Skip to content

Future Ideations: Erdős-Banger Roadmap & Tool Integration

Document Status: Non-normative roadmap (accurate-by-construction)

Last Updated: 2026-01-23

This document intentionally avoids factual claims about external products/benchmarks. For implementable, testable work, the SSOT is docs/_specs/README.md.

What This Is

  • A map of future integrations and how they fit the current architecture.
  • A pointer index to the corresponding specs (ironclad, test-driven).

Architectural Principles (Stable)

  1. Canonical state is repo-local text artifacts (mergeable, reviewable):
  2. research/ workspace (v3)
  3. Derived stores are regenerable (not SSOT):
  4. index/erdos.sqlite search DB
  5. literature/cache/, literature/extracts/ caches
  6. logs/ run and loop logs
  7. Vendor-neutral LLM integration:
  8. LLMs are invoked via external commands (ERDOS_LLM_COMMAND)
  9. Task-level routing is done by selecting different commands/scripts (SPEC-032)
  10. Good redundancy only:
  11. Add a new research API only if it contributes new information not already in OpenAlex/Crossref/arXiv.

Spec-Backed Roadmap

Spec Theme What It Adds
SPEC-029 Exa Research Integration Agentic literature synthesis → research leads
SPEC-030 Semantic Scholar Integration Citation intent/context (“why cite”)
SPEC-031 zbMATH Integration Math-native metadata (MSC, reviews excerpt policy)
SPEC-032 Multi-Model Routing Task → LLM command routing (no SDK lock-in)
SPEC-033 Lean Copilot Integration In-editor tactic suggestions via external API server
SPEC-034 Progress Dashboard Terminal dashboard + JSON snapshot mode

Reference Architecture (End-to-End)

Problems dataset + metadata APIs
  - Problems YAML (local)
  - OpenAlex / Crossref / arXiv (existing)
  - (future) Exa / Semantic Scholar / zbMATH
            │
            ▼
Canonical research state (SSOT)
  research/problems/<id>/
    meta.yaml
    SYNTHESIS.md
    leads/*.yaml, attempts/*.yaml, hypotheses/*.yaml, tasks/*.yaml
            │
            ▼
Derived retrieval index
  index/erdos.sqlite (FTS5 + optional embeddings)
            │
            ▼
CLI orchestration
  erdos research | search | ask | loop run | logs | ...
            │
            ├─► LLM via external command(s) (SPEC-032)
            │      ERDOS_LLM_COMMAND=...
            │      ERDOS_LLM_COMMAND_MATH=...
            │      ERDOS_LLM_COMMAND_CODE=...
            │
            └─► (future) Lean Copilot external API (SPEC-033)
                   erdos lean copilot serve  (localhost)

Example Workflows (Aligned With Current CLI)

Research → Ask

erdos research init 6
erdos research lead add 6 --title "..." --notes "..."
erdos research synthesize 6
erdos --json ask 6 "What have we tried so far?" --no-llm

Research → Loop

erdos research init 6
erdos research synthesize 6
ERDOS_LLM_COMMAND=./scripts/llm.sh erdos loop run 6 --max-iter 10

Summarize Activity

erdos logs --since 7d --summary

Open Questions (Pre-Spec / Needs Design)

  1. Should we add a “paper identity” module (DOI/arXiv/S2/zbl normalization) before integrating more APIs?
  2. If we add vector search, do we want it for:
  3. literature only, or also research artifacts?
  4. For Lean Copilot /encode, should embeddings be:
  5. local (sentence-transformers), or
  6. remote (API), or
  7. optional with a clear degraded mode?