flat-disc branch carried the figure, so writing out the surface —the better statement— gave a fuller answer and a poorer screen.

September 19, 2026

Account control on every platform. Web, desktop, Android and iOS now expose an always-visible Account entry with permanent self-service deletion.
Safer deletion. Active renewals are cancelled before identity and server data are removed; Apple authorization is revoked where applicable.
Independent web request. Account deletion can also be requested without installing or opening Eikos.
Vector Calculus expanded. Content, exercises and cross-platform presentation were updated.

September 17, 2026

The Pro price advertised on the Somatryx site was out of date. It said COP 269,900 and US$ 119.99 a year; the amount in force since 1 September is COP 299,900 and US$ 149.99, which is what this page has always said and what the payment gateway holds. Also corrected: that site claimed Pro “will only raise the included-AI allowance” and that no feature would depend on it, when since 16 September Pro is exactly what locks the manuscript and the presentation on web and desktop.
The pricing card contradicted itself. It attributed the peso amounts to “the Android and iOS apps” three lines before stating that Pro is not sold on Android or iOS. Pesos are the Colombian base market, not a platform.
The English version did not say who charges. The Spanish card warns that on web and desktop the charge is in US dollars, with a seven-day trial, and who processes it; the English one skipped that paragraph entirely.

September 16, 2026

AI in Free is yours. Without a subscription, the tutor and writing work with your own Gemini, Claude, OpenAI or DeepSeek key, or a local model, uncapped by Eikos and never through our server. Included AI becomes part of Eikos Pro, together with the manuscript and the presentation. The Founder programme is withdrawn: Eikos has no launch promotions.
Settings, finally in one place. Your account, theme, density, language, how much explanation you want, performance, privacy, help and the version, on one screen, as in Horme and Epistemis. Until today the account lived inside the AI assistant and you had to know it was there.
Signing in is visible from the first screen. The header has an always-visible button that says “Sign in” or “Account”, and it takes you straight to your account. Deleting the account is in plain view in Settings, not folded inside another section.
A welcome for people new to Eikos. Five skippable steps: what makes Eikos different, where you want to start, how much explanation you prefer, your account if you want one, and how to get around on your device. The tips change with where you use it: the bottom bar and the Files app on an iPhone, the Back gesture on Android, ⌘K and the keychain on a Mac, and installing it from the browser.

September 13, 2026

Sign in with Google or with Apple. An account is only needed for the included AI, and until today the only route was five steps across two apps: invent a password, leave to find an email, tap its link and come back. With Google or Apple it is two taps, and the address arrives already verified. Email and password stay for anyone who would rather not involve a third party, and Eikos asks them only for your address: not your name, not your photo, not your contacts. Neither of them learns what you analyze.
The account form, rebuilt. Pressing Enter now submits —it did nothing before—, the password has a show button and a strength meter, you pick between creating an account and signing in instead of facing both forms at once, the buttons lock while the request is in flight (two clicks used to mean two sign-ups), and the terms and privacy policy sit where you accept them.
An invisible space kept people out who had typed their address correctly. Phone keyboards add a space when autocorrecting, and that space travelled to the server as-is, which answered “that email is not valid”. It is now trimmed before sending.
Three sentences on that screen stayed in Spanish with the interface in English. They had the email address glued inside them, so the dictionary could never match them. Nineteen more strings on the same screen were fixed alongside.

September 10, 2026

Mathematics is now typeset like a book. Six measured differences between what the browser does and what TeX’s spacing table says: inside a script, relation and operator spacing is not inserted —0.3888 em too much in the k=1 under every summation, and the corpus has 127,081 such scripts—; the ellipsis in a list carried relation spacing in all 15,823 of its appearances; so did the absolute-value bar, the period ending an equation and the slash; and sin x came out without the thin space that separates a function name from its argument.
Theorems look like theorems. The material was written as a treatise —theorem, definition, proof, example— and the screen flattened it into running prose: 8,057 environments, about fifteen per guide. They are now typeset with the convention of any mathematics book: statement in italics, “Proof.” run into the first line and ∎ at the end.
One ink and one typeface. Display equations were printed darker than the paragraph introducing them, and exercises read in the typeface of the controls while their mathematics came out in serif. Guide headings had the letterform of a button.
Formulas are set in Computer Modern, and look the same on all five platforms. They used to be rendered with whatever math font the system had —a different one on macOS, Windows and Linux, and none at all on Android and iOS, where a parenthesis never stretched: 18.7 px next to a 62.8 px matrix—. Latin Modern, the TeX face, now travels with the app, and prose moves to sans with it: two serifs 1.3 % apart in cap height is the worst possible distance between two faces, and the eye keeps trying to merge them.
Theorems look like theorems. Statement on a panel, its own label, and ∎ at the end of every proof.
And 3,279 raw macros in the English library. The English tree still carried vector-calculus shorthands the Spanish corpus no longer uses, and they showed up underlined in the middle of the formula. Fifteen remain, all residue from the source material.

