# Execution audit — Infinite Players 3.0.0

Date: 18 October 2026. The uploaded source was copied to `verification_run` before any execution. The original `source/Infinite_Players_Final` directory was not edited. Fresh output is in `verification_output`.

## Validation outcome

All requested numerical checks completed successfully. This is fresh local source execution, not external replication, peer review, real-player validation, or proof-assistant verification.

| Check | Fresh result |
|---|---:|
| Supplied research file manifest | 82/82 files verified |
| New unit tests | 50/50 passed |
| New deterministic reference assets | 32/32 SHA-256 hashes matched |
| Preserved v2 unit tests | 22/22 passed |
| Preserved v2 reference assets | 14/14 SHA-256 hashes matched |
| Standalone `RUN_DEMO.py` | Passed |

Exact command from `verification_run`:

```bash
python VERIFY_RELEASE.py
python REPRODUCE.py --output /workspace/scratch/2533268896e4/verification_output --legacy
python RUN_DEMO.py --output /workspace/scratch/2533268896e4/standalone_demo_output
```

## Environment and dependency qualification

Fresh runtime: Python 3.12.14; NumPy 2.3.5; SciPy 1.17.0; Matplotlib 3.10.8; Linux x86-64/glibc 2.39; single-threaded numerical libraries; no accelerator. The supplied environment records Python 3.13.5/glibc 2.41. All 32 deterministic assets match despite that difference. The generated upstream reproduction report uses a hard-coded “same-environment” description; this audit records the actual environment.

**Recommended Hugging Face quickstart: Python 3.11+ for `requirements-locked.txt`.** Both locked NumPy and SciPy declare Python >=3.11. The broad Python >=3.10 declaration can use older compatible dependencies from `requirements.txt`, but Python 3.10 was not exercised here. Dependencies were already installed; no fresh pip installation was needed or claimed.

## Primary finite arena

Mean normalized discounted value is averaged over six initial ground states and 16 arena seeds. Semantic total variation is exact enumeration of the six-step semantic law from ground state zero. The paired intervals are descriptive t intervals over 16 seeds; they are not evidence for a larger real-game population.

| Controller | Mean normalized value | Gain versus source | Paired 95% gain interval | Mean six-step semantic TV |
|---|---:|---:|---|---:|
| source_anchor | 0.511568389 | 0.000000000 | [0.000000000, 0.000000000] | 0 |
| path_shield_LP | 0.642582819 | 0.131014430 | [0.116648674, 0.145380186] | 4.30286e-16 |
| conditional_class_only | 0.731438499 | 0.219870110 | [0.209962867, 0.229777353] | 0.483658 |
| unconstrained | 0.823908873 | 0.312340484 | [0.293457438, 0.331223529] | 0.856473 |
| nullspace_24_updates | 0.605199312 | 0.093630923 | [0.083067074, 0.104194772] | 2.30614e-14 |

The constrained LP optimum improves value within the protected semantic-law fiber. Weaker constraints allow higher return and substantially more semantic drift. The 24 projected-Q updates remain below the constrained LP ceiling.

## Sequence inference

| Events per record | Records | Operator-tree accuracy | Full-HMM replay accuracy | Paired-product accuracy |
|---|---:|---:|---:|---:|
| 8 | 1000 | 93.6% | 93.6% | 91.0% |
| 16 | 1000 | 97.5% | 97.5% | 96.2% |
| 32 | 1000 | 99.9% | 99.9% | 99.7% |

The maximum posterior difference between tree and independently implemented full replay is 5.3290705182007514e-15 across 3,000 records. The tree changes update cost; it does not improve on the same exact Bayesian inference model.

