Ideation: Diverse Methods With Rotation

Diversity is a selection problem, not a volume problem.

Generating ten ideas from one mental frame yields ten variations of the same idea. The fix is to ideate from different categories of method, then rotate methods across passes so the next round does not repeat the last. This is the documented lever against LLM mode collapse, not a brainstorming ritual.

When NOT To Use

• Surveying what already exists (use tome:research)
• Cross-domain analogies specifically (use tome:triz)

The one design rule

Structure the reasoning, not the output.

"The Price of Format" (arXiv 2505.18949) found that rigid output templates collapse output diversity, while reasoning scaffolds raise it. So each method hands you a reasoning prompt. Do not force the ideas into a fixed schema while generating; impose structure only when reporting the final set.

Catalog

Nine methods that port to technical problem-solving, each tagged with an honest evidence grade (see Sources below):

Method	Category	Evidence
SCAMPER	transformation	weak
SIT task unification	transformation	mixed
Morphological analysis (Zwicky)	decomposition	mixed
SIT subtraction	decomposition	mixed
Cross-domain analogy (TRIZ, Synectics)	analogical	mixed
Inversion (reverse brainstorming, pre-mortem)	inversion	anecdotal
Constraint provocation	perturbation	anecdotal
Devil's advocacy	adversarial	mixed
Multiple working hypotheses	hypothesis	mixed

The grades are deliberately conservative. Individual-method evidence is thin; the value is the diverse-selection-plus-rotation pattern, not any single method. For the structured analogical form, use Skill(tome:triz). For finding an implied-but-empty cell in a concept space, map the field and its hidden optimization axis first, then apply generation methods to fill the gap.

Workflow

1. Pick how many approaches you want (3 to 5 is the useful range).
2. Select methods spanning distinct categories: select_methods(n) from tome.channels.ideation returns methods from different categories by construction.
3. Apply each method's reasoning prompt to the problem. Generate ideas freely; do not normalize them into a schema yet.
4. On a second pass, rotate: select_methods(n, exclude=used_ids) or rotation_plan(passes, n_per_pass) so no method repeats until the catalog is exhausted.
5. Score the candidates with score_idea (weighted criteria per Skill(leyline:evaluation-framework)). Novelty claimed without evidence is capped, so "novel" must be earned, not asserted.
6. Report the final set with rationale and scores.

Scoring

score_idea(scores, novelty_evidence=False) applies weighted criteria (novelty 0.25, fit 0.20, feasibility 0.20, simplicity 0.15, reversibility 0.10, impact 0.10) and an anti-inflation rule: a novelty score of 8 or higher without supporting evidence is capped to 7. Pass novelty_evidence=True only when you have checked for prior art and found the idea genuinely new.

Anti-goals

• Do not generate many ideas from one method and call it diverse.
• Do not impose a rigid output schema during generation; it collapses diversity.
• Do not present anecdotal-evidence methods as proven; cite the grade.
• Do not import the full 20-method bundle; the catalog is curated on purpose.

Sources

The diversity rationale and method evidence grades draw on:

• Verbalized Sampling (arXiv 2510.01171): distributional prompting for 1.6 to 2.1x diversity gains, training-free.
• The Price of Format (arXiv 2505.18949): rigid output schemas collapse generation diversity. This is the basis for the reasoning-prompt-not-schema rule above.
• Barriers to Diversity in LLM-Generated Ideas (arXiv 2602.20408): chain-of-thought reduces fixation; diverse personas restore collective diversity.
• The originality and SCAMPER / morphological-analysis education literature underpins the per-method evidence grades.
• Innovation Paradox (Shi et al., Design Science 2024): patent analysis of 4M+ concepts found originality declined 31% from 1981-2016 as concept space expanded. Cognitive burden drives clustering near familiar terrain rather than the edge.
• Devil's advocacy: Schwenk, OBHDP 47(1), 1990, a meta-analysis in which devil's advocacy beat the expert-plan approach. Graded mixed because one model arguing both sides is weaker than two people.
• Multiple working hypotheses: Lord, Lepper and Preston, JPSP 47(6), 1984, where considering the opposite corrected biased judgment. The prompt stays light on purpose: in Dhami, Belton and Mandel, Applied Cognitive Psychology 33(6), 2019, the full hypothesis matrix (ACH) did not improve analyst accuracy.

Methods surveyed and not catalogued, with grades, are in ADR-0024.

Exit Criteria

• At least 3 candidate approaches were generated, each from a method in a distinct category (verifiable: the methods used span >= 3 categories from list_categories()).
• A second pass, if run, used exclude or rotation_plan so no method repeated before the catalog was exhausted.
• Each candidate carries a weighted score from score_idea, with any uncapped novelty backed by an explicit prior-art check.
• The final report structures the output; generation did not.