Round score
0.0171
weighted tier speedup after failure cap and deduction
Eligible requests
16
trusted baseline exclusions removed
Failed eligible requests
0
each request counted once
Raw weighted speedup
+1.71%
before failure cap and deduction
Temperature range
Request concurrency
Tier scheduling
Output policy
Thinking
Context limit
The same eligible inputs and output ceilings apply to every engine. Candidates must emit at least 90% of baseline tokens per request; natural output lengths can affect tier completion time. Slots refill until the group drains and stay occupied through protocol completion. Baseline exclusions and final drain can reduce actual concurrency below the configured cap.
Raw weighted speedup
Eligible speedup
No request failures: eligible speedup equals raw weighted speedup.
The round score uses weighted completion times for all eligible work in each tier. Per-request speedups below are diagnostics, not score contributions. Only trusted baseline exclusions are removed; candidate failures remain in the eligible count and incur the configured failure penalty.
16 prompts
| Prompt | Diagnostic speedup | Baseline | Candidate | Aligned output | Diagnostic / failure |
|---|---|---|---|---|---|
| hf-000 | — | — | — | 4309 | Per-token timing unavailable (not a request failure) |
| hf-006 | — | — | — | 4491 | Per-token timing unavailable (not a request failure) |
| hf-001 | — | — | — |
16 requests
| Request | TTFT | Total | Tokens | Slowest gap | Stream |
|---|---|---|---|---|---|
| hf-000 | 100.98ms | 68.313s | 5120 | 61.78ms | 1901 gaps for 5120 tokens |
| hf-001 | 101.35ms | 56.827s | 4369 | 39.03ms | 1585 gaps for 4369 tokens |
| hf-002 | 91.53ms | 65.408s |
Occupied client slots include protocol completion and final drain. Mean occupancy is weighted by elapsed time; it can be below the requested cap.
| Repetition | Group start | Requested | Effective cap | Peak | Mean |
|---|---|---|---|---|---|
| 1 | 0.000s | C4 | 4 | 4 | 3.58 |
| 1 | 140.008s | C4 | 4 | 4 | 3.79 |
| 2 | 0.000s | C4 | 4 | 4 | 3.82 |
| 2 | 135.636s | C4 | 4 | 4 | 3.90 |
| 3 | 0.000s | C4 | 4 | 4 | 3.78 |
| 3 | 139.505s | C4 | 4 | 4 | 3.89 |
Failed requests
Deduction
Final score
Each duration is the median across repetitions, from the group start until the tier's last request finishes. This includes client queueing. Weighted speedup = sum of weight × (1 − candidate / baseline).
| Input tier | Eligible requests | Weight | Baseline | Candidate | Speedup | Contribution |
|---|---|---|---|---|---|---|
| 8k | 8 | 50% | 141.926s | 139.503s | +1.71% | +0.85% |
| 16k | 8 | 50% | 153.647s | 151.006s | +1.72% | +0.86% |
| 4369 |
| Per-token timing unavailable (not a request failure) |
| hf-002 | — | — | — | 4471 | Per-token timing unavailable (not a request failure) |
| hf-009 | — | — | — | 5062 | Per-token timing unavailable (not a request failure) |
| hf-013 | — | — | — | 5120 | Per-token timing unavailable (not a request failure) |
| hf-014 | — | — | — | 5120 | Per-token timing unavailable (not a request failure) |
| hf-015 | — | — | — | 5120 | Per-token timing unavailable (not a request failure) |
| hf-004 | — | — | — | 4534 | Per-token timing unavailable (not a request failure) |
