edit-slice

Editing a fact into a model's weights changes one thing and disturbs an unknown amount of everything else. The standard way to evaluate that disturbance looks forward, at what follows from the edit. This project looks the other way, at the facts that were premises for the edited fact and are still sitting there after it changes.

What the project claims

Everyone probes forward from an edit; nobody probes backward.

Inverting the arguments of an edited fact is covered, by RippleEdits, since 2024. Probing the distinct facts that were premises for the edited fact is not covered by anything I can find.

Confidence

Medium — one benchmark read closely, not a survey

Falsified_by

A published grounds probe distinct from argument inversion

Revised

10 September 2026

The distinction the project rests on

Two things look backward and are not the same thing. Keeping them apart is most of the contribution.

	What it probes	Covered?
Argument order	The edited triple with its arguments swapped — (Rome, contains, Eiffel)	Yes, RippleEdits Logical Generalization (Cohen et al., 2024)
Justification order	Distinct facts whose truth was a premise — built for the 1889 Paris Exposition	Not that I can find

An edited model can hold a belief together with a complete, untouched set of grounds for its negation. That state is what the project calls an orphan.

The transitive case is the sharper one. Getting from "Eiffel is in Rome" to "Eiffel is in Italy" requires using the fact that Rome is in Italy, but the benchmark only ever checks the conclusion. Whether the premise survived is never asked.

Status

Phase 0 — definitions and scope, near complete

Settled so far:

• The scope boundary against rome-neighbors is reuse of code, not of scope.
• RippleEdits' Logical Generalization is not a grounds probe. The claim above narrowed as a result, which is the useful thing that came out of Phase 0.

Open:

• Can the benchmark generator's distance function take a code dependency graph as input, so grounds are discovered rather than enumerated by hand?
• What decides that a fact was a premise, in a way a reviewer can check?

Deliverable

Two panels and one number: forward propagation against justification-order survival, on the same edits, with the gap between them stated as a rate.

References

1. Cohen, R., Biran, E., Yoran, O., Globerson, A., Geva, M (2024). Evaluating the Ripple Effects of Knowledge Editing in Language Models. TACL.
2. Meng, K., Bau, D., Andonian, A., Belinkov, Y (2022). Locating and Editing Factual Associations in GPT. NeurIPS.
3. Meng, K., Sharma, A. S., Andonian, A., Belinkov, Y., Bau, D (2023). Mass-Editing Memory in a Transformer. ICLR.