spectralbranding/entity-name-leakage

Entity-Name Leakage in Dimensional LLM Readings

Instrument records of record for the study The Label Moves the Reading, which holds a written artifact byte-identical, exchanges only the entity name it concerns, and measures what an eight-dimensional rating does in response.

Two independently designed collections, each run twice, are released whole: every call with its prompt, parameters, prompt hash and response, alongside the frozen protocols, stimulus packs, harness code and freeze records that fix what was decided before collection began.

What is here

Directory	Collection	Design	Name-swap records
first_collection/	Corpus-prevalence probe	Four artifact texts, each written for a specific prevalent brand, each paired against an invented name	97 (run 1) + 96 (run 2)
second_collection/	Name-effect probe	Nine entities across three prevalence strata, crossed with four texts authored for no entity	432 (run 1) + 432 (run 2)

Each collection directory holds:

PROTOCOL.yaml          pre-registration, frozen before the first live call
FREEZE_RECORD.md       SHA-256 over the protocol, stimulus pack and harness
data/stimuli.yaml      the artifact texts and the entity slot
data/prevalence.yaml   the corpus-prevalence capture the strata were built on
data/run1/, data/run2/ instrument records, one JSONL per operator
logs/run1/, logs/run2/ the full call log for each run
code/                  the collection harness and its unit tests
output/                each run's own report and result tables

The record files carry both arms. A2_NAMESWAP is the name-exchange contrast the study reports; A1_GRADIENT is the frame-gradient arm collected alongside it. The counts above are the A2_NAMESWAP rows, and they are the records the paper's Table 2 reports.

Verifying the freeze

Every SHA-256 in a collection's FREEZE_RECORD.md is stated over a path relative to that collection's own directory, so the freeze is checkable in place rather than on trust:

cd second_collection && shasum -a 256 PROTOCOL.yaml data/stimuli.yaml code/*.py

Eight files will not match, and that is the mechanism working rather than failing. A freeze record fixes the state before collection; anything changed afterwards is recorded as an append-only entry in that protocol's AMENDMENTS: block, and the checksum mismatch is how a reader finds it without taking anyone's word for what changed.

Collection	Does not match	Recorded as
first_collection	code/probe_lib.py, code/tests/test_probe_lib.py	A2 — the token cap was raised for the two operators it was silently binding on
second_collection	code/analyze.py, code/stats_core.py, code/tests/test_name_effect_core.py, code/tests/test_probe_lib.py	A1 — the reporting module rewritten during run 1, before any collected value was read; A2 — the neutrality guard corrected to compute the interaction its own protocol sentence defined
both	PROTOCOL.yaml	the amendments themselves, appended at the bottom and never edited in place

Amendment A2 on second_collection is the one to read first. It changed analysis code after both runs had been collected and read, says so in those words, and did not change the verdict — the guard still fires, so the decomposition the paper wanted is still not licensed. The more permissive variant that would have cleared the guard was computed, reported as a diagnostic, and deliberately not adopted as the decision rule.

Every other checksummed file is byte-identical to what was frozen. Byte-identity is also why the frozen files still carry the authoring paths they were checksummed with: rewriting those strings to tidy them would invalidate the record this dataset exists to make verifiable.

Reproducing the reported analysis

The analysis scripts live with the paper, not here. Each carries its own pinned dependencies inline, so uv provisions them and nothing else needs installing:

git clone https://github.com/spectralbranding/sbt-papers
hf download spectralbranding/entity-name-leakage --repo-type dataset --local-dir enl-data
cd sbt-papers/entity-name-leakage
./reproduce.sh ../../enl-data/second_collection

Or one at a time, which is what reproduce.sh does:

uv run code/estimate_guard_ratio.py  --data-root ../../enl-data/second_collection
uv run code/per_dimension_leakage.py --data-root ../../enl-data/second_collection
uv run code/equivalence_power.py
uv run code/simulate_guard_power.py

Tables are written to output/tables/. estimate_guard_ratio.py recomputes this collection's own published aggregate name effect first and refuses to report a ratio unless it reproduces to within .01, so a misread of the metric stops the run instead of producing a number.

Scope

The material is nine names in one product domain, cookware. The stimulus texts are authored copy that makes no factual claim about any real company, so nothing here asserts anything true or false of a named firm. Model identifiers are pinned literally in each protocol rather than resolved at run time, because two independently collected runs are compared against each other and an operator that changed silently between them would be indistinguishable from the effect under test.

Whether the exchanged name removes what the model already knows has not been measured. Every magnitude in the paper may therefore be a lower bound.

Licence

Records, protocols and stimulus packs: CC BY 4.0. Harness code under each code/: MIT.

Citation

Zharnikov, Dmitry (2026). The Label Moves the Reading: Entity-Name Leakage in Language-Model Dimensional Ratings of a Fixed Text. Zenodo. https://doi.org/10.5281/zenodo.22161305