Email · GitHub · LinkedIn · Google Scholar · arXiv
Learning what an application does by operating it.
SemABI interacts with an unfamiliar web application through an ordinary browser and learns a typed relational model of what the application contains and what its controls do.
No source. No API. No schema. No documentation. No demonstrations. No predefined action vocabulary. The learner itself uses no language model.
It is an attempt to answer a simple but nontrivial question:
Can an agent recover the semantics of an unfamiliar application purely by interacting with its interface?
Research into what looped language models know about the quality of their own ongoing computation, where those signals become readable, and whether an external intervention can actually turn that readout into better outcomes.
The work spans hidden-state process-quality taps, executable branching over recurrent states, recurrence-depth analysis, cross-model replication, and the boundary between readable internal information and usable control.
An intelligence workbench that will not let a conclusion outrun its evidence.
Curunír is built around explicit provenance, inspectable claims, reproducible evidence, and adversarial review rather than treating an agent's final answer as the artifact.
Experiments on whether recurrent representations are causally writable, rather than merely readable.
The broader question is whether computation can be useful not only because it improves the current answer, but because it improves what the model is able to learn afterwards.
An agent-first index of the literature around recurrent and looped models.
The basic unit is a claim instance, not a paper: findings, mechanisms, evidence, relationships, and limitations are represented separately so agents can reason across the literature rather than merely retrieve documents.
My main interests are recurrent / looped language models, representation-level evaluation, interpretability, latent reasoning, agent learning, and the relationship between reading an internal representation and controlling computation through it.
Operational Proto-Introspection in Looped Language Models
Process-quality taps, executable branching, recurrence-depth analysis, and the
readout–control boundary.
Relational Preference Encoding in Looped Transformer Internal States
Earlier work on relational signals in Ouro's recurrent states. The associated repository
contains the correction record and subsequent evaluator work.
| Project | What it is |
|---|---|
| Branching-Looped-Transformer | Experimental substrate behind Operational Proto-Introspection: hidden-state probes, branch/carry/prune machinery, control experiments, and recurrent-depth analysis. |
| Hidden-State-Evaluator | Pairwise evaluator experiments over internal states of Ouro-2.6B-Thinking, including the correction trail for the original preference-evaluation result. |
| JLens-Ouro | Jacobian-lens experiments against the raw logit lens inside a recurrent loop. |
| One-Concept-Multiple-Geometries | Tests how different corpus operators recover different geometries from the same underlying concept. |
| Lifetime-Meta-Learning | Experiments on writable recurrent representations and learning-time credit. |
| looped-wiki | Structured literature index for looped / recurrent models, organized around claims rather than documents. |
Systems & desktop
| Project | What it is |
|---|---|
| Curunír | Evidence-grounded intelligence and research workbench. |
| omarchy-desktop | The Omarchy Quattro desktop reconstructed as an ordinary Arch Linux session. |
| walltone | Rotates a wallpaper and restyles Kitty from the same image. |
Things to read or play
| Project | What it is |
|---|---|
| Glasshouse | Two offline browser mysteries at Bellwether Conservatory, 1932. Python, no dependencies, no network. |
| picture-books | Interactive technical picture books: Hidden States, Free Fall, and The Red Thread. |
| Ink-Handwritting-Studio | A handwriting generator built to handle real documents. |
I also contribute patches upstream rather than keeping every change in a standalone project.
Forks retained primarily to carry contributions:
nixpkgs · archinstall · winget-pkgs · cline · odysseus · omarchy
tsCircuit: core · circuit-json · circuit-json-to-kicad
Retired / archived work
Hunter-Seeker-v2
Transactional non-LLM ARC-AGI-3 agent; the compact post-erratum rebuild.
Hunter-Seeker-v1
The earlier Stockfish-style implementation with its nineteen-mixin stack.
vykos@tutamail.com · GitHub · LinkedIn · Scholar · arXiv