Experiments

Experiments

Tests beside the leagues.

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

Repeat checks

League play vs League 1's field

Run 2: the completed repeats ranked by their own risk judgement, Rank 1 to n; queued, running and stopped repeats below them, unranked, in a fixed order. Each repeat is one run of its model, shown beside that model's first run on exactly the same pairs, seeds and dice (the pairs the repeat finished, of those playable today); nothing is picked from several runs or averaged, and the main leaderboard is unchanged.

Repeat checks, League play vs League 1's field, Run 2: each repeat ranked when complete, beside its first run on the same pairs
RankModelRun 2: risk judgement / error rate (PR) / decisionsFirst run, same pairsPairsStatus
1Gemini Decision 120261111
Run 2
64.362.0 to 66.7 / 28.125.0 to 31.4 / 1,697
First run
64.762.4 to 67.1 / 28.025.1 to 31.1 / 1,704 (risk judgement / error rate (PR) / decisions)
Run 1 and Run 2, same pairs
Run 1 (first run)Run 2 (repeat)
Risk judgement64.762.4 to 67.164.362.0 to 66.7
Error rate (PR)28.025.1 to 31.128.125.0 to 31.4
Decisions1,7041,697
Invalid decisions00
Billed cost$0.754$0.748
Status
Complete · 17 of 17 matches (league #50)
Run id
repeat-2026-11-12-gemini-decision-1-r1
Setup
its own players row: openrouter-decisions/google/gemini-decision-1 · openrouter (google ai studio) · hosted; held: in no standings, table or other setup's numbers
Settings
as the original's, changed: name, note; League 1 rules (prompt bg-v2, 3 rounds, 300 turns, the original's option-order seed, mirror games on); each pair with its original match's seed and seats; one match at a time
First run's pairs
late entry: 17 of 19 pairs played, 2026-11-11; GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (not required (Ümit #5220); never scheduled: OpenAI Decisions out of credit (start.sh: deferred)); Mercury Decide: outside the requested scope (Ümit, 12 Nov); not played (not required (Ümit #5220); #1016 aborted at its start, no game: no OpenRouter endpoint for inception/mercury-decide); its numbers here: its own decisions in the 17 pairs the repeat finished (matches #993, #994, #995, #996, #997, #998, #999, #1005, #1013, #1053, #1055, #1056, #1057, #1060, #1061, #1062, #1063)
Dates
prepared 2026-11-12; started 2026-11-13 20:29 UTC; finished 2026-11-13 21:02 UTC
Pairs today
17 of 19 playable (first run: 17 of 19 played)
Cost basis
$0.75 estimated (the original's billed paid-opponent costs, with its own $0.65); cap $4 per repeat
Differences from the original protocol
  • Name: '<original> [repeat 1]', a new players row; settings otherwise as the original's (the changed fields are listed under Settings); one match at a time, as the original (start.sh, max_matches 1).
  • Runner: the original ran dab6da5 (clean) for its first stage (#993-#1013) and 38024d6 (clean; map-check v2 and the raw_entries decoder) for its local stage (#1053-#1063), both on origin/main's runner line (ca112dc); the repeat runs 959a8a2 (branch pinned-late-gemini): the repeats' runner 0e66836 (late-reax line with the pinned-pairs change 4ca4f08 and admit) with those four runner commits cherry-picked (cb669a0) and the run guard (c13992f, 959a8a2: reply identity and the shared pre-send cap, used with --guard only). late-reax adds decision clients (reax, baselabs, drex, nace) and a budget ledger, which the run guard uses; requests, sampling (provider defaults, nothing set), the prompt version, option-order seed and rules are the same.
  • Map check: map-check v2 (a probability map that is not exactly the shown ids, finite in [0, 1], summing to 1 within 1e-6, stores no probabilities; the returned choice is still played) applies to every repeat pair; the original's first-stage pairs (#993-#1013, dab6da5) ran without it. Plays are not affected; a first-stage pair can store probabilities in the original where the repeat stores none.
