The evidence base for the Eidolons team. Papers, production patterns, citations.
Every design decision in every Eidolon traces back to something in this library. This is the nexus's aggregation of what was previously scattered across five separate DESIGN-RATIONALE.md files.
| File | Contents |
|---|---|
references.bib |
BibTeX file — every paper we cite, properly formatted |
production-patterns.md |
Non-paper evidence — documented patterns from Claude Code, Cursor, Aider, DocAgent, SWE-Agent, etc. |
papers/ |
One Markdown summary per paper — what it says, what we took from it, which Eidolon uses it |
context-lifecycle-survey-2026-07.md |
Tri-cycle survey ingest (session caching, KV compression, compaction, session lifecycle, autonomous policies) — the evidence base for ECM (docs/specs/ecm/spec.md) |
This is the inverse index. Given an Eidolon, which research backs it?
- AgentFold — phase-boundary context compaction → ATLAS's AgentFold at phase boundaries
- Scatter-gather subagents — operator pattern → ATLAS's subagent fan-out for independent sub-questions
- Bounded ACI — Aider, Claude Code read primitives → ATLAS's 7-tool read-only surface
- Evidence anchoring — SWE-Agent, DocAgent → ATLAS's
path:line+ confidence tier
- Plan-mode agents — Claude Code Plan Mode, Cursor Plan → SPECTRA's Explore phase
- Decision theory — classical MAUT scoring → SPECTRA's scoring rubrics
- Information theory — entropy-based validation → SPECTRA's 6-layer validation
- CorrectBench 2025 — evidence for bounded self-correction → APIVR-Δ's 3-attempt cap on Reflect
- Reflexion — verbal reinforcement learning → the Reflect phase itself (but bounded)
- SWE-Agent — tool design for software engineering → APIVR-Δ's tool surface
- Context engineering (Anthropic, Sept 2025) → layered loading
- DocAgent (Meta/FAIR, arXiv 2504.08725) — CHT verification framework → IDG's Gate phase
- Topological ordering of documentation dependencies → IDG's skeleton-first drafting
- TBD — candidates: Chain-of-Thought, Tree-of-Thoughts, Self-Consistency, Debate, Counterfactual reasoning literature
Patterns that appear in multiple Eidolons:
Research: Anthropic Context Engineering (Sept 2025) + APIVR-Δ's production evidence. Applied to: all Eidolons. Entry point ~900 tokens, skills load on-demand per phase.
Research: CorrectBench (2025) + Reflexion critique literature. Applied to: IDG (1 pass), APIVR-Δ (up to 3), ATLAS (three-strike halt).
Research: Harness AI incident conventions + DocAgent's CHT framework.
Applied to: IDG (four markers), ATLAS ([FINDING-NNN]), all Eidolons via Prime Directive D7.
Research: SWE-Agent tool design + DocAgent citation enforcement.
Applied to: ATLAS (path:line), IDG (source artifact refs), SPECTRA (pattern + rejected alternatives).
Research: context-lifecycle-survey-2026-07.md — context rot thresholds (arXiv:2601.15300, EMNLP'25), lossy-compression instruction loss (ACL 2026 pitfalls), prefix-cache economics (LMCache, provider caching APIs), server-side compaction convergence, Mem0-class harness-injected memory.
Applied to: ECM (docs/specs/ecm/spec.md) — zone ladder, pin set, externalize-before-compact, handoff brief, mechanical decision table. Touches all Eidolons via the harness kernel; CRYSTALIUM via the session_handoff convention.
Non-paper evidence. See production-patterns.md for full details.
- Claude Code — Plan Mode, skill activation, subagents
- Cursor — MDC rules, agent-requested skills,
AGENTS.mdsupport - Aider — file-aware editing, test-driven repair loops
- Copilot Workspace — multi-step agent coding
- SWE-Agent — bounded ACI for software engineering
- DocAgent — multi-role documentation pipeline
- OpenCode — permission-scoped agents
Each pattern lists: source, what it does, how the Eidolons borrow from it, how the Eidolons differ (important — we don't just copy).
If an Eidolon's design evolves and the rationale shifts:
- Update the paper summary in
papers/<slug>.mdif the finding is new - Update the production-patterns entry if the pattern is new
- Update
references.bibif adding a citation - Update this INDEX's "Research → Eidolon mapping" section
- Commit with message:
research: add/update <slug>
Do not duplicate rationale that already lives in an individual Eidolon's DESIGN-RATIONALE.md. The nexus library is for cross-cutting evidence. Eidolon-specific rationale stays in the Eidolon's repo.
This library is in an early state. Priority extraction order:
- Consolidate citations from existing
DESIGN-RATIONALE.mdfiles in each Eidolon repo - Build BibTeX from those citations
- Write one paper summary per major citation
- Cross-link back into each Eidolon's rationale
This is tracked as Open Thread §8 — Research aggregation in the project instructions.