pjev-2b (Kailune AI, frozen MiniCPM5-2B + merged LoRA r64, letter-logit readout, T=1 stub)

Composite (secondary)

15.8 · Composite rank #117

Cost and measurement conditions

$0.016 per 1,000 decisions (estimated)

ESTIMATE (*est.), size-class reference, carried over unchanged from PREREG-R36.md section 3.2. The UNMODIFIED frozen pricing_v15 path is asserted: openbmb/MiniCPM5-2B is in UNLISTED_BASE_MODELS - explicitly reviewed against both 25 Sep 2026 catalogs and not sold - so reference() returns no market reference rather than raising, and no exact base-model floor applies. The board already prices a row on THE SAME BASE this way: decision-2b (FlyMy.AI, openbmb/MiniCPM5-2B + LoRA and pointer head) carries the DeepInfra 2-4B size-class reference list price of USD 0.03 per 1M input and USD 0.0 per 1M output from the frozen 25 Sep 2026 snapshot (sha256 84eb7926acd4e93b71d695be24898e2ca999b20011d1b66cc0348ce7c6c3546f). pjev-2b is the same shape: 2.52B, one forward pass with logits_to_keep=1, NOTHING GENERATED - the author's serve.py reports usage.output_tokens as a literal 0, not a decision count - so the output term is 0.0, not 0.15. Multiplied by this row's own measured input tokens. No new release-lane pricing line is created. Free weights are never cost 100 and never cost 0. 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 0.016438 per 1,000 decisions (0.254x the official cost cap USD 0.064595), against USD 0.016274 over all 1,500 items (v1.6.29).

x2 + 0.15 s (assumption, not measured)

Sandy, the evaluator's own shared CPU host (AMD Zen 2, 12 threads, AVX2, no AVX-512 and no AMX), NOT a dedicated machine: the 1-minute load average was logged every 60 s into logs/CONDITIONS-pjev-2b.jsonl and is reported with the row. This is the evaluator-controlled local deployment the submission asks for ('please run it on your own machine, one request at a time. We do not operate an endpoint for it and do not want held-out or sealed items sent to us'). The author's own serve.py / decision_core.py / letter_adapter.py, source-reviewed at the pinned revision in run 36 before anything was executed, run as `python serve.py --model <pinned snapshot> --device cpu --host 127.0.0.1 --port 11474` at their own defaults (--max-tokens 16000), OMP_NUM_THREADS=10, inside a loopback-only network namespace with an outbound probe that failed as required, offline (HF_HUB_OFFLINE=1, TRANSFORMERS_OFFLINE=1) and credential-free. DEVIATION, DISCLOSED, AND THE ONLY ONE: letter_adapter.load() hard-codes dtype=torch.bfloat16; run 36's pre-registered one-line change reads the dtype from PJEV_DTYPE and DEFAULTS TO THE AUTHOR'S bfloat16, so without that variable the file behaves exactly as published (diff r36/receipts/PJEV-DTYPE-PATCH-R36.diff, patched sha256 6d15788caef5457cfd4690c3fe576d291217276416ed2bbf3bc0312db5e44390). float32 is a strict UPCAST of the stored bfloat16 weights - exactly representable, never lossy - so it cannot flatter accuracy or calibration; it can only cost Speed. serve.py and decision_core.py are byte-identical to the author's. letter_adapter.dist() RAISES rather than truncating when a prompt exceeds the author's own 16,000-token contract, and serve.py maps that to HTTP 422, so the draw's 15 items of 106,804-122,936 characters are honest refusals, not silent truncations - the same 15 items that the Uniphore-4B-Decider (HTTP 422), Attest-31B (HTTP 500) and Standard One 3B SH (HTTP 422) rows of this draw also failed on - verified this run against the published v1.6.16 raw runs, where the three failure sets are identical and julia-1's 17 refusals are a strict superset of them. Torchcast-Decision-4B-v1.1 and both MeTask-Rain rows of this draw answered all 1,500, so run 107's entry text naming them as peers was wrong and is corrected here. Speed is a loaded-shared-CPU float32-on-AVX2 figure and a declared lower bound against the author's own H100 reference of 20-62 ms per decision; the standard self-hosted x2 + 0.15 s adjustment is applied on top and the raw p50/p95 are kept. SPLIT WINDOW, DISCLOSED: this row was measured across four sittings under different host load - 2026-10-13 09:07:45-13:16:14Z (996 answers, stopped by the pressure guard at load 57), 18:15:34-18:21:41Z and 18:27:04-18:29:07Z (two short attempts stopped by the same guard for swap left behind by another job) and from 18:36:35Z to the end. The driver resumes by task id, so no item was answered twice. Its Speed therefore carries the same kind of interference disclosure as reflex-instinct-0.6b under #18761; the per-minute 1-minute-load trace of every sitting is in r107/logs/CONDITIONS-pjev-2b.jsonl and the guard decisions in r107/logs/PRESSURE-r107-pjev*.jsonl.