What does this research mean for the field?

The Combination Method, a hybrid LLM workflow that delegates combinatorial search to deterministic algorithms and semantic evaluation to lightweight models, reduces generation costs by approximately 90% while improving accuracy and reproducibility in precision design tasks. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

What question did this study set out to answer?

This research aims to develop a new methodology for precision design tasks through an LLM workflow.

June 7, 2026Open Access

Combination Method (Ko Method) — A Discovery.

Key Points

This research aims to develop a new methodology for precision design tasks through an LLM workflow.
Conducted four-stage methodology involving data acquisition, decomposition, combinatorial sampling, and refinement.
Employed Python algorithms for numeric filtering and Haiku models for semantic evaluation.
Human confirmation at the final stage for enhanced accuracy.
Cost reduced by approximately 10% compared to single-pass generation models.
Achieved improved accuracy, reproducibility, and debuggability in design outputs.
Validated method on a case involving a stair design with precise specifications.

Abstract

An LLM workflow pattern for precision design tasks, discovered 2026-06-01 during deLight house Phase 03 CAD work. This deposit contains bilingual versions of the same canonical document: · Korean original (...A Discovery.pdf, 13 pages) · English translation (...A Discovery (EN).pdf, 15 pages)Both versions are content-identical and use the formal academic tone. Core principle: Aggregation + Combination. The methodology compresses scattered domain knowledge into structured categories (Phase A: Aggregation), then algorithmically combines those categories to explore candidate solutions filtered by a numeric-then-semantic compound threshold (Phase B: Combination). The LLM is used as aggregator + refiner only — combinatorial search is delegated to deterministic algorithms (Python) and semantic evaluation to lightweight models (Haiku). Four stages + Human gate:1. Data Acquisition (Sonnet)2. Decomposition into categories (Sonnet → Haiku when ontology cached)3. Combinatorial Sampling + Compound Threshold (Python numeric → Haiku semantic)4. Refinement (Sonnet)+ Human Confirmation Gate Six derived sub-principles:- LLM as aggregator-refiner, not generator- Complexity-to-capability matching (model routing)- Determinism where possible- Compound threshold (cheap-first filtering)- Cumulative memory (ontology caching, amortized inference)- Human-at-confirmation, not generation First case validated: deLight house interior stair (14 risers × 178.57 mm rise × 280 mm tread). Cost ~10% of single-pass Opus generation, with improved accuracy, reproducibility, and debuggability. This is v1.1 (2026-06-05) — formal academic tone with theoretical foundation, case-study trace, and six derived sub-principles. Document is also publicly anchored via:- GitHub commit 27ee91e (2026-06-05) at https://github.com/vycanisko/DSI- OpenTimestamps Bitcoin blockchain anchor (commit 3c539c0)- dsi.run/research/combination-method- dsi.run/research/combination-method-en.pdf (English version)

Read Full Paperexternally

Demander à l'IA

Bookmark

View Full Paper