Same Problem, Different Field: cross-domain solution import via domain-stripped computational fingerprints (reproduction package)

The same underlying computational problem is independently re-solved across unrelated fields under different names: a recursive Bayesian estimator is a "Kalman filter" in control, "Bayesian forecasting" in pharmacokinetics, and "data assimilation" in geoscience. Topical and citation-based scientific embeddings cannot see this kinship, because they cluster papers by the vocabulary and field that differ.

This package reduces each paper to a domain-stripped, method-name-stripped computational fingerprint (one cached LLM call per paper) and ranks cross-field twins (papers solving the same computational problem) with a cheap embedder over those fingerprints. It reproduces the paper's two headline tables.

It runs offline: no model, no API key, no network after one small online step: five abstracts are not redistributable and are rebuilt locally by make restricted-data (sha256-verified; see the redistribution section). Everything else, fingerprints and licensed abstracts, is bundled under data/, so the headline reproduces from scikit-learn alone.

Average precision (AP): how cleanly the true cross-domain twins sit at the top of the ranked list (1.0 = every twin above every non-twin; about 0.04 = random at this 3.6% prevalence).

The faceted fingerprint embedded with a plain bag-of-words (TF-IDF) is the strongest method, more than double the abstract. The dedicated scientific embedders (SPECTER, SciNCL, SemCSE, and SPECTER2, abstract AP 0.182 via src/specter2_local.py) all fall below the plain abstract+TF-IDF baseline, because they encode topical and citation similarity, the wrong signal for cross-domain computational matching. Full AUROC and recall@K columns, plus the extended/wild-run table, are written to reproduce_out/.

Benchmark composition. The headline above is the 109-paper benchmark: 18 method families, among them the documented cross-domain imports of the paper's introduction (inverse Ising/Potts, self-exciting Hawkes, and optimal transport as families; matrix completion and data assimilation as members of the low-rank and Kalman families). Future versions are released as tags of this repository and as versions of the RDR dataset under the same DOI. All three distillers cover the full 109 papers, so reproduce.py's faceted Haiku / Opus / qwen3-14b rows (AP 0.557 / 0.533 / 0.396) are the three-distiller robustness comparison reported in the paper. Labels follow the label-by-core audit (FAMILY_AUDIT.md; every benchmark paper is labelled by the computational core its fingerprint actually computes, not by its field or its planted role), giving the canonical 210 twins; the full dataset specification is DATASET.md.

Python 3.12+ (developed and verified on 3.14; .tool-versions pins the verified interpreter for mise/asdf users, but any recent CPython works).

make setup        # one time: create .venv and install requirements
make reproduce    # print both tables, write reproduce_out/

or without make:

pip install -r requirements.txt
python reproduce.py

The first run downloads the SOTA embedder checkpoints (SPECTER, SciNCL, SemCSE, Qwen3-Embedding) from HuggingFace to print their baseline rows; any that cannot be loaded are skipped and the rest of the table still prints. The "ours" rows (the headline) need only scikit-learn, no download.

• Curated benchmark (complete labels within a closed candidate pool): the place to compare methods. AP is the headline; AUROC is kept only for continuity (at low prevalence a high AUROC is nearly free).
• Extended / wild run (incomplete labels): this is detection, not a scored benchmark. The honest metric is recall of the known twins (it ignores the unlabelled background); measured precision there is a lower bound (an unlabeled true pair counted as wrong can only lower it), but AP is not a score, since missing positives shift both the ranked hits and the normalizing positive set; the extended table is shown for relative ranking only.

same-problem-different-field/
  reproduce.py                   one command -> both tables
  data/                          the BUNDLED, reproducible corpus (derived artifacts only):
    skeletons_faceted_haiku/       fingerprints, our config (Haiku distiller)          [501]
    skeletons_v1_opus/             fingerprints, Opus distiller (comparison)           [501]
    skeletons_faceted_qwen_v3/     fingerprints, Qwen3-14b distiller (comparison)      [501]
    md/                            abstract-only Markdown per paper (title + abstract) [496 bundled;
                                   5 rebuilt locally by make restricted-data, see data/restricted.csv]
    skeletons_keepdomain_qwen/     keep-domain ablation arm (qwen3 distiller)          [109]
    vanco/                         the vancomycin PK example data (gendata.csv, genmodel.txt)
    manifest.jsonl                 per-paper provenance: source URL, license, field, family, role
  datasets/                      the link lists, to rebuild the full-text corpus:
    mode_a_seed_families.csv       curated benchmark (labelled families)
    mode_b_discovery.csv           the wild run (unlabelled arXiv background)
  src/                           fetch + convert + distill + embed code
  reproduce_out/                 the expected tables (regenerated by reproduce.py)
  DATASET.md  FAMILY_AUDIT.md   the benchmark specification and the label-by-core audit
  MODELS.md                      model identifiers, prompt versions, and run dates per distiller/annotator arm
  requirements.txt  Makefile  example.env  SHA256SUMS

