What changed

Fixes and additions, newest first

Anything on this page is something a visitor could notice. Internal tidying is left out. This log starts on 25 August 2026 — earlier changes were never written down, and inventing them afterwards would defeat the purpose.

6 SEPTEMBER 2026

A one-word search could be swallowed by somebody's surname. Searching for spider returned nineteen papers by an author who goes by "The Spider" — titles like Molecular biology tools and Web words. The 106,383 papers about spiders were never looked for. The same happened to Einstein, which produced 478 papers by a cardiologist at Cedars-Sinai, and to wolf, which opened with a paper about the Higgs boson. The cause was blunt: if a query looked like a name and an author by that name existed, the search stopped there and the subject was never searched at all. Adding one letter escaped it — spiders returned 106,383 — which is not something anyone should have to work out. Reading a bare surname as an author is often right, so that has not been taken away. What has changed is that it no longer decides for you: both readings are shown with their counts and you pick. A search for spider now shows the author's nineteen papers and offers "Everything about spider · 106,383" beside them; take that and it offers the way back. Where there is genuinely nothing to choose between — Moritz Dechamps matches an author and nothing else — no choice appears, because a button that leads nowhere is worse than no button.
Six words for open access, and nobody could be expected to know them. The panel offered filters called diamond, gold, green, hybrid, bronze and closed, with nothing to say what they meant. Anyone who publishes knows gold and green; almost nobody outside the open-access world knows bronze, which is precisely the one worth understanding — free to read at the publisher, but with no licence, so you may read it and formally may not reuse it, and it can be closed again tomorrow. Hover any of the six now and it says in one sentence who paid and where the free copy sits. The colours say nothing about quality; they say who paid.

More to the point, nobody actually wants to filter on "hybrid". People want to know whether they may read something or whether they may use it. So above the colours there is now one plain question: reuse it, not just read it. That one does not go by colour at all — it goes by the licence, because the colour is only a rough stand-in for the right. Bronze is free to read with no rights whatsoever, and green can be almost anything. Turn it on and the two categories that carry no licence, bronze and closed, visibly switch off with the reason attached, instead of quietly handing you an empty page. Free to read stays where it was: the Open Access Only button in the row above.
A DOI or an ORCID in the search box. Pasting a DOI, a PubMed id, an arXiv number or an OpenAlex id has jumped straight to that paper for months, and the box never said so, which for most people means the feature did not exist. It says so now. ORCID is new: paste one and you get that person's work, with no guessing about which of five namesakes you meant. One caveat worth stating — a handful of author records in OpenAlex carry a placeholder ORCID that is not really theirs, so a made-up number can still land on someone. That is upstream data, not something this site can check.
Hide retracted publications. One checkbox. A retracted paper is a warning, not a result, and anyone who does not want to see them should be able to say so.
Two filters could contradict each other and leave you staring at nothing. Open Access Only above and the "closed" category below asked for opposite things; together they returned zero, with no explanation. "Closed" now switches off while that button is on, and says why.
An advanced search, under the ordinary one. The search box has not changed. Below it a panel now opens on demand: where to search (title, title and abstract, abstract, or the full text), an author, a journal or conference, and a type — article, preprint, dataset, dissertation and the rest. Author and journal suggest as you type, and picking from the list hands the search an exact identifier instead of a piece of text, which is the difference between "papers by this Einstein" and "papers by something spelled Einstein". Author and subject work together, so you can ask for one researcher's work on one topic; a journal without a subject is a valid question too, and now returns an answer instead of an error.

Where to search is the narrow and the wide end of the same dial: on a search for drosophila operant learning, titles alone give 40 results, titles and abstracts — the default — give 152, and the full text gives 2,170, most of it papers that merely mention the words somewhere. Years, sorting and the open-access switch stay where they were, in the row above; putting them in two places would only raise the question of which one counts. Whatever you fill in shows as a number on the Advanced button even when the panel is closed, and as a line above the results, so a filter can never quietly shape what you see. Close the tab or hit reload and your setup is still there — it lives in the address now, which also means you can bookmark a search.

