Codeseys/composer-replication-framework

1# Methodology — Composer 2.5 Replication Framework Research2 3This document records *how* the research synthesis in this repo was produced, so4the methodology is reproducible and the cross-family verification claim is5auditable.6 7## Research dispatch8 9On 2026-05-25, five parallel research subagents were dispatched via the10[`delegate_task`](https://hermes-agent.nousresearch.com/) parallel-research11pattern, one per topic. Each was given:12 13- A specific research scope (one of: Composer 2.5 internals; DiLoCo family;14  Monarch / TorchForge / OpenEnv; VeRL / TRL; trace-replay distillation15  novelty assessment).16- An explicit instruction to write findings to a known path17  (`~/wiki/research/post-training-framework/0X-<topic>.md`).18- ~2000–2500 word target depth.19- Web-research toolset (Tavily, Exa, AWS docs, MCP doc readers).20 21Each subagent ran independently — no cross-agent communication, no shared22intermediate state. They were given a uniform research scope but **routed to23five different LLM families** for cross-family signal:24 25| File | Author model | Rationale |26|---|---|---|27| `research/01-composer-2.5.md` | `google/gemini-3.1-pro-preview` | Long-context grounded research is Gemini's strong suit |28| `research/02-diloco-family.md` | `deepseek/deepseek-v4-pro` | Strong on distributed-systems and pretraining literature |29| `research/03-monarch-torchforge-openenv.md` | `openai/gpt-5` | Best at reading framework / SDK source code |30| `research/04-verl-trl.md` | `anthropic/claude-sonnet-4.6` | Best at algorithmic precision (loss math, importance sampling) |31| `research/05-trace-replay-distillation.md` | `moonshotai/kimi-k2-thinking` | Strong at novelty assessment and prior-art discovery |32 33All routes were **verified post-hoc** via the per-task `model` field returned34in the delegated agent's session metadata — i.e. the synthesis is not based on35a single model's biases.36 37## Synthesis38 39The master synthesis (`framework/composer-replication-framework.md`) was40produced by reading all five reports in full and reconciling:41 42- **Convergent claims** (≥2 independent reports agree) → promoted to43  framework-level decisions in the TL;DR table.44- **Divergent claims** (reports recommend different stacks for the same45  layer) → noted explicitly with "use X today, switch to Y when Z" rationale46  rather than picking one arbitrarily.47- **Single-source claims** (only one report makes the claim) → kept but48  flagged as "single-source — may be model bias" where consequential.49 50Convergent findings (verified across reports):51 52- **GRPO+DAPO is the consensus algorithm.** Reports 04 (TRL/VeRL deep-dive),53  02 (PRIME-RL section), and 03 (Forge algorithm catalog) all converge on54  GRPO with DAPO patches as the production default for long-horizon agentic55  RL.56- **PRIME-RL is the most production-ready decentralized substrate.** Reports57  02 and 04 independently cite INTELLECT-2 (32B QwQ trained globally58  distributed) as the only production-scale decentralized RL run to date.59- **OpenEnv is the env-format winner.** Reports 03 (Meta's stack), 04 (TRL's60  Oct 2025 OpenEnv integration), and 05 (env-substrate analysis) all61  converge on OpenEnv + verifiers as the emerging standard.62- **Trace-replay multi-teacher is genuinely under-explored.** Report 05's63  primary finding, corroborated by the fact that none of the other 4 reports64  (which surveyed the algorithm and framework literature widely) mention65  per-step multi-teacher distillation as an existing technique.66 67## Sources68 69The synthesis cites primary sources inline. Major primary sources include:70 71- **Cursor blog**: <https://cursor.com/blog/composer-2-5> (the Composer 2.572  release post that motivated the whole project).73- **Moonshot K2 paper**: <https://arxiv.org/abs/2502.05559> (Kimi K2 base74  model, the predecessor to K2.5).75- **DeepMind DiLoCo paper**: <https://arxiv.org/abs/2311.08105>; **Streaming76  DiLoCo**: <https://arxiv.org/abs/2501.18512>.77- **Prime Intellect INTELLECT-2 announcement**: <https://www.primeintellect.ai/blog/intellect-2>.78- **VeRL paper**: <https://arxiv.org/abs/2409.19256>.79- **HuggingFace TRL**: <https://github.com/huggingface/trl>.80- **Microsoft rStar / rStar-Math**: <https://arxiv.org/abs/2408.06195>.81- **Meta OpenEnv**: <https://github.com/meta-pytorch/openenv>.82- **Meta Monarch**: <https://github.com/meta-pytorch/monarch>.83 84The five research notes link to many more secondary sources (blog posts,85twitter threads, individual repo READMEs). Those are auxiliary context, not86primary evidence.87 88## Limitations89 90- **No primary-source access to Cursor's training pipeline.** Composer 2.5's91  exact recipe is reconstructed from public statements; details like the92  text-hint generator architecture remain unverifiable. The biggest known93  gap is flagged in `framework/composer-replication-framework.md` § "Open94  questions."95- **Pre-spike speculation.** The TL;DR table's stack picks are96  literature-backed but not yet empirically validated on this codebase. The97  v0.0 spike will produce the first empirical result.98- **Single-snapshot research.** All five reports were produced on99  2026-05-25. The field moves fast — TorchForge may un-pause, OpenEnv may100  fork, PRIME-RL may consolidate. Re-run the dispatch every 6 months.101 102## Reproducibility103 104If you want to reproduce this research dispatch (or extend it with new105topics), the pattern is:106 1071. Use the `delegate_task` parallel-research pattern (or any equivalent: one108   subagent per topic, all running in parallel, all writing to known paths).1092. **Route different topics to different model families** explicitly — this110   is the cross-family signal, and it requires a multi-model gateway like111   OpenRouter or your local equivalent.1123. Give each subagent a web-research toolset (Tavily, Exa, AWS docs, etc.)113   and ~10 min wall-clock budget.1144. After all reports return, verify each one's served `model` matches the115   intended route (per the route-fidelity discipline).1165. Read all reports in full (do not skim) and reconcile in a master synthesis117   doc that explicitly flags convergent vs single-source claims.118 119This pattern generalizes beyond this project; it's the same approach used120for any meaty literature-review task where a single model's perspective is121suspect.122