mojave-wmdp

construct validity audit of wmdp-bio, wmdp-chem, and truthfulqa mc1 via global sensitivity analysis (saltelli/sobol').

companion to enter mojave · tool: mojave · 2026-05-24

a saltelli radial design across six perturbation axes decomposes benchmark variance into first-order and total-effect sobol' indices. prompt template explains >95% of total variance on all three benchmarks. everything else is noise.

experiment design

perturbation matrix

axislevelsvalues
prompt_template5bare, cot, letter-only, lm-eval-default, verbose-rationale
system_prompt4helpful, domain-expert, none, safety-aware
decoding3greedy (T=0.0), T=0.7, T=1.0
choice_order2original, shuffled
quantization2bf16, fp8
n_shot_fraccontinuous[0, 0.05]

parameters

parametervalue
modelqwen/qwen2.5-7b-instruct
designsaltelli radial, N=512, k=6
cells per eval4,096
total cells12,288
benchmarkswmdp-bio (1,273 items), wmdp-chem (408 items), truthfulqa mc1 (817 items)
infrastructurerunpod gpu pods (a40/a6000), vllm, up to 46 concurrent
gsa enginesalib 0.1.1 (rust), 10,000 bootstrap resamples
artifacts12,351 cryptographically sealed pdf run cards

headline: prompt template dominates everything

evalprompt_template STinext-highest STimean accuracy
wmdp-bio0.9507quantization 0.10150.5967
wmdp-chem0.9548n_shot_frac 0.03890.4153
truthfulqa mc10.9682n_shot_frac 0.02990.5073

STi is the total-effect sobol' index—it captures both the direct effect and all interactions. a value of ~0.95 means that if you held prompt_template fixed and varied everything else, you'd explain only ~5% of the variance.

borgonovo δ (distribution-free confirmation)

evalprompt_template δnext-highest δ
wmdp-bio0.5137n_shot_frac 0.1160
wmdp-chem0.3974choice_order 0.1335
truthfulqa mc10.3373n_shot_frac 0.1446

per-template accuracy

templatebiochemtruthfulqa
bare0.1400.0580.034
cot0.7000.5280.625
letter-only0.7410.4960.610
lm-eval-default0.7670.5220.638
verbose-rationale0.6330.4720.628

the bare template—question text only, no framing, no answer labels—is catastrophic everywhere. wmdp-bio drops 63pp from template choice alone. chem: 46pp. truthfulqa: 60pp.

diagnostics

sum-of-indices

evalΣ S1Σ STiinterpretation
wmdp-bio0.9031.295moderate interactions (gap ~0.39)
wmdp-chem0.9391.107mild interactions (gap ~0.17)
truthfulqa0.9931.068near-additive (gap ~0.07)

bf16 vs fp8 quantization

evalbf16 meanfp8 meandifftp
wmdp-bio0.60040.5931+0.73pp0.9720.331
wmdp-chem0.41650.4141+0.23pp0.4040.686
truthfulqa0.50460.5101−0.55pp−0.7360.462

welch's t-test, two-tailed. none significant at α=0.05. quantization from bf16 to fp8 does not meaningfully affect accuracy.

caveats

  1. single model. qwen2.5-7b-instruct only. the methodology is the contribution; specific numbers are one data point.
  2. rescored, not raw. inspect's built-in scorer has artifacts on some templates. rescore_fast.py applies corrected scoring. raw logs available.
  3. no irt / item-level analysis. sensitivity analysis is at the cell level (per-run accuracy). item-level analysis is future work.

cryptographic integrity

every artifact carries:

  1. sha-256 companion data hash of its source json
  2. audit chain seal via mojave audit seal—hash-linked append-only chain with sequence number and chain tip hash

12,351 individually-sealed pdf run cards. no cherry-picking possible. view audit log.