REVIEWED PUBLIC SOURCE / HISTORICAL SNAPSHOT
Indexed tensors: exact retrieval and compute findings
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/WORKSTATION_TENSOR_FINDINGS.md- Commit
b234cff1697714776dcceeaf9897e29216e4fd4c- SHA-256
77681f949fbc1738af59981d9d7b3f8bbe221c868386d91b3b599a54c9f319b6- Size
- 9,164 bytes
Download the exact original bytes · Read the source
# WORKSTATION-001 / TENSOR-001 findings The frozen four-stage pipeline completed locally on 2026-10-06: exact CPU memory, separate CUDA calibration, held-out indexed contraction, and trusted CPU replay. The [protocol](WORKSTATION_TENSOR_SPEC.md) and prior studies remain unchanged. This is a bounded synthetic study, not a supercomputer-equivalence claim. ## Execution and control Run: `G:/RenigmaLLM_Artifacts/workstation-runs/workstation-f79a7f127305`. - Implementation commit: `7f9db7666946fba6b224a45d97f41869dbe1ce1b`. - Evaluated source SHA-256: `d5a7ac57ae7fb86cee45c9e5dc84e05b24d82ff875f9232a9758ca016091f729`. - Configuration SHA-256: `6aa50225194d46a8c8e05ec5129a17f5b5af5a1d4ae425c266bf389e5d5d1a09`. - Receipt SHA-256: `d02149f216bcaedb8ab4e0c62c6c2b208f80c47e980aec1897c46d89701731fb`. The first memory attempt was deliberately canceled after 21.25 seconds. Its nonzero exit, incomplete status and sealed outputs remain. Explicit resume created memory attempt 002 and completed the remaining stages; no attempt was overwritten. A separate live CUDA control launched a checked batched matrix operation and then verified that cancellation terminated its owned worker. That control is not another scientific evaluation. An earlier observer reached calibration after it had completed and correctly sent no cancellation request. | Stage | Completed worker wall seconds | Supervised peak process-tree RSS bytes | |---|---:|---:| | Memory attempt 002 | 50.750 | 100,356,096 | | Calibration | 21.922 | 1,025,581,056 | | Evaluation | 21.797 | 1,082,437,632 | | Trusted replay | 49.265 | 153,374,720 | Recorded cumulative run time, including the incomplete attempt, was 166.422 seconds. Recorded output including source snapshots and caches was 252,377,121 bytes. Peak CUDA allocation was 319,702,016 bytes, reservation 394,264,576 bytes, and minimum observed free device memory 11,102,322,688 bytes. Monitored limits passed; these are not OS isolation guarantees. The CPU executes numerical work with one math thread (cap four); the selected device is the RTX 5070, IEEE float32, TF32 disabled. CPU reference is float64 and the timing baseline is float32. The GTX 1660 Ti is inventory-only; NPU, multiple GPU execution, WebGPU and cloud are unscheduled here. The product engine and cloud services are unchanged. ## Exact memory outcome Every arm covered 64 episodes, 7,744 snapshots, 2,802 explicit reads and 158,144 all-key probes. Miniature controls exhaust all 625 four-key states and their single writes/reset, with 45,000 reads and 45,000 probes per arm. Baselines and full-rank controls passed. Lower-rank failures are retained scientific results. | Arm | Wrong reads | Wrong probes | Mean shared-256 bytes per snapshot | Mean standalone bytes | |---|---:|---:|---:|---:| | Dense | 0 | 0 | 50.6641 | 220.0000 | | Sparse | 0 | 0 | 40.6604 | 211.9885 | | TT rank 1, natural | 489 | 107,369 | 112.7070 | 548.0000 | | TT rank 1, reversed | 492 | 107,892 | 112.7109 | 549.0000 | | TT rank 2, natural | 452 | 94,636 | 227.6017 | 662.8946 | | TT rank 2, reversed | 452 | 94,742 | 227.9527 | 664.2417 | | TT rank 4, natural | 0 | 0 | 392.1450 | 827.4380 | | TT rank 4, reversed | 0 | 0 | 392.1490 | 828.4380 | Four of eight arms qualified for exact observations. No tensor arm earned storage merit. The cheapest exact TT arm used 9.6444 times the better exact baseline's complete shared storage. Full-rank TT had 544 rank-growth mutations and no truncations. Rank-1 miniature failure includes `[0,0,0,1]`, then writing key 0 to 1: the compressed state loses the existing value at key 3. These are live updates from each codec's own serialized record, with no oracle repair. Shared-256 storage amortizes descriptors across snapshots; it is not a claim about a tensorized joint database. Updates materialize at most 32 float64 values. Keep the compact lossless memory representation: this float64 TT codec does not improve storage in this world. ## Indexed CUDA outcome All 18 cases and 419,328 query outputs passed the declared full-output parity test. Maximum absolute CUDA error was 1.8654559668096482e-6, within `1e-5 + 1e-4 * abs(reference)`. Independent dense, basis and sum controls passed; the all-zero substitute failed as required. Cores were supplied directly in TT form. A 20-axis virtual index has 1,048,576 possible addresses; it does not mean that every