AR6

Arsis

One person, agents, and measures to know when we are wrong.

A workshop that builds its own tools with agents and follows questions wherever they lead, with a dated record of what was measured, corrected or dropped.

Open questions

1 Mar 1 Oct 1 Jan 4 Feb 1 Mar 7 May 1 Jun 3 Jul 1 Aug 22 Sep 69 Oct
Entries per month, March to October 2026. August: no commit on any personal project.

Register · 111 entries · March 2025 to October 2026

m measured 97e estimated 1i interpreted 8c corrected 9 to re-check 0

October 2026

  1. correction LLMarium

    The local model launcher is reframed and renamed LLMarium (6 October). The launcher was reframed on 6 October and renamed LLMarium on 10 October. c

  2. correction LLMarium

    Models in memory are listed with the process that launched them (13 October). Listing each model in memory with the process that launched it was committed on 9 October. c

  3. question Research notes

    Opened: what does power do to language, and how do we read what it forces to hide? Newspeak, Klemperer, Strauss, with corpora and models as instruments. i

  4. question Research notes

    Opened: how does a concept travel, and when does it start to disqualify? A first check of three 2023 answers from a general model already suspects one error. i

  5. question Research notes

    Opened: what can a local model do on a personal machine? Working formulation: the method around the model matters more than the model. i

  6. pause Syntype

    Version 0.1.1 was announced on 30 September to fix the four blocking issues; sixteen days later the repository has no new commit. m

  7. decision arsis.dev

    The form was fixed as the register in a bibliography style with a timeline layer, with one status per claim, a stack of dated determinations, re-check dates and visible corrections. m

  8. milestone arsis.dev

    The register prototype became the site: the older pages that named clients were retired, a build guard now fails on any client name, and Source Serif 4 is served from the site itself. m

  9. rename Squatcat

    Ron-Ron, the training-session prototype, is renamed Squatcat. m

  10. decision arsis.dev

    The register was adopted as the form of the site: every entry dated and stating how it is known. m

  11. decision arsis.dev

    The site will show the workshop and its work, not the person behind it. m

  12. measure Spicilège

    Intentions: for each saved item, why it was kept, proposed by a local model and measured against examples. m

    1. det. 15 Oct · measured57 % with method v3, which merges two verbs, on 186 choices.
    2. det. 14 Oct · measured38 % with a local model on the Mac, same examples.
    3. det. 13 Oct · measured45 % with Julien’s profile and 69 of his choices as examples.
    4. det. 13 Oct · measured26 % of reasons guessed right from the content alone, on 138 of Julien’s choices.
  13. question Research notes

    The convergence inquiry is reframed: from geometric alignment between models to causal invariance, tested in worlds that look the same but work differently. i

  14. measure Rhizomes

    Claude Code deletes sessions older than 30 days by default: of 306 local sessions, none predated 10 June, so the archive keeps compressed copies. m

  15. release Rhizomes

    A searchable archive of Claude Code, Codex, ChatGPT and claude.ai conversations (3,831 at the first run) ships with a read-only MCP server for agents. m

  16. pause arsis.dev

    After two rounds of mockups, seven pages in all, the first release planned for 16 October was postponed without a new date. m

  17. milestone LLMarium

    Models in memory are listed with the process that launched them. m

  18. milestone Spicilège

    A local model proposes bookmarks for notebooks; tags that evolve. m

  19. decision Method

    The posture is written down: contribute by what can be verified; publish raw numbers and failures; correct in public. m

  20. question Research notes

    Kept from 58 candidate questions, the only one judged fundamental: does a typology of beliefs hold against ethnographic data? m

  21. rename Futon Fighters

    The couch-game collection becomes Futon Fighters, with Ippon and Fourtress; a new icon. m

  22. decision Tales of Chinatown

    Four drawn banners chosen for the store page. m

  23. correction Futon Fighters

    Two Steam Deck crashes were traced to the executable being replaced during play, not to the game, which corrects an earlier suspicion about material and interpolation switches. c

  24. decision Squatcat

    After building eight interface directions on one engine, the author picked a daily comic-strip home with a sun-shaped timer as the final interface. m

  25. milestone Tales of Chinatown

    The trailers are rebuilt in code from real captures: 60 and 30 seconds, French and English, sound normalised to −16 LUFS. m