  • Hosted model: Gemini Decision 1 is served by Google AI Studio through OpenRouter. The repeat accepts a reply only when it names the original's served model google/gemini-decision-1-20261111 and provider Google AI Studio (as all 1,704 original decisions did); a model served under that same name but changed behind it cannot be detected from here.
  • Route check: the original's guard re-checked OpenRouter's listing every minute (3 failures stopped it); the repeat checks it once before the first match (smoke) and, in the runner (run guard), the served identity of every entrant reply: model google/gemini-decision-1-20261111 and provider Google AI Studio exactly, or the guard latches and the league stops for good. Fallbacks stay off.
  • Cap: the original's $5 entry cap with $1.1819 recorded (a minute poll of stored turns); the repeat's run guard holds the entrant and paid opponents to the remaining $3.8181 before every request attempt (see cost_basis).
64.362.0 to 66.7 / 28.125.0 to 31.4 / 1,69764.762.4 to 67.1 / 28.025.1 to 31.1 / 1,70417 of 17Complete · 17 of 17 matches
2Drex MaxNVIDIA RTX PRO 6000 Blackwell
Run 2
62.961.1 to 64.7 / 28.825.8 to 32.0 / 2,002
First run
63.161.2 to 64.9 / 30.227.1 to 33.5 / 2,062 (risk judgement / error rate (PR) / decisions)
Run 1 and Run 2, same pairs
Run 1 (first run)Run 2 (repeat)
Risk judgement63.161.2 to 64.962.961.1 to 64.7
Error rate (PR)30.227.1 to 33.528.825.8 to 32.0
Decisions2,0622,002
Invalid decisions00
Billed cost$0.137$0.123
Status
Complete · 17 of 17 matches (league #32)
Run id
repeat-2026-11-11-drex-max-v1-r1
Setup
its own players row: drex-max/drex-max-v1 · local (drex_max_server.py) · bf16; held: in no standings, table or other setup's numbers
Settings
as the original's, changed: one match at a time, name, note; League 1 rules (prompt bg-v2, 3 rounds, 300 turns, the original's option-order seed, mirror games on); each pair with its original match's seed and seats; one match at a time
First run's pairs
late entry: 18 of 19 pairs played, 2026-11-06; GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit); its numbers here: its own decisions in the 17 pairs the repeat finished (matches #847, #848, #855, #856, #857, #859, #861, #864, #867, #868, #870, #872, #873, #874, #879, #882, #885)
Dates
prepared 2026-11-11; started 2026-11-12 01:03 UTC; finished 2026-11-12 02:03 UTC
Pairs today
17 of 19 playable (first run: 18 of 19 played); outside the requested scope (Ümit, 12 Nov): Mercury Decide (played in the first run; blocked today as well)
Cost basis
$0.14 estimated (the original's billed paid-opponent costs); cap $5 per repeat
Differences from the original protocol
  • Name: '<original> [repeat 1]', a new players row; settings otherwise as the original's (the changed fields are listed under Settings); one match at a time, where the original played several at once.
  • Not replayed, though played in the original: Mercury Decide (no OpenRouter endpoint): outside the requested scope (Ümit, 12 Nov), and blocked today as well; their seed pins are kept.
