Source Relations Under Recursive Transformation: Provenance Path Dependence and External Re-grounding in Causal Language Models
Authors/Creators
Contributors
Researcher:
Description
This preprint reports a controlled empirical study of source attribution and provenance under single-pass and recursive language-model transformation. The experimental program first narrows a source-boundary effect in Pythia-410M to source-token prediction, first occurrence, boundary distance, and literal repetition. It then tests recursive provenance in a preregistered Qwen2.5-7B-Instruct protocol with 32 records, five conditions, ten recursive passes, and 1,600 model generations.
The primary result is a strong dissociation between factual persistence and correct provenance. At generation 10, correct alias self-relay retained 66.8% correctly attributed facts, while a condition initialized with systematically corrupted provenance retained only 5.2%, a preregistered paired difference of 61.6 percentage points (95% bootstrap CI: 53.0–69.4; 32/32 records positive). Yet the corrupted-origin condition still retained 87.7% of entity-value facts. Thus, monitoring factual persistence alone can miss severe provenance failure. Correct alias relay also exceeded untagged relay by 45.5 points, and immutable true-source re-grounding exceeded self-relay by 15.1 points.
The archival package contains the complete raw 1,600-request matrix, fixtures, immutable origin ledger, preregistration and design files, parser and transition code, inference and analysis scripts, environment capture, logs, derived tables, and cryptographic result locks. Claims are intentionally scoped to the frozen models and protocols tested.
Files
sweeney_recursive_provenance_preprint_v1.0(1).pdf
Files
(2.6 MB)
| Name | Size | |
|---|---|---|
|
md5:2d5b1bf051316d165daa8971db40b317
|
2.2 MB | Download |
|
md5:fb06b4466ce7c8b73f838bf80d8b0d10
|
110 Bytes | Download |
|
md5:a90b4e2f2194744fc12d2b5788c98ab4
|
3.4 kB | Preview Download |
|
md5:584a65aebb1065c048981b48a3eba9c2
|
444.3 kB | Preview Download |
|
md5:710c3e768ca8fc10ca6bfb78e683a6cd
|
2.6 kB | Preview Download |