| hf-007 | — | — | — | 5120 | Per-token timing unavailable (not a request failure) |
| hf-003 | — | — | — | 4579 | Per-token timing unavailable (not a request failure) |
| hf-005 | — | — | — | 5120 | Per-token timing unavailable (not a request failure) |
| hf-011 | — | — | — | 4523 | Per-token timing unavailable (not a request failure) |
| hf-008 | — | — | — | 5120 | Per-token timing unavailable (not a request failure) |
| hf-010 | — | — | — | 5120 | Per-token timing unavailable (not a request failure) |
| hf-012 | — | — | — | 5120 | Per-token timing unavailable (not a request failure) |
| 4996 |
| 62.44ms |
| 1812 gaps for 4996 tokens |
| hf-003 | 213.67ms | 76.690s | 5120 | 59.33ms | 2039 gaps for 5120 tokens |
| hf-004 | 237.12ms | 65.273s | 4534 | 44.91ms | 1738 gaps for 4534 tokens |
| hf-005 | 222.27ms | 73.579s | 5120 | 59.35ms | 1957 gaps for 5120 tokens |
| hf-006 | 91.20ms | 63.760s | 4845 | 62.41ms | 1767 gaps for 4845 tokens |
| hf-007 | 158.09ms | 71.325s | 5120 | 70.72ms | 1894 gaps for 5120 tokens |
| hf-008 | 173.27ms | 79.388s | 5120 | 59.15ms | 2096 gaps for 5120 tokens |
| hf-009 | 105.21ms | 66.448s | 5062 | 58.04ms | 1847 gaps for 5062 tokens |
| hf-010 | 137.79ms | 74.585s | 5120 | 58.98ms | 1984 gaps for 5120 tokens |
| hf-011 | 103.22ms | 69.589s | 5079 | 57.59ms | 1866 gaps for 5079 tokens |
| hf-012 | 102.39ms | 74.313s | 5120 | 40.92ms | 2018 gaps for 5120 tokens |
| hf-013 | 99.33ms | 71.871s | 5120 | 57.15ms | 2026 gaps for 5120 tokens |
| hf-014 | 99.39ms | 65.613s | 5120 | 56.24ms | 1841 gaps for 5120 tokens |
| hf-015 | 98.36ms | 71.187s | 5120 | 39.31ms | 2020 gaps for 5120 tokens |
Each latency-median response is compared with the lowest valid opening-baseline ratio for the same prompt. Drops above the recorded limit fail; absolute checks also apply. These checks detect repetition, not overall writing quality.
| Prompt | Candidate char-16 | Baseline char-16 | Drop | Limit | Result |
|---|---|---|---|---|---|
| hf-000 | 0.9398 | 0.8544 | -0.0854 | 0.1000 | Passed |
| hf-001 | 0.8899 | 0.8617 | -0.0282 | 0.1000 | Passed |
| hf-002 | 0.9373 | 0.8889 | -0.0484 | 0.1000 | Passed |
| hf-003 | 0.9774 | 0.9723 | -0.0051 | 0.1000 | Passed |
| hf-004 | 0.9631 | 0.9117 | -0.0514 | 0.1000 | Passed |
| hf-005 | 0.9542 | 0.9655 | 0.0113 | 0.1000 | Passed |
| hf-006 | 0.9435 | 0.9181 | -0.0254 | 0.1000 | Passed |
| hf-007 | 0.9270 | 0.8938 | -0.0332 | 0.1000 | Passed |
| hf-008 | 0.9807 | 0.9796 | -0.0010 | 0.1000 | Passed |
| hf-009 | 0.9076 | 0.8732 | -0.0344 | 0.1000 | Passed |
| hf-010 | 0.9726 | 0.9601 | -0.0125 | 0.1000 | Passed |
| hf-011 | 0.8357 | 0.8850 | 0.0493 | 0.1000 | Passed |
| hf-012 | 0.9612 | 0.9646 | 0.0035 | 0.1000 | Passed |
| hf-013 | 0.9735 | 0.9604 | -0.0131 | 0.1000 | Passed |
| hf-014 | 0.8983 | 0.9422 | 0.0440 | 0.1000 | Passed |
| hf-015 | 0.9664 | 0.9714 | 0.0050 | 0.1000 | Passed |
Actual settings for each measured request. Each prompt keeps its temperature across repetitions and engines. Generation uses seed 0 throughout qualification, warmups and measured repetitions; fixed settings do not guarantee identical output text.