  26. measure Futon Fighters

    Measured against classic shooter maps, blocking the narrowest passage lengthened routes by only 4 to 6 percent in generated maps against a 51 percent median in the references, which led to a revised map generator. m

  27. rename Squatcat

    The training-session prototype gets its name, Ron-Ron. m

  28. correction Palimpseste

    Palimpseste’s rule, everything deduced from history, measured against itself: 17 of 82 quarrels and 15 of 91 alliances had no cause. Removed in version 17. m

  29. measure Futon Fighters

    A first Steam Deck session with two players and bots ran 17 matches and 135 rounds in 49 minutes, with a median round of 5.8 to 8.1 seconds and no crash. m

  30. decision LoopHuntr

    After a state-of-the-art study read against the code, four generator changes are planned and will be measured one at a time: a climb floor, diverse candidates, local improvement, a second climb calculation. m

  31. rename LLMarium

    The local model launcher, until then “Modèles locaux” with the command lml, becomes LLMarium after four waves of names. m

  32. decision LLMarium

    LLMarium keeps to three MLX engines and takes five ideas from neighbouring tools: a GPU memory ceiling, a speed estimate, an idle stop, a model trash and a radar sorted by use; Ollama and llama.cpp stay out. m

  33. milestone Squatcat

    All 81 cat animations were validated one by one by the author over five review passes, after automatic measures and a three-gate protocol replaced eyeballing, and are bundled in the app. m

  34. decision Squatcat

    Of five features proposed from a research dossier, four were adopted and built the same day, adjusting or setting aside today's session, logging activity done without the app, an easier entry level and a comeback after ten days, and the safety notice was deferred. m

  35. correction Research notes

    Corrected in the framing report: a strong convergence between models would not, by itself, favour Levin’s non-physicalism; data, objectives or optimisation could produce it. m

  36. milestone LLMarium

    A local model opens in four coding agents from the app, Claude Code, Codex, Pi and OpenCode, each finishing a tool-call task; the fifth, DeepSeek Harness, is configured but not yet run. m

  37. decision LLMarium

    A local entry point is built without any dependency to replace a LiteLLM gateway: it rewrites only the model name and streams answers through untouched; the switch itself is not made. m

  38. decision Squatcat

    After testing a puppet rig, a video model and generated poses on the same squat, the cat's movement is settled as two to five generated poses per exercise, set on the feet, with 81 animations planned. m

  39. decision Method

    Decisions move from chat to follow-up pages, where each proposal carries its evidence. Six repositories use one within a week. m

  40. milestone Futon Fighters

    The duel and a first-person game now share one application and one set of menus, so the same players move from one to the other without picking seats again. m

  41. decision Futon Fighters

    After a first playtest, the first-person game switches to health points by default and keeps one-hit kills as an option, with sound, bot hearing and round endings added the same day. m

  42. rename Futon Fighters

    The collection is first named Irori and its first-person game Hitome, names that last five days. m

  43. measure Squatcat

    Three open voice models read the same eight phrases of the app and a transcription check found 4, 11 and 56 percent of errors; Qwen3-TTS was kept as the coach voice, in two versions of 325 phrases each. m

  44. decision Slyfari

    Saving is switched off on eleven large video and music platforms whose terms forbid it; a list of blocked sites is chosen over a list of allowed ones. m

  45. decision tralala

    Re-orchestrating a generated song with a general soundfont is set aside after listening tests; only separating stems with Demucs stays on the list. m

  46. release Appothèque

    Appothèque 0.5.0, first release, open source under MIT. m

  47. decision LLMarium

    The local model launcher is reframed and renamed LLMarium. m

  48. decision Futon Fighters

    The one-hit duel is named Ippon, after the judo point that wins in a single stroke, over the drier Estoc. m

  49. decision Appothèque

    The launcher is redesigned as a native macOS 26 app: eight windows and panels become one launcher and one window, and the nine custom themes give way to eight tints. m

  50. release Appothèque

    Appothèque 0.5.1 follows the first release by 53 minutes, with launch progress in the launcher rows and the list of changed files. m

  51. measure LLMarium

    Claude Code and Codex each finish a tool-call task, reading a secret file and answering with its token, on a local oMLX model and on a local MTPLX model: four pairs out of four. m

  52. milestone Syntype

    A research model reviews the field for Syntype: missing species traits are a real need that phylogenetic imputation already serves; the open question is transfer to poorly documented species. i