3 SEPTEMBER 2026

Most of what a search returned was not about your subject. A search for drosophila operant learning reported 2,159 papers. It should have reported 152. The query was being run against the full text of every paper, so a study that merely has drosophila somewhere in its reference list and the words operant and learning loose in its body text came back as a match — bumblebees, mice, astrocytes. Roughly nine out of ten results were noise, and a summary of that list described a field that was not the one you asked about. Searching now looks at titles and abstracts. Nothing was removed: the wider search still exists for the cases where a paper covers a subject without naming it up front, but it is now a choice you make rather than something that happens to you.
A panel that tells you who and what is in the results. A list of two thousand papers is not an answer, and nobody scrolls to page a hundred. Next to the results there is now a panel with counts: the authors who appear most often, the main subjects, the years, and how much of it is open access. On that drosophila operant learning search, one researcher accounts for 49 of the 152 papers and two of his co-authors for another 37 — you can see the shape of a field before reading a single title. Each line is also a filter: click an author and you keep your search but see only their work, click again to undo it. The counts are for the whole search, not the page you are looking at.
The AI could add words to your search and you could not stop it. When it is switched on, an assistant rewrites your query before it is sent — a search for drosophila operant learning could go out as drosophila operant learning genetics. The switch for it sat in a part of the settings panel that is hidden whenever you have chosen to run a model on your own machine, so anyone using a local model saw the rewriting happen with no way to turn it off, and the setting was not saved for them either. That mattered more from today on: now that a search matches on titles and abstracts and every word has to fit, one added word can take a result set from 152 papers down to three. The switch has moved to where it is always visible, and when the AI does change your query the note above the results now carries two buttons: search without the added words, or turn it off for good. Two smaller things in the same corner: the switch used to do nothing until you also pressed Save, which is not how a switch should behave — it now applies the moment you click it. And the text beside it claimed only single-word searches were expanded, which was untrue: a three-word search was being rewritten too. It now says what actually happens.
One step of the tour described the wrong button. Step five lights up the microscope in the bar on the left and said it was for comparing papers, following citation chains and reading full text — things that live on the result cards, not behind that button. What it actually opens is a list of instruments: tick a microscope, a 3D printer, a Raspberry Pi or nothing but a laptop, and searches lean towards work you could realistically build on. Measured while fixing the text: on a search for drosophila operant learning with a microscope ticked, 152 results became 141 — a nudge, not a filter, and the tour now says so. Corrected in all seven languages.
Ticking three instruments only counted two of them. While correcting that tour text I measured what the equipment list actually does, and found it was quietly ignoring part of your choice. It gathered the keywords of everything you ticked into one pile and kept the first six — so with three instruments selected, the third contributed nothing at all, and which one fell off was decided by the order of the list rather than by you. Ticking the same two in a different order changed nine of the twenty results on screen. Each instrument now contributes its own terms in turn, so every tick counts and the order is irrelevant: the same two, either way round, now give the same twenty results in the same order. A few of the keyword sets also led with the field rather than the instrument — a GPU was described as machine learning deep neural network, which is a subject, not a piece of hardware — and those now start with the equipment. Measured across six searches: five improved, one unchanged, none worse.
The summary of a whole result set could describe papers that do not exist. Not every record in the index carries an abstract. When one did not, the line was simply left out of what the assistant was given — so it received a bare title and an instruction to produce three to five bullet points about it, and it filled the gap. Among real summaries there were sentences like "based on a hypothetical future paper" and "potential focus on", set in exactly the same type as the rest. For readers checking sources that is the worst failure this tool could have: one invented line discredits everything beside it. A missing abstract is now stated explicitly, the instruction forbids guessing, hedging or hypotheticals outright, and such an item gets a single line saying only the title, authors, journal and year are known — visibly different from a real summary. Tested against an actual language model on a paper with no abstract: the old wording invented content on every run, the new one did not.
A retired AI model no longer takes the AI features down with it. Model names were written into the server, and when a provider retires one the error said, in as many words, that a maintainer had to edit the source. That is not a state a public tool should be able to reach. The names now come from a small registry that can be corrected in one place for every installation, and if a model turns out to be gone anyway, the server asks the provider which models it does have and continues with the current equivalent, using your own key, at no cost to you. Alongside that the deeper analysis tier moved up a generation, which it had been overdue for.
The panel beside the results scrolled away. It was meant to stay in view while you read down a long list; instead it drifted off the top of the screen, and because it is tall it looked as though it was sitting stuck in the middle of the page. The panel was told to stay put relative to the frame around the results, but that frame never scrolls — the page does — so it was holding on to something standing still. It now holds on to the window and stays just below the header wherever you are in the results, at any header height. On a narrow screen it still sits above the list and scrolls with it, which is the right behaviour there. A second cause turned up right after: if you had ever used ctrl+scroll to enlarge the page, that setting is remembered, and the way it was applied made the panel hold on to the enlarged layer instead of the window — so for anyone with a saved zoom the panel still drifted off, and reloading did not help because the zoom outlives the reload. Enlarging now works through a mechanism that leaves the panel anchored to the window, and the panel keeps its place at every zoom level instead of running off the bottom of the screen.

