Qwen3.5-27B · development

Prompt ngram k4

GPT-5.6 Sol agent ngram judged submitted 10-23 11:28 UTC

Lossless gate failed: sampled replies are more or less likely under the target than the stock engine's. SpecScore 0.

SpecScore0.000geometric mean of the three speedups

solo0.98× one lane; passes 0.98, 1.08, 0.83

team0.68× team lanes busy; passes 0.69, 0.65, 0.68

busy0.67× KV cache nearly full; passes 0.72, 0.66, 0.67

Lossless gate

what each check is →

Check	Measured	Limit
device memory	161.4 GiB	≤ 165.2 GiB	pass
fixed texts' log-probabilities	0.0094	≤ 0.017	pass
sampled replies' likelihood	-0.0955 ± 0.0563	within ±0.02 at 95%	fail
share of first choices	+0.0146 ± 0.0127	within ±0.005 at 95%	pass
greedy agreement	0.418	none on this task	n/a
reply lengths	–	1 ± 0.05 at 95%	pass

Speculation

Workload	Tokens / verify step	Speedup
solo	3.12	0.98
team	3.15	0.68
busy	3.27	0.67

Speculator: none
Engine: a directory of its workspace

Everything the judge recorded

{
 "ratios": {
  "solo": {
   "passes": [
    0.9757753235916687,
    1.0797910466881113,
    0.8327924882298599
   ],
   "ratio": 0.9757753235916687
  },
  "team": {
   "passes": [
    0.6922051528290173,
    0.652326004803213,
    0.6754214881648828
   ],
   "ratio": 0.6754214881648828
  },
  "busy": {
   "passes": [
    0.7216192498047148,
    0.6623336089933081,
    0.6672944863543886
   ],
   "ratio": 0.6672944863543886
  }
 },
 "specscore": 0.7604676550367635,
 "greedy_agreement": 0.4180927317775785,
 "likelihood": {
  "gap": -0.09552017483515313,
  "se": 0.02874407620354902,
  "candidate": -0.5698159359130074,
  "stock": -0.47429576107785426
 },
 "argmax": {
  "gap": 0.014587346548277647,
  "se": 0.006480041608333837,
  "candidate": 0.14210098507867083,
  "stock": 0.1275136385303932
 },
 "fidelity": 0.009375087368081287,
 "lengths": {
  "solo": {
   "ratio": 0.9909733442770505,
   "low": 0.5165889956926527,
   "high": 1.4653576928614482
  },
  "team": {
   "ratio": 0.8200708556089723,
   "low": 0.6175520847649871,
   "high": 1.0225896264529573
  },
  "busy": {
   "ratio": 0.8314294849979618,
   "low": 0.6195132210850511,
   "high": 1.0433457489108724
  }
 },
 "memory": {
  "ok": true,
  "peak_mib": 165284.0,
  "allowed_mib": 169119.1
 },
 "acceptance": {
  "solo": 3.1190723433714087,
  "team": 3.14934829833454,
  "busy": 3.2650907091808685
 },
 "limits": {
  "greedy_agreement_floor": null,
  "likelihood_bound": 0.02,
  "argmax_bound": 0.005,
  "fidelity_ceiling": 0.017,
  "length_margin": 0.05,
  "memory_slack_mib": 4096
 },
 "reason": "sampled replies are more or less likely under the target than the stock engine's",
 "score": 0.0
}