Making combinations change the research structure
Fusion investigates how structural relationships between questions or methods can produce a testable combination. Each parent must contribute something necessary, and the cross-domain bridge must have an explicit role. Question fusion can retain a valuable tension; idea fusion must resolve or reject hard conflicts that prevent execution. Separate problem and idea entry points preserve this distinction and prevent shared vocabulary or longer prose from being mistaken for a scientific mechanism.
Scientific graphs, bridges and constrained operations
The prototype compiles inputs into typed scientific graphs, searches for cross-parent mappings and shared structure, then establishes a fusion contract. An explicit bridge is connected to a constrained FusionDSL programme. Verifiers check types, scope, logic, interfaces and evidence, alongside parent and bridge necessity at each tree stage. Accepted candidates enter an archive; one result is selected while mappings, failures, lineage and verification traces remain available through command-line, API and local interfaces.
Structural validity still needs scientific evidence
The release implements a deterministic structural kernel, recorded failures and a multi-parent fusion workflow. Structural benchmarks test contracts and invariants, but do not establish scientific importance, empirical truth or superiority to the strongest precedent. Real prior-art search, domain experiments, trained semantic providers, independent review and human sign-off require external integration. Grades express a judgement under the supplied evidence policy. Bridge ablations, matched-budget controls and replication are the next route to testing substantive gains.
HOW IT WORKS
From several inputs to one traceable candidate
The local v0.2.0 package contains minimal text inputs, a three-parent output and full audit artefacts. The example demonstrates structural combination and evidence obligations without a real learning experiment.
Inputs
- Two or more concrete questions or ideas of the same type
- A title, content and optional internal grade for each item
- A knowledge cutoff, output language and search budget
Retained artifacts
- One selected question or idea card
- Parent contributions and the staged fusion tree
- Bridge, mappings, constrained operations and verification records
- Failure codes, further tests and evidence requirements
A four-step walkthrough
- 01
Choose question fusion or idea fusion
The user chooses whether to develop a better question or combine executable ideas, then submits concrete parent content. A question may retain a useful tension; an idea must address hard conflicts that prevent execution.
- 02
Identify each parent’s contribution
Each input is independently compiled into a typed scientific graph before cross-parent mappings are explored. An explicit bridge names the states, constraints or mechanisms it connects and records preconditions, rather than simply merging prose.
- 03
Check structure and preserve failures
Constrained operations generate candidates, while verifiers check types, scope, logic, interfaces and necessity. Multiple inputs are combined through recorded stages. Incompatible branches retain failure codes, conflicts and repair leads.
- 04
Select one card without discarding its history
The product returns one selected question or idea card while retaining parent contributions, the bridge, operation path, ablation obligations and verification traces. Later external evidence can reassess the same candidate without silently changing its scientific structure.
EXAMPLE
Constructed package demonstration: memory, consolidation and drift gating
The three inputs address preserving decisive evidence in long contexts, protecting important knowledge during continual learning, and adapting the policy when the distribution drifts. The stored output places the memory objective and drift constraint in a shared optimization structure, then requires consolidation to remain an indispensable channel. It also specifies controls that remove each parent or the bridge, concatenate the parents, or give a strong parent the same budget. The output is a mechanism candidate awaiting scientific validation, rather than a model already shown to perform better.
Conversations & next steps
I welcome discussions about the design of this tool and where it may be useful.