Dead Circuits: Activation Profiling and Script Invisibility in Large Language Models
This companion paper profiles activation behavior when a model is asked to process N'Ko. The thesis is that some scripts do not merely perform poorly at the output layer; they fail to activate stable internal circuits because the training distribution never made them visible.
Paper workspace
Live draft structure
Artifacts
Draft PDF
Activation-profiling companion draft. Public for reading as live manuscript copy.
Open artifactRelated split paper: script invisibility
Later split-paper render that consolidates the script-invisibility line.
Open artifactEditable source
LaTeX source and draft PDF exist. The page should stay tied to the exact activation-profiling snapshot.
Source anchors
nko-brain-scanner/paper/current/paper1_dead_circuits.tex
nko-brain-scanner/paper/archive/paper1_dead_circuits.pdf
nko-brain-scanner/paper/final/01-script-invisibility/paper.tex
Method tags
Ingest intersections
Status
Drafted; part of the flagship's companion set.
Key claims
01
Script failure can be an internal representation failure.
02
Activation profiling makes script invisibility measurable.
03
A low-resource script can reveal model blind spots hidden by benchmark averages.
Public reading note
Drafted, not yet release-ready.
Standard skeleton
What this paper must keep proving
problem
Output failure on N'Ko may reflect inactive or unstable internal circuits rather than only bad decoding.
method
Profile activation behavior when models are presented with script inputs that were structurally underrepresented in training.
implementation
Activation-probe manuscripts and N'Ko text prompts, tied to the script-invisibility paper set.
data
Controlled N'Ko script probes and model-family outputs. Public release should preserve prompt/data provenance.
evaluation
Activation differences, output behavior, and cross-checks against tokenization coverage.
references
Activation analysis, tokenizer coverage, multilingual representation learning, low-resource script benchmarks.
openQuestions
How much of the observed darkness belongs to tokenizer fragmentation versus pretraining absence.
Checkpoints and references
Proof chain
Draft artifact exists
paper1_dead_circuits PDF and LaTeX
The paper is not just a listing entry; a rendered manuscript and source exist.
Cross-model structural claim
paper3_cross_model companion
The stronger structural claim belongs to the cross-model companion and should not be overclaimed here.