OpenRSI boundary

Type: agentic-system-analyses/types/agentic-system-boundary.md

Boundary and evidence

This run examines the OpenMLE implementation shipped by OpenRSI: construction of executable machine-learning tasks, production and selection of training experience, operator training, and test-time program search and evaluation. Its intended use is to establish the stack's source-native control and retained-state mechanisms, not to reproduce its benchmark scores or establish general recursive self-improvement.

The target is a builder or improvement plane. Model-dependent program generation is central rather than incidental: the search solver constructs separate model clients for Draft, Improve, Debug, Crossover and Analyze, then binds them to operators. The boundary is whole-system at the source-release level, covering the separately runnable stages and their shipped drivers, vendored runtimes, prompts, evaluators, persistence and consumers. It is not a claim that one autonomous process closes the entire task-building-to-model-training-to-search cycle. SRC-1, OpenMLE-Evo/third_party/aira-evo/src/dojo/solvers/evo/evo.py, README.md.

OpenRSI contains several independently runnable stages rather than one synthetic top-level command. --- README.md

draft_llm = GenericLLM(self.cfg.operators["draft"]) improve_llm = GenericLLM(self.cfg.operators["improve"]) debug_llm = GenericLLM(self.cfg.operators["debug"]) crossover_llm = GenericLLM(self.cfg.operators["crossover"]) analyze_llm = GenericLLM(self.cfg.operators["analyze"]) --- OpenMLE-Evo/third_party/aira-evo/src/dojo/solvers/evo/evo.py

Frozen evidence and coverage

The sole frozen source is SRC-1 at full commit 71ae803a035d5e3b78c19fa49ed9f67d0550cbaa; the checkout HEAD was checked against the supplied acquisition object. The applicability cutoff is 2026-10-14, not a claim about subsequent revisions or live services. Overall evidence is code-grounded. Repository documentation supplies declared design and attributed operational claims; reading it does not establish observed execution or causal experiments.

Every top-level tracked entry is classified below. Included directories remain available for later analysts to inspect at this same commit. Initial inspection traced representative shipped entry points and state transitions; it is not a line-by-line audit. Available but unread files are not access gaps.

State-loop inclusions

The Gym CLI is not merely an API wrapper outside the target: it invokes task building, metadata overview, metric checks and evaluation. Its worker dispatches those operations and writes machine-readable results. The graph and generated task packages are therefore inside the boundary, as are the sandbox's scoring and lifecycle mechanisms. SRC-1, OpenMLE-Gym/openmle_gym/cli.py, OpenMLE-Gym/openmle_gym/task_worker.py, OpenMLE-Gym/builder_core/design.py.

Experience production and its training consumers are also included. The SFT pipeline collects scored rows, applies token and validation/test-gap filters, selects per-task examples and writes an alignment dataset. Its training launcher consumes prompt data through an SFT rollout function. The RL generation/reward module conditionally admits programs into its database and exposes database snapshots. These source paths authorize analysis of the mechanisms, not a claim that the released corpora or checkpoints were regenerated here. SRC-1, OpenMLE-ERL/SFT/tts_search/data_produce/pipeline.py, OpenMLE-ERL/SFT/slime_scripts/common/run_slime_sft.sh, OpenMLE-ERL/RL/generate_mle.py.

The inference side includes persistence and later use, not just operator APIs. The solver saves and reconstructs its journal, gathers related memory nodes and ensures their summaries before operator use. Shipped prompt templates explicitly consume this memory. Both benchmark adapters write configurations and launch a shipped runner; that runner checks retained outputs for resume. SRC-1, OpenMLE-Evo/third_party/aira-evo/src/dojo/solvers/evo/evo.py, OpenMLE-Evo/third_party/aira-evo/src/dojo/configs/solver/operators/mlebench/aira_operators/improve_experience.yaml, OpenMLE-Evo/third_party/aira-evo/examples/mle_bench/runner.py.

self.journal = Journal.from_export_data({"nodes": journal_export}) --- OpenMLE-Evo/third_party/aira-evo/src/dojo/solvers/evo/evo.py

Use horizontal sibling memory to avoid repeating nearby failed alternatives and to borrow useful details. --- OpenMLE-Evo/third_party/aira-evo/src/dojo/configs/solver/operators/mlebench/aira_operators/improve_experience.yaml

External dependencies and prevented conclusions

Model serving and weights, full training corpora and historical checkpoints, the separate OpenMLE Tasks release, upstream benchmark environments and datasets, credentials, and deployment-specific infrastructure are outside this frozen source. Repository-bundled smoke packages, evaluation tools and source-native example reports remain in scope; they do not stand in for the full external releases. The pinned RL SLIME dependency documented by the project has not been acquired into this evidence set. No second repository, model card, paper, web page or live endpoint is registered. SRC-1, docs/release.md, OpenMLE-ERL/RL/README.md.

Training corpora, model weights, checkpoints, service credentials, private infrastructure configuration, generated outputs, and deployment-specific endpoints are deliberately kept outside this code repository. --- docs/release.md

Those gaps prevent verification of historical training lineage, exact numerical reproduction, full task-release coverage and deployed service behavior. A source-level whole-system boundary does not establish that external dependencies work together in any particular deployment. The target, tests and examples were not executed; no fresh observed-run or causal-experiment evidence is claimed. Source-native reports of runs can be examined later as attributed evidence under their actual retained conditions.

Source register