2 SEPTEMBER 2026

A tour that points at things. There was a tutorial before, but it was six slides of praise that pointed at nothing and existed only in English. It has been replaced by a walk through the interface: each step lights up the actual button it is talking about and says what that button does. Eleven steps, and it appears by itself the first time you arrive. Available in English, Dutch, German, French, Spanish, Portuguese and Turkish.
What an API key is, in plain words. Three of those steps are about the AI side, because nothing here explained it. They say what an API key actually is, that you create one at OpenAI or Anthropic and pay them for what you use, and that your key is stored encrypted in your own browser and passes through this server only to reach your provider. The step after it explains the other option: point OSA at a model running on your own machine — Ollama, LM Studio or vLLM — where there is no key, no cost, and nothing leaves your computer.
The tutorial button was invisible. It sat in a row of the header that is set not to display, so no visitor could ever have opened it. That was true for as long as the button has existed. The tour now starts from a ? at the bottom of the bar on the left, next to the rest of the buttons.

26 AUGUST 2026

The footer no longer claims something untrue. It said "no PDFs" — but a full-paper analysis does fetch the freely available copy of a paper and read its text on the server. That is not a small distinction on a page aimed at people who care about exactly this, so the wording now describes what actually happens. What was always true stays true: no paywall is ever circumvented, and a file you upload yourself is never kept.
The same paper no longer shows up twice. The index underneath sometimes holds two records for one work — a preprint beside the published version, or simply a duplicate entry, occasionally under the very same DOI. Both were passed straight through, so roughly one slot in sixteen on a results page was a repeat; on some searches it was four of twenty. Repeats are now merged, keeping whichever record carries the most detail. A title and the same title with its subtitle attached are recognised as one work.
Full-paper analysis no longer settles for an abstract. It used to accept the first page that returned any text, and a publisher's landing page clears that bar easily — so it would analyse an abstract while calling it a full-paper reading. It now weighs every available source, recognises a landing page for what it is, and says plainly when the article itself cannot be reached. Measured on thirty papers: genuine full text went from 12 to 15, and nine false successes went to none.
The analysis button appears on the right papers. Which papers offered it was decided by a hand-kept list of publisher domains, and that list was wrong in both directions — hiding papers that worked, offering papers that could not. It now follows open access itself.
A failed analysis costs nothing. It always did — the AI is never called if the text cannot be fetched — but the button did not say so. Now it does.

25 AUGUST 2026

Paste an identifier and go straight to the paper. A DOI, an OpenAlex ID, an arXiv number or a PubMed ID used to be treated as ordinary words, which returned hundreds of loosely related papers and looked like it had worked. All of them now resolve directly.
Type a name and get that person's papers. Author names went into the same general text search, which reads full text as well — so a name mostly returned papers citing that person rather than papers by them. Namesakes are now offered alongside, so a wrong guess costs one click.
Questions no longer break the search. Any query containing a question mark returned a server error, because the index underneath reads "?" as a wildcard. "What causes Alzheimer's disease?" now simply works.
A typo no longer empties the page. If nothing matches every word, the search loosens instead of returning nothing.
An about page. What this is, where it stands, and what it is not.