Docs

justin-napolitano/jaysearch / docs/research-hypothesis-branching-v1.md

Research Hypothesis Branching V1

Synced 2026-10-23 · Updates are made in the source repository.

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_base fields
  • hypothesis_id
  • problem_id
  • hypothesis_family
  • hypothesis_statement
  • approach_outline
  • evaluation_focus
  • evidence_refs
  • branch_rank
  • prune_conditions

Recommended optional fields:

  • source_question_refs
  • source_problem_ref
  • risk_notes
  • expected_advantages
  • expected_tradeoffs

Initial Hypothesis Families

V1 should support a small fixed family set:

  • reuse_direct
  • reuse_hybrid
  • novel_synthesized
  • baseline_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:

  1. preserve the original problem_id
  2. derive stable hypothesis_ids
  3. keep each hypothesis materially distinct
  4. map evidence and constraints into each branch
  5. 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: 2 branches
  • default: 3 branches
  • maximum: 4 branches

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_packet can produce explicit branches
  • each branch is typed and traceable
  • the output is machine-readable
  • the output is bounded enough for later candidate generation