| version | 0.1.0 |
|---|---|
| created | 2026-03-09 |
DINOv2 frozen embeddings can detect "wrong artist" misattributions (p=2e-11) but cannot distinguish master from skilled imitator — the commercially valuable question. Ceiling confirmed at 63.7% balanced accuracy after exhaustive 3-tier search. Project complete.
Feb 26 – Mar 9, 2026 (~12 days)
A CLI pipeline (scan.py) that fetches paintings from 6 museum APIs + Wikidata SPARQL, computes DINOv2 vision embeddings, and runs unsupervised + supervised classification probes. 1,311 paintings across 4 attribution levels (autograph, circle, pupil, dutch_other). 78 tests, CUDA/MPS portable, GCE T4 deployment scripts.
| Task | Result | Verdict |
|---|---|---|
| Autograph vs pupil/other Dutch | p=2e-11, p=3e-10 | Strong — reliably detects different artist |
| Autograph vs circle | p=0.80 (unsupervised), 63.7% (supervised) | Weak — can't separate master from skilled imitator |
| Hard negatives (top 25% hardest circle) | 55.1% balanced accuracy, p=0.228 | Dead — no edge on commercially relevant cases |
| Tier | What we tried | Best result |
|---|---|---|
| Tier 1: Frozen features | ViT-B mean (59.0%), ViT-L (59.5%), entropy (63.7%), CLIP (62.9%), concat (62.3%), per-tile voting (60.2%), fine-tune (60.0%) | 63.7% |
| Tier 2: Diagnostics | Confounder audit, robustness, calibration, hard-negatives, domain-shift, multimodal | Signal real but marginal; no step change |
| Tier 3: LoRA + transfer | LoRA Rembrandt (60.9%), LoRA leave-artist-out (crashed), curriculum transfer (59.0%) | All underperform frozen 63.7% |
Art Recognition and Hephaestus train custom supervised models per artist on brushstroke-level features. They don't use frozen foundation model embeddings. Our approach traded training time for signal strength — fast validation, but the ceiling is real.
Training a custom model would require: months of work, significant GPU compute, tile-level brushstroke annotation, and competing head-to-head with companies that have a 7-year head start + art world relationships + (in Hephaestus's case) an insurance product.
- Compute: <$100 (GCE Spot/on-demand T4, ~10 GPU-hours total)
- Data: $0 (all museum APIs are CC0, no keys)
- Time: ~12 days part-time
Complete. Technical exploration exhausted. The question "can frozen vision embeddings authenticate art?" is definitively answered: no, not at commercially useful accuracy. The research, pipeline, and 33-entry decisions log have reference value for future vision projects.
| File | What |
|---|---|
docs/decisions.md |
33-entry chronological decisions log |
docs/scan-results.md |
Full experimental results with statistics |
docs/research.md |
Market opportunity landscape |
docs/execution.md |
Competitor deep-dive, pricing benchmarks, risk analysis |
SPEC.md |
Retroactive build contract |
docs/graduation-tracking.md |
Gate documentation |