“The winning answer is always a superset. You’re not picking the best model. You’re mining each one for what the others couldn’t see.” — Kasimir Hedstrom
Session context: 2026-01-15_Mastermind — emerged from Kasimir’s live multi-model synthesis experiment, where he observed that the common synthesis failure mode is summarising away unique signal from each model’s response.
Core Idea
When combining outputs from multiple AI models — or from multiple passes of the same model — the instinctive operation is to summarise: find the common thread, express it cleanly, discard the rest. This instinct produces outputs that feel polished but are shallower than any of the inputs. The unique contributions from each model — the edge cases Claude flagged, the human factors ChatGPT surfaced, the structural constraints Gemini identified — all get averaged away.
The correct synthesis operation is the inverse of summarising. Instead of finding what the models agreed on, find what each model said that the others didn’t say. Those unique contributions are the golden nuggets — the signal that would be lost in a summarised output. The synthesis document is a superset of these unique contributions, not a distillation. Nothing that appeared in any input gets discarded; only what is genuinely redundant is removed.
This rule applies beyond multi-model synthesis. It applies to any situation where you’re combining AI outputs: multiple drafts of the same document, multiple analytical passes over the same data, multiple research threads on the same question. The failure mode is always the same — the synthesiser discards signal in the name of clarity. The golden nugget rule prevents this by making the unique contribution, not the common ground, the primary object of the synthesis operation.
Practical Application
The Unique Contribution Prompt: When synthesising multiple AI outputs, use this prompt with a neutral fourth model:
“I have three responses to the same question from different AI models. Your task is NOT to summarise. Instead: for each response, identify every claim, consideration, or angle that does NOT appear in the other two responses. List only those unique contributions, organised by source. Do not discard or combine anything that is unique to one source.”
Then review the unique contributions and build your synthesis as a superset — adding each unique element to a master document. Compare the result to what a summarised version would look like. The delta is the signal the summary would have lost.
Related Insights
- Insight - Run Your Prompt Through Multiple Models and Synthesize at the Top — the multi-model workflow this synthesis rule governs; the golden nugget rule is what makes synthesis work rather than just average the inputs
- Insight - Multi-Model Debate as a Quality Control System for High-Stakes Work — applies the same multi-model principle specifically to quality control; the golden nugget rule applies equally there
- Insight - Multi-Model Debate as a Decision-Making Accelerator — earlier articulation of the multi-model deliberation pattern; this insight adds the synthesis mechanics
- Insight - Design AI Systems for Maximum Composability and Minimum Context Pollution — composable systems need clean handoffs between components; the golden nugget rule ensures information isn’t lost at synthesis boundaries
Evolution Across Sessions
This establishes the baseline for AI synthesis methodology in the vault. The multi-model synthesis pattern has been discussed across multiple sessions, but the explicit failure mode (summarising away unique signal) and the corrective operation (only add, never omit) were first articulated precisely in the January 15 session. Future sessions should test and refine this synthesis protocol as the group uses multi-model workflows more systematically.