Sigil — a message code that fades instead of breaking

A short message, turned into a small glyph by a neural encoder and read back by a shared neural decoder — trained together through simulated real-world damage. The result is a visual code that fades instead of breaking: cover part of it, blur it, add glare, and most of the message still comes through. It's the Latent Radio joint-coding principle applied to the channel a printed or photographed code actually suffers.

Why this is new. Today's tool for this is the QR code, whose algebraic error-correction has a hard wall — past its limit, a damaged QR decodes to nothing (a cliff). Sigil has no wall: because sender and receiver share a learned code that spreads the message across the whole glyph, accuracy slides down smoothly. Honest scope: it carries a short message (32 bits — a code, an ID, a few characters), not an essay. Information theory still applies; what changes is how it fails.

It fades, it doesn't break

93%of the message still readable at 70% damage

65%a decoder not trained through damage (≈ coin-flip)

The same Sigil at 0%, 30%, 50%, 70%, 90% damage (noise + blur + occlusion + brightness shift). Even heavily corrupted, the decoder still reads it.

Graceful degradation — bit accuracy vs. damage

Sigil (channel-aware)no-channel ablation (trained on clean glyphs)

The Sigil's accuracy declines smoothly across the whole damage range. The ablation — an identical network trained only on clean glyphs — falls apart as soon as any real damage appears, which proves the robustness comes from training through the channel, not from the architecture. This is the same joint source-channel coding idea behind Latent Radio, now protecting a human-shareable code.

Whole-message recovery

Stricter test: every one of the 32 bits correct (a perfectly recovered message). Sigil recovers the full message reliably up to ~40% damage, then degrades — while the ablation never survives any damage.

Sigilno-channel ablation

What this is / isn't

Is: a real, trained-from-scratch visual code that degrades gracefully where a QR code cliffs — a new, honest building block for getting a short message through a hostile visual channel (bad prints, glare, partial cover, low-quality photos). Isn't: unlimited compression or a replacement for all of QR (QR carries far more bits and is a mature standard). It's one specific superpower — meaning that survives damage — proven on 32-bit messages with small models.

Reproduce: python3 experiments.py && python3 make_report.py. Next: an in-browser demo — type a message, watch the Sigil form, damage it with a slider, watch it still decode, side-by-side with a QR code that dies.