Meta-n boundary

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

Boundary and evidence

This run analyses the shipped Meta-n framework for model-generated improvements to task-solving agents. Its primary target class is builder or improvement plane: the framework selects evaluated candidate chains, asks a model to generate additional improvement code, and evaluates the resulting solver. Native solving and external-agent orchestration are included constituent responsibilities, not excluded hosts. The intended use is to explain the implementation's improvement loop, retained state, injection channels and evaluation boundaries; this is not a verification of the paper's performance claims.

The boundary is whole-system at Git commit b7081843d3c7b0e0f418ca10aaf2ccbff856e7f8, with applicability cutoff 2026-10-14. The acquired checkout's HEAD matches that commit and its tracked working tree is clean. The overall tier is code-grounded. Documentation supplies declared design, and executable source supplies implementation evidence. No target code, test, benchmark or model call was executed. Tests and fixtures are source-level contracts, not newly observed runs.

The shipped caller belongs to this target. The installed meta-n entry point maps to meta_n.main:main; the CLI constructs the improvement engine, executor and orchestrator, invokes the search with resume state, and saves its results (SRC-1, pyproject.toml, meta_n/main.py). The classification is grounded by the model-generation and admission path, not just the repository's title (SRC-1, meta_n/core/evolutionary_orchestrator.py):

injected, omega_tokens = await self.omega.generate( traces=parent.traces, context_stack=parent.injected_codes, tasks=tasks, depth=child_depth, temperature=temperature, --- meta_n/core/evolutionary_orchestrator.py

if self.config.verified_code: barred = self._nonadopting_verified_tasks( child, injected, sandbox_verified_names ) self.archive.add(child, bar_from_best=barred) else: self.archive.add(child) --- meta_n/core/evolutionary_orchestrator.py

Functional inclusion covers task/configuration loading; native and external base solvers; fixed prompts and prompt rendering; generated preprocessing and helper libraries; candidate selection, gates and optional repair paths; execution/scoring adapters; archive and checkpoint writing/rebuilding; later reuse of traces and code; budget/concurrency guards; telemetry; and shipped post-run analysis tools. Feature existence does not imply activation in every configuration. In particular the bundled configuration and explicit CLI choices are part of the same boundary (SRC-1, meta_n/configs/benchmark_features.yaml, meta_n/main.py).

Persistence and subsequent consumption are inside the boundary. Candidate traces and complete injected chains are written by RunPersistence, reconstructed by Archive.rebuild_from_disk, and used by resumed search. Archive traces and prior injections also enter subsequent Ω calls; MetaLayer consumes generated context and libraries when invoking its inner solver. Analysis tools read the saved run summaries (SRC-1, meta_n/core/run_persistence.py, meta_n/core/archive.py, meta_n/core/meta_layer.py, meta_n/analysis/best_of_k.py). Minimum implementation anchors for the write/reload/consume chain are:

for i, ic in enumerate(candidate.injected_codes): with open(cand_dir / f"injected_code_d{i+2}.json", "w") as f: json.dump( ic.model_dump(exclude={"raw_omega_prompt", "raw_omega_response"}), f, indent=2, ) --- meta_n/core/run_persistence.py

self._orch.archive = Archive.rebuild_from_disk( archive_dir, barred_from_best=checkpoint.get("barred_from_best"), **self._orch._archive_kwargs(), ) --- meta_n/core/run_persistence.py

script, reasoning, tokens = await self.inner_solver.solve(task, context) script = self._prepend_library(self._maybe_deploy_verified_helper(script)) return script, reasoning, tokens, context, False --- meta_n/core/meta_layer.py

Repository coverage

Every top-level tracked file or directory is classified below. Included paths remain available for deeper inspection at this same commit; this table records initial boundary inspection, not an exhaustive audit or an evidence allowlist.

Dependencies and unavailable evidence

External dependencies include model providers/OpenAI-compatible endpoints, optional OpenHands and Terminus 2/terminal-bench SDKs, Harbor, Docker, benchmark datasets and upstream evaluator packages. Meta-n's wrappers, subprocess drivers, environment provisioning and result parsing are included; upstream implementations and live services are not independently frozen sources in this run (SRC-1, pyproject.toml, README.md, data/README.md, meta_n/core/external_agents/solver.py, scripts/t2_runner.py). This prevents attributing upstream internals to Meta-n or establishing end-to-end operation from wrapper text alone.

The source tree supplies no complete frozen empirical campaign or model-provider execution trace for this boundary. The README's paper link is not a registered source and was not followed. These limits prevent conclusions about measured improvement, emergent roles in actual runs, causal benefit of depth, external-backend parity in operation, or safety under real generated code. They do not block a source-grounded account of the shipped design and implementation.

Source register