When Attention Sinks Enter the Index

Computational showcase for a first-author study of patch-level evidence attribution in late-interaction visual retrieval.

Scientific claim

We audit whether attention sinks remain benign after visual patches enter a late-interaction retrieval index. Among nine checkpoints with an index-side verdict, two indexed an attention-confirmed sink; only one also formed a separate similarity attractor and showed a localization deficit. In that affected checkpoint, the sink absorbed 89.8% of its image block’s EOS-attention mass and the sink–attractor pair occupied 37.8% of a four-pick attribution budget. True deletion improved RegionHit@4 from 0.1875 to 0.2677 on the industrial support (+0.0802; query-clustered bootstrap 95% CI [0.0685, 0.0923]) and from 0.4086 to 0.5188 on the gate-first pharmaceutical support (+0.1102; [0.0858, 0.1357]). Post-hoc candidate suppression overstated the deletion effect by 0.0293–0.0827 across the audited supports. All 18 official-pool nDCG@5 intervals included zero; this is evidence of no detectable change, not equivalence or zero cost.

Frozen-paper Figure 2. True deletion improves RegionHit@4 by 8.0–11.0 percentage points on the annotation-selected supports. The horizontal nDCG@5 intervals in this figure use those same supports; the separate official-pool analysis reports all 18 intervals containing zero. Neither result establishes non-inferiority.

What is included

• Late-interaction retrieval, attribution, evidence-budget, evaluation, and true-deletion implementations under src/ceb/.
• Stable embedding and retrieval entry points under scripts/.
• A reference environment in requirements.txt, package metadata in pyproject.toml, and model/configuration declarations in configs/default.yaml.
• Synthetic and regression tests under tests/.
• Claim-level result summaries under results/; the final publish tree also preserves the existing public derived-result payloads, figures, and freeze provenance as historical evidence.
• A claim-level experiment-family manifest in experiments/manifest.csv.

The repository also retains previously public experiment modules and historical curation records. They remain available for provenance and review, but they do not expand the nine-row clean canonical claim surface.

Headline results

Scope	Result	Reproduction status
Checkpoint screen	9 index-side verdicts; 2 attention-confirmed sinks; 1 sink–attractor localization deficit	documented-only
Affected checkpoint	89.8% EOS-attention share; 37.8% of a four-pick attribution budget	documented-only
Industrial support	RegionHit@4 0.1875 → 0.2677; +0.0802, 95% CI [0.0685, 0.0923]	documented-only
Gate-first pharmaceutical support	RegionHit@4 0.4086 → 0.5188; +0.1102, 95% CI [0.0858, 0.1357]	documented-only
Intervention audit	Post-hoc candidate suppression overstated the true-deletion effect by 0.0293–0.0827	documented-only
Official-pool ranking	All 18 nDCG@5 intervals included zero; no detectable change, without evidence of equivalence or zero cost	documented-only

The machine-readable values are in results/headline_claims.csv. Their public provenance and availability status are in results/claim_provenance.csv.

Reproduction status

Pathway	Status	Meaning here
Repository inventory, static Python parsing, and relative-link checks	verified	Completed against this tree without installing dependencies
Dependency installation	verified	The pinned environment installed in an isolated local environment
Synthetic/regression test suite	verified	The retained publish-tree suite completed with 108 passed and 4 skipped
CLI import and help paths	smoke-tested	embed.py --help and retrieve.py --help returned successfully
Paper-facing module help paths	smoke-tested	The four documented module signatures below returned help successfully; no experiment body ran
Included derived-result evidence	documented-only	Public JSON/CSV summaries and frozen figures are mapped to claims; they were inspected, not regenerated
Headline-result rerun	unavailable	Required data, model artifacts, caches, and checkpoints are not distributed here

The final publish tree preserves previously public derived-result payloads for evidence and provenance, but it still omits the inputs required for a headline-result rerun.

Cold-clone validation

The declared runtime is Python 3.11 or 3.12. This setup path was verified in an isolated local environment and is documented here for a fresh environment:

python3.11 -m venv .venv
.venv/bin/python -m pip install -r requirements.txt
.venv/bin/python -m pip install -e .

The repository exposes embedding and retrieval entry points:

.venv/bin/python scripts/embed.py --backbone colqwen2.5 --data pcb_pilot
.venv/bin/python scripts/retrieve.py --backbone colqwen2.5 --data pcb_pilot

These are generic Phase-1 wrappers configured for the legacy pcb_pilot data block; they are not the frozen paper's ViDoRe V3 reproduction path. The two entry points were smoke-tested only through --help, and full execution remains unavailable because the configured data and model artifacts are not bundled.

Paper-facing modules

The retained paper-facing modules include:

.venv/bin/python -m ceb.evidence --backbone colqwen2.5
.venv/bin/python -m ceb.phase3.attention --backbone colqwen2.5
.venv/bin/python -m ceb.phase15.existence_domain \
  --domain pharmaceuticals --backbone colqwen2.5
.venv/bin/python -m ceb.revision.truedel_run \
  --corpus industrial --run-id showcase-audit-run

Status: smoke-tested for the four --help paths and documented-only for the command bodies. These commands identify the relevant code paths, not a complete single-command replication workflow. No experiment body ran for this refresh, and the required ViDoRe pages, model artifacts, embeddings, attention caches, and checkpoints are absent. See docs/reproducing.md before attempting them.

Repository map

Path	Role
src/ceb/attribution/	Counterfactual attribution primitives
src/ceb/budget/	Evidence-budget selection
src/ceb/retrieval/	Late-interaction MaxSim retrieval
src/ceb/revision/	True-deletion implementation and exports
src/ceb/eval/	Retrieval and localization metrics
tests/	Synthetic and regression tests
results/	Clean claim table plus retained public derived-result evidence
figures/	Retained public figures derived from the audited results
provenance/	Retained public freeze manifest and artifact hashes
docs/	Reproduction, availability, limitations, and scope

Boundaries

• This is a computational showcase, not a complete replication package.
• Raw data, model weights, caches, checkpoints, logs, and manuscript files are not distributed.
• Existing public derived-result payloads and figures are retained as evidence; they do not make this a complete replication package.
• Headline results are documented-only, not rerun from this tree.
• The final publish tree preserves the public repository's existing LICENSE and CITATION.cff; no license or citation metadata is silently replaced.

See docs/data-and-models.md, docs/limitations.md, and docs/repository-scope.md for the full boundary statement.