September 8, 2026

The solver was audited against outside references. It was entered through the same door the screen uses, with 443 hand-written problems, and every answer was checked against a value computed outside Eikos with scipy and statsmodels. That produced 44 defects, all fixed; 35 of them returned a correct calculation… of a different problem: the second derivative returned the first, and an expansion “about 1” was done at 0.
Twelve mathematical branches with their own figure. Analytic geometry, optimisation, numerical analysis, graph theory, multivariable and vector calculus, series and modular arithmetic are no longer solved without a drawing: each has its own native SVG rendering, with no external libraries and under the same content security policy as the rest of the application.
Analyze no longer produces impossible figures. With zero variance or identical values, the scales divided by zero and the SVG came out with non-finite coordinates; in tables with zero rows or cells, chi-square, Fisher, Cramér’s V and the risk difference returned meaningless values instead of declaring that they cannot be computed.

September 6, 2026

“Solve” was reorganised by measuring the screen, not by reading the code. One single list was doing three different jobs —a catalogue of shortcuts, a one-of-many control and an action bar— so six destinations read wrong: the selected depth was painted like the button that runs the calculation, and Back and Next appeared one below the other.
Two layout defects closed. A tab group that did not fit on one line wrapped below its label, so the second row read as an unnamed group; and the screen scrolled 14 pixels horizontally at 360 px because one long tab would not shrink.
An inference keyboard. The contextual keyboard had calculus, linear algebra and probability banks, and none for inference: anyone typing “test whether the mean is 72” found no key for H₀, x̄, α or chi-square.
The numbers in the statement are now controls on the mathematical route too. Moving a coefficient and watching the solution move only worked when the expression matched a recognised shape; it now works on an ordinary exercise, and bifurcations are detected by measurement rather than declared.

September 1, 2026

The included AI requires an account. Eikos ships its own assistance allowance, and using it requires an account: that is what makes the usage countable. With AI turned off or running locally, the application still works in full and data does not leave the device.
The whole Gemini route was fixed. The client asked for reasoning effort using the previous generation’s scale; in current models that field no longer exists and the request was rejected outright, both with your own key and through the proxy. The suites were green because none of them exercised a current-generation model.
The verification figure is read, not copied. The test run writes down the count of its last complete pass and this page publishes it from there, in both languages; before, it was a hand-transcribed number that aged without anyone noticing.

August 25, 2026

Engines now reach the interface: census analysis, spatial regression, difference-in-differences, interrupted time series, synthetic controls, hierarchical models, record linkage and risk scores now have visible product workflows.
Operational refusals: all 18 safeguards that stop unsupported calculations now have a real entry point; no refusal is inert and no public function remains without both a route and a test.
Survival and clinical data: Gray's test for competing risks, ICD-10 indexes and DOCX export are verified local capabilities rather than external blockers.
Accessible figures: JavaScript-generated colour ramps are tested against their real dark-mode surface and under colour-vision simulations.
Signed and notarised desktop build: the macOS build is signed with a Developer ID and passes Apple's notary service; Gatekeeper accepts it. Public installers remain undistributed until the update channel is decided.
Real progress on large files: opening a table larger than memory reports its two passes —hashing and profiling— separately, instead of a single bar with an invented split.
Complete verification: 106 suites, 48,908 checks, 132 journeys across three browsers, 27 core and desktop tests, and 569 site URLs before publishing.

August 24, 2026

Causal and population inference: new workflows for regression discontinuity, doubly robust methods, complex surveys, census data and stepped-wedge designs.
Defensible formatting: tables, numbers, intervals and exports are independently checked against R and Python.
Reproducible research: notebooks, reports and DOCX files retain methods, decisions and replication code.

August 22, 2026

Public library: each page contains an introduction and a LaTeX-rendered mathematical sample; full content remains inside Eikos.
Instrumental variables: causal diagrams lead into 2SLS, first-stage strength and Anderson–Rubin inference.
Missing data: patterns, complete-case cost, multiple imputation, Rubin pooling and declared-δ MNAR sensitivity.
530 guides and 2925 exercises: counted from the live corpus.

Status

Preview under validation. The Windows and macOS downloads show each installer’s actual version and automatic-update status on the Downloads page.

Last full verification: 221 suites and 100,959 checks passing (2026-10-08). The figure is written by the test run itself; it is not transcribed by hand.