Fig. 2 — Dispatches from the workbench
News
Dated dispatches: releases, results — negative ones included — and policy changes. Newest first; every entry links the record that backs it.
-
2026-10-02
Kapteeni v1.1c released — the multimodal line
kapteeni-v1.1c extends the wire format with
state.image: the same noul/choice/score questions can now be asked about an image alongside text, on one Qwen3.5-4B backbone (en/zh). The pack — merged model, heads, serving constants, card, and a self-contained demo site with 15 pre-configured cases including real photographs — is pinned at revision17affb8. Public-half run of record on the released pack: composite 63.72 — inside run noise of v1's 65.71 — with the separable gains in hard tier (0.469 → 0.559), multi_hop (0.278 → 0.667), and long_policy (+10.5pt); Intelligence 68.33 is the best measured in the repository, top-label ECE 0.0984 between the two v1 variants. Bench request open for both benchmarks: fstandhartinger/jevbench#178 · JEVBENCH.md section. -
2026-09-30
The Kapteeni operating manual is live
The manual — overview, wire API, benchmarks, organized from a user's viewpoint — is served at triusai.github.io/Kapteeni from that repository's
docs/folder by GitHub Pages (written 2026-09-26; live as of this entry). This news column opens on the frontpage alongside it. -
2026-09-26
Kapteeni v1 released as two variants
kapteeni-v1-meticulous — the default; conservative confidence — and kapteeni-v1-intuit — sharper on well-formed numeric, temporal, and multi-step traffic — ship as self-contained distribution packs (merged bf16 model, heads, serving constants, model card). Cards carry numbers and caveats only, and link the repository and each sibling variant.
-
2026-08-27
Blacksmith-1 checkpoint released
Blacksmith-1-20260827 is on Hugging Face: a small language model trained from scratch on pre-tokenized UltraChat and Cosmopedia v2; the repository's own README introduces the result as "an artificial idiot which can barely answer any questions at all". Repository.
-
2026
Spectre, tarha, OpenPixie archived
The three agent frameworks were closed to development and moved to archived; the repositories remain public in the org catalogue.