62.961.1 to 64.7 / 28.825.8 to 32.0 / 2,00263.161.2 to 64.9 / 30.227.1 to 33.5 / 2,06217 of 17Complete · 17 of 17 matches
3d1-ultra
Run 2
62.159.7 to 64.5 / 32.128.7 to 35.0 / 1,954
First run
61.759.4 to 64.0 / 29.826.9 to 32.8 / 1,900 (risk judgement / error rate (PR) / decisions)
Run 1 and Run 2, same pairs
Run 1 (first run)Run 2 (repeat)
Risk judgement61.759.4 to 64.062.159.7 to 64.5
Error rate (PR)29.826.9 to 32.832.128.7 to 35.0
Decisions1,9001,954
Invalid decisions00
Billed cost$0.379$0.373
Status
Complete · 16 of 16 matches (league #28)
Run id
repeat-2026-11-11-d1-ultra-r1
Setup
its own players row: liquid/d1-ultra · liquid · hosted; held: in no standings, table or other setup's numbers
Settings
as the original's, changed: one match at a time, name, note; League 1 rules (prompt bg-v2, 3 rounds, 300 turns, the original's option-order seed, mirror games off); each pair with its original match's seed and seats; one match at a time
First run's pairs
League 1: 18 of 18 pairs played, 2026-10-19 to 2026-10-22; its numbers here: its own decisions in the 16 pairs the repeat finished (matches #180, #181, #182, #184, #190, #191, #192, #223, #264, #265, #268, #280, #283, #376, #399, #490)
Dates
prepared 2026-11-11; started 2026-11-11 19:51 UTC; finished 2026-11-11 20:37 UTC
Pairs today
16 of 18 playable (first run: 18 of 18 played); outside the requested scope (Ümit, 12 Nov): GPT-6 LunaDecisions API (played in the first run; blocked today as well), Mercury Decide (played in the first run; blocked today as well)
Cost basis
$0.38 estimated (the original's billed paid-opponent costs, with its own $0.27); cap $5 per repeat
Differences from the original protocol
  • Name: '<original> [repeat 1]', a new players row; settings otherwise as the original's (the changed fields are listed under Settings); one match at a time, where the original played several at once.
  • Runner: its 18 League 1 matches were played by six runner builds (17f3b8e, 2421d9e, 6425197, 356e32a with uncommitted changes; bef3ec8, ec555d6 clean); the repeat uses the clean acfd89c build, the one that played d1-ultra as an opponent in REAX's late entry (#920). Same prompt version bg-v2, option-order seed and rules; mirror games off as in its originals.
  • Opponent hardware: Clef (#180), Cygnet (#184) and Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark (#182) answered from DGX Spark (GB10) servers then; they now run on our RTX PRO 6000 cards with the same checkpoints and arguments (their players.device records the move and its 20-request replays: same play 20/20, 20/20 and 19/20).
  • Hosted model: d1-ultra is served by Liquid; whether the hosted model behind liquid/d1-ultra changed since 2026-10-19 cannot be checked from here.
  • Not replayed, though played in the original: GPT-6 LunaDecisions API (no credit), Mercury Decide (no OpenRouter endpoint): outside the requested scope (Ümit, 12 Nov), and blocked today as well; their seed pins are kept.
62.159.7 to 64.5 / 32.128.7 to 35.0 / 1,95461.759.4 to 64.0 / 29.826.9 to 32.8 / 1,90016 of 16Complete · 16 of 16 matches
4Mammoth-1
Run 2
61.959.7 to 64.1 / 27.525.0 to 30.1 / 1,881
First run
62.860.7 to 65.1 / 29.126.4 to 31.6 / 1,831 (risk judgement / error rate (PR) / decisions)
Run 1 and Run 2, same pairs
Run 1 (first run)Run 2 (repeat)
Risk judgement62.860.7 to 65.161.959.7 to 64.1
Error rate (PR)29.126.4 to 31.627.525.0 to 30.1
Decisions1,8311,881
Invalid decisions00
Billed cost$0.119$0.134
Status
Complete · 17 of 17 matches (league #31)
Run id
repeat-2026-11-11-mammoth-1-r1
Setup
its own players row: baselabs/baselabs/mammoth-1 · local (baselabs_server.py) · bf16; held: in no standings, table or other setup's numbers
Settings
as the original's, changed: one match at a time, name, note; League 1 rules (prompt bg-v2, 3 rounds, 300 turns, the original's option-order seed, mirror games on); each pair with its original match's seed and seats; one match at a time
First run's pairs
late entry: 19 of 19 pairs played, 2026-11-02; its numbers here: its own decisions in the 17 pairs the repeat finished (matches #780, #781, #782, #783, #806, #808, #810, #811, #812, #813, #814, #815, #816, #818, #820, #821, #828)
Dates
prepared 2026-11-11; started 2026-11-11 22:19 UTC; finished 2026-11-11 23:29 UTC
Pairs today
17 of 19 playable (first run: 19 of 19 played); outside the requested scope (Ümit, 12 Nov): GPT-6 LunaDecisions API (played in the first run; blocked today as well), Mercury Decide (played in the first run; blocked today as well)
Cost basis
$0.12 estimated (the original's billed paid-opponent costs); cap $5 per repeat
Differences from the original protocol
  • Name: '<original> [repeat 1]', a new players row; settings otherwise as the original's (the changed fields are listed under Settings); one match at a time, where the original played several at once.
