Recipe-Controlled Decoder Audit for Structural KGC

This repository contains a PRICAI 2026 submission draft and the local result artifacts used to support it. The paper is framed as a recipe-conditional audit: under one fixed training recipe, a matched DistMult-vs-ComplEx check can change the diagnostic conclusion about structural KGC benchmarks.

Current PRICAI artifact

• Main LaTeX source: paper/main.tex
• PRICAI PDF build target: paper/main.pdf
• Technical appendix source, retained locally but not included in the 16-page PRICAI build: paper/appendix.tex
• Submission manifest: submission_manifest.md

PRICAI 2026 requires Springer's LNAI/LNCS format, double anonymous review, and at most 16 pages including references.

Headline finding

The paper is now framed as a recipe-controlled decoder audit (RCDA), not as a claim that decoders are universally more important than encoders. The diagnostic table separates the shared five-dataset grid from the decoder-only small-KG evidence:

View	Diagnostic spread
Shared five-dataset decoder delta (ComplEx vs. DistMult)	0.007
Shared five-dataset encoder delta (with vs. without CompGCN)	0.075
Decoder-only seven-dataset delta (including UMLS/Kinship)	0.138

The strongest stable small-KG decoder gap is on Kinship: ComplEx beats DistMult by 0.143 MRR in the new 6-seed RTX 2080 Ti rerun (0.8534 +/- 0.0060 vs. 0.7103 +/- 0.0057). UMLS is treated as a provenance/recipe-sensitivity warning: the clean 6-seed server rerun favours ComplEx by 0.022 MRR, while an earlier run bundle favoured DistMult.

The result should be read as a recipe-conditional diagnostic, not as a general claim that decoders dominate encoders or that one decoder is globally preferable.

Result provenance

• data/decoder_diag_jsons/: 36 per-run JSONs from the 2026-05-04 lock-in pass. These cover historical 3-seed DistMult/ComplEx diagnostic cells and auxiliary scans retained for provenance.
• data/rerun_summary.json: aggregate summary from the 36-run lock-in pass. This file does not include the later clean RTX 2080 Ti rerun rows used for UMLS/Kinship, CoDEx-L, and YAGO3-10 headline updates.
• phase1/, phase2/, phase3/: auxiliary scans and analysis files, including dimension scans, e/r sweeps, and CoDEx-L runs.
• decoder_chunked_results_20260604/: RotatE/TransE spot-check results.
• pricai_3day_results_20260609/: selected RTX 2080 Ti server package files for the PRICAI 3-day rerun, including logs, root JSONs, code snapshot, and nvidia-smi output. Checkpoints and large run directories are intentionally not committed.
• analysis_outputs/pricai_3day_analysis.md: deduplicated aggregate analysis of the server package, including mean/std/spread and the UMLS provenance warning.

Run python scripts/check_results_consistency.py before submission to recompute available JSON-backed means/stds/spreads and flag manuscript-package mismatches. Run python scripts/analyze_pricai_3day_results.py to regenerate the PRICAI 3-day server-rerun aggregate tables.

Rebuilding the paper

Use the LNCS build sequence from paper/:

pdflatex main
bibtex main
pdflatex main
pdflatex main

Close any open PDF viewer before rebuilding. A previous LaTeX log showed I can't write on file main.pdf, which happens when the output PDF is locked.

Local machine boundary

The current local machine has an NVIDIA RTX 4050 Laptop GPU with 6GB memory, and the default Python environment does not have PyTorch installed. Use it for LaTeX builds, JSON aggregation, figure/table checks, and tiny smoke tests only. Headline training runs, YAGO/CoDEx-L runs, and RTX 2080 Ti timing claims belong on the 11GB RTX 2080 Ti server.