Source
Adams, Fernandez & Witkowski (2022), LIT-1034, read in NOTE-725: the CNN section, Figs. 4–6, Discussion and Conclusion.
What was actually shown
A convolutional network takes a 3×3 patch of a 10-minute snapshot of the 2017 canvas and outputs the centre pixel 10 minutes, 1, 3 or 6 hours later.
- Prediction worsens as the horizon lengthens (Fig. 4).
- One frame ahead, the reconstructed final canvas is nearly the real one, with mean SSIM 0.95. Six hours ahead, the EU dove, the objects inside the Blue Corner and the central US flag are missing. Objects “growing linearly over time such as the blue corner or the German flags” are partly recovered (Fig. 5, Discussion).
- Short-horizon accuracy dips in the middle of the event and recovers toward the end, while activity stays high (Fig. 6).
The authors give a mechanism that is a reading of the case. Some objects are defined by a simple local rule (recolour a neighbour blue, or black). Others are copies of an external image, whose pixels depend on their position in a design. A local model can learn the first kind. The second kind needs “extra inputs”.
What this does not say
- It is barely tested. There is no copy-the-last-frame baseline, so SSIM of 0.95 at ten minutes may be persistence. There are no error bars, and the split into predictable and unpredictable objects rests on one figure. This is a hypothesis with an illustration.
- It does not measure participants’ reach. The exchange proposed this paper as a start on how far different participants’ reach extends. It models no participants. “Locality” here is a pixel’s spatial and temporal context.
- It does not show that coordination caused the missed objects. The authors infer it from what the missed objects are (flags, emblems, famous images). No coordination data were used.
- The mid-event dip is unexplained. The authors’ reading, that rules are more mixed when objects are forming and competing, is not tested.
Connections
One relation is declared: it corroborates CLAIM-273’s imported-order half, which does not rest on it. The evidence is too thin to support or answer anything else. What it points at:
- QUESTION-084. If local dynamics miss exactly the template-copied designs, then the canvas alone does not reveal coordination by planning. That is consistent with the question’s second horn, unmeasured.
- QUESTION-093. This is the flat model that A8 §5 proposed to beat with latent coalitions. Its authors predict per-coalition models would do better, and they leave it untested. Wu & Silva’s recovered coalitions (THEORY-216) would supply the units.
- CLAIM-279. The Blue Corner and the Void share a rule type and oppose their neighbours. That fits the claim and adds nothing to it.
- Litherland & Mørch (LIT-1037), whose Mona Lisa case this paper invokes, show the off-canvas design work directly (THEORY-218).
Note of 2026-10-29: what holds it back
Read in full (NOTE-725), but its promote_when is unmet. The paper has no persistence baseline, no error bars and no coordination data, and its split into two kinds rests on one illustrated six-hour reconstruction. No test on a crowd-built canvas has compared a local model with a copy-the- last-state baseline and a coordination-aware model. Narrowing to what Adams et al. report would free nothing, since CLAIM-273 now cites it as corroboration.