Skip to content

Repository files navigation

DecPort

DecPort tested whether input-specific probabilistic decision behavior can be transferred across heterogeneous frozen language-model backbones without target labels, and whether a particular learned frozen DecisionCore contributes reusable functionality after the backbone change.

Final scientific outcome: NO-SHIP for the original reusable learned-core hypothesis.

The preregistered nonlinear final gate did not support reusable learned DecisionCore portability in the tested setting. Separately, label-free input-specific decision-behavior distillation remains supported within that setting. Choice and Score carry most of the positive transfer result; Noul did not recover input-specific source behavior in either accepted stage. OOD effects are smaller and calibration is weaker.

Paper

Distinguishing Cross-Backbone Decision-Behavior Transfer from Frozen-Core Portability

Lorenzo Suffritti
Technical research article, 24 September 2026

DOI: 10.5281/zenodo.22947927

Research conducted with the support of artificial intelligence (AI).

The article covers the full experimental progression through decision 0007. Citation metadata is available in CITATION.cff, and the paper-specific overview is in paper/README.md.

Evidence at a glance

Question Result
Input-specific cross-backbone behavior transfer (RQ1) Supported within the tested setting
Reusable learned DecisionCore portability (RQ2) Not supported
Final preregistered ship gate NO-SHIP
Audit Passed
Same-environment reproduction Bit-exact
Independent or cross-hardware replication Not performed

RQ1 compares correct-teacher transfer with a decision-type- and width-matched mismatched teacher. RQ2 asks the stronger question: whether the particular learned frozen core contributes reusable functionality beyond a matched arbitrary frozen core under the same teacher signal. Evidence for RQ1 does not imply evidence for RQ2.

Accepted evidence

Decision 0006 — Open-Jev linear DecisionCore

The first accepted stage used the real released Open-Jev 2B source system, its scalar Linear(2048 -> 1) core, five seeds, and three heterogeneous frozen targets. Label-free behavior transfer was positive overall and was carried by Choice and Score. Noul failed. The matched random core comparison also exposed the linear-core identifiability problem: the experiment provided no evidence that the particular pretrained rank-one core mattered.

Decision 0007 — nonlinear final ship gate

The preregistered final gate used a learned nonlinear 1,118,977-parameter source DecisionCore, a weak rank-128 target adapter, a scale-matched random nonlinear core, mismatched-teacher and untrained controls, and a target-specific distilled baseline. Criteria A–E tested whether the learned core mattered, transfer was input-specific, effects survived distribution shift, external behavior cleared the JevBench public-subset threshold, and the shared-core route had practical utility. Criteria A, D, and E failed, producing NO-SHIP. The audit passed and a second execution in the same hardware/software environment was bit-exact.

The run and reproduction provenance both record the frozen protocol commit d0aa96f4defc3ce202996be1f75d669b791d57fc and git_worktree_dirty:true. The accepted archive verifies the frozen repository configuration and data hashes, but it is not described as a clean-checkout guarantee.

Research and evidence map

RESEARCH.md maps decisions 0001–0007 and makes the progression from supervised transfer to the final falsification gate explicit. The repository preserves three evidence levels:

Acceptance certifies the archived measurement, including its negative findings; it does not extend the claim beyond the recorded models, datasets, tasks, and training budget.

Inspect and reproduce

Install the development and Open-Jev dependencies and run the ordinary package checks:

uv sync --extra dev --extra openjev
uv run ruff check .
uv run pytest

Real-model tests are opt-in because they download checkpoints:

DECPORT_RUN_REAL_MODELS=1 uv run pytest tests/test_real_backbones.py tests/test_real_openjev.py

The final report and verdict can be inspected without rerunning the expensive experiment. The accepted final-gate archive README records the exact data preparation, full run, audit, comparison, JevBench public-subset, and mechanical-verdict commands. The archived evidence includes configs, data manifests, per-seed results, aggregate tables, artifact hashes, audit failure-path probes, and reproduction results. Base-model checkpoints and downloaded datasets are intentionally not committed.

The v0.1 Python package remains a research implementation. Its basic API exposes frozen backbone wrappers, adapters, typed Choice/Noul/Score decision cores, experiment runners, and audit scripts. No DecPort v0.1 software release was made.

Scope and provenance

The conclusion is bounded to three target families, one Open-Jev/Qwen source representation boundary, the archived datasets and budgets, one nonlinear core design, and the preregistered criteria. It does not prove mathematical equivalence between learned and random cores, and it does not imply that nothing transferred.

Historical evidence remains in Git so the commit SHAs cited by the article stay valid. Future large publication artifacts should preferably use release assets or an archival deposition rather than rewriting existing provenance.

The code is licensed under the Apache License 2.0.

About

[RESEARCH ARTIFACT REPO] Reproducible research on cross-backbone probabilistic decision transfer and learned DecisionCore portability. DOI: 10.5281/zenodo.22947927

Topics

Resources

Stars

1 star

Watchers

1 watching

Forks

Releases

Contributors

Languages