Editing Openai/691a41cd-2efc-800c-9eff-de439224a90d (section)

==== We can also look at the semantic compression ratio—how much meaningful output you get per byte of LoreToken. ====
* LoreToken size: 81 bytes
* Suppose the expanded documentation is ~25,000 characters of output (easily reached with a long article like the one above). That’s ≈ 25 KB of text.

Semantic expansion ratio (output bytes per input byte):

Ratio≈25,00081≈308:1\text{Ratio} \approx \frac{25{,}000}{81} \approx 308:1Ratio≈8125,000≈308:1
So in this example:
* An 81-byte LoreToken guides the model to generate on the order of hundreds of times more human-readable, structured medical content.
* In token terms, ~25 tokens of input are unlocking thousands of tokens of carefully targeted medical explanation.

If the same LoreToken is reused many times (across different questions, contexts, or agents), the effective semantic compression ratio skyrockets, because:
* You only store and transmit the 81 bytes
* The model continuously regenerates rich, domain-specific structure on demand

That’s the core idea: a tiny, stable, machine-native seed driving large, precise expansions of meaning.