  53. question Layon

    A literature review commissioned from a research model finds the need for missing species traits real but already served by established methods: Layon must first show a measurable gain. i

  54. measure Daymon

    Window tracking degrades: one or two applications seen per day, against sixteen to twenty-five before. m

  55. measure NamerKit

    Two automatic ways of predicting the author’s taste, a base model’s probabilities and a language-model judge, do no better than chance on pre-registered tests; the studio moves to unranked waves sorted by hand. m

  56. measure NamerKit

    Prompting for distant territories cuts obvious junk names from 16% to 3% but the author keeps 0.2 names per 10 instead of 1.8, so the approach is dropped. m

  57. decision Daymon

    Daymon is limited to macOS on Apple Silicon; the Windows and Linux trackers stay experimental. m

  58. rename Spicilège

    Herbier becomes Spicilège: from the Latin spicilegium, gleaning ears of grain. m

  59. decision Rhizomes

    The analysis layer is set aside on its own branch; the project restarts from an automatic daily journal. m

  60. measure Rhizomes

    A local model, Qwen 3.8 Flash, summarised a full working day into the same eight episodes as a manual reading, in about four minutes. i

  61. milestone Futon Fighters

    First commit of a local duel prototype where one hit kills, built to test the core fight before any generated weapons. m

  62. decision Appothèque

    Before publication the launcher drops every personal default, a shipped list of the author's projects and fixed search folders, and starts empty. m

  63. rename Appothèque

    The launcher called Dev Apps is renamed Appothèque, a play on apothecary. m

  64. milestone Appothèque

    A first working version, then called Dev Apps, runs within an hour of the first prompt: one click builds an app only if its files changed, then opens it. m

  65. rename tralala

    YuE Atelier becomes tralala, "variations et ritournelles", a name that depends on neither the model nor the final shape of the app. m

  66. decision tralala

    The app also becomes a library of old public-domain songs as starting points: 18 French songs from a 1910 anthology are added as recipes, with English translations. m

  67. milestone BackShelves

    BackShelves can restore only chosen tables from a PostgreSQL archive, a feature first written into its roadmap on 10 Oct 2025; it was checked on disposable databases. m

  68. milestone tralala

    The first version of tralala generates full songs on a Mac with YuE2, with no server: at 32 steps a song takes about as long as it lasts. m

  69. milestone arsis.dev

    By 1 October 2026 the site had gone through 46 homepage and design prototypes, none published, the last 23 built from rules rather than from references. e

    re-check by 16 Nov · now estimated

September 2026

  1. contribution NVIDIA SDM

    Apple Silicon support sent to NVIDIA’s structured-data-models: mps and mps:0 were treated as two devices. Merged (#1032). m

  2. correction Syntype

    A second model reviewed Syntype v0.1 and found four blocking issues. Forest results declared invalid, trait results preliminary, the same evening. m

    1. det. 30 Sept · correctedA second model’s review finds four blocking issues: forest results invalid, trait results preliminary.
    2. det. 30 Sept · measuredv0.1 published: the foundation model wins on species traits, not on forest inventories.
  3. milestone Layon

    Layon starts: an open ecological data platform. Early research, no usable code yet. m

  4. measure Syntype

    Light fine-tuning of Kumo Tabular (small) on three trait databases, tested on the fourth, brought no measurable gain: better than the untuned model on 2 of 8 traits (preliminary). m

  5. decision Syntype

    Values that the source databases imputed themselves are removed from the targets, along with features derived from the target; the audit found three such leaks before any result was published. m

  6. correction Layon

    Layon’s first public vision note cited figures nobody else could check; an hour later a public benchmark replaced them, and by evening its results were marked provisional. c

  7. decision Layon

    No platform before a confirmed pain: Layon’s first step is to ask ecologists which of four problems is real. m

  8. decision Layon

    Layon is published under Apache-2.0, chosen the day the repository was created. m

  9. decision Method

    Work is planned in numbered lots, each closed by an independent reviewer: Daymon, LoopHuntr, NamerKit. m

  10. milestone Daymon

    Tests written to fail first: replay deadlock, corrupt journal rows, duplicate effects after a crash. Then an independent review of the lot. m

  11. milestone LoopHuntr

    Release preparation is run in numbered lots with independent review. The first of five gates, traceable work, is passed. m

  12. measure Method

    A second model reviews the first one’s work before each stop: 41 passes in a day, 6 blocked. m

