Research Hypothesis Branching V1
Research Hypothesis Branching V1
Objective
Add a first bounded branching layer inside research-tool so one research_problem_packet can produce multiple explicit hypothesis branches before candidate generation.
This is the first concrete improvement slice for the research system.
Why This Slice Is First
Hypothesis branching is upstream of:
- candidate family diversity
- evaluation comparability
- recommendation traceability
If research still starts from one flattened problem interpretation, the system will keep producing narrow and weakly differentiated candidate sets.
Hard Boundary
This slice must not become:
- full research execution
- planner logic
- evidence search
- governance
It only does:
- branch the research problem into explicit bounded hypotheses
- preserve traceability back to the source problem and evidence
New Packet
research_hypothesis_packet
Purpose:
- one explicit research branch derived from a
research_problem_packet
Required fields:
- all
packet_basefields hypothesis_idproblem_idhypothesis_familyhypothesis_statementapproach_outlineevaluation_focusevidence_refsbranch_rankprune_conditions
Recommended optional fields:
source_question_refssource_problem_refrisk_notesexpected_advantagesexpected_tradeoffs
Initial Hypothesis Families
V1 should support a small fixed family set:
reuse_directreuse_hybridnovel_synthesizedbaseline_reference
The brancher should emit at least:
baseline_reference- one non-baseline branch
when the problem has enough information to support that.
Branching Rules
The brancher should:
- preserve the original
problem_id - derive stable
hypothesis_ids - keep each hypothesis materially distinct
- map evidence and constraints into each branch
- define what would cause the branch to be pruned later
V1 does not need to score hypotheses deeply. It only needs to materialize a bounded and explicit branch set.
Minimum V1 Output
For a normal bounded technical problem, the brancher should target:
- minimum:
2branches - default:
3branches - maximum:
4branches
The brancher should not emit large branch populations in V1.
Example Flow
research_problem_packet
-> materialize_research_hypotheses
-> research_hypothesis_packet[]
-> candidate generation per surviving branch
Acceptance
V1 is successful when:
- one
research_problem_packetcan produce explicit branches - each branch is typed and traceable
- the output is machine-readable
- the output is bounded enough for later candidate generation