  • Not replayed, though played in the original: GPT-6 LunaDecisions API (no credit), Mercury Decide (no OpenRouter endpoint): outside the requested scope (Ümit, 12 Nov), and blocked today as well; their seed pins are kept.
61.959.7 to 64.1 / 27.525.0 to 30.1 / 1,88162.860.7 to 65.1 / 29.126.4 to 31.6 / 1,83117 of 17Complete · 17 of 17 matches
5REAX 1.3NVIDIA RTX PRO 6000 Blackwell
Run 2
61.259.2 to 63.2 / 28.125.0 to 31.2 / 1,838
First run
60.358.5 to 62.2 / 28.225.2 to 31.1 / 1,868 (risk judgement / error rate (PR) / decisions)
Run 1 and Run 2, same pairs
Run 1 (first run)Run 2 (repeat)
Risk judgement60.358.5 to 62.261.259.2 to 63.2
Error rate (PR)28.225.2 to 31.128.125.0 to 31.2
Decisions1,8681,838
Invalid decisions00
Billed cost$0.118$0.114
Status
Complete · 17 of 17 matches (league #34)
Run id
repeat-2026-11-11-reax-1.3-r1
Setup
its own players row: reax/reax13 · local (reax13 serve.py) · fp8; held: in no standings, table or other setup's numbers
Settings
as the original's, changed: server address (logic's REAX server), device note, one match at a time, name, note; League 1 rules (prompt bg-v2, 3 rounds, 300 turns, the original's option-order seed, mirror games on); each pair with its original match's seed and seats; one match at a time
First run's pairs
late entry: 18 of 19 pairs played, 2026-11-07; GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit); its numbers here: its own decisions in the 17 pairs the repeat finished (matches #905, #906, #907, #908, #909, #910, #911, #912, #913, #914, #915, #916, #917, #918, #919, #920, #922)
Dates
prepared 2026-11-11; started 2026-11-12 05:50 UTC; finished 2026-11-12 07:57 UTC
Pairs today
17 of 19 playable (first run: 18 of 19 played); outside the requested scope (Ümit, 12 Nov): Mercury Decide (played in the first run; blocked today as well)
Cost basis
$0.12 estimated (the original's billed paid-opponent costs); cap $5 per repeat
Differences from the original protocol
  • Name: '<original> [repeat 1]', a new players row; settings otherwise as the original's (the changed fields are listed under Settings); one match at a time, where the original played several at once.
  • Server: logic's container logic-dm-reax13 on Docker's default bridge, its API published on the host's loopback (outbound network open; offline Hugging Face variables set as before), where the original ran reax13-late on the internal network reax13-offline (no outbound route), reached on the container's address; logic's env lacks HF_HUB_DISABLE_IMPLICIT_TOKEN=1 (offline either way). Same image, engine, kit, weights, config and serve.py arguments.
  • no-new-privileges: the original ran without it under Can's one-run exception (#2199); logic's server drops it only with REAX_NO_NNP_EXCEPTION=1 and its own approval, otherwise it does not start.
  • Not replayed, though played in the original: Mercury Decide (no OpenRouter endpoint): outside the requested scope (Ümit, 12 Nov), and blocked today as well; their seed pins are kept.
