Coding / AI Engineering / Context Engineering / billion-context
CHAPTER 10 / 11 · CONTEXT ENGINEERING
Growth Gate
When to ask is not what to compress
The normal gate decides when to ask the model to evaluate compression. The model still decides what is consumed; the kernel validates the resulting operation.
INTERACTIVE STUDY / 10
Local teaching modelTIMING / THEN JUDGMENT
The gate asks.
The model decides.
Follow one normal cycle in a 200K window. Passing the two checks invites evaluation; it does not identify useful information.
At least 45% of this window
PASSSince baseline 90K
WAITCandidate already consumed
Add growth. A normal nudge still leaves semantic judgment to the model.
This isolates floor + the pinned 50K growth interval. Production also checks compressible mass (including the lower effective-growth path), tier mass and hard budgets. The model can initiate compression without a nudge. All displayed counts are scenario values.
THE NOTEBOOK
Concise learning notes · source-grounded mechanismsSeparate the responsibilities
| Kernel / gate | Model + doctrine |
|---|---|
| Track context growth and candidate ranges | Judge whether the current step still needs the detail |
| Ask for an evaluation | Fold a consumed range, defer, or initiate a fold |
| Validate and execute the chosen operation | Write the digest’s useful meaning |
The normal path combines a context floor with growth since the last compression. The pinned kernel uses a 50K growth interval, with a lower effective requirement when pending compressible mass is small, plus tier-specific checks.
A nudge is not an overflow warning and not a command to compress everything. The model may defer during active debugging. After a successful fold, the gate updates its baseline; emergency budget handling is a separate path.
CHECK YOUR INTUITION
A nudge fires in the middle of a critical task. Must the model compress?
Choose an answer to reveal the reasoning.
FOLLOW THE SOURCE
Reviewed source: billion-context 0.1.180 / 03d27f9. The notes are Lawson’s learning interpretation of the reviewed repository.