UCI Option Reference¶

Every option Kirin advertises in response to the uci command, with the shipped default. The authority for this table is kirin/src/engine/uci.cpp — if the engine’s own uci output disagrees with what follows, the engine is right and this table is a bug.

Note

Defaults are the shipped configuration, not a recommendation. Several capable techniques default to false because they have not yet cleared an SPRT run; see Ch. Measuring Strength. The option surface exists primarily so those tests are possible.

Search features¶

All are check (boolean) options.

Option	Default	Effect
`Use PVS`	`true`	principal variation search: null-window the non-first moves
`Use Null Move`	`true`	null-move pruning, reduction $2 + \text{depth}/6$
`Use LMR`	`true`	late move reductions past move 4, depth $\geq 3$
`Use Aspiration`	`true`	aspiration windows around the previous score
`Use History`	`true`	history heuristic in quiet-move ordering
`Use SEE`	`true`	sets both SEE options below together
`Use SEE Ordering`	`true`	bucket winning/losing captures when ordering
`Use SEE Pruning`	`true`	skip $\text{SEE}<0$ captures in quiescence
`Use Rich Eval`	`true`	full positional term set rather than the legacy subset
`Use Futility`	`false`	prune quiets that cannot raise alpha
`Use Singular`	`false`	extend a move much better than all alternatives
`Use ProbCut`	`false`	shallow wide-margin search to prove a cutoff
`Use IID`	`false`	internal iterative deepening when no hash move exists
`Use ContHist`	`false`	continuation history (ordering by follow-up)

Use SEE is a convenience that writes both underlying flags. Setting the two individual options is what a gauntlet does, because the halves have very different risk profiles — see §The pruning and reduction surface.

Evaluation¶

Option	Type	Default	Effect
`Use NNUE`	check	`false`	enable neural evaluation
`EvalFile`	string	empty	path to a `.knnue` network
`NNUE Blend`	spin	`0` (0–100)	master gain, percent
`NNUE Residual`	check	`false`	treat the net as a correction, not an evaluator
`NNUE Residual Clamp`	spin	`0` (0–2000)	bound the correction, centipawns; 0 = unbounded
`NNUE Classical Limit`	spin	`0` (0–3000)	drop the correction when $
`NNUE Phase Scale`	check	`false`	fade the correction out as material leaves

Warning

NNUE Blend defaults to 0, and it gates both evaluation modes. Use NNUE true on its own produces exactly the classical score — the network is loaded, evaluated, and then multiplied by zero. See §The one thing that catches everyone.

Time management¶

Option	Type	Default	Range
`Move Overhead`	spin	`25`	0–5000 ms
`Slow Mover`	spin	`85`	10–200 %
`Max Managed Move Time`	spin	`0`	0–60000 ms (0 = no cap)

These apply only to the managed clock form (go wtime/btime/winc/binc). The default of 0 for the absolute cap is deliberate and was worth about 17 Elo — see D-16 in Ch. Time Management.

General¶

Option	Default	Effect
`Skill Level`	`2` (0–2)	0 = Easy, 1 = Medium, 2 = Hard (full strength)
`Hash`	platform maximum	transposition table size in MB

Hash advertises its default and maximum from getMaxHashMegabytes(), so both are platform-dependent rather than literal constants.

Note

Skill Level is the engine’s three-tier setting and is not the user-facing difficulty ladder. The appliance presents named opponents backed by option presets and node budgets (the personality model, D-51), and the seven-band ladder design (D-10, D-11) is a separate weakening layer. Do not conflate the three.