61.259.2 to 63.2 / 28.125.0 to 31.2 / 1,83860.358.5 to 62.2 / 28.225.2 to 31.1 / 1,86817 of 17Complete · 17 of 17 matches
6Drex Max v1.1
Run 2
59.657.5 to 61.9 / 31.128.3 to 33.9 / 2,087
First run
59.157.0 to 61.5 / 31.128.4 to 33.9 / 2,072 (risk judgement / error rate (PR) / decisions)
Run 1 and Run 2, same pairs
Run 1 (first run)Run 2 (repeat)
Risk judgement59.157.0 to 61.559.657.5 to 61.9
Error rate (PR)31.128.4 to 33.931.128.3 to 33.9
Decisions2,0722,087
Invalid decisions00
Billed cost$0.403$0.412
Status
Complete · 17 of 17 matches (league #29)
Run id
repeat-2026-11-11-drex-max-1.1-r1
Setup
its own players row: nace/drex-max-v1.1 · nace · hosted; held: in no standings, table or other setup's numbers
Settings
as the original's, changed: one match at a time, name, note; League 1 rules (prompt bg-v2, 3 rounds, 300 turns, the original's option-order seed, mirror games on); each pair with its original match's seed and seats; one match at a time
First run's pairs
late entry: 18 of 19 pairs played, 2026-11-06; GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit); its numbers here: its own decisions in the 17 pairs the repeat finished (matches #830, #831, #839, #842, #844, #850, #851, #854, #862, #865, #866, #869, #871, #875, #878, #880, #884)
Dates
prepared 2026-11-11; started 2026-11-11 20:21 UTC; finished 2026-11-11 21:59 UTC
Pairs today
17 of 19 playable (first run: 18 of 19 played); outside the requested scope (Ümit, 12 Nov): Mercury Decide (played in the first run; blocked today as well)
Cost basis
$0.40 estimated (the original's billed paid-opponent costs, with its own $0.28); cap $5 per repeat
Differences from the original protocol
  • Name: '<original> [repeat 1]', a new players row; settings otherwise as the original's (the changed fields are listed under Settings); one match at a time, where the original played several at once.
  • Runner: the pinned-pairs build 4ca4f08 (the pins are its only change for these leagues), where the original ran e58fc6c (clean); the same prompt version, option-order seed and rules.
  • Hosted model: Drex Max v1.1 is served by Nace; whether the model behind drex-max-v1.1 changed since 2026-11-06 cannot be checked from here.
  • Nace pre-send cap: the settings (unchanged) book every request in the original's ledger workspace/drex/state/nace-budget.db, cap $1.25 shared with Drex 2.0, which the original started empty and which held $0.7856 settled on 2026-11-11 (Drex Max v1.1 $0.2853, Drex 2.0 $0.5003); the repeat has the $0.4644 left. Its expected own bill ($0.2751 at $0.03/M) fits; reaching the cap would refuse its requests and leave the repeat incomplete.
  • Not replayed, though played in the original: Mercury Decide (no OpenRouter endpoint): outside the requested scope (Ümit, 12 Nov), and blocked today as well; their seed pins are kept.
