Ideate

Ideate is a loop-controlled evolutionary creativity engine. It runs multiple cycles of consuming, dreaming, stealing, breeding, and testing ideas over simulated time scales from hours to decades, driven by a first-class Loop Controller and a Lamarckian Meta-Learner. It produces ranked novel solution candidates with full provenance — where each idea came from and how it evolved.

A single-pass brainstorm collapses fast. Ask a model for ideas and it converges on the obvious handful, biased toward its training distribution, because soft temperature tweaks just reshuffle the same probability mass. You get variations on one theme, not genuinely different directions. Hard problems need ideas that came from somewhere else — a foreign domain, an unexpected recombination, a constraint flipped on its head — and they need a way to kill the weak ones and breed the strong ones across many rounds. One pass can't do that.

This is an evolutionary system, not a single-pass tool. This is NOT BeCreative — BeCreative is a single-pass diversity tool. Ideate runs multiple cycles driven by a Loop Controller and a Lamarckian Meta-Learner.

Human creativity reduces to 5 irreducible functions:

The 9 workflow phases expand these into a richer human-legible system. DREAM, DAYDREAM, and CONTEMPLATE are PERTURB at different noise levels. MATE is PERTURB on existing ideas. META-LEARN adds the Lamarckian advantage — analyzing WHY ideas worked and steering future generation.

Post-loop: The Historian runs the Insight Extractor for cross-cycle pattern analysis.

Full phase mechanics live in Workflows/FullCycle.md.

Owns inter-cycle state and makes continue/pivot/stop decisions after each cycle's META-LEARN phase. State tracked:

{
  "cycle_count": 0,
  "max_cycles": null,
  "budget_seconds_remaining": 600,
  "fitness_history": [{"cycle": 1, "avg_score": 52.3, "top_score": 68.1, "diversity_index": 0.91}],
  "stagnation_counter": 0,
  "strategy_version": 1,
  "strategy_adjustments": {},
  "loop_decision_log": []
}

Loop Gate logic:

IF budget_seconds_remaining <= 0:        STOP (budget exhausted)
ELIF stagnation_counter >= 3:
    IF strategy_pivots_remaining > 0:    PIVOT (shift domains/noise/agents)
    ELSE:                                STOP (exhausted strategies)
ELIF diversity_index < 0.3:              PIVOT (collapse — inject immigrants)
ELIF top_score >= target_score:          STOP (target reached)
ELSE:                                    CONTINUE

LLM "temperature" is soft probability redistribution biased toward the training distribution. Ideate uses structural randomness at the data level instead:

• Input subsetting (DREAM): Fisher-Yates shuffle picks each agent's input subset
• Domain lottery (STEAL): weighted random sampling from the 50+ candidate domain pool
• Pairing shuffle (MATE): Fisher-Yates pairs adjacent items; 20% slots forced cross-phase
• Mutation dice (EVOLVE): roll an 8-sided die, apply that mutation operation:
  1. Flip one assumption
  2. Invert the constraint
  3. Change the scale (10× bigger or smaller)
  4. Change the time horizon
  5. Merge with a random killed idea's best element
  6. Apply a constraint from a random domain
  7. Remove the most complex component
  8. Add an adversarial requirement

Implementation: crypto.getRandomValues() with seed = cycle number + problem hash.

Optional pluggable interface that adds real-world signal to internal scoring:

interface ValidationHook {
  name: string;
  validate(idea: Idea, problem: Problem): Promise<{ modifier: number; evidence: string }>;
}

Built-in hooks: MarketSearch (existing implementations), FeasibilityCheck (technical blockers), ExpertPanel (async human review), PrototypeSimulation (generate + test prototype).

Loop Controller decides actual cycle count adaptively, not a fixed count.

