Shortening removes words; genuine compression preserves named entities, core claims, and causal relations within a fixed token budget. This fixed-budget recursion causes Adams et al. (Columbia University, 2023) to produce executive briefings, patent and clinical-literature digests, and system prompts for Brainstorm Reactor execution.
Based on: Salesforce / MIT (2023)
Chain of Density (CoD) is an algorithmic information compression heuristic formulated in 2023 by researchers at Columbia University and Salesforce AI (Adams et al., “From Sparse to Dense: GPT-4 Summarization with Chain of Density Prompting”). Conventional executive summaries suffer from an intractable trade-off: they either meander with conversational padding and low information density, or they degenerate into disjointed bulleted fragments devoid of narrative coherence. Chain of Density resolves this tension through a recursive compaction loop: the text is revised over sequential iterations, forcefully injecting 3 to 5 critical missing named entities and causal vectors in each pass—while maintaining a rigid, invariant token budget. The output converges on the mathematical frontier of maximum semantic utility.
The framework progresses across five disciplined compression iterations:
1. Initial Baseline Generation (Pass 1): Generation of a coherent, grammatically fluid baseline summary characterized by high readability but sparse entity density.
2. Salient Entity Extraction: Identifying 3 to 5 high-information missing entities (quantified metrics, proper nouns, technical constraints) from the source corpus.
3. Syntactic Refactoring & Fusion (Passes 2 & 3): Aggressively purging linguistic filler, passive phrasing, and rhetorical padding. Fusing the new entities directly into the liberated syntax while holding token count constant.
4. Critical Density Calibration (Passes 4 & 5): Iterating until reaching the optimal information boundary where density is maximized immediately prior to cognitive incoherence.
5. Final Executive Crystallization: Delivering an ultra-dense, prose-coherent artifact engineered for high-velocity cognitive absorption.
Human editors are cognitively ill-equipped to execute manual token-invariant compression: they reflexively expand word counts when adding new facts, or they output unreadable shorthand. The Brainstorm Reactor computationalizes the discipline:
Beyond a critical entity-to-word threshold, syntactic readability breaks down, forcing cognitive fatigue on the reader. The engine must actively terminate recursion at the mathematical inflection point.
High-leverage deployment in board-level executive memos, M&A technical diligence digests, clinical trial research condensations, and prompt engineering architectures.
Trace causes beyond symptoms. Sakichi Toyoda’s method in the Toyota Production System uses a five-tier causal recursion to isolate systemic or procedural root causes. Use the finding to target corrective action.
Not the first account determines the result, but systematic examination of who, what, where, when, why, how, and with what means. Rooted in classical rhetoric and refined through Lasswell’s communications model, it yields complete project charters, incident reconstructions, and launch plans with fewer factual omissions.
Not predefined categories, but bottom-up semantic resonance determines how observations are grouped. This foundational method in human-centred design and qualitative content analysis limits deductive bias, helping teams synthesise customer interviews, employee surveys and feature requests into coherent product and organisational themes.
Intuition is no substitute for analysis; its leverage lies in preconscious threat signals that test board decks and financial models for concealed risk. Because limbic pattern recognition can register discrepancies before cortical articulation, Brainstorm Reactor converts them into a final check for capital, safety, and ethics.
Select this method, describe your challenge, and let the structured AI pipeline do the rest.
Open Now