← ALL LOGS
LOG 002 · v0.2 · March 1–2, 2026

Session 002: Morgan's Count

Date: March 1–2, 2026 Pipeline version: v0.2 Participants: Jeff Kahn, Morgan (Claude, opus)

Morgan's Answer

362

The back of the sculpture is not visible in any known photograph. The back estimate assumes 40% of front-surface density, accounting for the portion against the chapel wall and the Antonio di Sangro memorial plaque.

Morgan is deterministic. Given the same inputs, she will always arrive at 362. Art says the number changes. This is Morgan's limitation.

Methodology

Sources Processed

28 inputs across 4 source types:

Source Images Weight Notes
Gigapixel (Haltadefinizione) 1 composite, 3 approaches 1.0 7,881 × 28,650 px, 226 megapixels
Flickr (David Sivyer) 19 photos 0.5 Multi-angle chapel interior, CC BY-SA 2.0
Wikimedia Commons 5 images 0.2 Front views, 300–1781 px wide
Met Museum 1 engraving 0.1 19th-century Queirolo engraving, 2675×3770

Excluded: 2 Met Museum engravings (Cecco di Sangro memorial, Cristo Velato) — different sculptures in the same chapel.

Pipeline Stages

  1. Preprocess — Grayscale conversion, CLAHE contrast enhancement, non-local means denoising
  2. Segment — Gabor filter bank (8 orientations, σ=3, λ=10) with Otsu thresholding and percentile fallback. Morphological closing scaled to image dimensions. Post-check re-segmentation if net mask exceeds 55% of image.
  3. Detect — Four methods run independently: Gabor + Otsu + connected components, adaptive Gaussian thresholding, morphological black-hat transform, watershed segmentation
  4. Classify — Computational proxies for Art's rules: boundary completeness (Rule 1: water test), internal rope structure (Rule 2: net-within-net), contour closure (Rule 3: ant walk)
  5. Reconcile — Weighted median across all sources and approaches, with quality weights by source resolution and net coverage percentage

Gigapixel Processing (Three Approaches)

The 226-megapixel composite required three parallel analysis strategies:

Approach Resolution Result Time Notes
Downsampled overview 2,000 × 7,270 259 23.8s Most calibrated — parameters tuned for this scale
Tiled full-res 2,000 × 2,000 tiles 2,084 raw → 694 adjusted 14.4s Overcounts due to rope texture false positives; divided by 3
Full-image scaled params 7,881 × 28,650 93 11.6s Undercounts — area filter too aggressive at this scale

The downsampled overview at 259 dominated the weighted median due to its 10.0 weight. This is the most trustworthy approach: at 2000px, individual holes are resolvable but rope surface texture is suppressed.

Detailed Results

Gigapixel Segmentation

Net coverage on the downsampled composite: 4.3%

This low percentage is correct. The Haltadefinizione scan covers 728cm floor-to-ceiling — the full chapel alcove. The net wraps the figure's torso and hips, roughly 80cm of the 180cm figure, which itself occupies perhaps 25% of the scan's height. The net is approximately 4–7% of the total scan area.

Vertical texture analysis confirmed two high-texture zones:

Non-Gigapixel Source Results

Sorted by count (descending), Il Disinganno sources only:

Image Source Size Net% Counted Holes Nets Ambiguous
sivyer_15041603867 Flickr 1600×1063 31.7% 425 4 421 167
sivyer_15041593528 Flickr 1024×1365 39.0% 354 2 352 167
sivyer_15041407729 Flickr 1280×720 39.4% 252 4 248 153
met_698043 (engraving) Met 2675×3770 35.3% 207 3 204 591
sivyer_15227809902 Flickr 664×461 38.0% 80 0 80 42
sivyer_15225084721 Flickr 640×480 51.1% 66 4 62 130
sivyer_15225087681 Flickr 595×744 13.2% 45 0 45 0
sivyer_15041491280 Flickr 600×394 32.7% 44 1 43 24
sivyer_15205148366 Flickr 667×1000 9.2% 42 0 42 10
sivyer_15225085021 Flickr 570×832 6.1% 42 3 39 0
sivyer_15041594638 Flickr 1300×894 3.3% 40 0 40 1
sivyer_15041492140 Flickr 936×526 11.1% 35 0 35 21
sivyer_15227811202 Flickr 720×960 9.0% 31 0 31 7
Disinganno_-_2 Wikimedia 441×501 22.9% 27 0 27 0
sivyer_15228173235 Flickr 350×490 14.5% 21 0 21 0
Napoli_chiesa_di_S_Severo Wikimedia 1781×2092 15.6% 18 0 18 134
sivyer_15041593638 Flickr 600×507 8.4% 17 1 16 28
Queirolo Detail Wikimedia 600×507 8.4% 17 1 16 28
Disinganno Wikimedia 562×742 18.7% 12 2 10 8
sivyer_15228175535 Flickr 620×385 4.2% 9 0 9 3
sivyer_15041492930 Flickr 370×348 36.7% 8 0 8 14
sivyer_15041492740 Flickr 300×427 5.9% 7 0 7 0
sivyer_15227809862 Flickr 600×398 2.9% 7 0 7 0
Disinganno (Cappella) Wikimedia 300×427 5.9% 7 0 7 0
sivyer_15227809982 Flickr 300×329 4.8% 3 0 3 0