  13. correction NamerKit

    A 74% win rate recorded in July is set aside: the name format gave the source away, the Namelix sample was reused and the test briefs were the tuning briefs. c

  14. milestone LoopHuntr

    Offline data goes live: France is cut into 1° cells, signed, and served from the project's own server. The manifest now lists 99 cells, about 5.7 GB. m

  15. measure Method

    From 25 to 29 September, about 390 commits on Daymon and LoopHuntr, most of them written by agents. i

    re-check by 1 Dec · now interpreted

  16. measure LoopHuntr

    Share of attempts that yield an accepted loop, measured on a Mac: 61 % for a lowland 35 km gravel loop, 13 % for 60 km and 1,200 m, 0 to 7 % in the Cévennes. Per attempt, not per request. m

  17. measure Slyfari

    Reading a 400 MB file in 1 MB blocks peaked at 407 MB of memory; releasing each block as it is read brought the peak down to 7 MB. m

  18. milestone Slyfari

    Vault files now authenticate the order of their blocks and their end, so truncation or reordering is detected. m

  19. milestone Spicilège

    A desktop app over the Python core. m

  20. decision Method

    One instruction file per repository, AGENTS.md, read by every agent. m

  21. milestone Spicilège

    Started, as Herbier: ingestion contracts checked against real exports, 6,479 X bookmarks and 1,985 Raindrop links. m

  22. question Research notes

    Opened: Michael Levin argues that minds are non-physical patterns; do independently trained models converge on shared structures, and what would that prove? m

August 2026

  1. pause Method

    No commit on any personal project from 6 to 31 August. m

July 2026

  1. measure Slyfari

    Exporting two passages of a 1,178 MB video gave a 1.4 GB file, because the export re-encoded everything at highest quality; it now copies the original stream. m

  2. measure NamerKit

    A blind scoring by one person gives 46% wins against Namelix on 10 briefs, up from 34% in May; the 50% target is missed. m

  3. milestone Slyfari

    First numbered milestones, U1 to U14. m

June 2026

  1. rename Rhizomes

    Pulsed becomes Rhizomes, and is no longer meant to stay within software work. m

May 2026

  1. milestone Research notes

    Wittgenstein and Davidson, turned into a small game to test what a model does with meaning. m

  2. milestone Research notes

    An annotated bibliography on language and language models. m

  3. question Research notes

    Opened: frictional realism. Meaning is born in continuation; truth is born in friction. m

  4. decision LoopHuntr

    A spike accepts 19 % of 512 generated loops, 52 % in its best run on one zone and 8 to 32 % on two other zones (Cévennes, Provence). The project moves to a native iPhone app anyway, with a measured quality gate. m

  5. decision NamerKit

    The on-device model is dropped for a naming studio: local language models propose names, and filters and scores sort them. m

  6. correction NamerKit

    Scored again by an agent that had not seen the data, the first blind benchmark against Namelix falls from a ratio of 0.84 to 0.74 and fails; a report that printed a pass is fixed the same day. m

  7. milestone NamerKit

    NamerKit starts as a plan to match Namelix with a small naming model trained from scratch and run on the device. m

March 2026

  1. question Research notes

    Reading Bloom’s Western Canon: how far does a canon travel from English into French, and who decides it? m

February 2026

  1. milestone BackShelves

    In one day BackShelves gains a SQL reader for PostgreSQL, MySQL, SQLite and DuckDB, an MCP server that can list and restore backups, and a French and English interface. m

  2. rename Slyfari

    Vaultware becomes Slyfari the day after the first commit. m

  3. milestone arsis.dev

    The site was started on 11 February 2026 with Astro, Tailwind and a Docker image, and had its first pages, themes and content collections within the day. m

  4. rename BackShelves

    Backup Explorer is renamed BackShelves, and its repository is created under the arsis-dev organisation the same day. m

January 2026

  1. milestone Rhizomes

    First commit: Pulsed, a command-line tool that reads Claude Code sessions; topics, research and daily briefings follow the same day. m

October 2025

  1. milestone BackShelves

    BackShelves begins as Backup Explorer, a desktop app to browse backups on Backblaze B2 and SFTP and download them. m

March 2025

  1. milestone LoopHuntr

    A first web prototype, built on Replit as TrailGenius, shows GPX files and Strava rides on a map. It has no loop generation. m