|
| 1 | +# Provenance optimization: v2 qualification and performance |
| 2 | + |
| 3 | +The optimized H4 implementation passed a separate, bounded synthetic |
| 4 | +requalification. It preserves the v1 numerical results while reducing repeated |
| 5 | +metadata work. **The six neural pilot fits and three ancillary ridge fits remain |
| 6 | +held.** The primary 180-fit scope and $150 combined ceiling are unchanged; these |
| 7 | +receipts do not authorize scientific fits or paid cloud jobs. |
| 8 | + |
| 9 | +## What was verified |
| 10 | + |
| 11 | +The [v2 qualification](../reports/gap-pilot-qualification-v2.json) used exactly |
| 12 | +16 new synthetic optimizer updates: clean and augmented arms each ran a |
| 13 | +four-update baseline, a two-update prefix, and a two-update fresh-process |
| 14 | +continuation. It finished in 21.84 seconds. The consumed v1 allowance and all v1 |
| 15 | +artifacts were preserved. The previously qualified ridge artifact was reused, |
| 16 | +with zero new ridge fits. |
| 17 | + |
| 18 | +Both arms matched v1 exactly on update traces, masks, losses, model and optimizer |
| 19 | +state, sampler order, saved H2 transforms, CPU RNG, selected best state and |
| 20 | +physical predictions. Fresh-process restart and label-free best inference also |
| 21 | +passed. The new checkpoints have different code fingerprints, as expected; old |
| 22 | +checkpoints were compared as authenticated reference data, not restored under |
| 23 | +changed code. This establishes engineering equivalence on the fixed synthetic |
| 24 | +fixture, not forecast skill or convergence on scientific data. |
| 25 | + |
| 26 | +The optimization removes repeated JSON roundtrips and reuses validated, |
| 27 | +immutable original provenance. It preserves source IDs, ordered qualifiers, |
| 28 | +original observation masks, clocks, and case mappings. It does not cache Q |
| 29 | +values, targets, prepared fills or model outputs. Dynamic input values, masks, |
| 30 | +gap endpoints and fill admissibility still receive their existing checks. |
| 31 | + |
| 32 | +## Measured preparation and inference |
| 33 | + |
| 34 | +The [v2 profile](../reports/gap-pilot-preparation-profile-v2.json) made 120 timed |
| 35 | +calls: 20 calls in each of six stages, with zero optimizer updates, scaler fits |
| 36 | +or ridge fits. It finished in 11.32 seconds with approximately 323 MiB observed |
| 37 | +process RSS. CPU, one Torch thread, FP32, B64 and L168 were fixed. |
| 38 | + |
| 39 | +| Path | v1 mean | v2 mean | Interpretation | |
| 40 | +|---|---:|---:|---| |
| 41 | +| Clean adapter | 86.24 ms | 50.18 ms | v2 uses the validated snapshot view used by training | |
| 42 | +| Augmented adapter | 76.90 ms | 53.03 ms | Same snapshot path plus deterministic artificial gaps | |
| 43 | +| Full provenance digest | 211.40 ms | 62.72 ms | Actual optimized digest helper; identical logical bytes | |
| 44 | +| Model-only forward | 14.88 ms | 14.82 ms | Prepared tensors; no preparation or inverse transform | |
| 45 | +| Repeated-batch full prediction | 335.92 ms | 144.09 ms | v2 includes one initial cache miss and 19 hits | |
| 46 | +| New-origin full prediction | Not measured | 175.06 ms | Twenty distinct batches, all actual cache misses | |
| 47 | + |
| 48 | +The model-only rate is essentially unchanged. The observed improvement is in |
| 49 | +work surrounding that model call. These are separately timed paths, not an |
| 50 | +additive trace that precisely assigns every millisecond. Workspace/runtime |
| 51 | +residency and cache conditions also changed, so the table is not a controlled |
| 52 | +ablation attributing every gain to one code edit. |
| 53 | + |
| 54 | +The original training provenance snapshot was built once outside the adapter |
| 55 | +timers; its measured setup was 31.68 ms for 64 cases. New-origin fixtures were |
| 56 | +also generated outside their timers, and their output-identity checks ran after |
| 57 | +timing. Each new batch shifted its 64 origins by another hour within the |
| 58 | +declared outer interval. Cache diagnostics confirmed 20 additional misses and |
| 59 | +zero hits, with at most one retained entry. The batches overlap in calendar |
| 60 | +content; this measures metadata-cache misses, not cold disk reads or actual data |
| 61 | +extraction. |
| 62 | + |
| 63 | +The 19 warm-hit calls alone averaged 142.18 ms. A repeated-batch result must not |
| 64 | +stand in for ordinary new-origin inference. Runtime planning below therefore |
| 65 | +defaults to 175.06 ms. A batch-major scenario is shown separately and requires |
| 66 | +one cold request followed by 16 warm requests for each outer batch; it does not |
| 67 | +assume all 17 requests are warm or that this schedule has been adopted. |
| 68 | + |
| 69 | +## Full training-source constructor check |
| 70 | + |
| 71 | +The [constructor qualification](../reports/gap-pilot-constructor-qualification-v2.json) |
| 72 | +passed at **17,869 synthetic source rows**: 15,717 gradient-fit and 2,152 inner |
| 73 | +rows, matching the largest proposed CV1 gauge counts. It retained **zero outer |
| 74 | +or unassigned rows**. Synthetic placement and values were used; actual scientific |
