Xnhyacinth/OPERATE-Benchmark

OPERATE

The parent v0.61 admission ledger records 2,476 terminal candidate decisions with zero unresolved. v0.62 introduces zero newly mined candidates; it preserves that historical lineage while qualifying corrected contracts for the same 769-row, 502-source denominator. Historical admission evidence is not relabelled as newly executed evidence.

Benchmarking Persistent Operational Agency in Source-Grounded Executable Systems

This private dataset distributes the manifest-backed artifacts for the active operate_v0_62_0 release in Xnhyacinth/OPERATE-Benchmark, retaining private maintenance history. The public current distribution is Xnhyacinth/OPERATE. The promoted Core contains 769 scenarios grounded in 502 canonical physical sources.

The v0.62 manifest selects 701 corrected contracts under v0.62, 48 under v0.58, 8 under v0.59, 11 under v0.60 and 1 under v0.61. Inherited source paths do not authorize resuming historical provider runs.

OPERATE evaluates persistent LLM supervision and control in seeded executable systems. The environment evolves independently of the model, observations are partial, events and actions are typed, and every score is linked to recorded evidence.

The Git repository contains the promoted Core definition, source locks, scenario contracts, and compact redistributable source assets. This HF runtime companion carries the matching release manifest, hash-bound formal replay and diagnostic evidence, and declared private runtime archives needed by a fresh clone. Upstream assets that cannot be redistributed retain their original URLs, hashes, licenses, and acquisition instructions.

The repository code is MIT-licensed, but this runtime companion is a mixed- license collection. Redistributed backend assets retain their upstream terms; their exact license files and per-root bindings are included in the bundle manifest. The dataset-card other value therefore applies to the companion as a whole and must not be read as relicensing upstream data under MIT.

Primary results use logical_persistent with live scoring 0.18.0 (wait_relative_outcome_v1). Qualification artifacts remain bound to 0.15.0 and are not rewritten. realtime_persistent produces a separate supervision scorecard under native_dt_v1 / realtime_persistent.v3 for proactive monitoring, correct silence, latency, cancellation, supersession, action lifecycle, and safety takeover. The default speed card is the 37 Core rows whose native tick is at most 60s and whose plant wall is at most 30 minutes. Exact model/provider/harness/context settings are bound by a treatment hash. Requested and effective concurrency are separately immutable in the shard run scope and output namespace. Use the immutable private HF commit recorded in the verified publication receipt; an older release revision is not a valid runtime binding for this release. Full and Lite are also provided as reversible Parquet tables. Lite retains exact Core-locked rows and applies a quality-gated development compression; it is an efficiency/development track, not a Full leaderboard denominator or a statistically representative sample. The current Lite contains 141 rows from 82 physical sources. Lite v11 freezes the prior model-informed 144-row selection. The routine budget is 192 ticks, with the shortest longer candidate per backend/family/difficulty-mode stratum retained: Dyna moo (298) and breakdown recovery (400), alongside Sweep (111). Three longer alternatives remain in Core/diagnostics. The interim v10 uniform cap is withdrawn. The new set retains 204/204 declared parent features across six domains, not complete Core or scientific coverage. Native ticks do not bound wall time or API calls; the cutoff is not a definition of long-horizon agency. Historical zero/ceiling scores await repaired-measurement and valid-baseline retests, not automatic exclusion. The review ledger records decisions. Private Parquet retains complete maintenance metadata; the public export omits redundant display fields.

Historical qualification retains its original implementation identity. Maintenance fixes use affected-scope tests, while new runs and their finalized results bind their actual code separately. Old evidence is never relabelled.

The promoted manifest is formal-evaluation ready, but formal provider runs are not complete. This maintenance archive remains private independently of the already-public code/data distribution; public_release_ready and leaderboard_eligible are both false. Formal logical-persistent and realtime-persistent evaluation and runtime-evidence distribution remain the three explicit release blockers: formal_logical_persistent_evaluation_pending, formal_realtime_persistent_evaluation_pending, and formal_runtime_evidence_distribution_pending.