Main idea

A document about a competitor arrives on Friday. To update its record, the skill rereads the analyses already produced on that competitor — 102,000 input tokens — when the only new material weighs five thousand.

Spending then follows the size of the history rather than the quantity of new information. Nothing changes in the usage, and the same gesture costs more every month. An incremental mode breaks that coupling: you start from the existing synthesis and from the added items alone. It has a condition — knowing what has already been integrated, and therefore tracing it somewhere, for example in the frontmatter of the record produced.

Why it matters

This separates two cost regimes that are easily confused: proportional to what's new, or proportional to the stock. A skill of the second kind becomes unplayable without anyone having changed anything, by simple accumulation.

The diagnosis therefore has to be made early: this is a property of the architecture, not a setting.

Nuances and limits

Rereading everything is sometimes the work itself: a synthesis that has to rebalance the whole can't make do with the additions.

And incremental mode drifts: each pass inherits the previous one's approximations without ever confronting them with the original material. It calls for a full regeneration from time to time.

Open questions

• How often is a full regeneration needed to catch up on the drift accumulated by an incremental mode?