Fixed-source, binary64 sampled trajectories through the public CSL::controllers replay API. The estimator and controller are caller-defined callbacks. These examples check one supplied initial state and error sequence each; they do not establish policy nonanticipativity, state-estimate accuracy, all-initial-states feasibility, all-disturbance performance, optimality, family completeness or validated rounding enclosure.
| Fixture | Physical states | Released observations | Actions |
|---|---|---|---|
| Partial, noisy, delayed observation | (1, 7) → (2, 7) → (9, 7) | 4 before(0); 7 measured after(0), released before(1) | 1; 7 |
| Post-decision feedthrough | 2 → 3 → 10 | 7 = x₀ + 2u₀ + v₀, released after(0) | 1; 7 |
| No observation channel | 5 → 7 | None; no implicit observation | 2 |
The noisy partial fixture checks sample channel/index and delayed release identity; the post-decision fixture includes the control-feedthrough term; the channel-free fixture checks that no full-state reading is invented. Decision stage, each available observation identity, state, estimate, action and retained inputs are checked inside every measured call.
One excluded warmup per fixture, then seven alternating measured batches per fixture with 5,000 fresh and fully checked calls per batch. All 24 rows passed with zero observed failures. These are verified-batch durations in milliseconds, not isolated controller latency. Scope: batch of fresh public LinearSystem snapshots and replay_observed_policy calls, estimator/controller callbacks, full literal state/history/release/input oracle checks, and per-call result JSON construction and assignment/replacement; excludes process startup, JSONL serialization, Python saved-row validation, persistence and rendering.
| Fixture | Calls/batch | Successful batches | Median ms | Range ms |
|---|---|---|---|---|
| partial noisy delayed | 5,000 | 7/7 | 254.442 | 250.066–261.357 |
| post feedthrough | 5,000 | 7/7 | 160.891 | 157.320–166.016 |
| no channel | 5,000 | 7/7 | 68.216 | 67.760–76.103 |
Performance interpretation is exploratory and inconclusive. There is no comparator, source-mapped limiter profile, application-scale workload, speedup or throughput claim. Internal phase durations and observed native thread count are unavailable.
Frozen algorithm source c23aaf13c1c1e5480cc22dd5bac5cd3e79562609; reviewed harness 959755ed5f69641ae643e852c00efe718fa9eb86; accepted collection ext04-df62c7c6d295e135. Raw JSONL SHA-256 7c6c33024ffa93eb37065bf1a66df8a8f02827e0e0e039ec93a9090db1b900d8; native manifest SHA-256 13950e8c24b689b18109b9da182d575ab5d55f3f0dc5749f6510b9f438811343. The later artifact commit carries this page and is distinct from the source and harness revisions.
Release x64 MSVC v143, C++20, /O2 /Ob2 /DNDEBUG; one sequential driver requested. Windows host Windows-11-10.0.22631-SP0, 12 logical processors. Pre/post five-second host utilization was 1.1%; source cache, compiler record, build log and host snapshots are retained. The collection status records acceptance for report review and a then-pending publication gate.
Versioned publication ext04-observed-replay-eb0b06e560360611d2f2a16fba9b1204e0415b4ba5cd0d5210427c6349963b34. Report envelope · Publication manifest.