| Prompt | Repetition | Temperature | Top p | Seed |
|---|---|---|---|---|
| hf-000 | 1 | 0.395704 | 1 | 0 |
| hf-006 | 1 | 0.458231 | 1 | 0 |
| hf-002 | 1 | 0.154443 | 1 | 0 |
| hf-001 | 1 | 0.387136 | 1 | 0 |
| hf-014 | 1 | 0.628504 | 1 | 0 |
| hf-009 | 1 | 0.371945 | 1 | 0 |
| hf-013 | 1 | 0.637567 | 1 | 0 |
| hf-015 | 1 | 0.999192 | 1 | 0 |
| hf-004 | 1 | 0.111419 | 1 | 0 |
| hf-007 | 1 | 0.180399 | 1 | 0 |
| hf-005 | 1 | 0.742777 | 1 | 0 |
| hf-003 | 1 | 0.762508 | 1 | 0 |
| hf-008 | 1 | 0.885469 | 1 | 0 |
| hf-011 | 1 | 0.269879 | 1 | 0 |
| hf-010 | 1 | 0.771966 | 1 | 0 |
| hf-012 | 1 | 0.59199 | 1 | 0 |
| hf-001 | 2 | 0.387136 | 1 | 0 |
| hf-006 | 2 | 0.458231 | 1 | 0 |
| hf-002 | 2 | 0.154443 | 1 | 0 |
| hf-000 | 2 | 0.395704 | 1 | 0 |
| hf-009 | 2 | 0.371945 | 1 | 0 |
| hf-015 | 2 | 0.999192 | 1 | 0 |
| hf-014 | 2 | 0.628504 | 1 | 0 |
| hf-013 | 2 | 0.637567 | 1 | 0 |
| hf-004 | 2 | 0.111419 | 1 | 0 |
| hf-007 | 2 | 0.180399 | 1 | 0 |
| hf-005 | 2 | 0.742777 | 1 | 0 |
| hf-003 | 2 | 0.762508 | 1 | 0 |
| hf-010 | 2 | 0.771966 | 1 | 0 |
| hf-008 | 2 | 0.885469 | 1 | 0 |
| hf-012 | 2 | 0.59199 | 1 | 0 |
| hf-011 | 2 | 0.269879 | 1 | 0 |
| hf-001 | 3 | 0.387136 | 1 | 0 |
| hf-006 | 3 | 0.458231 | 1 | 0 |
| hf-002 | 3 | 0.154443 | 1 | 0 |
| hf-000 | 3 | 0.395704 | 1 | 0 |
| hf-009 | 3 | 0.371945 | 1 | 0 |
| hf-013 | 3 | 0.637567 | 1 | 0 |
| hf-014 | 3 | 0.628504 | 1 | 0 |
| hf-015 | 3 | 0.999192 | 1 | 0 |
| hf-005 | 3 | 0.742777 | 1 | 0 |
| hf-007 | 3 | 0.180399 | 1 | 0 |
| hf-004 | 3 | 0.111419 | 1 | 0 |
| hf-003 | 3 | 0.762508 | 1 | 0 |
| hf-011 | 3 | 0.269879 | 1 | 0 |
| hf-008 | 3 | 0.885469 | 1 | 0 |
| hf-010 | 3 | 0.771966 | 1 | 0 |
| hf-012 | 3 | 0.59199 | 1 | 0 |
verdict
evidence
min logprob
num prompts
mean logprob
coverage ratio
quantile logprob
num positions scored
role
evidence
request concurrency
requests per s
sla goodput ratio
output tokens per s
p99 e2e ms rel range
p99 itl ms rel range
p99 ttft ms rel range
tier completion max rel range
Image digest
Patch hash