The 42 KB context audit
The reader was serving a quarter of the context the model was trained for, and the plan said there was nothing above it worth reaching for. One unit error, one constant, and 12 GB of headroom.
Labels. [MEASURED] — from the live endpoint, a receipt or the audit in the pack. [FACT] — a file you can check. [OPINION] — mine.
The audit is 42 KB: 684 lines, five findings, one recommendation, and a section titled “What I
did not verify” [FACT, llamacpp-context-audit.md]. It exists because one number in the
deployment looked wrong against another number in the code, and the gap between them was 4×.
The reader is a 35B model served by llama.cpp on an L40S. The live endpoint said:
| field | value |
|---|---|
default_generation_settings.n_ctx |
65,536 — per slot |
total_slots |
4 |
n_ctx_train |
262,144 |
| model | Qwen3.6-35B-A3B, Q4_K_M, 20.4 GB |
[MEASURED, the live /props and /v1/models]
This piece is part of the members' archive — $8 a month or $80 a year.
$8 / month · $80 / year
Sign in if you are already a member, or create an account.