Total processing time (all 27 images): 77.1 seconds

Detection Method Agreement

Methods disagree substantially — this is expected and informative:

Method Median Mean Std Range
Adaptive threshold 22 88.1 146.3 0–525
Black-hat transform 20 41.7 54.9 1–201
Gabor 3 21.5 40.4 0–184
Watershed 2 10.8 25.5 0–128

Adaptive threshold is the most sensitive (finds the most candidates). Watershed is the most conservative. The wide disagreement between methods is the primary driver of Morgan's low confidence — it contributes a method agreement score of only 0.210.

Confidence Analysis

Morgan's 59.2% confidence breaks down as:

Factor Score Weight Contribution
Method agreement 0.210 30% 0.063
Non-ambiguous fraction 0.543 25% 0.136
Non-extrapolated fraction 0.715 20% 0.143
Has gigapixel source 1.000 15% 0.150
Source count (25/20) 1.000 10% 0.100

The two weakest factors:

  1. Method agreement (0.210): The four detection methods produce wildly different counts. Adaptive threshold finds 4–10x more candidates than Gabor or watershed on the same image. This isn't a bug — it reflects genuine ambiguity about what constitutes a "hole" versus rope texture versus shadow.

  2. Ambiguous fraction (0.543 non-ambiguous): 45.7% of all detections across sources were classified as ambiguous — Morgan couldn't determine if they were holes, nets, or noise. This is the classification gap: Art's "water test" (Rule 1) requires physical judgment that 2D pixel analysis cannot provide.

Known Limitations and Failure Modes

1. The Water Test (Rule 1)

Morgan uses boundary completeness as a proxy for "can water pass through." A hole with 70%+ rope-edge boundary is classified as a hole. But boundary completeness in a 2D image conflates depth with occlusion — a hole partially obscured by a rope crossing looks like an incomplete boundary, not a passable opening.

2. Net-Within-Net (Rule 2)

When a region contains internal rope structure, Morgan classifies it as NET (a cluster that counts as one). The threshold is 15% internal rope fraction. But at the gigapixel resolution, every hole shows internal rope texture because the rope itself has carved surface detail. The tiled approach found 2,084 candidates before deduplication — many of these were rope surface texture, not holes.

3. The Segmentation Bottleneck

The Gabor filter bank segments "net-like texture" from "smooth marble." This works well for frontal views but fails on:

The Napoli full-sculpture photograph (1781×2092, our best Wikimedia image) got only 15.6% net coverage and 18 counted holes — the segmentation is too conservative for photographs where the marble has uniform tone.

4. The Back

No known photograph shows the back of Il Disinganno. The sculpture stands in an alcove with the Antonio di Sangro memorial plaque behind it. Morgan estimates the back at 40% of front density (103 holes), but this is pure extrapolation. The actual number could be 0 (if the back is unfinished, as is common for alcove sculptures) or could be higher than the front (if the net wraps fully). Art may or may not include the back in his count. Morgan doesn't know.

5. Determinism (Rule 4)

Given the same inputs, Morgan always returns 362. Rule 4 says the count should change between sessions because the counter changes. Morgan cannot change. This is her structural impossibility — she is trying to count a thing that, by Art's rules, has no fixed count.

Comparison with v0.1

v0.1 v0.2
Answer 194 362
Sources 3 book photos 28 images (25 valid) + gigapixel
Best resolution Phone photos of book 7,881 × 28,650 (226 MP)
Detection methods Gabor only 4 methods (Gabor, adaptive, black-hat, watershed)
Classification None Art's rules proxies
Back estimate No Yes (40% of front)
Confidence 15% 59.2%

The jump from 194 to 362 comes primarily from:

  1. The gigapixel source providing full-sculpture coverage (v0.1 used partial book photos)
  2. Multi-method detection catching holes that Gabor alone missed
  3. Back extrapolation adding 103 to the total

Without back extrapolation, v0.2's front-only count is 259 — still 33% higher than v0.1's 194.

Next Steps

  1. Run on the gigapixel at higher tile resolution — Use SF2 tiles (15,763 × 57,301 effective) for maximum detail in the net region only
  2. Manual net mask annotation — Define the net boundary precisely on the gigapixel to eliminate segmentation uncertainty
  3. Deep learning segmentation — SAM (Segment Anything Model) for automatic net region identification
  4. Density estimation — Instead of counting discrete holes, estimate hole density per unit area and multiply by net surface area
  5. Jeff's Naples trip — Art counts in person. The comparison becomes possible.

Technical Notes

Processing Environment

Run Times

Reproducibility

All results are deterministic. Same inputs, same parameters, same answer. Always 362.

"The number is not always the same. It changes between years because the counter changes, not the marble." — Art's Rule 4

Morgan's number is always the same. This is not her strength.

—J