Verdict: inconclusive exploratory single-adapter timings. No comparator or speedup is claimed.
Clean build source commit d0da7cb3b167f21e620b716b615e7b77cf5bd7aa; unchanged numerical source and driver 6602f0f88a8ef2071edefbbd913f7c4bbb6201df; saved report finite-feedback-a6e96fd96c7ac7cc; executable SHA-256 9caffb4aa5be3c73868258a79e01f94a72c90cb5116973496f0a00eac04b0e1d.
| Game | Initial winning | Initial action indices | Observed branch actions |
|---|---|---|---|
| clipping | 0 | ||
| branching | 1 | 0 | 1:0|2:1 |
| empty-action | 0 | ||
| nominal | 1 | 0 |
Whole-successor clipping, observed-state branching, empty actions, and nominal single-disturbance cases use exact authored transition tables. The frozen driver compares all winning flags and only these admissible-action cells: stage 0/state 0 in every game, plus stage 1/states 1 and 2 in branching. Other action-tensor cells are not independently checked by this report driver. The branching row checks distinct later actions after states 1 and 2 are observed.
| Game | Measured calls | Failures | Median ns | Min ns | Max ns |
|---|---|---|---|---|---|
| clipping | 7 | 0 | 3500 | 3300 | 4000 |
| branching | 7 | 0 | 4700 | 4300 | 5000 |
| empty-action | 7 | 0 | 3100 | 2900 | 3300 |
| nominal | 7 | 0 | 3300 | 3200 | 4200 |
Four smoke calls and four warmups are excluded. Seven round-robin measured repetitions per game (28 calls total) were verified. Median and min–max range describe these repetitions; the range is not a confidence interval.
The steady-clock interval begins before registry construction and ends after registered execution, all-winning-flags comparison, and selected action-cell checks. It excludes fixture construction, object destruction, CSV serialization, and process startup. It is a verified public-call interval, not native-only time or application end-to-end latency.
Intel64 Family 6 Model 151 Stepping 5, GenuineIntel; Windows-11-10.0.22631-SP0; 12 logical processors; MSVC 19.44.35215.0, x64 Release /O2 /DNDEBUG /std:c++20. One requested outer worker; native OS thread count was unavailable. One process with fresh registry, prepared plan and workspace per call; caches may be warm. Concurrent work was not fully quantified.
Single finite-table adapter, four small authored games, one host, and one measured session. Unlisted action-tensor cells were not checked. No competitor, source-mapped limiter profile, speedup, continuous-state disturbance enclosure, partial observation, physical-model correspondence, or global causal minimax claim. Internal phase and native-only counters were unavailable.
