Research that never touches reality converges on plausible

Graph Engineering pattern 11 of 15: the Research–Synthesis Graph. Assigned territory, central synthesis under critique, and an experimenter loop that turns claims into ground truth.

Research that never touches reality converges on plausible, not true. An agent organization can read everything ever written about two databases and still recommend the one that fails your workload — unless something in the graph is allowed to run the benchmark.

Context for new readers: pattern 11 of 15 in the Graph Engineering series. Research inverts the production patterns: the raw material is breadth, the deliverable is convergence, and the truth test is an experiment. The topology reflects all three.

The problem

The organization needs a grounded answer to an open question: which technology to adopt, whether an approach is feasible, what the state of an area actually is. The familiar failure modes of doing this badly: surveys broad but shallow, recommendations confident but unverified, conclusions that quietly reflect the first source read.

The topology

  Researcher 1 --+
  Researcher 2 --+-> Synthesizer <-> Experimenter
  Researcher N --+        ^              |
                          |              |
                       Critic     new questions
                                  re-seed research

Research questions fan out to N researchers, each covering assigned territory — different sources, different candidate technologies, different sub-questions. Findings converge on a synthesizer, which integrates them into candidate answers, open contradictions and confidence levels. A critic challenges the synthesis. And the synthesizer's bidirectional edge to an experimenter is what separates this from a literature review: claims that can be tested cheaply are tested — a prototype, a benchmark, a compatibility check — and results feed back as ground truth. Experiments that raise new questions re-seed the researchers.

Three disciplines carry the pattern

Territory is assigned. Without it, N researchers return N copies of the most findable sources — the research analog of pattern 6's correlated reasoning.

Synthesis is not concatenation. Contradictions between researchers are the most valuable signal the synthesizer can surface, not noise to smooth over. A synthesis with no surfaced contradictions is usually a stack of summaries wearing a conclusion.

The critic asks the falsification question. "What would change this conclusion?" If nothing the experimenter could do would change it, the loop is decorative and the recommendation is faith.

Where it breaks

• Shared ground. Researchers without assigned territory duplicate coverage and miss the corners.
• Unfalsifiable conclusions. See above — the critic's standing question is what keeps the pattern honest.
• Experiments as theater. Benchmarks run after the conclusion is written confirm rather than test. The experiment edge feeds into synthesis, not out of it.

The dimension answers

Diverge, converge, verify, repeat: the scientific method, drawn as a graph — and an organization your agents can actually run, with the experimenter as the only node whose output nobody gets to argue with.

Next: pattern 12, the Intelligent Router Graph — a classification node in front of specialist organizations.

Building this with design partners → rysh.ai/design-partner