⚠ no full-vocab projection for this model — run heinrich mri-vocab — showing 2,000-token sample only
scripts ▸vocab reveal: on
Tokens
click A · shift / right‑click B
A—
tap or click a token to pin
B—
right-click or long-press to compare
projecting direction…
Discover Features
global PCs ranked by bimodality at this layer
find candidate opposing token clusters for the current layer
Attention
Token A · current layer
Direction Analysis
on demand · support-window approx.
Max separation PCs
Steer
inject A−B into the live forward
layerα+4.0
Cross-Model Compare
same pair, second model
pin A and B; pick a second model; Run
Live Session
loaded backend for run/chat
session idle
Live Forward
run a prompt through the model
set HF id above; model load on first run ~5-30s
Homing
live → pinnedlogjson
◉ T clear
◉ T2 clear
Model Chat
talk to the loaded model directly
uses the same live backend as prompt MRI
🔓 Unlock live mode
1 · run heinrich locally (prints a pairing code)
pipx run heinrich companion
2 · connect (Chrome) — stay right here, go live
or · open this exact view in a local tab (non-Chrome fallback)
Connect keeps you on this origin and reroutes the instrument to your local companion over Chrome Local Network Access. The link is the fallback: same model · layer · pins carry over.