# Sorso View, Issue 009 Computation Verification v1.0 ``` Document class: computation verification, Protocol 1 artifact Issue: SVCI Issue 009 Prepared: 2026-08-24 Method: Python Decimal, precision 34, statistics by explicit median function. NO MENTAL MATH. NO MEMORY ASSERTION. Every figure below is recomputed from the sixteen filed observations in this session. Status: PROVISIONAL. This is NOT a workbook confirmation. The locked workbook remains the seal-grade authority and must reproduce these values before publication. NOT COMPUTED, and deliberately so: Composite Coverage Score and the Compute Opacity Index. Both require dispositions for all 27 panel members. The 11 deferred entries have none, pending the F09-02 founder ruling. Publishing a coverage figure without that ruling would be a Core Rule 3 violation. ``` --- ## 1. Inputs ### 1.1 Locked constants, from the methodology, not recomputed ``` Base medians T1 10.5306 T2 3.5500 T3 4.6200 SVCI weights T1 0.45 T2 0.45 T3 0.10 ``` ### 1.2 Observed rates, Issue 009, from filed frames ``` T1 E001 AWS 6.88 frame I009-F0002 T1 E002 Azure 12.29 frame I009-F0006 T1 E003 GCP 11.0613 frame I009-F0020 T1 E004 Oracle 10.00 frame I009-F0028 T2 E020 Jarvislabs 2.69 frame I009-F0076 T2 E022 Atlas Cloud 2.95 frame I009-F0083 T2 E010 Hyperstack 3.20 frame I009-F0059 T2 E011 Verda 3.25 frame I009-F0062 T2 E008 Nebius 3.85 frame I009-F0044 T2 E006 Crusoe 3.90 frame I009-F0036 T2 E007 Lambda 3.99 frame I009-F0041 T2 E009 Together AI 3.99 frame I009-F0055 T2 E012 Voltage Park 4.50 frame I009-F0073 HELD, F09-19 T2 E005 CoreWeave 6.155 frame I009-F0033 T3 E023 RunPod 3.29 frame I009-F0086 T3 E025 Paperspace 5.95 frame I009-F0090 ``` ## 2. Derivations recomputed, the four that require division ``` E001 AWS p5.48xlarge 55.04 / 8 = 6.88 E002 Azure ND96isr H100 v5 98.3200 / 8 = 12.2900 E003 GCP a3-highgpu-8g 88.490000119 / 8 = 11.061250014875 E005 CoreWeave HGX H100 49.24 / 8 = 6.155 Locked display precision applied: E003 quotient 11.061250014875 quantized to 11.0613 E005 quotient 6.155 is exact and terminates; carried at three decimals, never 6.16 The other twelve entries publish per-GPU rates directly. No division. ``` ## 3. Vendor-published cross-checks recomputed ``` E025 Paperspace vendor formula and worked example: 8 x 5.95 = 47.60 vendor Multi-GPU table row H100 x8: 47.60 agreement: True E020 Jarvislabs vendor banner: 1 GPU 2.69 or 8 for 21.52 recomputed 2.69 x 8 = 21.52 agreement: True E004 Oracle table A 10.00 vs table B 10.00, difference 0.00 E008 Nebius nebius.com 3.85 vs docs 3.85, difference 0.00 E009 Together three surfaces all 3.99 ``` ## 4. Account-state reproduction checks ``` E003 GCP, 15:32 signed-in versus 16:13 logged out, six figures: Price 88.490000119 identical, difference 0E-9 DWS Flex-start 38.3200 identical, difference 0.0000 DWS Calendar 41.6000 identical, difference 0.0000 Current Spot 53.093624548 identical, difference 0E-9 CUD 1 Year 61.383674231 identical, difference 0E-9 CUD 3 Year 38.864383195 identical, difference 0E-9 E022 Atlas Cloud, 21:54 signed-in versus 22:00 logged out, eight values: 2.95, 3.50, 80 VRAM, 117.19 GiB, 12 vCPU, 8 max, 141 VRAM, 234.38 GiB all identical. CONCLUSION, evidenced not assumed: list pricing at both providers is invariant to account state. ``` ## 5. Tier medians and index, both scenarios ### SCENARIO A, E012 INCLUDED at 4.50 ``` T1 median (n=4) 10.53065 T2 median (n=10) 3.875 T3 median (n=2) 4.62 pooled non-hyperscaler (n=12) 3.875 T1 index = 100 x 10.53065 / 10.5306 = 100.0004748067536512639355782196646 T2 index = 100 x 3.875 / 3.5500 = 109.1549295774647887323943661971831 T3 index = 100 x 4.62 / 4.6200 = 1E+2 SVCI = 0.45 x T1i + 0.45 x T2i + 0.10 x T3i = 104.1199319728982979983484749875815 SVCI public display 104.1 Hyperscaler Premium = T1 median / pooled