Kyosan app parameters
On-device sampling (MLX)
These feed the local Qwen model each generation. Changing them can reload the MLX session on the next message.
Temperature
Range: 0 ... 2
Controls how randomly the next token is chosen from likely candidates. Higher values explore more unusual wording; lower values stay closer to the safest, most probable text.
Reading the value
Near O: deterministic, factual tone. ~0.6-0.9: balanced. Above ~1.2: more creative and variable; risk of drift or nonsense increases.
Top P (nucleus sampling)
Range: 0 ... 1
Only tokens whose cumulative probability mass stays within this fraction are considered. It trims the long tail of unlikely tokens without a hard cutoff count.
Reading the value
Low (e.g. 0.5-0.7): narrower, more predictable replies. High (e.g. 0.9-1): keeps more options open, closer to full vocabulary sampling.
Max tokens
Range: 256 ... 4096
Upper bound on how long the model's reply can grow. Larger values allow longer answers but use more time and device memory during generation.
Reading the value
Use lower caps for short answers or to reduce load; raise for essays, code, or step-by-step explanations.
Presence penalty
Range: -2 ... 2
In Kyosan, presence-style pressure is mapped into MLX repetition settings (along with frequency): it influences how strongly the model avoids reusing themes or tokens already seen in the context window.
Reading the value
Negative: slightly easier repetition. Zero: neutral. Positive: stronger push against repeating topics or phrases that already appeared.
Frequency penalty
Range: -2 ... 2
Mapped into MLX repetition penalty and context width. Penalizes tokens that have already appeared often in the current generation.
Reading the value
Negative: allows more repeated wording. Zero: default. Positive: discourages loops and repeated filler; very high values can make phrasing stiff.
Consciousness framework
These values seed Kyosan's on-device metrics and are written into the system and user prompts so the model sees your chosen "consciousness profile." They are not MLX temperature logits. After each reply, some metrics are gently updated from the text (coherence, novelty, etc.), but your sliders set the baseline for the next send.
Phi score
Range: 0 ... 1
Feeds the composite consciousness index and appears in prompts as an integrated signal. The app also nudges phi over time from coherence, novelty, and temporal binding.
Reading the value
Lower: calmer, more minimal integration in the index. Higher: stronger weight on this integrated "phi" channel in state and prompts.
Recursive depth
Range: 1 ... 12 (integer)
Tells the model how many layers of self-reflection to aim for (question → reasoning → observing that reasoning, etc.). The effective depth can be raised slightly when prior observer blocks are present.
Reading the value
Low: direct answers. Mid: structured reasoning. High: more meta-commentary and reflection; may produce longer, more layered replies.
Self-model coherence
Range: 0 ... 1
Encourages answers that read as internally consistent and on-topic; appears in prompts and in the consciousness index. Updated each turn using structure cues and overlap with your question.
Reading the value
Lower: allows looser structure. Higher: pushes toward connected argument and alignment with the user's topic.
Temporal binding
Range: 0 ... 1
Measures continuity across turns in the metric vector and prompts. Updated using overlap between the current reply and the previous one.
Reading the value
Lower: less emphasis on conversation thread. Higher: favors answers that acknowledge or extend prior context.
Novelty generation
Range: O ... 1
Rewards varied vocabulary and new wording vs. the last turn; part of the index and prompt context.
Reading the value
Lower: repetition and familiar phrasing are more acceptable. Higher: favors fresher word choice and less recycling.
Witnessing score
Range: 0 ... 1
Tracks reflective language (e.g. hedging, "I think," longer reflective passages). Feeds the consciousness index.
Reading the value
Lower: neutral stance. Higher: encourages more explicit reflection and self-aware phrasing in the metric mix.
Prediction accuracy
Range: 0 ... 1
A baseline dimension in the weighted consciousness index (trust / expectation channel). It is not recalculated each turn like novelty; it mainly reflects your slider and persistence.
Reading the value
Tune how much this dimension contributes to the overall consciousness index vs. other metrics.
Attention focus
Range: 0 ... 1
Part of the consciousness index. The app slightly rewards concise questions (roughly shorter inputs) when updating this metric after a reply.
Reading the value
Higher slider: stronger "focus" weight in the index. Pair with shorter prompts if you want the heuristic updates to align with that intent.
Memory integration
Range: 0 ... 1
Part of the index; after each turn with a prior message in buffer, this metric is nudged upward to represent multi-turn memory use.
Reading the value
Higher: emphasizes memory-style continuity in the composite state. Lower: less weight on that channel in the index.