REVIEWED PUBLIC SOURCE / HISTORICAL SNAPSHOT
Geometry: preserved operations and missing-context controls
This is the exact UTF-8 source text at the recorded commit. The implementation repository is private. Older records may describe earlier access, hosting or release decisions; their original wording is preserved. Links and commands inside the record remain inert source text and do not require repository access to read this record.
- Original path
docs/GEOMETRY_CONTROL_FINDINGS.md- Commit
b234cff1697714776dcceeaf9897e29216e4fd4c- SHA-256
d261d949d70897e30813fa81f58b8a9e642e4c8d2ae1bcbab0659608ae54544c- Size
- 8,244 bytes
Download the exact original bytes · Read the source
# GR-CTRL-001 findings and delivery evidence Both supervised version-1 executions completed and trusted CPU replay passed. The final integrated source verification is recorded below. See the accepted [protocol](GEOMETRY_CONTROL_SPEC.md). All results are local, synthetic controls. ## First supervised execution Run `geometry-3ce10d4c8981` completed on the NVIDIA GeForce RTX 5070 using the designated GPU environment. All 13 CPU controls and both CUDA formulations on 18 rays (256 steps each) passed. The erased-context witness retained equal local features with different finite boundary outcomes: flat outward-outer-boundary, curved inner-boundary. Restoring the metric reproduced both outcomes. | Measurement | Recorded result | Acceptance / interpretation | |---|---|---| | Maximum CUDA vs CPU terminal-state error | 8.881784197001252e-16 for each formulation | <=1e-9 absolute +1e-9 relative | | Maximum paired CPU grid error | 1.4406253967536031e-12 | <=2e-7 | | Maximum null residual, including long/crossing cases | 1.7516404060557988e-7 | <=2e-7 | | Maximum conserved-quantity drift | 1.4043114848760752e-7 | <=2e-7 | | CPU short-grid reference time | 0.0796995 seconds | Descriptive implementation timing | | CUDA compute time | 13.1747718 seconds | This small dispatch-heavy control is slower on CUDA | | Host to device / device to host | 0.7967951 / 0.0001798 seconds | Synchronizations included | | Peak CUDA allocation / reservation | 14,848 / 2,097,152 bytes | Bounded parity grid, no scale claim | | Supervised aggregate peak RSS | 1,045,975,040 bytes | Below 12 GiB | | Total retained execution wall time | 20.0 seconds | Below registered resource limits | Receipt: `G:/RenigmaLLM_Artifacts/geometry-runs/geometry-3ce10d4c8981/receipt.json`. Receipt SHA-256: `d72159230d46c3d8385c4cf9b39d954eb27d88a87bb6f85c418fffe7a01a3578`. Evaluated source SHA-256: `24791573da74e65d1f00e619a58b5aeab5a916eb65c0ae01b2c5dcf357f341bc`. Trusted replay receipt: `G:/RenigmaLLM_Artifacts/geometry-validation/first-verify.json`. The code had not yet been committed at this freeze; the source manifest binds the evaluated bytes rather than claiming the recorded parent commit contains them. Internal code review then identified two monitoring robustness gaps inherited from the shared supervisor: a later telemetry sample could overwrite an earlier monitor failure, and a later memory recovery could hide a previously sampled reserve violation. The final integrated run will include sticky monitor failure and minimum observed free-memory retention. No such failure or reserve violation was observed in this first run. It remains retained under its original source; matching trusted code is required to replay it. Numerical cases/tolerances remain version 1; the subsequent run evaluates the hardened supervision and updated map. ## Interpretation boundary These tests compare numerical representations of prescribed GR backgrounds. They do not solve Einstein-Maxwell dynamics, independently validate gravity, establish a material medium, certify floating-point intervals, or establish language understanding. The context-erased representation is deliberately insufficient; detecting its failure is an expected successful control. The curated local map is available at `/transformations` on the existing loopback workbench. Evidence excerpts are packaged for offline inspection; full external references open only by explicit navigation. No published website, cloud service, private workbench document or model engine is changed by this slice. ## Final integrated execution Run `geometry-ecd4cccad6f0` evaluated implementation commit `d43e7ed2650d54a687e7321cb83950bd766485a9`, including monitoring hardening and map revision TRANSFORM-MAP-001.2. It completed all 13 