SOTA: provenance-bound environment reconstruction from agent trajectories

[ruvnet]

Finding

Terminal Universe, arXiv:2609.04148, submitted 2026-09-03, reconstructs executable terminal environments from recorded agent trajectories and re-queries those environments to synthesize new verified tasks. The originating team reports 37.3k task-sufficient environments, plus 11.9 absolute points on Terminal-Bench 2.1 and 13.8 points on EvoCode-Bench v2 MT@4 after supervised fine tuning of Qwen3.5-27B.

Evidence class: originating-team measured, public paper, RuV reproduction pending.

Opportunity Score: 4.655.

RuV mapping

Dream Machine can turn completed agent sessions into reusable evaluation and training environments. MetaHarness can use the same manifests for frozen reproduction. RVForge can package reconstructed environments. Core Memory can index the receipts. Ruflo can contribute trajectories. RVM remains the authority boundary for any later execution.

Missing primitive

Introduce a deterministic EnvironmentReconstructionManifest and receipt with explicit provenance for every reconstructed artifact.

Required properties:

1. Source trajectory digest is immutable.
2. Every recovered artifact has a normalized relative path, content digest, source class, required flag, and verification state.
3. Synthesized or completion-agent artifacts are never represented as recovered ground truth.
4. Missing required artifacts make the environment not task sufficient.
5. Path traversal, absolute paths, duplicate paths, malformed digests, and resource exhaustion fail closed.
6. The receipt contains no raw file content or secrets.
7. Receipt authority is always none. Any execution still requires RVM authorization.
8. Canonical digest must be stable across artifact ordering.

Benchmark

Compare raw trajectory replay, reconstruction manifest plus deterministic replay, and a completion-assisted candidate on at least 100 held out Dream Machine or MetaHarness trajectories.

Report reconstruction success, task verifier pass rate, false reconstruction rate, missing dependency rate, time, tokens, storage, environment size, variance, failures, and downstream task success.

Promotion gate: at least 90 percent exact artifact provenance coverage on reconstructable files, zero path escape, zero synthesized-as-recovered misclassification, and at least 20 percent more reusable verified tasks per source trajectory at no more than 10 percent additional orchestration cost.

Governance

No autonomous merge, deployment, credential use, source workspace mutation, or execution of reconstructed commands. Completion artifacts remain explicitly inferred until verified. Raw trajectory and source repository evidence remain authoritative roots.

[ruvnet]

2026-09-08 update: CUA-Universe (arXiv:2609.05374, submitted 2026-09-04) materially extends the environment-generation direction from terminal-only reconstruction to shared GUI plus CLI state. The originating team reports a 9B model improving CUA-Verse by 39.3 points with 37% fewer steps and 60% fewer tokens, OSWorld success by 16.8 points with 57% fewer steps and 44% fewer tokens, and OSWorld-MCP by 7.84 points with 27% fewer steps and 30% fewer tokens.

Reuse this issue rather than creating a parallel environment subsystem. Add a future benchmark arm after the current provenance reconstruction gate: reconstructed terminal surface versus reconstructed hybrid GUI plus CLI surface under the same app state, task, model, seeds, verifier, and budget.

Additional gate for hybrid promotion: at least 20% fewer agent steps or tokens at matched task success, zero unverified CLI surface generation treated as recovered ground truth, deterministic mapping between GUI state and command effects, no privilege expansion, and complete rollback to terminal-only reconstruction. The largest uncertainty is whether generated command surfaces preserve real application semantics outside the 16 adapted applications; cross-application heldout transfer is mandatory before any general claim.