This package ships derived artifacts plus abstracts only where a license covers them:

• the fingerprints (stored in the skeletons_faceted_* directories, named so for historical reasons): domain-stripped, method-name-stripped LLM re-descriptions of each paper's computation. They do not reproduce the source text. (Terminology: the full fingerprint is a mechanism skeleton, the MECHANISM field alone (AP 0.513), plus its controlled facets (together AP 0.557); 'skeleton' on its own always means the MECHANISM field, not the whole fingerprint.)
• an abstract-only Markdown per paper (title + abstract; never the full body), bundled for 496 of the 501 papers under the licenses that permit it: abstracts of arXiv-hosted papers are descriptive metadata under arXiv's CC0 1.0 grant (https://info.arxiv.org/help/api/tou.html, which defines descriptive metadata to include the abstract and permits storing and sharing it, independent of the e-print's own license), and a small PMC subset is CC BY (each file's source URL and license are in data/manifest.jsonl; attribution travels with those records).
• five abstracts are not redistributable and do not ship (two publisher pages with no license, one IEEE paper, and two PMC NC/ND entries handled conservatively). data/restricted.csv records their ids, source URLs, and the sha256 of the frozen bytes; make restricted-data rebuilds them locally (the one online step) and verifies the checksums, and reproduce.py refuses to run on a silently smaller corpus. If a publisher page has changed, the frozen files are available from the author for verification.
• data/vanco/ is LAPKB's own NPAG example (distributed with their open-source Pmetrics solver, https://github.com/LAPKB): the example data (an oral-rifapentine study, 20 subjects, data credited by LAPKB to Chuck Peloquin) and model, plus NP_RF0001.TXT, LAPKB's completed NPAG run report on that data (engine 1.8, Oct 2015), from which pk_npml.py reads NPAG's population means and support count for the comparison row. Included unchanged; the folder name reflects the paper's dosing case, not the example drug.

The full-text corpus is not redistributed. Rebuild it locally from the link lists with make data (fetches arXiv + open URLs; paywalled or non-arXiv rows skip cleanly). Each paper's source URL and license are recorded in data/manifest.jsonl.

The curated benchmark is 109 papers, all bundled as fingerprints, so the curated headline (faceted fingerprint + TF-IDF, AP 0.557) reproduces exactly once the five restricted abstracts are rebuilt; the extended/wild run reproduces from the bundled 501-paper fingerprint set.

Each surfaced import is carried end to end by a self-contained, offline script (this is the paper's import section, made reproducible); expected output is in reproduce_out/:

python src/pk_npml.py            # clinical NPAG dosing            -> open nonparametric-MLE / EM (vancomycin)
python src/causal_mc.py          # causal-panel fixed effects      -> recsys SoftImpute completion (69% lower RMSE)
python src/kriging_gp.py         # geostatistics kriging           -> ML Gaussian process (same BLUP + uncertainty)
python src/needleman_wunsch.py   # comp.-linguistics edit distance -> bioinformatics Needleman-Wunsch (wild pair)
python src/import_surfacing.py   # the retrieval half: the fingerprint links each pair across fields

Four imports across four computational cores (sparse-mixture MLE, low-rank completion, Gaussian-process regression, sequence alignment), each reproducible with numpy/scikit-learn only. import_surfacing.py shows the fingerprint ranks the recsys factorization paper 3rd of 500 cross-field neighbors of the ideal-points paper (abstract cosine 0.03), the retrieval a topical embedder misses.

You do not need either of these to reproduce the tables.

