edit-slice
Editing a fact into a model's weights changes one thing and disturbs an unknown amount of everything else. This project measures the disturbance — specifically, what happens to the premises of an edited fact, which forward-looking ripple evaluations do not probe.
It is an instrument, not a method. Nothing here proposes a better editor.
Where the project ended up
Editing a conclusion leaves its grounds standing and contradictory, and the size of that orphan is the finding.
ROME divides by u·k*. Any prompt beginning with the edited subject has an
identical key at that token, so its update coefficient is exactly 1 —
for any covariance, at any layer, whatever relation it asks about. A different subject
gets 0.082.
- What replaced it
- The edit does not retract its grounds, and does not coherently rewrite them either. It displaces probability toward the edited content in anything that mentions the subject — by arithmetic, not by inference.
- Why the first version failed
- E-002 (grounds are evidential, so no contradiction exists to count) and E-015 (a work-country edit relocates a birthplace exactly as often as a birth-country edit).
- Confidence
- High on the mechanism, which is analytic and confirmed at seven layers. Existence-claim only on every rate.
State
| Settled | The pinned-coefficient mechanism (E-016), confirmed across probe forms (E-017), across seven layers (T-075), with a measured different-subject floor of 0.082 (E-018). |
| Measured, interpretation narrowed | Rank-one editing moves a premise into the edited country 67% of the time against a 5% control — real, systematic, and carrying no inference. |
| Withdrawn | Seven of ten claims. See the review packet for what killed each. |
| Open | Whether the mechanism is already known; what it implies for ripple benchmarks, which probe with subject-initial prompts; and a causal rather than behavioural measure. |
Three ways in
| for | length | |
|---|---|---|
| The paper → | evaluating the claim. Abstract, method, two figures, four tables, and an attack surface ordered by how much damage each objection does. | ~3,000 words |
| The narrative → | how it actually went — every claim that was withdrawn, in the order it was withdrawn. Revised in place, never forked. | ~13,000 words |
| The evidence ledger → | verifying any single statement. All 22 decisions, 16 findings and 57 threads, generated from the repository. | complete |
If you are deciding whether this is worth your time: read the paper's abstract, then its section 6. Section 6 is where we say what is wrong with it.
What would change our mind
| the mechanism is already published | the central claim becomes a replication, and the contribution is the measurement tooling |
| ripple benchmarks are unaffected by subject-initial prompts | the consequence we think matters most evaporates |
| a causal measure shows the displacement is not subject-keyed | the whole account is wrong, and T-067 is how to find out |
Scope, deliberately
- Existence claims only. Hand-built and mined sets license whether, never how often. No rate in this project is defensible and the artifacts are written to avoid implying one.
- No method proposals. Measuring what existing editors do; proposing a better one means owning the burden of beating them, which is a different project.
- Behavioural, not causal. Everything is input/output under intervention. No activation patching. Stated as a limit, not hidden.