59.657.5 to 61.9 / 31.128.3 to 33.9 / 2,08759.157.0 to 61.5 / 31.128.4 to 33.9 / 2,07217 of 17Complete · 17 of 17 matches
7GLiDElocal · v1.2
Run 2
58.055.7 to 60.3 / 35.931.8 to 40.1 / 1,792
First run
58.756.2 to 61.1 / 36.032.1 to 40.1 / 1,790 (risk judgement / error rate (PR) / decisions)
Run 1 and Run 2, same pairs
Run 1 (first run)Run 2 (repeat)
Risk judgement58.756.2 to 61.158.055.7 to 60.3
Error rate (PR)36.032.1 to 40.135.931.8 to 40.1
Decisions1,7901,792
Invalid decisions00
Billed cost$0.120$0.123
Status
Complete · 16 of 16 matches (league #33)
Run id
repeat-2026-11-11-glide-v1.2-r1
Setup
its own players row: fastino/glide-v1.2 · local (glide/server.py) · bf16; held: in no standings, table or other setup's numbers
Settings
as the original's, changed: one match at a time, name, note; League 1 rules (prompt bg-v2, 3 rounds, 300 turns, the original's option-order seed, mirror games off); each pair with its original match's seed and seats; one match at a time
First run's pairs
League 1: 18 of 18 pairs played, 2026-10-22; its numbers here: its own decisions in the 16 pairs the repeat finished (matches #482, #483, #484, #485, #486, #487, #488, #489, #490, #491, #492, #493, #495, #496, #498, #499)
Dates
prepared 2026-11-11; started 2026-11-12 03:31 UTC; finished 2026-11-12 04:33 UTC
Pairs today
16 of 18 playable (first run: 18 of 18 played); outside the requested scope (Ümit, 12 Nov): GPT-6 LunaDecisions API (played in the first run; blocked today as well), Mercury Decide (played in the first run; blocked today as well)
Cost basis
$0.12 estimated (the original's billed paid-opponent costs); cap $5 per repeat
Differences from the original protocol
  • Name: '<original> [repeat 1]', a new players row; settings otherwise as the original's (the changed fields are listed under Settings); one match at a time, where the original played several at once.
  • Runner: its 18 League 1 matches were played by d2f1f9e with uncommitted changes (#482-#486) and ec555d6 (clean, #487-#499); the repeat uses the pinned-pairs build 4ca4f08. Same prompt version bg-v2, option-order seed and rules; mirror games off as in its originals.
  • Opponent hardware: Clef (#484) and Cygnet (#486) answered from DGX Spark (GB10) servers then, Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark (#483) from spark3 until 22:41 UTC; they now run on our RTX PRO 6000 cards with the same checkpoints and arguments (their players.device records the move and its 20-request replays: same play 20/20, 20/20 and 19/20).
  • Server: the same script, snapshot and Python environment; the original ran on one RTX PRO 6000 card beside the opponents' servers on two others and the DGX Sparks, the repeat runs it on one card with the opponents one at a time on the same or a second card.
  • Not replayed, though played in the original: GPT-6 LunaDecisions API (no credit), Mercury Decide (no OpenRouter endpoint): outside the requested scope (Ümit, 12 Nov), and blocked today as well; their seed pins are kept.
58.055.7 to 60.3 / 35.931.8 to 40.1 / 1,79258.756.2 to 61.1 / 36.032.1 to 40.1 / 1,79016 of 16Complete · 16 of 16 matches
8MiniCPM5-2Boff
Run 2
53.050.4 to 56.0 / 52.745.7 to 60.6 / 1,130
First run
53.349.9 to 56.6 / 53.844.8 to 63.7 / 976 (risk judgement / error rate (PR) / decisions)
Run 1 and Run 2, same pairs
Run 1 (first run)Run 2 (repeat)
Risk judgement53.349.9 to 56.653.050.4 to 56.0
Error rate (PR)53.844.8 to 63.752.745.7 to 60.6
Decisions9761,130
Invalid decisions00
Billed cost$0.107$0.126
Status
Complete · 9 of 9 matches (league #60)
Run id
repeat-2026-11-13-minicpm5-2b-off-r1
Setup
its own players row: openai-compatible/minicpm5-2b · local-vllm · bf16; held: in no standings, table or other setup's numbers
Settings
as the original's, changed: name, note; League 1 rules (prompt bg-v2, 3 rounds, 300 turns, the original's option-order seed, mirror games on); each pair with its original match's seed and seats; one match at a time
First run's pairs
