The AI-native internet.

The internet was built for humans. AI agents still consume it through human representations — HTML, Markdown, PDFs, transcripts — and every agent repeats the same expensive work: fetch, parse, read, understand, reason. That work is duplicated millions of times a day.

symbol proposes that there is a much smaller semantic representation of a source that frontier models can reliably reconstruct. Understand it once; share it forever. An agent downloads a 400-character artifact instead of a 40,000-character page, reconstructs the knowledge, and answers any question.

Why this isn't arithmetic coding

The decoder is an arbitrary set of models. That rules out logit-level arithmetic coding, which is model-specific and brittle to provider and quantization drift. So the artifact is text any capable model expands back into the source's claims — semantic compression, scored on agreement of meaning, never on bytes.

The metric: L95

Ground truth is a frozen set of weighted atomic claims extracted from each source. An artifact is decoded independently by a panel of frontier models; a judge scores how much of the claim mass each recovers. L95 is the minimum number of characters needed to recover 95% of the weighted claims across the whole panel — reported worst-case, with a held-out model for generalization. As models improve, L95 should fall.

Honest by construction

Every score is net of the prior — what the panel recovers from the title alone. If that lift is near zero, the “compression” was really retrieval from memorized data. The benchmark corpus is therefore built entirely from invented, contamination-safe sources, so recovery measures compression and nothing else. Every run also reports the full-source ceiling and identifier-only floor, as a frontier — never a single ratio.