Late entry, not in standings · League 1 field

KnowLine 4B Gen5.5

Published

Results f24d1f41, updated 16 November 2026, 21:24 UTC; scored by GNU Backgammon 1.08, 2-ply.

Late entry, not in standings

Late entry
Late entry, not in standings: its games count in no standings table, no league table, no Hype Cup input and no other setup's numbers
Entry date
2026-11-11first match 11:48 UTC
League field played
League 1 (original table)19 setups
Matches played
17 / 19GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)Mercury Decide: outside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)
Record
5–1–11won–drawn–lost
Error rate (PR), 95% interval
53.648.8 to 58.6
Decisions
2,153 (2,153 scored)
Risk judgement, margin 0.001
49.747.0 to 52.3
Ability d′
0.590.47 to 0.72 · risks taken 448 / 778 Risk pays · needless 281 / 813 Safety pays
Caution c (+ cautious, − bold)
+0.10+0.02 to +0.18 · cautious
Engine's best play
36.9%
Median s per decision
0.28
$ per 1,000 decisions
0.054estimate · DGX Spark
Invalid answers
0of 2,153 decisions
Cut off at the output limit
00 of 2,153 decisions sent an output limit; 0 over 32,768 output tokens without one
Route and effort sent
local (knowline_server.py)knowline/m · effort not in the requests

A late entry joins an original league table after it was played and plays every setup of that table once, in matches of its own. Its games count in no standings table, no league table, no Hype Cup input and no other setup's numbers. The numbers here come from its own late-entry games only, computed as for a setup in a league.

Error rate and risk judgement: 95% intervals, whole games resampled. Ability d′ and caution c at the 0.001 margin: rates corrected by adding 0.5 to the counts and 1 to the totals; 95% intervals, blocks of two seat-swapped games resampled. Cut off: answers that stopped at an output limit the request sent (finish reason length, stop reason max_tokens, finishReason MAX_TOKENS, incomplete max_output_tokens), in the stored response or error. Effort: the reasoning effort the requests sent. Engine's best play: the share of its scored decisions that chose the engine's best play (equity loss under the best-play threshold). Time: median request seconds per decision (waits between retries not included). Cost per 1,000 decisions: the recorded tokens at the price shown, as labelled there (list or billed price; for a setup on our GPUs the hosted model's list price, not billed).

Late entries on the Leagues page · JSON

How we called it

Model ID
PelaAI/KnowLine-4B-Gen5.5 @ 779944d2cc8ffa19cef386aee83c83bcb967ade2 (Hugging Face, Apache-2.0)
Provider or endpoint
our GPU (NVIDIA DGX Spark, GB10, spark2), FP8 at load, through the repo's own serve_knowline.sh (SGLang 0.5.21 with FP8 at load, as the card serves it, then its knowline_server.py front end), one request at a time
Date run
11 and 14 Nov 2026
Price
no price per token; estimate: GPU time, at most; source content/own_cost.json (wall clock on our machines × hourly rate); price date not recorded
Model type
decision model
Weights
open
Runs on
our GPUs
Reasoning setting
none (decision model)
Confidence source
decision-model probability
Rules text
long
Option groups
up to 255 options per request; more legal plays are split into equal groups, a winner per group, then among the winners; in this data 1 decisions were split into groups of 136
Sampling settings
not sent (decision request)
Notes
Late entry to the original League 1 table (not in standings, not eligible for the Hype Cup), late-entry league #22. PelaAI/KnowLine-4B-Gen5.5 @ 779944d on our DGX Spark (spark2, the whole GB10, handed over by logic one server at a time), FP8 at load, through the repo's own serve_knowline.sh (SGLang 0.5.21 with FP8 at load, as the card serves it, then its knowline_server.py front end), one request at a time; never on a hosted API. First stage 11 Nov 2026: its 9 hosted and program opponents (#957–#965, runner b727222); local stage 14 Nov 2026: the 8 setups we run ourselves, one at a time on an RTX PRO 6000 Blackwell of ours (#1371–#1378, runner b727222). It played 17 of its 19 opponents: its pair with Mercury Decide was not played (Mercury Decide is no longer served on OpenRouter; match #966 aborted at its warm-up; outside the requested scope, Ümit, 12 Nov) and its pair with GPT-6 Luna was not played (GPT-6 Luna's API was out of credit; outside the requested scope, Ümit, 12 Nov). FP8 at load (--quantization fp8, as the card and serve_knowline.sh serve it), --mem-fraction-static 0.72. Its cost is an estimate from its time on the DGX Spark (GPU time, at most).

Matches

Swipe for more columns →

KnowLine 4B Gen5.5 against each opponent, its score first: 5 won, 1 drawn, 11 lost; 2 outside the requested scope (Ümit, 12 Nov), not played, not a loss; none counts in a standings table
OpponentScoreResult#Finished
GLiDEAPI4–5Lost9572026-11-11
celeris-1-decision2–6Lost9582026-11-11
GNUbg weak0–9Lost9592026-11-11
Random8–2Won9602026-11-11
Safest2–5Lost9612026-11-11
d1-ultra0–5Lost9622026-11-11
Jev 1.140–5Lost9632026-11-11
Jev 1.132–5Lost9642026-11-11
SubQ Snap PreviewThe opponent was requested as SubQ Snap Preview; under that name the API returned answers identical to SubQ Snap 1.1 on 1,000 paired positions, so the opponent's model identity is not verified.6–6Draw9652026-11-11
Cygnet6–2Won13712026-11-14
Winnow-12B3–5Lost13722026-11-14
Perplexity Decider Flash v1.24–2Won13732026-11-14
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell1–6Lost13742026-11-14
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell1–6Lost13752026-11-14
Clef0–7Lost13762026-11-14
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell4–2Won13772026-11-14
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark4–2Won13782026-11-14
GPT-6 LunaDecisions APINo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)N/A
Mercury DecideNo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)N/A