address was queried or that a million arbitrary records were compressed. Seed-0 calibration retained all three repeats for each candidate. Median end-to-end seconds were 0.648653 (chunk 256), 0.0412379 (4,096), and 0.0175282 (65,536). All passed parity. Chunk 65,536 was sealed before seed-303 evaluation. | Held-out case (axes / rank / queries) | CPU float32 seconds | CUDA end-to-end seconds | CPU / CUDA ratio | Maximum absolute error | |---|---:|---:|---:|---:| | 10 / 4 / 256 | 0.000533 | 15.181667 | 0.000035 | 3.947e-7 | | 10 / 4 / 4,096 | 0.002122 | 0.002337 | 0.9083 | 5.244e-7 | | 10 / 4 / 65,536 | 0.030238 | 0.005928 | 5.1014 | 5.244e-7 | | 10 / 16 / 256 | 0.000549 | 0.002377 | 0.2309 | 5.671e-7 | | 10 / 16 / 4,096 | 0.002527 | 0.002287 | 1.1048 | 6.595e-7 | | 10 / 16 / 65,536 | 0.074840 | 0.080651 | 0.9279 | 6.595e-7 | | 10 / 32 / 256 | 0.000632 | 0.009794 | 0.0646 | 7.017e-7 | | 10 / 32 / 4,096 | 0.003239 | 0.004738 | 0.6836 | 8.420e-7 | | 10 / 32 / 65,536 | 0.102478 | 0.012570 | 8.1523 | 8.420e-7 | | 20 / 4 / 256 | 0.001007 | 0.003744 | 0.2691 | 5.295e-7 | | 20 / 4 / 4,096 | 0.004470 | 0.003817 | 1.1710 | 7.844e-7 | | 20 / 4 / 65,536 | 0.062690 | 0.004625 | 13.5544 | 9.117e-7 | | 20 / 16 / 256 | 0.001158 | 0.003795 | 0.3051 | 8.124e-7 | | 20 / 16 / 4,096 | 0.005219 | 0.003620 | 1.4419 | 1.062e-6 | | 20 / 16 / 65,536 | 0.128169 | 0.008192 | 15.6464 | 1.256e-6 | | 20 / 32 / 256 | 0.001317 | 0.003787 | 0.3478 | 1.242e-6 | | 20 / 32 / 4,096 | 0.006557 | 0.003558 | 1.8428 | 1.507e-6 | | 20 / 32 / 65,536 | 0.218299 | 0.017502 | 12.4727 | 1.865e-6 | A ratio above one favors CUDA. These are single held-out timings against this one-thread NumPy baseline, not stable service-latency estimates or comparison against every optimized CPU implementation. CUDA was faster in 9 of 18 cases. The first invocation in the fresh evaluation worker cost 15.182 seconds; its cause is not isolated by this protocol. Across all 18 cases, CPU time totaled 0.646046 seconds and CUDA end-to-end time 15.354990 seconds. The initial cost dominates this aggregate; omitting it would give an incomplete cost claim. Per-case preparation, H2D, contraction and D2H timings remain in the raw receipts. The next performance gate should preregister repeated cold/warm measurements, CPU thread comparisons and worker reuse, choosing a CPU/CUDA routing policy using calibration only before new held-out inputs. Do not tune that policy against these evaluation cases. Useful GPU work is established for some measured inputs; universal acceleration and production routing are not established. ## Validation and local use - Frozen-source CPU suite: **1,809 passed, 27 subtests passed, 30 skipped**. Skips cover unavailable CUDA and optional SciPy/Ripser in the original CPU environment. Log: `G:/RenigmaLLM_Artifacts/workstation-validation/pytest-frozen.txt`. - Focused GPU-environment suite: **77 passed**, including tensor compute, ownership, crash recovery, tampering, resource and cancellation checks. Log: `G:/RenigmaLLM_Artifacts/workstation-validation/pytest-gpu.txt`. - `renigma math-check`: passed, with retained JSON in the same validation folder. - Actual Chromium map checks passed: all filters/source links, keyboard, touch, no-JavaScript use, 320px layout, reduced motion and no external requests. Narrow screenshot inspected. Receipt: `browser/transformation-map-browser-LqT8zU/receipt.json` under the validation folder. - Launcher CheckOnly reported inventory without work; missing artifact root rejected without fallback. Real cancel/resume is in the scientific run. - Separate live CUDA cancellation: `cuda-cancel-live/receipt.json` under the validation folder. Fixed test scripts and their hashes are retained beside it. - Full trusted replay: `official-verify.json` under the validation folder. Replay uses matching installed source in a bounded fresh CPU interpreter, not archived executable code; it does not independently attest CUDA timing. From this checkout, run `Launch-Workstation.cmd`, or inspect first with `scripts/start-workstation.ps1 -CheckOnly`. The launcher selects the existing G: GPU environment. Ctrl+C or `renigma workstation-cancel RUN` interrupts owned work; `renigma workstation-run --resume RUN` validates retained attempts before continuation. Use that same numerical environment for canonical TT byte replay. Source/protocol changes reject resume; changed research settings need a new study. This delivery is local and committed. No remote publication, third-party model, customer record, NPU execution, independent replication, or general language understanding is claimed. The transformation map cites the frozen protocol as a protocol record; this findings document supplies its separate execution evidence.