| 75 | +case membership and source bodies were not read. |
| 76 | + |
| 77 | +Recorded work was: |
| 78 | + |
| 79 | +- Synthetic fixture generation: 6.48 seconds. |
| 80 | +- Trainer construction, including one synthetic original-H2 scaler fit: |
| 81 | + 9.01 seconds. |
| 82 | +- One clean and one augmented 64-case batch: 114.46 and 112.65 ms. |
| 83 | +- Total after imports: 15.73 seconds. |
| 84 | +- Peak observed parent-plus-child RSS: **1.832 GiB**, under the 4 GiB limit; |
| 85 | + wall allowance: 120 seconds. |
| 86 | +- Zero optimizer construction, neural updates, model forward calls or ridge fits. |
| 87 | + |
| 88 | +The original NumPy payload was 97,779,508 bytes and provenance covered 3,001,992 |
| 89 | +history slots. Payload bytes, retained metadata size and total process RSS are |
| 90 | +different quantities. This check does not isolate their individual peak costs. |
| 91 | + |
| 92 | +The future pilot assembler must keep this fit/inner footprint and load outer |
| 93 | +cases separately. The optional 26,425-row outer-inclusive constructor was not |
| 94 | +run. Nor does this check qualify the optional final-development constructor, |
| 95 | +whose largest fit/inner source would contain 42,696 cases, or full-cohort |
| 96 | +backpropagation memory. Those larger workloads require their own qualification |
| 97 | +or a separately reviewed streaming design. |
| 98 | + |
| 99 | +## Updated conditional runtime |
| 100 | + |
| 101 | +The [v2 reassessment](../reports/gap-pilot-runtime-reassessment-v2.json) retains |
| 102 | +the per-gauge admitted counts and B64 batch ceilings. It separates each arm's |
| 103 | +first update from its next three samples: |
| 104 | + |
| 105 | +| Arm | First update | Mean of next three | v1 mean of next three | |
| 106 | +|---|---:|---:|---:| |
| 107 | +| Clean H4 | 654.89 ms | 151.16 ms | 356.46 ms | |
| 108 | +| Augmented H4 | 545.17 ms | 155.27 ms | 324.11 ms | |
| 109 | + |
| 110 | +These update timings include preparation, hashes, copies, model work and the |
| 111 | +qualification worker's durable status writes. The first update also constructs |
| 112 | +the optimizer lazily. Four samples per arm do not establish stable throughput, |
| 113 | +convergence or a confidence interval. The JSON includes later-sample |
| 114 | +minimum/mean/maximum scenarios without interpreting those extremes as bounds. |
| 115 | + |
| 116 | +The following table uses the later-three means. All figures are **modeled hours, |
| 117 | +not measured scientific-run durations**. |
| 118 | + |
| 119 | +| Proposed work | 20 epochs, cold default | 50 epochs, cold default | 20 epochs, conditional batch-major | 50 epochs, conditional batch-major | |
| 120 | +|---|---:|---:|---:|---:| |
| 121 | +| Six-fit CV1 pilot | 2.03 | 4.06 | 1.91 | 3.94 | |
| 122 | +| Optional 18-fit confirmation, excluding unknown constructor cost | 11.71 | 26.47 | 11.38 | 26.14 | |
| 123 | +| Combined modeled portions | 13.74 | 30.53 | 13.30 | 30.09 | |
| 124 | + |
| 125 | +The pilot estimate includes the largest-CV1 measured synthetic constructor time |
| 126 | +once per fit as a setup proxy for all three gauges. It does not substitute the |
| 127 | +synthetic fixture-generation time for actual source extraction. Final-stage |
| 128 | +constructor cost is explicitly unknown and omitted, not assigned a zero cost. |
| 129 | + |
| 130 | +For each fit, training time is the observed first update plus the remaining |
| 131 | +update count multiplied by the selected later rate. Clean-inner time uses the |
| 132 | +new-origin full-predictor rate as an **unmeasured surrogate** in both schedules; |
| 133 | +it is neither a measurement of the inner-loss loop nor a proven upper bound. |
| 134 | +Outer time is either 17 cold calls per batch, or one cold plus 16 warm calls. |
| 135 | +The older H2 per-epoch checkpoint-write placeholder remains separately labeled; |
| 136 | +the cost of growing H4 journals is not yet measured. |
| 137 | + |
| 138 | +The 17 requests are evaluation opportunities, not guaranteed successful scored |
| 139 | +panels. Mixed-recent ridge still correctly rejected its 64-case qualification |
| 140 | +batch because only 56 cases had admissible original lag context. Failures may |
| 141 | +not be silently dropped to improve coverage or metrics. |
| 142 | + |
| 143 | +Scientific ridge fitting, actual data assembly, scoring/resampling, larger |
| 144 | +constructor costs, long-run thermal effects and failed attempts remain outside |
| 145 | +these measured costs. The six neural pilot fits plus three ridge fits, and any |
| 146 | +optional expansion, still require their separate protocol and execution review. |
| 147 | + |
| 148 | +The [v1 reassessment](../reports/gap-pilot-runtime-reassessment.json) and |
| 149 | +[original proposal](../reports/coverage-sensitivity-resource-proposal.json) |
| 150 | +remain unchanged. The primary 180-fit estimates are not revised by this optional |
| 151 | +H4 timing work. Synthetic checks also do not certify that every development |
| 152 | +corpus file is resident; data restoration has separate evidence and completion |
| 153 | +criteria. |
0 commit comments