NRR-Transfer: Cross-Domain Transfer of Phase 1.5 Operators Under Fixed Interface Constraints

NRR-Transfer is a cross-domain package for testing ambiguity-preserving inference outside a single benchmark domain. It evaluates when to avoid premature commitment in LLM decoding, how to reduce semantic collapse across scenarios, and how to apply a practical defer vs commit policy under fixed interface constraints. This repository includes manuscript artifacts together with code/schema/scripts for reproducibility. The goal is measurable transfer behavior across models and temperatures without inflating claims or hiding boundary conditions. The emphasis is controlled comparison and transparent limits, so users can distinguish transportable behavior from setup-specific effects.

Quick links

• arXiv: pending (pre-submission; no public URL yet)
• Positioning (NRR vs related approaches)
• Search Keywords and Weekly Rank Log

EN/JA query terms

• early commitment = 早期確定
• ambiguity-preserving inference = 曖昧性保持推論

Part of the Non-Resolution Reasoning (NRR) research program.

NRR Series Hub (Start here)

For the cross-paper map and current series links, start here:

• NRR Series Hub

NRR is not an anti-LLM framework. NRR does not replace standard LLM use. NRR optimizes when to commit and when to defer, under explicit conditions. Series numbering policy: paper3 is permanently skipped and never reused.

DOI

• Concept DOI (always points to the latest version): 10.5281/zenodo.18793344
• Latest version DOI (as of 2026-02-28): 10.5281/zenodo.18796093

Publication handling

• This repository includes manuscript and reproducibility assets.
• Current manuscript snapshot is stored under manuscript/current/.
• Primary experiment logs for the 324-run protocol are included at data/results/transfer_3trial_results.json.
• Generated final figure PNGs from local reruns are not committed by default.

Current protocol shape

• Protocol: 18 scenarios x 3 models x 2 temperatures x 3 trials
• Re-run script: experiments/run_transfer_3trial.py
• Scenario schema: data/transfer_scenarios.json

Repository structure

nrr-transfer/
|-- README.md
|-- LICENSE
|-- requirements.txt
|-- reproducibility.md
|-- manuscript/
|   `-- current/
|       |-- paper5-nrr-transfer-v38.tex
|       |-- paper5-nrr-transfer-v38.pdf
|       |-- paper5_fig1_horizontal_v2.png
|       |-- paper5_fig2_all_domains.png
|       |-- paper5_fig3_structural_similarity.png
|       |-- paper5_fig4_operator_heatmap.png
|       `-- checksums_sha256.txt
|-- data/
|   |-- transfer_scenarios.json
|   `-- results/
|       `-- README.md              # output policy (pre-submission)
|-- experiments/
|   `-- run_transfer_3trial.py
|-- figures/
|   |-- generate_fig1_horizontal.py
|   |-- generate_fig3_structural_similarity.py
|   |-- generate_figures_from_results.py
|   `-- README.md                  # output policy (pre-submission)

Reproducibility

See reproducibility.md for fixed settings, run commands, and artifact mapping.

Stable review-package entrypoints:

• bash scripts/generate_manuscript_figures.sh
• bash scripts/build_current_manuscript.sh
• bash scripts/verify_current_package.sh

Current figure provenance:

• figures/generate_fig1_horizontal.py regenerates the conceptual horizontal-flow diagram paper5_fig1_horizontal_v2.png with no data-file dependency.
• figures/generate_figures_from_results.py regenerates paper5_fig2_all_domains.png and paper5_fig4_operator_heatmap.png from the bundled 324-run log.
• figures/generate_fig3_structural_similarity.py regenerates the conceptual structural-similarity diagram paper5_fig3_structural_similarity.png with no data-file dependency.
• The stable wrapper writes reruns to /tmp/nrr-transfer_current_figures by default so the tracked review package under manuscript/current/ remains fixed.

Related repositories

• https://github.com/kei-saito-research/nrr-core
• https://github.com/kei-saito-research/nrr-phi
• https://github.com/kei-saito-research/nrr-ime

License

CC BY 4.0. See LICENSE.

Collaboration Style

I support written technical Q&A, concept clarification, and small evaluation design.

Typical flow:

1. you send questions and context,
2. I return a structured technical response,
3. if needed, I provide an English-ready version for external sharing.

Scope: research interpretation and evaluation planning.
Out of scope: production integration, implementation outsourcing, ongoing operations, and SLA/deadline commitments.

Contact

Kei Saito Independent Researcher, Japan ORCID: https://orcid.org/0009-0006-4715-9176 Email: kei.saito.research@gmail.com