non-hyperscaler median = 10.53065 / 3.875 = 2.717587096774193548387096774193548 display 2.7176x Neocloud Discount = 1 - T2 median / T1 median = 0.6320265130832379767630678068305375 display 0.6320 Price Floor, locked Section 9.12, K = MAX(1, ROUNDUP(N x 0.25)) N = 16, K = 4 bottom K rates 2.69, 2.95, 3.20, 3.25 median = 3.075 ``` ### SCENARIO B, E012 EXCLUDED ``` T1 median (n=4) 10.53065 T2 median (n=9) 3.85 T3 median (n=2) 4.62 pooled non-hyperscaler (n=11) 3.85 T1 index = 100 x 10.53065 / 10.5306 = 100.0004748067536512639355782196646 T2 index = 100 x 3.85 / 3.5500 = 108.4507042253521126760563380281690 T3 index = 100 x 4.62 / 4.6200 = 1E+2 SVCI = 0.45 x T1i + 0.45 x T2i + 0.10 x T3i = 103.8030305644475937729963623115251 SVCI public display 103.8 Hyperscaler Premium = T1 median / pooled non-hyperscaler median = 10.53065 / 3.85 = 2.735233766233766233766233766233766 display 2.7352x Neocloud Discount = 1 - T2 median / T1 median = 0.6344005355794751511065318854961470 display 0.6344 Price Floor, locked Section 9.12, K = MAX(1, ROUNDUP(N x 0.25)) N = 15, K = 4 bottom K rates 2.69, 2.95, 3.20, 3.25 median = 3.075 ``` ## 6. Scenario A reproduces Issue 008 on all seven metrics The Issue 008 Claude-side verification note of 2026-08-16 recorded a recomputation with E012 included. Compared against Issue 009's entirely independent capture run: ``` Issue 008 note Issue 009 Scenario A T1 median 10.53065 10.53065 T2 median (n=10) 3.875 3.875 T3 median (n=2) 4.62 4.62 pooled non-hyp (n=12) 3.875 3.875 Hyperscaler Premium 2.7176x 2.7176x Neocloud Discount 0.6320-class 0.6320 Price Floor (K=4) 3.075 3.075 site-published SVCI 104.1 104.1 ``` This RETIRES the open item in the Issue 008 note that the published 104.1 was "not independently re-derivable here". It is now derivable, from Issue 009 evidence, under Scenario A. ## 7. Provenance caution on Scenario B ``` Scenario B lands at an SVCI internal value that is identical to available precision to the SEALED ISSUE 004 figure of 103.8030305644476. This is a coincidence of composition. It is not an error and not a repeat. The canonical archive already records exactly this pattern for Issues 003 and 004, which independently landed on the same value from separate capture runs. IF SCENARIO B IS ADOPTED, the issue text and the archive increment MUST state plainly that Issue 009 is a distinct capture run that coincides with Issue 004, in the same terms the archive uses for the 003/004 case. Left unstated it reads as a duplication fault. ``` ## 8. Movement statement, recomputed entry by entry ``` Every one of the sixteen captured rates equals its Issue 008 baseline. Sixteen differences computed, sixteen zeros. E001 6.88 - 6.88 = 0.00 E010 3.20 - 3.20 = 0.00 E002 12.29 -12.29 = 0.0000 E011 3.25 - 3.25 = 0.00 E003 11.0613 -11.0613 = 0.0000 E012 4.50 - 4.50 = 0.00 E004 10.00 -10.00 = 0.00 E020 2.69 - 2.69 = 0.00 E005 6.155 - 6.155 = 0.000 E022 2.95 - 2.95 = 0.00 E006 3.90 - 3.90 = 0.00 E023 3.29 - 3.29 = 0.00 E007 3.99 - 3.99 = 0.00 E025 5.95 - 5.95 = 0.00 E008 3.85 - 3.85 = 0.00 E009 3.99 - 3.99 = 0.00 ZERO REPRICINGS. ZERO NEW INCLUSIONS. ZERO EXITS. BASELINE CORRECTION APPLIED AT E023. The Issue 007 sealed value was 2.99 and the Issue 008 note records a +10.03 percent repricing. Recomputed: 2.99 x 1.1003 = 3.29. The Issue 008 baseline is therefore 3.29, and today's 3.29 is movement 0.00. Stating a RunPod repricing this issue would be a Protocol 1 failure. ``` ## 9. Counts recomputed ``` Panel 4 + 18 + 5 = 27 Scheduled this issue 4 + 10 + 2 = 16 Deferred 11 Check 16 + 11 = 27 Frames filed 92 Frames discarded 0 Copies verified 184 Hash disagreements 0 Providers with movement 0 of 16 ``` ## 10. Em dash declaration This document contains zero em dash characters and zero en dash characters. End of Issue 009 computation verification v1.0.