## Continuous Markov lag study

| Lag | Synthetic trajectories | Accuracy | Wilson 95% interval |
|---|---:|---:|---|
| 0 | 16000 | 100.00000% | [99.97600%, 100.00000%] |
| 1 | 16000 | 99.57500% | [99.46161%, 99.66459%] |
| 2 | 16000 | 96.46250% | [96.16493%, 96.73776%] |
| 3 | 16000 | 89.07500% | [88.58222%, 89.54902%] |
| 4 | 16000 | 80.01875% | [79.39200%, 80.63109%] |
| 6 | 16000 | 66.08125% | [65.34389%, 66.81089%] |
| 9 | 16000 | 56.13750% | [55.36723%, 56.90482%] |

Each trajectory has 64 events. These 112,000 trajectories use a fixed synthetic known-model design.

## Retrospective query memory

| Retained prefix bits / 256 | Trials | Observed accuracy | Analytic accuracy |
|---|---:|---:|---:|
| 0 | 16000 | 49.81250% | 50.00000% |
| 8 | 16000 | 51.39375% | 51.56250% |
| 32 | 16000 | 56.09375% | 56.25000% |
| 64 | 16000 | 62.56875% | 62.50000% |
| 128 | 16000 | 75.10625% | 75.00000% |
| 256 | 16000 | 100.00000% | 100.00000% |

## Local evidence replacement

| Archived events | Tree compositions per replacement | Replay steps per replacement | Tree array bytes | Raw two-sign bytes | Largest centered logscore discrepancy |
|---|---:|---:|---:|---:|---:|
| 256 | 8 | 256 | 81,920 | 512 | 3.97904e-13 |
| 2048 | 11 | 2048 | 655,360 | 4,096 | 6.82121e-12 |
| 16384 | 14 | 16384 | 5,242,880 | 32,768 | 6.18456e-11 |

The work units differ (batched matrix composition versus forward-vector step); the arithmetic also depends on four hypotheses and two hidden states. Tree arrays use 160 times the two-sign raw payload on these fixtures, excluding Python object overhead. These are update accelerators, not compression gains.

| Archived events | Tree build (s) | 16 tree updates (s) | 16 replay updates (s) | Observed replay/tree ratio |
|---|---:|---:|---:|---:|
| 256 | 0.007675 | 0.006572 | 0.035252 | 5.36 |
| 2048 | 0.073419 | 0.010919 | 0.293609 | 26.89 |
| 16384 | 0.512575 | 0.011217 | 2.205227 | 196.59 |

**Timing ratios exclude tree construction and are one local diagnostic run.** Do not present them as a general hardware, GPU, or commercial-game benchmark.

## Additional controls and audits

- Positive-circuit approximation: 1,000 trials, zero bound violations; maximum posterior TV 0.010810353214991988; maximum observed TV/bound ratio 0.2204985843214893.
- Calibration fixture: protected reward goes from 0.4 before calibration to 0.85 in world 0; the negative world control remains 0.4. The no-headroom arena changes value by at most 6.661338147750939e-15.
- Lifetime promotion study: 10,000 histories per condition, 40 versions, 256 audit sessions/version, lifetime budget 0.05. Zero-gain condition: 60/10,000 histories accepted any promotion (0.6%). Positive gain 0.3 condition: 10,000/10,000 accepted at least one promotion; mean promotions 29.5762.
- Exact arithmetic audits pass for 16 lag cases/2,550 atoms, 16 hidden operator paths, 276 deterministic encoders, 2,046 semantic words, and 72 persistent-composition cases. These remain internally authored audits, not external mathematical review.

## Standalone demo

Normalized source value: 0.5378491294755662. Descendant value: 0.6666691635919242. Six-step semantic TV: 3.2289488872397007e-16. Maximum signature residual: 3.885780586188048e-16. One 1,000-step execution trace has reward mean 0.676; that single trace is not a separate population performance estimate.

## Issues and unexecuted checks

No runtime failure or supplied deterministic reference mismatch was found. The Python requirement distinction above should be made explicit in the Hugging Face release. Native Windows execution, TeX/PDF rebuild, real-player testing, learned world models, commercial-engine execution, local-LLM integration, and GPU benchmarking were not performed. The title’s beyond-human-expertise aspiration must not be represented as an achieved evaluation result.

Machine-readable command outcomes, environment, full metrics, and exact audits are in `execution_audit.json`.