PurpleMIST-Mini-1.0 (Olaverse)

Composite (secondary)

0.7 · Composite rank #158

Cost and measurement conditions

$0.0094 per 1,000 decisions (estimated)

ESTIMATE (*est.), base-model market reference via the frozen module's Base-to-listed-model mapping (Qwen/Qwen3.5-2B-Base -> deepinfra:Qwen/Qwen3.5-2B), from the frozen 25 Sep 2026 snapshot through the UNMODIFIED pricing_v15 path (sha256 dae86740...): declared base Qwen/Qwen3.5-2B-Base -> deepinfra:Qwen/Qwen3.5-2B USD 0.02/M input, USD 0.1/M output. The author's server reports NO token usage, so the input tokens were counted offline with the author's own pinned tokenizer and their own unchanged encode() on the exact prompts their server builds from the runner's request bodies (CoS deputy decision #18974; receipt r109/receipts/TOKCOUNT-PURPLEMIST-R109.json; independent recount PASS, r109/review-tokcount/REVIEW.json): 700,492 input tokens over the 1,485 answered items. Nothing is generated (one forward pass, pointer-head option probabilities), so output is 0 under the 9 Oct missing-output rule. Cost on the common 1,485-item basis. Estimated, not a bill. COMMON COST BASIS (v1.6.30): cost per 1,000 decisions is computed on the common 1,485 items of this draw, leaving out the 15 very long items (at least 76,358 input tokens as counted by the rows that answered them). They are the 15 items the Jev reference refused with max_tokens_exceeded, which the other rows that did not answer them refused or failed on; their tokens count for no row, whether the row refused or answered them, and a refused item still counts as a wrong answer for Intelligence. This row: USD 0.009434 per 1,000 decisions (0.146x the official cost cap USD 0.064595), against USD 0.009340 over all 1,500 items (v1.6.29).

x2 + 0.15 s (assumption, not measured)

Sandy, the evaluator's own shared CPU host (AMD Ryzen 5 3600, Zen 2, AVX2, 12 threads), NOT a dedicated machine; the 1-minute load was logged every 60 s by the pressure guard (r109/logs/PRESSURE-r109-purplemist.jsonl). The author's own serve.py and purplemist.py, unchanged, with their own documented benchmark flags (--device cpu --dtype float32 --max-len 65536 --no-truncate); the single deviation is --no-warmup (their warmup compiles every length band up to max_len, hours on a CPU, and measures nothing). Offline, credential-free, loopback-only network namespace, MemoryMax 20G and no swap. With --no-truncate the author's code refuses rather than truncates a prompt over 65,536 tokens, so the draw's 15 longest items (76-80 k tokens) are honest HTTP 422 refusals scored as wrong; they are the same 15 items that Uniphore-4B-Decider, Attest-31B and Standard One 3B SH failed on. DISCLOSED: the pre-registration predicted 0 refusals from a character-based estimate; the offline token count and the live 422s both show 15. Speed is a loaded-shared-CPU float32 figure; the standard self-hosted x2 + 0.15 s adjustment applies and raw p50/p95 are kept. SPLIT WINDOW, DISCLOSED: measured in two sittings - 2026-10-13 19:56:04-21:19:35Z (870 answers; from about 21:00Z other jobs' processes, and this job's own build worker, raised the 1-minute load to 16-34; the last 13 answers of that sitting (items 857-869) took 34-192 s (8 of them 40 s or more) instead of 1-8 s, and the evaluator stopped the row by hand at 21:19:35Z; one earlier item took 66 s) and 22:41:42-23:30:36Z after the host was quiet again (load 7.2-11.3). The driver resumes by task id, so no item was answered twice. Load over all 133 per-minute samples: 7.24-34.24, mean 10.77 (r109/logs/CONDITIONS-purplemist-mini-1.0.jsonl; stop and guard events in r109/logs/PRESSURE-r109-purplemist*.jsonl). The disturbed items count in the Speed axis like every other item, so its Speed is a loaded-shared-host figure.