Experiments & Artifacts (BOSS Vision, roadmap6 revision)
Experiments & Artifacts (BOSS Vision, roadmap6 revision)
All paper numbers are generated from machine-readable files here; none are
hand-copied into LaTeX. The manuscript \inputs results/result_macros.tex and the
figure/table .tex files under figures/.
Evidence taxonomy (roadmap6 §1.2)
| Kind | What | Where |
|---|---|---|
| Construction validity | facts guaranteed by authored gold graphs (e.g. predicate-only / node-only ties, gold oracle 100%) | results/mapping_per_family.json (gold rows), shortlist_full_matrix.json (gold_exhaustive_oracle_p1) |
| Automatic end-to-end | independently extracted representations | results/mapping_per_family.json (canonical / free) |
| Synthetic feasibility | indexed-shortlist probe on seeded synthetic banks + calibrated noise | results/shortlist_full_matrix.json, scaling_fixed_targets.json |
| Future work | learned state Z, natural-corpus scale, consolidation, generation | not implemented (Vision paper) |
Reproduce (in order)
python3 experiments/analysis/run_full_shortlist_matrix.py # Table 4 matrix (+ gold oracle, exhaustive noisy)
python3 experiments/analysis/analyze_paired_deltas.py # relation-aware vs predicate-only paired CIs
python3 experiments/analysis/analyze_fixed_target_scaling.py # nested fixed-target scaling (sublinear test)
python3 experiments/analysis/mapping_per_family.py # frozen 40-item paired/per-family + baselines
python3 experiments/analysis/analyze_structured_noise.py # (optional) noise sensitivity sweep
python3 experiments/analysis/export_result_macros.py # -> result_macros.tex + figures + tables
python3 experiments/analysis/test_ceiling_invariant.py # guard: only the gold oracle may be a "ceiling"
Frozen inputs (do not retune): banks/queries under data/phase3/, the 40-item set
data/mapping_frozen_40.json, the refiner scripts/relation_aware_match.py. Seeds,
noise rates, signatures, grid, and checksums are pinned in
configs/shortlist_frozen.yaml and results/v12_audit.md.
Key outcomes (calibrated, synthetic)
- Relation-aware vs predicate-only shortlist is a small, uncertain gain: refined P@1 @N=5 Δ = +5.0 pts (1K), family-clustered 95% CI [0.0, +13.3], corrects 3 / breaks 0. Predicate abstraction supplies most of the shortlist signal.
- The exhaustive noisy refiner (0.83) is a baseline, not a ceiling — the shortlisted system reaches 0.90; the filter can regularize the noisy refiner. The gold structural oracle (1.00) is the only ceiling.
- Sublinear empirical growth holds only under the nested fixed-target construction with hard negatives: relation-aware log-log slope 0.66, CI [0.45, 0.84] (< 1).
- Frozen mapping test: canonical ranking 87.5 vs surface-matched foils, but edit localization 32.5% ≈ random 33.3% and causal-reversal localization 0% — an honest limit motivating counterfactual hard negatives and learned relational state.
Files
configs/— frozen run configs + checksums.results/— JSON result files,result_macros.tex,v12_audit.md.figures/— pgfplots/TikZ.texfigures & tables (compiled inline by pdflatex; matplotlib is not used).analysis/— the scripts above; each imports the frozen harness verbatim and only ADDS analyses.results/phase3/v12_snapshot/— preserved pre-revision artifacts.