Wald 26B-A4B (Gemma 4 26B-A4B)v1.7 · 10 s rule

Composite (secondary)

68.2 · Composite rank #21

Cost and measurement conditions

$0.048 per 1,000 decisions (estimated)

ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module (sha256 aa2afcd5a58d0c4cc0e3d6867d590c70b6ac58e971410c27faf9628f97bcf8b7, re-frozen 22 Oct 2026): base string google/gemma-4-26b-a4b-it (the card's base_model google/gemma-4-26B-A4B-it in the lower-case spelling that the module lists; the exact card spelling is in none of its decision tables and raises) priced as google/gemma-4-26b-a4b-it USD 0.09/M input, USD 0.3/M output. The same estimate on the same base is already carried by the published rows respan-span-02 and shisa-de-1. Estimated, not charged, not a quote from the author or any host. Tokens as reported by the usage block of the author's reader: input_tokens = the tokens of the one read per question (system prompt, state, options and chat template; no item has more than 26 options, so there is one read each and no knockout); output_tokens = 0 on every row (gate 0, nothing generated). Over all 1,500 items the server reported 1,982,239 input tokens. PRE-REGISTERED USAGE PREDICTION MISSED BY A CONSTANT +1: our offline count through a first-party transcription of the author's prompt path and the pinned tokenizer predicted 1,980,739 tokens (exact equality, item by item), and the server's usage.input_tokens was our count plus one on every one of the 1,500 items. This MISSES the pre-registration (exact equality, stop and report): the runner carried a +-2 token tolerance that was not pre-registered, and the +1 was already visible on all six smoke anchors before the first sealed byte, so the run did not stop (lead Amendment 1c in the pre-registration records this). The offset is constant and systematic (probably a leading BOS token added by the server; that reading is ours and was not verified); its cause is unknown; billing uses the server's usage like every other row. 0 refusals, as pre-registered. Cost effect of the +1: 0.19 % (USD 0.047939 vs 0.047850 per 1,000 on the common basis), no rank change in the scoring field. The 15 very long items (76,550-81,725 tokens) are answered in full and count for no row's cost on the common 1,485-item cost basis (v1.6.30). COST ON THE COMMON BASIS: USD 0.047939 per 1,000 decisions on 1,485 items, 0.742 times the official cost cap USD 0.064595; estimated, not charged.

x2 + 0.15 s (assumption, not measured)

Evaluator-owned rented Verda GPU VM, one NVIDIA RTX PRO 6000 Blackwell Server Edition 96 GB (97,887 MiB), FIN-03 (Finland, EU), pay-as-you-go non-spot, instance type 1RTXPRO6000.30V at USD 2.549/h. Stock vLLM 0.30.0 (127.0.0.1:8371) and the author's reader eval.systemone_vllm (127.0.0.1:8000) run in ONE docker container with network none; the reader was timed over loopback, one request at a time. Network cut proven from inside the container before any item hash or model load (loopback only, no default route, DNS and TCP fail); read-only code, read-only model and input, empty credential variables (names-only proof), HF offline flags. Measured 23 Oct 2026 (VM 06:50-07:07Z; measurement 338 s, total 981 s; scored run 204 s) on the frozen v1.6-regular-20261012 self-hosted input (sha256 dfd36720077ef7c41ace367524896507c7f44ce71306fb3e659e8b85069d995b, re-checked inside the container). The input was shredded from the node and its absence verified, and the VM was removed by exact id afterwards (teardown confirmed 07:07Z).

