JevBench · measurement note

The readout is the ceiling

On 231 public JevBench items, a single forward pass that never generates a token was the binding constraint on hard-tier questions — not model quality. Re-running the same prompts on the questions it got wrong, with a thinking budget, recovered most of them.

231 public items (easy 48 · standard 72 · hard 111) Mac Studio M3 Ultra 96 GB · Mac mini M4 16 GB github.com/vasanthsreeram/myNameisJevToo
82.0
intelligence axis, best 27B leg (Q8_0)
73.9%
hard-tier accuracy, 82 / 111
~95.0
upper bound if the 29 misses are reasoned
1.0
McNemar p between the 27B legs on hard

01Headline: the readout, not the model

The harness reads the model's probability mass on each option's letter in one forward pass. No token generation at all — the output is a typed decision with a probability. That design is what makes it cheap, and it is also what caps it.

29 hard items were answered wrongly by the one-pass readout. Re-running the same rendered prompts with a thinking budget (greedy, up to 4000 tokens) got 23 of those 29 correct.

0 25 50 75 100 82.0 one-pass readout MEASURED ~95.0 with thinking budget UPPER BOUND — NOT MEASURED 23/29 recovered intelligence axis
Read it as a bound. Only the 29 failures were re-attempted, so this assumes reasoning never breaks an item the readout already had right. 3 of the 29 still hit the 4000-token cap mid-reasoning, so 23 is a floor on the recoverable set — and ~95.0 is an upper bound on the reasoned score. Hard-tier chance-corrected moves 60.6 → 91.9 (82/111 → 105/111) under that same assumption.

02Two ways to get an answer

LANE A · ONE-PASS READOUT rendered prompt single forward pass 0 tokens generated read letter probs typed decision + p 82.0 LANE B · REASONED (APPLIED TO THE 29 MISSES ONLY) same prompt generate, thinking greedy · budget 4000 tok commit answer 23 / 29 corrected ~95*
The harness is deliberately generation-free. Lane B is the pilot that shows what the readout alone leaves on the table. *Upper bound, not a measured reasoned score.

03All legs, same 231 items

LegIntelCalibSpeedScoreEasyStandardHard
Qwen3.8-27B GGUF Q8_082.086.166.477.7100.0%97.2%73.9%
Qwen3.8-27B GGUF Q4_K_M82.677.265.974.9100.0%97.2%74.8%
Qwen3.8-27B MLX 4-bit77.686.967.476.9100.0%94.4%69.4%
Ternary-Bonsai-2-27B PQ2_0 (6.7 GB)67.281.266.371.2100.0%86.1%60.4%
MiniCPM5-2B MLX 4-bit27.440.273.6~1377.1%50.0%39.6%

Do not read a spread into the 27B rows. McNemar on paired hard-tier results: Q8_0 vs Q4_K_M p = 1.000; Q8_0 vs MLX 4-bit p = 0.267; Q4_K_M vs MLX 4-bit p = 0.238. With n = 111 on hard, one item is 0.9% raw ≈ 0.57 intelligence points — Q8_0 (82/111) vs Q4_K_M (83/111) is literally a one-item difference.

04Which hard questions break a one-pass readout

0%25%50% 75%100% temporal_numeric probability tradeoff long_policy judge_hard multi_hop trap ambiguous adversarial routing_hard 40.0% 6/15 60.0% 6/10 66.7% 4/6 73.7% 14/19 76.5% 13/17 77.8% 14/18 87.5% 7/8 100% 7/7 100% 6/6 100% 5/5 ← the multi-step-computation families
Hard-tier accuracy by family, Q8_0 leg. The two worst families — temporal_numeric and probability — are exactly the ones that need multi-step computation, which a single forward pass cannot perform.

05Three smaller findings

F3 · Calibration

Calibration tracks the quant scheme, not bit-width. MLX 4-bit scores 86.9 — about equal to GGUF Q8_0 at 86.1. The GGUF Q4_K_M leg (77.2) is the outlier. An earlier claim that calibration is "the precision-sensitive axis" was over-generalised from a single pair and is wrong.

F5 · Speed

Latency is size-independent. Bonsai is 6.7 GB, Qwen Q8_0 is 27 GB — p50 600 vs 589 ms, p95 8.7 s vs 8.5 s. The tail is long-context prefill (hard-tier states run ~3,700 tokens), not weight size. Q8_0 raw p95 8519 ms; adjusted per JevBench's own-server rule (raw ×2 + 0.15 s) p50 1.33 s / p95 17.19 s.

F6 · State-blind control

With the state replaced by a placeholder and the options kept: the 27B legs hold on to ~71–73% of their hard-tier accuracy, while the 2B scored better without the state — 45.9% vs 39.6% — which looks like an option-list prior rather than state grounding.

06Ruled out

07Method

Public JevBench task items only — easy 48, standard 72, hard 111. Scoring uses JevBench's own formulas: per-tier accuracy, chance-corrected (acc − chance) / (1 − chance), tier weights easy 0.14 / standard 0.28 / judge 0.28 / hard 0.30. The intelligence axis is the weighted mean of the chance-corrected tier scores. Other axes: calibration (hard-tier ECE) and speed (latency, using JevBench's own-server rule of raw ×2 + 0.15 s). All 27B legs ran on a Mac Studio M3 Ultra 96 GB; the 2B baseline on a Mac mini M4 16 GB.

What this does and does not say