CPU controls and 18 paired CUDA cases. Trusted replay passed; maximum CUDA discrepancy remained 8.881784197001252e-16 for both formulations. Numerical configuration was unchanged. Wall time was 18.578 seconds, aggregate peak RSS 1,013,456,896 bytes, and output including the frozen source was 92,869,730 bytes. GPU allocation peaked at 14,848 bytes; minimum observed free GPU memory was 11,530,141,696 bytes. The CPU reference took 0.0786101 seconds, CUDA compute 12.2974301 seconds, upload 0.7089500 seconds, and download 0.0001778 seconds. This establishes parity and execution, not speedup. - Receipt: `G:/RenigmaLLM_Artifacts/geometry-runs/geometry-ecd4cccad6f0/receipt.json` - Receipt SHA-256: `ddbe2e494c252a72390ec5a00d4a378efe24a19fc4fb728423a0ec16c7ea68e2` - Source SHA-256: `9d838678192646658d3d5265b1f1c317e109168f08f42c739ec5cb41152a5a59` - Trusted replay: `G:/RenigmaLLM_Artifacts/geometry-validation/final-verify.json` The map deliberately cites the first executed snapshot as a historical evidence record. Its first-run times and receipt remain accurate; the final run above establishes the integrated source after the evidence was added to the map. ## Interface verification Actual Chromium verification passed on the integrated source, with keyboard Enter/Tab/Escape, touch evidence expansion, all filter options and empty results, all source previews, JavaScript-disabled native forms/navigation, 320px layout, visible focus, reduced motion and no external requests. Screenshots were visually inspected. Receipt: `G:/RenigmaLLM_Artifacts/geometry-validation/browser/transformation-map-browser-ZnmaMz/receipt.json`. The first browser attempt failed because its selector expected “Transformation map” while the link read “Transformation research map”. The selector was corrected; the failed receipt remains in `transformation-map-browser-KANNUM`, and subsequent passes in `transformation-map-browser-0hRdPM` and `transformation-map-browser-ZnmaMz`. The final full CPU suite passed **1,732 tests and 27 subtests**, with **29 skips** for unavailable CUDA and optional SciPy/Ripser in that environment. This preserves the isolated CPU environment; skips are not passes. Log: `G:/RenigmaLLM_Artifacts/geometry-validation/pytest-final.log`. Focused GPU-environment regression checks passed 128 tests, including existing operator, spectral and distillation GPU checks and the telemetry regressions. The final math-check and geometry-check CLIs passed. The indexed delivery receipt is `G:/RenigmaLLM_Artifacts/geometry-validation/delivery-receipt.json`. No substantive CodeRabbit review or independent scientific replication is claimed; this local slice creates no PR and performs no deployment. ## Open and reproduce locally From this checkout, launch `./scripts/start-workbench.ps1` and choose **Transformation research map** on the starting screen. The existing launcher selects this checkout's source and the isolated external environment, and detects unavailable storage or occupied ports without terminating other processes. Direct local route: `/transformations`; manifest: `/api/transformations`. For research CLI checks, set `PYTHONPATH` to this checkout's `src` directory. Run `python -m renigma geometry-check` in the isolated CPU environment, `python -m renigma geometry-experiment` in the designated GPU environment, and `python -m renigma geometry-verify RUN_DIRECTORY` with matching trusted source. The current source verifies the final integrated run above. Verification checks saved outputs against trusted equations; it does not execute archived Python or repeat CUDA and is not independent scientific replication. ## Prefreeze development observation The first numerical redshift transport check used affine step 0.025 and had residual 2.937206033948314e-12 against an unchanged 1e-12 tolerance. The numerical resolution was refined to step 0.01 before freezing version 1. This was a development check, not an official experiment or held-out result; it is retained here rather than described as a pass. The final case is explicit in the frozen configuration. No official threshold was relaxed. ## Subsequent decision The next research gate remains a separately preregistered tensor-network factorization study with declared axes/ranks, complete storage accounting and live-write observation recovery. This geometry control supplies an example of context sufficiency; it supplies no new compression or language result.