Model and serving disclosure

THE AUTHORS' OWN READER, run unmodified, in front of stock vLLM. Author code: 10 Python modules of org2ai/Wald-26B-A4B at 4a5d0b94bff26a8bba359ef311b3e535cefae901 (eval/__init__.py, eval/systemone_vllm.py, eval/vllm_letter.py, kev/__init__.py, kev/api.py, midtrain/__init__.py, midtrain/encode.py, midtrain/letter.py, midtrain/rawprompt.py, midtrain/tokens.py; CODE_MANIFEST.json sha256 5ce8feb796d52add3dfa4520fa0fe418c4eaaf4551591ad7d2485aca72aefaa8), hash-verified, mounted read-only at /code. The unshipped modules midtrain/chat_lora.py (contains snapshot_download and torch.load(weights_only=False)) and midtrain/gemma_moe.py are NOT shipped and fail closed with ImportError if reached; the author's run.sh (which installs packages from the network) was NEVER run. Instead we reproduce its settings (EFFORT=none, checkpoint 05901-c33): vLLM 0.30.0 --served-model-name wald-26b-a4b --max-model-len 131072 --gpu-memory-utilization 0.90 --max-num-seqs 128 --max-logprobs 64 --seed 0 --language-model-only, and the author's reader python3 -m eval.systemone_vllm --vllm http://127.0.0.1:8371 --gate 0 --budget 512 --wide knockout --template chat --prompt-format plain --max-model-len 131072 --noul-decisive-floor 0.81 --host 127.0.0.1 --port 8000. Image: official vllm/vllm-openai:v0.30.0@sha256:8a69ffad015f138d7170c4ddc429e230a3bc1c1719f67e14324749df200a4b90 plus two hash-pinned wheels (transformers 5.17.0, huggingface-hub 1.33.0) built offline on the VM (image sha256:dc93fedeea119e31b524c2352f397326d3b315cf129d33c0a7001cfce7c8e56e); torch 2.13.0+cu130, tokenizers 0.23.2. DISCLOSED DEVIATIONS from run.sh: run.sh not executed; reader bound to 127.0.0.1 (run.sh uses 0.0.0.0); DO_NOT_TRACK and HF_HUB_DISABLE_TELEMETRY added; the source-pins check is replaced by sha256 verification of all 10 files before upload. The submitted code was reviewed before the run, with no blocking findings (no download, subprocess, env read, exec/eval, pickle, torch.load or telemetry in the 10 shipped modules). The source and runner were reviewed; findings concerning the model card and measurement-window guard were resolved in the preregistration amendments before the run. Weights: org2ai/Wald-26B-A4B at 4a5d0b94bff26a8bba359ef311b3e535cefae901 (Apache-2.0, public, not gated, 39 files, 11 safetensors shards totalling 51,612,008,604 bytes (51,644,629,065 bytes for all 39 files), no pickle; Gemma-4-26B-A4B-it after full-parameter decision training, base google/gemma-4-26B-A4B-it@4d7ae4984b7db7de8f8457170b3f1a419ee76d52), sha256-verified file by file against the Hub LFS oids (MODEL_MANIFEST.json sha256 cfbc591cc731191f99f0585f334a2735b5e984deb6b7808cf9a8357d89dcbe5e). FIRST MEASUREMENT (found by our own intake; Author submission, a submission by the authors; same family as the published wald-4b-v2-1 row, a different base). AUTHOR'S OWN READER, on our hardware, not the author's run.sh (see code_revision). BENCHMARK-AWARE OUTPUT POLICY (author-shipped, measured as shipped): the reader runs with --noul-decisive-floor 0.81 (eval/systemone_vllm.py:36-42, 488-502, and serving.json). It moves a Noul probability inside (0.19, 0.81) to the nearer edge; the author's own comment says 'JevBench v1.5/v1.6 (O1S) scores 1-f<P<f as an abstention'. It is a pure function of the model's own probability, with no item ids or item strings. EFFECT ON THIS RUN: 123 of 375 Noul answers (32.8 %) sit exactly on 0.81 or 0.19, no answer lies inside (0.19, 0.81), so the abstention rate is 0.0; the Noul calibration is measured as shipped and is not comparable to an unfloored run (the unfloored probabilities are not recorded). This is a disclosure per exposure ruling on file. AUTHOR EXPOSURE (self-reported, not verified): the README claims 219/231 on the JevBench public set (Jev 1.13: 200/231), run by the authors; training on public items is possible and was not checked by us (no overlap check; no JevBench use or denial declared). 10 s RULE: max measured decision latency 9.79 s (median 0.043 s, p95 0.058 s, 99th percentile 8.63 s); 0 decisions over 10 s, but the slowest decisions are close to the limit (in the smoke a 78.5k-token request took 8.8 s cold and 0.18 s warm). DETERMINISM: first 50 positions run twice, only 1/50 bit-identical, 0 argmax flips, max probability difference 0.042 (disclosed; the cause was not investigated). USAGE OFFSET: server usage.input_tokens = our offline count + 1 on every item; see cost_basis. CONTEXT: the author's limit is 131,072 tokens per read and a longer read is refused with HTTP 422, never truncated; the longest sealed item is 81,725 tokens, so nothing was refused; the smoke accepted 131,070 tokens. The scored configuration is EFFORT=none (no thinking); no sealed item had more than 26 options, so the knockout mode was never used. PRICING: the frozen module has no explicit listing decision for the card's exact base spelling; the estimate follows the published sibling rows (see cost_basis). No outside request for this model.