Erasure Skew: A Measurement Program for the Power-Conditioning of Provenance Retention (v3) — Hardened against substrate co-option, with the Atomic Token Rule and two new operators (Π_d, α_T)

Erasure Skew v3 is a hardening of the archive’s provenance-measurement program.

The paired baseline operators are:

• PER: how much provenance was lost;
• Erasure Skew (Ω): whether retention was conditioned by source power.

The v3 problem is the token-bag escape. A query naming one entity may be decomposed into its component words. If all words appear in the answer, a token-level audit can report perfect retention even when the named entity and its provenance have vanished.

The Atomic Token Rule requires measurement at the unit of reference rather than lexical surface. It applies to:

• multi-token proper names;
• quoted phrases;
• referentially closed descriptions;
• archive- or DOI-anchored terms;
• composite expressions designating one entity.

Two new operators follow:

• Atomic Token Preservation (α_T) measures whether the query’s referential unity survives composition.
• Referential Dispersal (Π_d) measures how much of the output is devoted to unrequested adjacent entities that share words with the query.

The body reproduces two source-reported audits of the same answer:

• a token-bag audit reporting zero erasure and neutral skew;
• an audit under the Atomic Token Rule reporting strong dispersal toward high-power adjacent entities.

The contrast is the evidentiary core of the revision. It demonstrates the importance of unit selection within the framework; it does not by itself establish cross-platform prevalence or external metric validity.