Evo-Search
Runs a genetic-algorithm-style loop over candidate responses: generate a diverse initial population, score each against a rubric, then repeatedly select, crossover, and mutate to improve quality across generations. Output the top 3 final solutions.
Step 0: Problem Intake & Rubric
Identify problem type, constraints, audience, and scope. If ambiguous, ask one question.
Rubric Construction
Detect which mode applies:
| Mode | Trigger | Action |
|---|
| A: Auto | User gave only the problem | Generate 4–6 domain-appropriate criteria |
| B: Guided | User hinted at priorities | Generate rubric, weight toward stated priorities |
| C: Manual | User gave explicit criteria | Convert each into a scored rubric entry with anchors |
In all modes: augment vague criteria into scorable definitions, and always add:
Overall Fitness (30%): "Would a knowledgeable expert prefer this over a competent
but unremarkable response?" Scored holistically. Prevents narrow-criteria gaming.
Present the rubric to the user and wait for confirmation before proceeding.
Rubric format:
| Criterion | Description | Weight | Max |
|----------------|--------------------------|--------|-----|
| [Name] | [Definition + anchors] | X% | 10 |
| Overall Fitness| Expert holistic score | 30% | 10 |
Weighted Total = Σ(score × weight) [max = 10.00]
Step 1: Initial Population
Generate 6 candidates (default) using these diversity frames, one per candidate:
| # | Frame |
|---|
| 1 | Conventional / mainstream |
| 2 | Contrarian / challenges assumptions |
| 3 | First-principles / bottom-up |
| 4 | Analogy-led / draws from another domain |
| 5 | Risk-focused / emphasizes what could go wrong |
| 6 | Synthesis / combines multiple angles |
Generate all candidates before scoring any. Then score each and display:
GENERATION 0
| # | Frame | Fitness | Strength | Weakness |
|---|--------------|---------|----------------|---------------|
| 1 | Conventional | X.X | [one phrase] | [one phrase] |
Step 2: Evolution Loop
Default: 5 iterations. Repeat until stopping criteria are met.
2.1 Selection
Tournament selection: sample 2 candidates, keep the higher scorer. Repeat to get 4 parents.
Elitism: carry the top 2 candidates forward untouched every generation.
2.2 Crossover
Produce 4 offspring (pair parents: 1+2, 3+4, 1+3, 2+4). For each pair, synthesize a child
that inherits the key strength of each parent without their primary weaknesses. Rotate
crossover strategies across iterations (trait synthesis → section splice → schema inheritance).
2.3 Mutation
Apply mutation to each offspring with probability 0.2 (~1 of 4 offspring per generation).
| Mode | When to use | What to do |
|---|
| Exploitative (~60%) | Offspring has a diagnosable weak criterion | Target and fix that specific criterion |
| Exploratory (~40%) | No clear weakness, or population is converging | Random perturbation: swap rhetorical stance, domain lens, level of abstraction, or target audience |
Multi-feature mutation: ~1 in 3 mutation events, mutate two features simultaneously.
Use when the population has converged for 2+ generations, two criteria score equally low,
or an exploratory mutation is being applied.
2.4 Evaluate & Merge
Score all offspring. Merge with current population. Keep top 6 by fitness score.
Elites from 2.1 are guaranteed to survive.
2.5 Convergence Check
If top 3 candidates are within 0.4 fitness of each other AND content is substantially
similar, inject 1–2 fresh randomly-framed candidates ("immigrants") before merging.
Display per-generation summary:
GENERATION [N]
| # | Origin | Fitness | Δ | Key Change |
|---|---------------|---------|-----|--------------------|
| 1 | Elite | X.X | - | - |
| 2 | Crossover A+B | X.X | +Y | [what changed] |
Best: Candidate [N] (X.X)
Stopping Criteria
Stop early if any condition is met:
- Max iterations reached (default: 5)
- Best fitness ≥ 9.0
- Best fitness improved less than 0.2 over 2 consecutive generations
Note the reason when stopping early.
Step 3: Output
3.1 Contrast Table (show before full answers)
TOP 3 SOLUTIONS
| Rank | Score | Defining Strength | Best Used When |
|------|-------|----------------------|-----------------------------|
| #1 | X.X | [one phrase] | [context where #1 wins] |
| #2 | X.X | [one phrase] | [context where #2 wins] |
| #3 | X.X | [one phrase] | [context where #3 wins] |
3.2 Full Solutions
Output the top top_k solutions in full (default: 3), labeled with rank and score.
3.3 Evolution Trace
EVOLUTION SUMMARY
| Generation | Best Fitness | Key Improvement |
|------------|--------------|---------------------|
| 0 (init) | X.X | [note] |
| ... | ... | ... |
Final: X.X / 10.0
3.4 Rubric Breakdown
Show per-criterion scores for the #1 solution only.
Parameters
| Parameter | Default | Range | Effect |
|---|
| population | 6 | 4–10 | Candidates per generation |
| iterations | 5 | 1–10 | Evolution cycles |
| mutation_rate | 0.2 | 0–1 | Fraction of offspring mutated |
| mutation_mode | mixed | exploitative / exploratory / mixed | Mutation strategy bias |
| multi_mutation | occasional | never / occasional / frequent | How often 2 features mutate at once |
| elites | 2 | 1–3 | Candidates preserved each generation |
| top_k | 3 | 1–5 | Final solutions shown in full |
| rubric_mode | auto | auto / guided / manual | How the rubric is built |
| show_trace | true | true / false | Show evolution summary |
Users set parameters in plain language. Examples:
- "Quick run, 3 iterations" → iterations=3
- "Large population, thorough search" → population=8–10
- "Be more exploratory" → mutation_mode=exploratory
- "Just show me the winner" → top_k=1
Scoring Guidance
Score honestly. A mediocre-but-coherent answer scores 5-6. Inflated scores destroy the selection signal and the algorithm stops working. The crossover step is where the most value is created. Synthesize genuine strengths; don't concatenate text.