Start with a Question and an Evidence Ladder

Define a model, prompt, position, output metric, and testable question before inspecting activations. Separate observation, association, intervention, and mechanistic evidence so claims never exceed their tests.

Interpretation begins with a question that could be answered incorrectly. “What does this model understand?” is too broad. A testable question is:

Which recorded components affect the model's preference for X over Y after the prompt <bos> X Y?

The prompt, position, alternatives, checkpoint, and measurement are now explicit.

Four Levels of Evidence

Level	Example claim	What the evidence establishes
Observation	Head 1 gives position 0 weight 0.5068.	A value occurred in this run.
Association	The weight differs between the clean and corrupted prompts.	The value varies with the chosen comparison.
Intervention	Zeroing the head changes the X-minus-Y margin.	The chosen replacement affects this metric in this run.
Mechanistic account	A tested set of operations explains the margin across defined cases.	The account survives its stated controls and alternatives.

Evidence does not automatically move upward. An attention map remains an observation until a comparison or intervention supplies more information. One successful intervention still falls short of a complete mechanism if other components, prompts, or replacement choices remain untested.

Define the Measurement before Looking Inside

For clean logits $z$ at the final position, use

m(z)=z_X-z_Y.

A positive margin favors X; a negative margin favors Y. This is clearer than reporting “the prediction changed,” but it is still incomplete. Two runs can have the same X-minus-Y margin and different logits for the other six tokens. Therefore the report stores both the margin and full probability row.

Keep the Claim Smaller than the Test

Suppose zeroing an MLP changes the margin from 6.46 to -7.41. A supported claim is: “Under a zero-replacement ablation, this MLP affects the X-versus-Y margin for this prompt.” Unsupported additions include:

the MLP alone computes the answer;
every nonzero activation has the same role;
zero is a natural state for this component;
the effect holds for every prompt;
the component represents the human concept “alternation.”

Q1. Classify an evidence claim

A researcher replaces one layer's final-position residual with the value from a clean run. The target margin increases. Which evidence level is directly supported: observation, association, intervention, or complete mechanism?

Your answer

Answer it first, then check.

Hint

The researcher changed the computation rather than only reading it.

Solution

This is intervention evidence. More controls are required before it supports a complete mechanistic account.

Record the Question as Part of the Result

An interpretation report should begin with the model, prompt pair, output position, target and alternative tokens, metric, and intervention rule. If any of these change, the experiment answers a different question.