This is the Elephants' graveyard where I put more or less big projects which have been tackled and, sometimes, even fulfilled beyond expectations, mainly based on Things to do with Claude on the Week-End.
This is the first one I completed, on 25 July (2026). That was the goal:
Find a way to map TeX-to-Unicode on the go, without passing by https://www.unicodeit.net/, e.g., instead of Ctrl+Shift+U opening a u and inviting a unicode character (for em-dash), do something like Ctrl+Shift+Alt Gr+U invites a TeX command that parses into unicode, such as g^{(2)}(\tau) turning into g⁽²⁾(τ).
and it resulted in the mighty TeX2unicode. This was better than I intended, so a big success for a first try, that went beyond expectations, although those were very high. The wonders of AI and the power of the tool!
On the next day, 26 July (2026), this one went down too:
Prettify the listing of my glosses in my diary, so that instead of, e.g., for July (2026):
* [[July 22|22]]: Gloss on {{Malygin85b}} * [[July 22|22]]: Gloss on {{Howell04a}} * [[July 22|22]]: Gloss on {{Edgar12a}} * [[July 23|23]]: Gloss on {{Moreau12a}} * [[July 23|23]]: Gloss on {{Lantz15a}}which renders as
- 22: Gloss on Spatiotemporal grouping of photons in spontaneous parametric scattering of light. A. A. Malygin, A. N. Penin and A. V. Sergienko in Sov. Phys. Dokl. 30:227 (1985).
- 22: Gloss on Realization of the Einstein-Podolsky-Rosen paradox using momentum- and position-entangled photons from spontaneous parametric down conversion. J. C. Howell, R. S. Bennink, S. J. Bentley and R. W. Boyd in Phys. Rev. Lett. 92:210403 (2004).
- 22: Gloss on Imaging high-dimensional spatial entanglement with a camera. M. P. Edgar, D. S. Tasca, F. Izdebski, R. E. Warburton, J. Leach, M. Agnew, G. S. Buller, R. W. Boyd and M. J. Padgett in Nature Comm. 3:984 (2012).
- 23: Gloss on Realization of the purely spatial Einstein-Podolsky-Rosen paradox in full-field images of spontaneous parametric down-conversion. P.-A. Moreau, J. Mougin-Sisini, F. Devaux and E. Lantz in Phys. Rev. A 86:010101 (2012).
- 23: Gloss on Einstein-Podolsky-Rosen paradox in single pairs of images. E. Lantz, S. Denis, P.-A. Moreau and F. Devaux in Opt. Express 23:26472 (2015).
Have something more styleish and tidy, e.g., with a manicule and some grouping:
- 22: Gloss on Spatiotemporal grouping of photons in spontaneous parametric scattering of light. A. A. Malygin, A. N. Penin and A. V. Sergienko in Sov. Phys. Dokl. 30:227 (1985).
- ☞ Realization of the Einstein-Podolsky-Rosen paradox using momentum- and position-entangled photons from spontaneous parametric down conversion. J. C. Howell, R. S. Bennink, S. J. Bentley and R. W. Boyd in Phys. Rev. Lett. 92:210403 (2004).
- ☞ Imaging high-dimensional spatial entanglement with a camera. M. P. Edgar, D. S. Tasca, F. Izdebski, R. E. Warburton, J. Leach, M. Agnew, G. S. Buller, R. W. Boyd and M. J. Padgett in Nature Comm. 3:984 (2012).
- 23: Gloss on Realization of the purely spatial Einstein-Podolsky-Rosen paradox in full-field images of spontaneous parametric down-conversion. P.-A. Moreau, J. Mougin-Sisini, F. Devaux and E. Lantz in Phys. Rev. A 86:010101 (2012).
Indentation should be neater, which is also part of the task.
I use Palmyrene numerals 𐡻, 𐡺, 𐡹 (they normally read from right to left) which are beautiful glyphs for Palmyrene Aramaic, the language of Palmyra (in present-day Syria) from roughly the 1st century BCE to the 3rd century CE:
I met with a network failure, so some files could not be uploaded:
laussy@azag:~$ !llw
llw2lw --doit
Running in write mode
Enter the start datetime (YYYY-MM-DD [HH:MM:SS], default: 2026-07-23 22:59:26):
== Pages to push ==
2026
July (2026)
Oemrawsingh02a
Template:Mostafavi25a
Things to do with Claude on the week-end
== Files to push ==
File:Screenshot 20260723 231737.png
File:Screenshot 20260723 235846.png
File:Screenshot 20260724 000205.png
File:Screenshot 20260724 001310.png
File:Screenshot 20260724 002100.png
Uploaded: File:Screenshot 20260723 231737.png
Uploaded: File:Screenshot 20260723 235846.png
== Failures ==
File:Screenshot 20260724 000205.png: Invalid CSRF token.
