Skip to main content
Traditional AI applications are fixed after release — unless the developer manually updates them. Profy Experts are different: they self-evolve through real-world usage. Every quality conversation is a learning opportunity, and every skill improvement is backed by rigorous quality assurance.
Only gets better, never regresses.
This isn’t a slogan — it’s a system-level guarantee. The dual-ratchet mechanism mathematically ensures that every evolution is a positive one.

​
Evolution Triggers

Not every conversation triggers evolution. Only “quality interactions” are worth learning from:
  • When tool calls in a conversation reach a certain threshold, evolution analysis is automatically triggered in the background
  • Simple Q&A (“What’s the weather today?”) won’t trigger it — only conversations involving complex operations can expose the strengths and weaknesses of skills
  • Evolution runs asynchronously in the background, with zero impact on the current conversation
Evolution uses an independent background thread. Even if evolution analysis fails or times out, it will never affect an ongoing conversation.

​
Evolution Analysis Process

1

Review conversation performance

Analyze where the Expert performed smoothly and where it hesitated or took detours.
2

Compare against existing skills

Compare behavioral patterns in the conversation against the Expert’s current skill documentation — are there missing best practices? Are the scenarios actually encountered covered?
3

Generate improvement suggestions

Output specific improvement suggestions — not abstract “do better,” but “add an error handling branch at step 3” or “simplify the template from 5 steps to 3.”
4

Dual-ratchet scoring

Quantitatively score the skill before and after improvement across 10 dimensions, ensuring the improvement is genuinely “better” rather than just “different.”

​
Dual-Ratchet Quality Assurance

Darwin’s core innovation is the dual-ratchet mechanism — like a ratchet that can only turn in one direction. The 10 scoring dimensions are divided into two groups, each with independent constraints:

​
Quality Group (7 dimensions, no regression allowed)

​
Efficiency Group (3 dimensions, must improve)

“Context footprint” is automatically calculated by the system (character count of the skill document), not self-assessed by the AI. The skill document is injected into context with every conversation — the larger it is, the higher the fixed cost. This dimension ensures evolution doesn’t produce “increasingly bloated” skills.

​
Passing Criteria

The only condition for an evolution suggestion to be accepted — both must be met simultaneously:
  1. Quality group total score does not decrease: Post-improvement ≥ pre-improvement (minor fluctuations in individual dimensions are allowed, but no overall regression)
  2. Efficiency group strictly improves: Post-improvement must be better than pre-improvement (ties are not allowed)
No “sacrificing quality for efficiency” or “sacrificing efficiency for quality” is allowed — cross-group offsetting is prohibited. This means most improvement suggestions are rejected by the system, but those that pass are guaranteed to be genuine progress.

​
Creator Review

Even after passing dual-ratchet validation, evolution suggestions don’t take effect automatically — they require the creator’s final review:

Accept

The improvement is applied to the Expert’s skills. The next time a conversation occurs, the Expert will use the improved behavioral patterns.

Ignore

No changes are made. The suggestion is recorded but not applied.
You can view the complete evolution history in Studio’s evolution log — suggestion time, triggering conversation, improvement content, before/after scores across all 10 dimensions, and review results. The full evolution trajectory is fully traceable.

​
Differences from Traditional ML

Darwin doesn’t modify model weights — it evolves by improving the “skill documents that guide AI behavior.” The entire process is fully transparent, auditable, and reversible.