• Rebuild the full-text corpus from the link lists (network needed):

  make data
  

• Regenerate the fingerprints with your own model (reproduces the method, not the exact bundled skeletons). reproduce.py reads the three fixed directory names below; re-distill into them with --force to compare your own distiller:

  # the eight-facet prompt that produced the bundled Haiku and Opus arms (see MODELS.md):
  python src/distill_v1.py     --backend ollama --model qwen3:14b --outdir data/skeletons_faceted_haiku --force
  # the twelve-facet prompt that produced the bundled qwen3 arms (adds four verifiability facets the distance ignores):
  python src/distill_faceted.py --backend ollama --model qwen3:14b --outdir data/skeletons_faceted_qwen_v3 --force
  python reproduce.py
  

  Distillation needs a model (local via ollama, or the Anthropic API via .env; see example.env); the bundled fingerprints need neither. The paper's Haiku/Opus skeletons were produced by agents, not a hosted API call, so the bundled fingerprints are the authoritative ones.

Each paper is distilled once (cached) into a faceted computational fingerprint: a free-text "mechanism skeleton" that says what the paper computes and how, with the domain vocabulary and canonical method names removed, plus a few controlled computational facets (structure or motif, data object, inference, problem form, outcome distribution, complexity). A tunable distance over the fingerprint retrieves candidate twins; a cheap TF-IDF over the whole fingerprint is already the strongest ranker. The distillation prompt is in src/, and the method and analysis are in the paper.

sha256sum -c SHA256SUMS                 # 2181 files, all OK
python src/name_free_audit.py           # reproduces the anti-cheat name-free rates

Each skeleton describes its computation without naming the canonical method, so a cross-field match is abstraction, not name recognition. name_free_audit.py ships the pinned name lists and reports the name-free rate under each: 93% against the method names the distiller is told to strip (Kalman, Gaussian process, SVM, EM, HMM, PageRank), 88% if every core's canonical name is added, 82% with generic labels (PCA, Fourier, MCMC) too. Expected output is in reproduce_out/name_free_audit.txt.

Every number reported in the paper is regenerated by a shipped script in src/: the AP grid by reproduce.py and full_grid.py, the bootstrap confidence intervals by ap_bootstrap.py (with the family-level cluster bootstrap by family_bootstrap.py), the precision/recall frontier by facet_ops.py, the domain-stripping ablation by abl_109.py (with distill_faceted.py --keep-domain regenerating the keep-domain arm), the E5-large-v2 baseline by e5_grid.py, the SPECTER2 baseline (local model + proximity adapter) by specter2_local.py, the held-out-family generalization test by held_out_family.py, the three-distiller comparison by distiller_subset.py, the STRUCTURE-none skip rule report by nonmath_reject.py, the curated AP with the ten no-method/no-method pairs dropped by curated_skip_ap.py (the paper reports the all-pairs, conservative figures), the interventional perturbation test by perturbation_score.py, the wild three-arm study (detection and discovery views) by wild_three_arm_score.py, the two-member-family robustness by perfam_sens.py, and the per-family / P@1-MRR / facet-agreement / DOMAIN-ablation diagnostics by perfam.py, p1mrr_boot.py, facet_agree_ap.py, and nodomain.py.

This repository is the public record of this work: the full method, every reported number, and everything needed to reproduce them offline (after the one-time make restricted-data rebuild of the five non-bundled abstracts). A frozen copy (release v1.0.0) is deposited in the KU Leuven Research Data Repository (RDR). Please cite both:

• Paper: Eryk Kulikowski. 2026. Same Problem, Different Field: Cross-Domain Solution Import via Domain-Stripped Computational Fingerprints. In The 2026 ACM/IEEE Joint Conference on Digital Libraries (JCDL '26). ACM. https://doi.org/10.1145/3805696.3846016
• Dataset: Eryk Kulikowski. 2026. Replication Data for: Same Problem, Different Field - Cross-Domain Solution Import via Domain-Stripped Computational Fingerprints. KU Leuven RDR. https://doi.org/10.48804/W3B9WC

Code: Apache-2.0 (the LICENSE file covers the code only). Derived data produced by this project (the fingerprints in data/skeletons*, the manifest, the link lists, the validity artifacts): CC-BY-4.0. Third-party text keeps its original terms; this package grants no rights over it: abstracts of arXiv-hosted papers are bundled under arXiv's CC0 metadata grant, the PMC CC-BY abstracts under CC BY with the source recorded per paper in data/manifest.jsonl, and five abstracts are not redistributed at all (see the redistribution section). data/vanco/ is LAPKB's NPAG example data, credited above.