File:Screenshot 20260724 001310.png: Invalid CSRF token.
File:Screenshot 20260724 002100.png: Invalid CSRF token.
2026: Invalid CSRF token.
July (2026): Invalid CSRF token.
Oemrawsingh02a: Invalid CSRF token.
Template:Mostafavi25a: Invalid CSRF token.
Things to do with Claude on the week-end: Invalid CSRF token.
Log file updated: /home/laussy/comp/lw/llw2lw-log.txt
When this is the case, on next attempt, they should be uploaded again, as opposed to filed as "backed-up" or "synced". Importantly, this must not concern files taken from syncy which also generate an error but of duplication.
Claude took care of this and it appears now to be robust to such failures.
There was some nice KIM service-menu way to get essential shortcuts at the tip of your ring finger but this became nightmarish to install and maintain, so I essentially gave it up. Claude could probably bring this back in an efficient/portable way, to do things like:
etc.
It did, and the list of things he's servicing is in LaussyMenu.
I have some (minor, yet still better if fixed) bibTeX processing issues, such as: Demonstration of nondeterministic quantum logic operations using linear optical elements. T. B. Pittman, B. C. Jacobs and J. D. Franson in Phys. Rev. Lett. 88:257902 (2002). dropped the initials (should be T. B. Pittman, B. C. Jacobs, J. D. Franson). The "Title Case" vs "sentence-case" BibTeX convention should be consistent.
Claude did it on 6 October (2026). The initials were lost because the given name was split on hyphens only, so CrossRef's T. B. came out as T.. doi2bib became v°2.0.0: every initial kept, compound surnames braced, titles in sentence case with the names and acronyms braced (the style lowercases everything else, so an unbraced QED printed as "qed"), and MathML turned into TeX. The same went over the whole of sci.bib: 1555 titles brought to sentence case, 424 entries given back their initials from CrossRef, 20 broken doi links mended, an entry that carried another paper's title (Single photon emission based on coherent state preparation. P. Ester, L. Lackmann, S. Michaelis de Vasconcellos, M. C. Hübner, A. Zrenner and M. Bichler in Appl. Phys. Lett. 91:111110 (2007).), a bare % that cut every bibliography short (Detecting telecom single photons with $99.5^{+0.5}_{-2.07}$% system detection efficiency and high time resolution. J. Chang, J. W. N. Los, J. O. Tenorio-Pearl, N. Noordzij, R. Gourgues, A. Guardiani, J. R. Zichi, S. F. Pereira, H. P. Urbach, V. Zwiller, S. N. Dorenbos and I. Esmaeil Zadeh in APL Photon. 6:036114 (2021).) and some ten others that stopped LaTeX—the whole file now compiles without a single error. The templates were updated silently, and the new initials got redirects to their author pages (P.-A. Moreau, M. G. Bawendi, D. G. Thomas).
Instead of something like
on a pdf file page, implement a previewer and other useful servives (what-links-here, etc.)
Wanted since 0N2pV (af).
Claude did it on 24 October (2026), as the gadget filepage, for images and pdfs alike. A file page now opens with the file's name as its title, then the file itself, centred and as large as the column allows: a pdf is drawn page by page in a scrolling box (with the same pdf.js as {{pdf|…|view}}), a video or a sound plays, and a click on a picture opens it full size. Under it, its size and two buttons, Download and Open. A paper's pdf named after its key comes with the citation from its template, e.g., File:Zubizarretacasalengua24b.pdf.
Then "About this file", for my own comments, written on the file page itself and shown only when there are some (the uploaders' stamps, "MsUpload", "Dropped into the edit box" and the like, which are 95% of the descriptions, are left out), and the pages where the file appears or, for a pdf, every page citing it. A small print says that the pdf is made available by the author, with a link to the version of record, when I am one of the authors, and offers to take it down otherwise. The history, metadata and other wiki matters are folded in a closed box, the history reduced to the last upload with its AF tag, e.g., File:Lw-0NMjV.png.
Shortcuts in Emacs don't work, e.g., \equiv (AltGr+Shift+=) reads
\eiuqv \iuevq \qveiu \equvibecause of some latency of some sort in the key deliveries. Surprising, that such a powerful editor gets badly broken in this way. I should see whether this can be fixed without inserting very long delays between letters (which end up being noticeable).
