Sample output, committed for illustration — produced by ./run.sh demo (interp explain ... --outdir reports --figdir figures). Normal runs write to the gitignored out/ directory instead.

Why did the model predict `Mary`?

Prompt: When John and Mary went to the store, John gave a drink to

1. Final prediction (logit lens)

token	probability
`Mary`	0.481
`them`	0.207
`the`	0.089
`his`	0.031
`their`	0.022

The answer first becomes the top guess at layer 10.

2. Which components write the answer (DLA)

Attribution to logit( Mary) - logit( John) (correlational):

component	type	attribution
`L9H9`	attn_head	+1.656
`L10H7`	attn_head	-1.455
`L9H6`	attn_head	+1.317
`L11H10`	attn_head	-1.102
`L10H0`	attn_head	+0.882
`L11H1`	attn_head	-0.838

3. Which heads carry the causal signal (head patching)

clean logit-diff = 3.17, corrupt = -3.53

Patching each head's clean output into the corrupt run — the causal counterpart of the DLA table above:

head	recovery
`L10H7`	-0.29
`L8H10`	+0.20
`L8H6`	+0.15
`L11H10`	-0.12
`L9H9`	+0.12
`L7H3`	+0.12

Residual-stream patching over (layer, position) is included as a sanity check: patching the token where the clean and corrupt prompts differ trivially restores the clean behavior (recovery ≈ 1.0), which validates the minimal pair rather than localizing the circuit.

4. Concepts active on the prompt (SAE features)

feature	peak token	explanation
12-res-jb/3700	`When`	triggering moments or events in narratives
11-res-jb/19004	`When`	instances of the word "When"
11-res-jb/6791	`John`	proper nouns, specifically names of individuals
11-res-jb/22481	`went`	instances of the word "went" in various contexts
10-res-jb/3926	`When`	specific trigger words like "When"
10-res-jb/2165	`John`	names of people

Generated by interp. Every number above is reproducible from the prompt.

FilesExpand file tree

explain.md

Latest commit

History