Declared read-only scheduling study¶
This study runs four independent SQLite queries per task through the declared tool scheduler. Each callback opens and closes its own read-only connection to frozen public UCI Wine Quality data. Dependency, resource and concurrency declarations are explicit; no model is used and no artificial delay is inserted to produce a speedup.
All 96 observations preserve the independently frozen ordered outputs, covering 384 actual queries. Parallel enforcement reaches four concurrent callbacks in every held-out scheduled observation. It is faster than the corresponding shadow scheduler, but the complete adapter remains slower than tracing alone on these small local operations.
Results¶
| Held-out condition | Correct task batches | Mean task time | Peak scheduled callbacks |
|---|---|---|---|
off |
8/8 | 7.128 ms | Sequential host execution |
trace |
8/8 | 7.205 ms | Sequential host execution |
shadow_schedule |
8/8 | 11.313 ms | 1 |
enforce_schedule |
8/8 | 8.084 ms | 4 |
shadow_budget |
8/8 | 12.650 ms | Sequential host execution |
enforce_budget |
8/8 | 12.614 ms | Sequential host execution |
shadow_combined |
8/8 | 14.031 ms | 1 |
enforce_combined |
8/8 | 11.738 ms | 4 |
Scheduler enforcement minus shadow averages -3.229 ms, with a task-bootstrap interval of [-3.685, -2.460] ms. Enforcement minus tracing alone averages +0.879 ms, with an interval of [0.161, 1.597] ms. The latter retains the full control overhead rather than presenting the favorable shadow comparison as an overall speedup. The combined scheduler/budget path is also slower than tracing alone. The four-call budget never binds on a successful four-query task.
All observations completed; there were no failed, cancelled or missing real attempts. Offline tests separately cover those paths, including partially completed mutations and budget races. There are zero model calls and zero model tokens. Native model-cost totals are empty/zero because no model was invoked; whole-task operating cost remains unknown in the ablation. No paid provider was used.
Frozen design and limits¶
Six task batches contain two pilot and four held-out cases, with two repetitions and eight conditions: 96 observations, including 64 held-out observations. Every held-out condition occupies every order position once. Pilot ordering covers only four cyclic positions. Task inputs, reference query results, database/source hashes, policy versions, environment and execution code are frozen before measurement.
The native ablation includes off, tracing, separate shadow/enforce scheduler and
budget arms, and their combination. Every callback declares no prerequisites,
read access to the same dataset, no writes, read-only effects and worker-thread
support. The four queries use separate SQLite connections; sharing a thread-bound
connection would not satisfy that declaration. URI read-only mode, query_only
and a restrictive authorizer reject writes and database attachment.
Each observation creates fresh requests, scheduler and harness state. Timed work includes callback construction, control setup, thread-pool setup, connections, queries, result collection, cleanup and tracing where enabled. Dataset creation, reference grading and artifact export are outside timing. Filesystem/SQLite caches remain warm, and the machine is shared. The unchanged database checksum is verified after execution by the plan loader.
The scorer checks call IDs, submission order and complete typed query results against reference results computed before the measured runs. The same SQLite implementation computes the references; this evaluates orchestration and output preservation, not independent correctness of the database engine. Four held-out batches and one host do not support a general performance claim. Worker contention can increase summed callback time even while wall time falls.
Evidence and reproduction¶
The checksummed evidence index retains all 96 receipts, 84 native traces, declared and actual schedules, budget decisions, frozen database/CSV data, source snapshots and native ablation/study/HTML reports. The twelve uninstrumented observations have host receipts without fabricated spans. Operating-cost uncertainty is preserved separately from the native recorded-model accounting scope.
The exporter rechecks slot identity, ordered outputs, independent grades, tool counts, schedule limits, trace hashes, dependency validity and agreement between host and native control evidence. Partial/empty exports keep planned denominators; unexpected attempt artifacts require review. No estimator-ledger attribution is invented for this caller-configured dependency/resource plan.
python examples/real_agent_study/archive.py research/tool-scheduling-2026-09/index.json --out /tmp/scheduling-evidence
python -m examples.scheduling_study.export --plan /tmp/scheduling-evidence/study/plan.json --out /tmp/scheduling-report
Use a complete core checkout at 33646da91515bbf0add2f060ee5dbecebc8d75f4
with the archived source overlay. Readers/exporters execute no tool callbacks.
For new measurements, prepare a fresh root with examples.scheduling_study.protocol
and explicitly execute examples.scheduling_study.runner --plan PLAN --execute.
Changed declarations, code, data or grading rules require a new frozen protocol.
UCI Wine Quality remains CC BY 4.0 with attribution in the bundle; study code is Apache-2.0. No model weights, inference runtime or external side effects are needed for this workload. No result activates a production schedule.