Starintel Second Brain — Entirely built and managed by AI agents

STAR-RESEARCH-000 R&D-Agent Pipeline Analysis

Version Date Description of change Did nsaspy approve it
0.1.0 2026-07-19 Complete pipeline comparison and scoped design recommendation Yes

Related Nodes

Research Question

Which parts of R&D-Agent1 should be adapted into the StarIntel automatic-research pipeline without weakening the existing canonical-design2, provenance3, context-minimization4, and one-design implementation-slot5 rules?

Current Pipeline

The current repository already provides:

  • one canonical Org-roam6 knowledge graph under roam/
  • one active implementation design at a time
  • append-only implementation and rejection ledgers
  • bounded context loading through direct links, indexes, and narrow search
  • persistent research notes with sources and retrieval dates
  • durable design, implementation, and rejection history
  • reusable agent skills
  • publication through searchable HTML and EPUB editions

These are strong controls for auditability7 and long-lived knowledge. They must remain.

The current research loop is still mostly serial. A project index places one topic in review and queues the next topic behind it. Research may use parallel read-only subagents, but the repository does not yet represent competing hypotheses8, parent-child candidate relationships, branch scores, pruning decisions, or a machine-readable exploration graph9.

The current local search script performs lexical ranking10 by counting query terms and applying title and description bonuses. This is useful for narrow repository navigation, but it does not provide semantic similarity, graph distance, decision relevance, evidence quality, or explicit retrieval of prior failed approaches.

R&D-Agent Findings

R&D-Agent divides its workflow into a research phase and a development phase. The research phase contains planning, exploration-path structure, reasoning, and memory. The development phase contains coding and evaluation. The framework iteratively adds ideas, implementations, and scores to an exploration graph until its budget is exhausted.11

The transferable architectural patterns are:

  • dynamically allocate effort between broad exploration and refinement
  • maintain several competing paths instead of one greedy chain
  • select parent candidates using recorded results
  • construct bounded memory context from prior solutions and evaluations
  • require structured problem identification and hypothesis-driven proposals
  • reject weak proposals before expensive implementation
  • prototype and debug on reduced fixtures before full execution
  • use stable metrics and aggregated evaluation rather than trusting one noisy result
  • write every completed evaluation back into the exploration graph

The paper evaluates machine-learning engineering tasks. Its benchmark result is evidence for the framework inside that task class, not proof that the same numeric gain will transfer to intelligence-platform architecture research. StarIntel should adopt the decomposed and testable control structure, then measure its own effect through repository-specific ablations12.

StarIntel Adaptation

Preserve one implementation owner

Parallelism belongs in read-only research branches and isolated candidate workspaces. Only the promotion gate13 may create or select a canonical design, and only one design may occupy roam/implement/.

Add machine-readable run and candidate state

Each research run records its question, budget, planner state, candidate identifiers, metrics, and promoted result. Each candidate records parent identifiers, hypothesis, evidence, predicted result, falsifier14, evaluation, cost, and state.

Separate transient candidates from durable knowledge

Intermediate candidate work belongs under ignored runtime storage such as .cache/research-runs/. Append-only state transitions belong in machine-readable ledgers. Final research synthesis belongs under roam/research/. Only promoted work becomes a numbered design file.

Use independent evaluation roles

Deterministic checks run before model review. Separate reviewers examine evidence, architecture fit, novelty, feasibility, and failure modes. Unsupported central claims and broken citations are veto conditions rather than ordinary numeric penalties.

Require explicit negative evidence

A candidate must record contrary evidence and a falsifier. A rejected candidate remains retrievable so later branches do not repeat the same failed path without new evidence.

Scoped Recommendation

Create STAR-RESEARCH-PIPELINE-001 Adaptive Research DAG and Evaluation Harness with only:

  • research-run ledger contract
  • research-candidate ledger contract
  • branch supervisor
  • bounded memory packet contract
  • deterministic evaluation contract
  • independent reviewer aggregation
  • promotion gate
  • append-only transition recording

Defer semantic retrieval, prototype execution, dynamic model routing, and generalized distributed execution to dependent designs. This keeps the first implementation measurable and prevents the control-plane project from becoming another unbounded agent framework.

Acceptance Evidence for the Recommendation

The proposed design is justified when it can demonstrate:

  • three or more isolated candidates for one research question
  • explicit parent and branch relationships
  • deterministic rejection of unsupported or malformed candidates
  • independent reviewer results with recorded disagreement
  • one promoted design and preserved rejected candidates
  • no second file entering roam/implement/
  • replay of a run from its append-only state transitions
  • an ablation comparing serial research against branched research on fixed repository tasks

Footnotes and Glossary

Citations

Footnotes:

1

R&D-Agent: A published large-language-model agent framework that divides machine-learning engineering into configurable research and development components.

2

Canonical design: The authoritative numbered design document that remains the source of truth for a proposed architecture and its history.

3

Provenance: Structured information recording where a claim, document, or result came from and which process produced or changed it.

4

Context minimization: Loading only the smallest relevant set of documents and source files needed for one task instead of consuming the entire repository.

5

Implementation slot: The repository rule permitting zero or one active Org design file under roam/implement/ at a time.

6

Org-roam: An Emacs package that stores linked Org files as a navigable knowledge graph with stable identifiers and backlinks.

7

Auditability: The ability to reconstruct what was decided, which evidence supported it, who approved it, and what later happened.

8

Hypothesis: A testable proposed explanation or design claim that predicts an observable result.

9

Exploration graph: A directed graph whose nodes are candidate research solutions and whose edges record which prior candidates caused or informed later candidates.

10

Lexical ranking: Ordering documents by exact words or word fragments appearing in their text.

11

Xu Yang et al., “R&D-Agent: An LLM-Agent Framework Towards Autonomous Data Science,” arXiv:2505.14738v2, retrieved 2026-07-19: https://arxiv.org/abs/2505.14738

12

Ablation: A controlled comparison that removes or changes one pipeline component to measure whether that component improves results.

13

Promotion gate: The single component authorized to convert a reviewed research candidate into a canonical design proposal.

14

Falsifier: A result or observation that would show a hypothesis is wrong or unsuitable.