Claude did it on 24 October (2026), without a single delay. It was not latency, nor the fault of Emacs, but ibus, the input-method server that sits between the keyboard and the applications. Ubuntu's im-config starts it in "asynchronous" mode (IBUS_ENABLE_SYNC_MODE=0), and Emacs is the one application here that talks to it through XIM, the old X input protocol: every key makes a round trip to ibus and back, and in asynchronous mode the answers can come back in any order. keymapper types the six characters of \equiv a few milliseconds apart, so they came back shuffled, while the same letters typed by hand, further apart, survived.
ibus brings me nothing anyway (its only engine is the French layout, and KDE does the switching between layouts), so Emacs now starts without it and uses X's own input handling, which keeps the order and still does the Compose key, the dead keys and keymapper's ñ. This is a small wrapper, ~/bin/emacs, that sets XMODIFIERS=@im=none before starting the real Emacs, with a copy of the Emacs launcher in ~/.local/share/applications that points to it (the system one calls /usr/bin/emacs directly, so Dolphin would bypass the wrapper). It cannot be set from .emacs: Emacs connects to ibus before it reads any init file.
Every option was tried in a private virtual X display, typing \equiv 120 times at full speed:
| Setup | Correct |
|---|---|
| ibus as it was (asynchronous) | 0/120 |
| ibus in synchronous mode | about 90% |
| ibus in "hybrid" mode | 0/120 |
| Emacs without ibus | 120/120 |
Wolfram has this nice touch of tearing out the sides of paper extracts, being mindful of which sides would have been actually cropped:
I once programmed something in Mathematica to do that. But it had to be run by hand. It could be integrated to Spectacle. Feature requested on 12 September (2026) (af).
Claude did it on 24 October (2026), in Spectacle itself, where a button in the viewer's toolbar tears the screenshot. Which sides it tears is decided the way Wolfram does it: a side cut out of something larger is torn, a side that lay along the edge of the screen is left straight. So a region grabbed from the middle of the screen is torn on all four sides, a window or a whole screen on none, and any combination can be set by hand afterwards.
The edge is rendered rather than merely jagged: the rip carries a fringe of fibres, the paper scuffs and lightens where it was pulled apart, and the edge casts a shadow on what is under it. Sliders set how deep the tear bites and, for each torn side, how far along the sheet it runs; the shadow can be switched off. Each capture gets its own seed, so a screenshot always tears the same way however the settings are played with, and one button puts every side back to the automatic choice.
What is saved or copied is cropped to what the tear leaves, so the file is the picture and not an empty margin around it. While the image is being annotated the tear stays on show, so an arrow or a label is placed against the edge it will be exported with.
It runs on both machines: built against the current Plasma in a container here, since the one shipped with this Ubuntu is too old, and carried to the other machine as a Flatpak, which brings its own KDE along and so does not mind that the system there is older still. Giving it back to KDE waits on an account with them.
Design a tailor-tuned Kanban (かんばん) service to organize my tasks, called kairos (Greek: the opportune moment).
Platform. Build it on the best available open-source Kanban/board framework, backed by an SQL database and running in the browser on localhost. Everything must be stored in the database.
Board. Four columns: To-do, In progress, Waiting and Done. An alternative calendar view should show the same tiles laid out in time.
Tiles. Each tile is identified by its AF tag and carries:
- Priority: a category (Now, Urgent, High, Normal, Low, Maybe) plus a score from 0 to 10 within that category. Priorities should evolve automatically as time passes and deadlines approach. Tiles are piled up according to priorities.
- Dates: a start date (or should-start date), an official deadline and a hard deadline.
- Group and colour: tasks belong to groups, each with its own colour, and groups can have sub-groups shown as shades of that colour. For example:
- Papers (green): each sub-group is a manuscript with its codename, e.g., Ergoli, and its own shade of green. Use pleasant shades only.
- Microcavities (purple): broken down into chapters or other tasks.
- Admin (black), Glosses (brown), Fun (yellow).
- Expanded view: clicking a tile reveals its attached content, links to llw and wl, notes, statistics, and so on.
Time tracking. The service also manages time. It records how much time is spent on each task and when, and time spent on sub-tasks adds up to the parent task. Time can be attributed automatically through the virtual desktop that has focus. For example, if Emacs is editing a paper on a given desktop, all time spent on that desktop counts toward the corresponding task, or even better, map Emacs buffer file paths or window titles to tasks, so that two papers open on the same desktop are told apart. Same with gmail, llw, twitter, etc.