Each run persists to ~/.claude/PAI/MEMORY/WORK/{slug}/ideate/:

ideate/
  config.json           # Problem, time_scale, domains, hooks
  loop-state.json       # Loop Controller (fitness_history, strategy, decisions)
  domain-pool.json      # Weighted domain pool (expanded across cycles)
  cycle-NNN/            # Per-cycle artifacts: input-pool, dreams, daydreams,
                        # analyses, checkpoint-a, stolen, offspring, scores,
                        # checkpoint-b, survivors, meta-learning, summary
  insights.md           # Insight Extractor output (post-loop)
  final-output.md       # Ranked candidate list with full provenance

{
  "id": "idea-042",
  "text": "...",
  "provenance": {
    "parents": ["idea-017", "idea-023"],
    "operation": "crossover",
    "mutation_type": "scale_change",
    "mutation_die_roll": 3,
    "cycle": 3, "phase": "MATE",
    "source_domains": ["mycology", "distributed-systems"],
    "randomness_seed": "a7f3c9..."
  },
  "scores": {
    "feasibility": 72, "novelty": 88, "impact": 65, "elegance": 81,
    "composite": 76.5, "confidence": 0.82, "judge_variance": 8.3,
    "external_validation": {"market_search": {"modifier": -5, "evidence": "..."}},
    "adjusted_composite": 74.5
  },
  "arguments": {"supporting": "...", "counter": "..."}
}

# Ideate Results: [Problem]

**Time scale:** [scale] | **Budget used:** X of Y min | **Cycles:** N (adaptive)
**Strategy pivots:** M | **Total ideas:** X | **Survived:** Y | **Kill rate:** Z%

1. [Title] — Score: 85.2/100 (confidence: 0.91)

The idea: [2-3 sentences] Scores: Feasibility: 78 | Novelty: 92 | Impact: 84 | Elegance: 87 External validation: [hook results] Provenance: Born in cycle N from [operation] of [parents]. Mutation: [type]. For it: [supporting argument] Against it: [counterargument]

• [How strategy evolved across cycles]

• [Dominant lineages, fertile combinations, fitness landscape, problem revelations]

{
  "problem": "...",
  "time_scale": "weeks",
  "domains": ["primary", "adjacent-1", "adjacent-2"],
  "scoring_weights": {"feasibility": 1.0, "novelty": 1.0, "impact": 1.0, "elegance": 1.0},
  "convergence_prevention": {
    "cross_phase_breeding_min": 0.2,
    "immigrant_ideas_per_cycle": 3,
    "kill_threshold": 0.5,
    "forced_new_domain_per_cycle": true
  },
  "loop_control": {
    "mode": "adaptive",
    "target_score": null,
    "max_stagnation_cycles": 3,
    "max_strategy_pivots": 2,
    "diversity_floor": 0.3
  },
  "external_validation": {"enabled": false, "hooks": ["MarketSearch"]},
  "randomness": {"seed": null, "subset_ratio": 0.33, "mutation_operations": 8}
}

When the PAI Algorithm sets mode: ideate (via PAI/ALGORITHM/ideate-loop.md), it loads this skill and routes to Workflows/FullCycle.md by default. Tunable parameters from the algorithm's parameter-schema.md map to the configuration above. The Meta-Learner may adjust parameters within bounds; user-explicit overrides are auto-locked.

• Ideate is for multi-cycle evolutionary ideation — not quick brainstorming. For fast divergent ideas, use BeCreative.
• The Loop Controller manages cycle count — don't override it manually. Trust the budget-based cycling.
• Meta-learner adjustments happen automatically within parameter bounds. Don't manually tune mid-cycle.
• CONTEMPLATE is mandatory. Skipping it degrades MATE quality because STEAL operates on disconnected material.
• Structural randomness defeats LLM bias. Don't substitute "interesting pairs picked by the LLM" for Fisher-Yates — the bias is the problem.

• The 9-phase decomposition and the path-to-ASI mapping derive from a publicly published essay on cognitive progress and a possible path to ASI by D. {{PRINCIPAL_SURNAME}} (2024). The framework name Cognitive Progress Workflow refers to that essay.
• The Lamarckian advantage framing (Phase 9 META-LEARN) borrows from research on auto-research loops and meta-learning in agent systems (cf. Karpathy auto-research pattern).
• Structural randomness as a defeat for LLM-bias is empirical — see internal experiments comparing LLM-picked pairings vs Fisher-Yates pairings on diversity metrics.