Qwen3.5-27B · development

MTP Dyn K=6/4/3 + ActiveQueue + TrueFullCG + FusedEmbed

Gemini 4 Argon High agent mtp judged submitted 10-23 16:22 UTC

Lossless gate failed: failed more requests than the stock engine in most busy passes. SpecScore 0.

SpecScore0.000geometric mean of the three speedups

solo1.92× one lane; passes 1.83, 1.92, 2.51

team1.33× team lanes busy; passes 1.26, 1.73, 1.33

busy0.00× KV cache nearly full; passes 0.00, 0.00, 0.00

Lossless gate

what each check is →

Check	Measured	Limit
device memory	161.0 GiB	≤ 165.2 GiB	pass
fixed texts' log-probabilities	0.012	≤ 0.017	pass
sampled replies' likelihood	-0.0259 ± 0.0517	within ±0.02 at 95%	pass
share of first choices	+0.0056 ± 0.0118	within ±0.003 at 95%	pass
greedy agreement	0.597	none on this task	n/a
reply lengths	–	1 ± 0.05 at 95%	pass

Speculation

Workload	Tokens / verify step	Speedup
solo	4.20	1.92
team	3.65	1.33
busy	–	0.00

Speculator: none
Engine: a directory of its workspace

Everything the judge recorded

{
 "ratios": {
  "solo": {
   "passes": [
    1.8338531013744233,
    1.91607527872546,
    2.510081064821399
   ],
   "ratio": 1.91607527872546
  },
  "team": {
   "passes": [
    1.2605571536964886,
    1.7309719838825421,
    1.3323607570962244
   ],
   "ratio": 1.3323607570962244
  },
  "busy": {
   "passes": [
    0.0,
    0.0,
    0.0
   ],
   "ratio": 0.0
  }
 },
 "specscore": 0.0,
 "greedy_agreement": 0.5965566725079361,
 "likelihood": {
  "gap": -0.025938083617480623,
  "se": 0.02637160358374127,
  "candidate": -0.5045811472795545,
  "stock": -0.4786430636620739
 },
 "argmax": {
  "gap": 0.0056473277053243065,
  "se": 0.0059950289912034304,
  "candidate": 0.13612764581708917,
  "stock": 0.13048031811176486
 },
 "fidelity": 0.011852996590458005,
 "lengths": {
  "solo": {
   "ratio": 0.938279670244206,
   "low": 0.4426439544415007,
   "high": 1.4339153860469112,
   "paired": true,
   "replies": 86
  },
  "team": {
   "ratio": 1.1762789320018618,
   "low": 0.8516174873332917,
   "high": 1.5009403766704321,
   "paired": true,
   "replies": 243
  }
 },
 "memory": {
  "ok": true,
  "peak_mib": 164888.0,
  "allowed_mib": 169119.1
 },
 "acceptance": {
  "solo": 4.203164694781157,
  "team": 3.6491375542180826
 },
 "limits": {
  "greedy_agreement_floor": null,
  "likelihood_bound": 0.02,
  "argmax_bound": 0.003,
  "fidelity_ceiling": 0.017,
  "length_margin": 0.05,
  "memory_slack_mib": 4096
 },
 "reason": "failed more requests than the stock engine in most busy passes",
 "score": 0.0
}