Idle detection: stop the clock when there is no input, and ask on return whether the gap was thinking time or a break.
Everything should allow for Manual corrections, since an automatic log will be wrong sometimes, so make it easy to edit after the fact.
Estimates vs. actuals: store an estimated duration per task and report my personal calibration ratio. Most people underestimate by a stable factor, which kairos can then apply to its suggestions.
Aggressive storing: every status change, move and time slice are logged as events, so that any statistic can be reconstructed later, including ones I haven’t thought of yet.
Attention entropy: a single daily number for how scattered time was across groups. A nice physicist’s metric for time-tracking, to watch over months.
Wanted since 0Qmsy (af).
Claude took it on from 30 October (2026), and it is now Kairos, whose versions are listed there.
I had my own suite of synchronising the online and local sites, but this was limited to 𝕐 (see syncY). It would be convenient to have it apply to the whole site.
Claude did it on 7 November (2026), with lw2llw, which llw2lw now runs first thing every time (since its v°3.22.0): whatever was changed online since the last run comes down to the local wiki before the local changes go up, for every page and file of the site, and whoever made the change. A page changed online only is brought up to date here. A page changed on both wikis is merged when the two changes are in different paragraphs; otherwise it is held back, with the three texts and the marked merge kept aside, until I settle it by hand, as happened with Welcome Matilde Workshop (2026) on 16 November (2026).
On 11 November (2026), llw2lw v°3.23 also learnt to carry page moves and deletions online, and a file's description changed without a new upload. What had drifted apart between the two wikis before (29 pages deleted here but still online, moves never carried, corrections that never reached the online wiki, an old post whose online edits had been lost) was sorted out the same day and checked once pushed. What is left are a few leftovers for me to decide on: old test pages, unused files, and pages left online under misspelt titles.
My listing of all the papers in bib in a laussy.org page Scientific Articles is problematic, as it stretches the limit of what is serverly do-able. But maybe AI can transcend human and server limits combined. Let it have an audit and then a go (ultracode).
Claude did it on 15 November (2026), starting from the fact that the page did not work at all on the local wiki: it had answered with a server error every day since at least the beginning of the month. Rendering its 5253 references took 77 seconds where a web request is allowed 30, and the nightly job that refreshes the page was throwing its stored copy away every night and then failing to make a new one.
The audit found that the cost was not in the number of references as such but in asking for them all at once: a single #dpl query costs more per reference the longer its list gets (4.6 ms each for 500 of them, 16 ms each for 4000). The list is therefore now made of twenty slices of 500 references, each one a call of {{bibliolist}}, inside the same single list. The page that comes out is the same to the last byte, in a third of the time:
| How the 5253 references are listed | Time to render |
|---|---|
| One #dpl query (as it was) | 77 s |
| Twenty #dpl slices of 500 (as it is now) | 24 s |
| A generated list of {{key}} calls, without #dpl | 16 s |
| The same as static text, without any template | 19 s |
These times were taken with the machine busy on all its cores, so they are the worst case. The last two rows are the floor for a single page: whatever makes the list, it remains 7 MB of HTML with more than 42000 links in it, and those take that long to produce. From now on the cost grows in proportion to the number of references and no faster.
On the local wiki a web request may now also run for 300 seconds instead of 30, so that a page which has lost its stored copy is slow once instead of broken until somebody repairs it, and the nightly job completes again there. What is left is in the list of things to do.
The support of hashtags for 𝕐 and ℤ should be more robust (several, suggestions, etc.)
Claude did it between 14 November (2026) and 16 November (2026). A #word in a post is now a link to Special:Hashtag, which gathers all the posts that carry it, newest first, e.g., #jc-bubbles, and {{Y|#jc-bubbles}} lists them on any other page, the same way as {{Y|0TDsw}} shows a single post anywhere. A post can carry as many as wanted, in its text or on lines of their own at its end, which are then gathered at the foot of the post, on the right, at the height of its comments. A post shared between 𝕐 and ℤ carries its tags in both copies. Since a line starting with #word is now a hashtag, a numbered list in a post needs a space after the #.
For suggestions, a # button in the editing toolbar of 𝕐 and ℤ—in the composer, once its toolbar is switched on, and when editing a month page—lists the hashtags already in use, the most used first or alphabetically, and adds the one picked, or a new one, to the post.