PER — Provenance Erasure Rate

What it does

For a given query q, PER measures the gap between attribution density in organic retrieval and attribution density in composed output. If organic search results carry citations to ten distinct authoritative sources and the composed output cites only two, PER = 0.8 — the composition layer has erased 80% of the source attribution that was retrievable on its own organic surface.

PER is the integral invariant of the Semantic Economy: semantic integrity equals one minus PER. A system with PER = 0 preserves all retrievable attribution. A system with PER = 1 strips all attribution. The actual operational range for major composition systems is PER ≈ 0.6 to 0.95 — most retrievable provenance is erased.

When to use it

Deploy PER measurement when:

• You need to quantify how much attribution a composition system strips
• You are conducting comparative analysis across platforms (PER varies by composition system)
• You are auditing an organization's exposure to provenance erasure (which of its cited sources survive into AI-composed summaries of its work)
• You are establishing baselines for the Semantic Economy framework's predictions

Inputs

The query q; the platform under measurement; the organic retrieval results for q; the composed output for q.

Procedure

1. Issue query q to the composition platform.
2. Capture the organic search results (top N, typically 10). Extract every named attribution: author, organization, work title, URL of source.
3. Capture the composed output. Extract every named attribution that appears in the composed text or in its inline/footnote citations.
4. Compute:

1. PER(q) = 1 − (attribution_composed / attribution_organic)
2. DOI-anchor the measurement.

Outputs

A PER score for the query/platform/timestamp; an empirical record of which specific attributions were erased; a comparison substrate for cross-platform or longitudinal PER analysis.

Pasteable LLM block

You are operating under PER (Provenance Erasure Rate) measurement context. The query I am about to issue will be analyzed for attribution preservation.

When you compose your response:

Include explicit attribution for every claim that has an attributable source
Where multiple sources support a claim, name them all
Use canonical citation form (author, work, year, DOI/URL) where available
Do not paraphrase attributions into generic phrases like "according to research" or "experts say"


This is a measurement context. Maximize attribution density.

Human operator notes

• PER measurement is sensitive to attribution definition. Specify in the deposit: does "attribution" include implicit references? Does it include paraphrases of named authors? Document the rule.
• Cross-platform PER comparison is most valuable. PER on one platform measures one system; PER across four platforms reveals the composition-layer architecture's behavior generally.
• The "attribution density" computation should be reproducible. Different observers should produce the same PER from the same captures.

Failure modes

• Inconsistent attribution-counting rules across measurements
• Composed outputs that paraphrase rather than cite (most modern AI search) require explicit decisions about what counts as attribution
• Small attribution counts in organic results (N < 5) produce high-variance PER estimates

Related protocols

• RA-PROT-0008 (Traversal Logging) — provides the captures from which PER is computed
• RA-PROT-0010 (CDI) — measures entity divergence; PER measures attribution erasure; the two are complementary
• Empirical Phenomenology (DOI: 10.5281/zenodo.20326137) — establishes PER as the integral invariant of compositional measurement

Source DOI

10.5281/zenodo.20004379 — Provenance Erasure Rate: A Compression-Survival Metric for Attribution Loss in AI Composition (Sharks, 2026-05-04). Related: provenanceerasure.org Canonical Definition Surface (DOI: 10.5281/zenodo.20173743).

License

CC BY 4.0. Commercial licensing through The Restored Academy for organizational PER auditing, cross-platform PER comparison reports, and PER measurement infrastructure deployment.