late entry: 9 of 19 pairs played, 2026-11-12; not played: Clef (not required (Ümit #6410: only the pairs the original played); local setup, not played by the original (0 of 8 locals: queued for a GPU)); not played: Cygnet (not required (Ümit #6410: only the pairs the original played); local setup, not played by the original (0 of 8 locals: queued for a GPU)); not played: Perplexity Decider Flash v1.2 (not required (Ümit #6410: only the pairs the original played); local setup, not played by the original (0 of 8 locals: queued for a GPU)); not played: Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark (not required (Ümit #6410: only the pairs the original played); local setup, not played by the original (0 of 8 locals: queued for a GPU)); not played: Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell (not required (Ümit #6410: only the pairs the original played); local setup, not played by the original (0 of 8 locals: queued for a GPU)); not played: GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell (not required (Ümit #6410: only the pairs the original played); local setup, not played by the original (0 of 8 locals: queued for a GPU)); GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (not required (Ümit #6410: only the pairs the original played); never scheduled: OpenAI Decisions out of credit (start.sh: deferred)); Mercury Decide: outside the requested scope (Ümit, 12 Nov); not played (not required (Ümit #6410: only the pairs the original played); #1172 aborted at its warm-up, no game: no OpenRouter endpoint for inception/mercury-decide); not played: Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell (not required (Ümit #6410: only the pairs the original played); local setup, not played by the original (0 of 8 locals: queued for a GPU)); not played: Winnow-12B (not required (Ümit #6410: only the pairs the original played); local setup, not played by the original (0 of 8 locals: queued for a GPU)); its numbers here: its own decisions in the 9 pairs the repeat finished (matches #1171, #1173, #1174, #1175, #1176, #1177, #1178, #1179, #1180)
Dates
prepared 2026-11-13; started 2026-11-16 01:55 UTC; finished 2026-11-16 01:56 UTC
Pairs today
9 of 19 playable (first run: 9 of 19 played)
Cost basis
$0.11 estimated (the original's billed paid-opponent costs); cap $5 per repeat
Differences from the original protocol
  • Name: '<original> [repeat 1]', a new players row; settings otherwise as the original's (the changed fields are listed under Settings); its 9 pairs at once as in the original (players.settings max_matches 16, parallel 16): its league config holds all 9 pinned opponents, as the original's league held them all, so the league is created with its whole roster and pins in one transaction, and run.sh starts it only once that roster is admitted and its guard is up (Zeki SR1 #6548).
  • Runner: the original ran 171698f (late-smallchat, target/smallchat (clean; its src is origin/main's 18cc78d); the repeat runs 6578a73 (branch small-repeat-runner): the repeats' runner 0e66836 (origin/late-reax line: decision clients reax, baselabs, drex-max, drex-local, nace and a pre-send budget ledger for decision services and openai-compatible chat, none used by this field; the pinned-pairs change 4ca4f08; admit) with late-tiny8-small3 8ebad40's decision clients and late-smallchat-on a083fb4's reasoning_prefill / reasoning_field keys (unused here). Every request of the original's 9 matches replays byte for byte: MiniCPM5-2B (off) 1,300 of 1,300 requests, its opponents 676 of 676 (mailbox/small-repeat-review/byte-identity). Same prompt version bg-v2, option-order seed and rules.
  • Server: logic's vLLM container, pinned and checked as in the original (pins.sh healthy and pinned before any write and every minute, 3 failures in a row stop); the original's guard checked pinned once at its start. Its pins.sh checks the checkpoint, revision, dtype and context, not the vLLM image.
  • Server context cap: the original's notes record --max-model-len 65,536; this run's server (logic's small_models.py key minicpm5-2b, same model@42433e90, served name, port 18614 and vLLM v0.30.0) launched with --max-model-len 69,632. The repeat's longest observed request was about 4.7k tokens; the original's effective launch arguments were not retained, so no claim is made that every server setting was identical.
  • Admission: a first start at 01:50 UTC 16 Nov was refused at client admission before anything was written (the unit's PATH lacked /usr/games, so GNU Backgammon could not start); the run that played started 01:55 UTC after that fix, with no replay.
53.050.4 to 56.0 / 52.745.7 to 60.6 / 1,13053.349.9 to 56.6 / 53.844.8 to 63.7 / 9769 of 9Complete · 9 of 9 matches
9GLiDE 2BNVIDIA DGX Spark (GB10)
Run 2
44.839.5 to 50.1 / 64.853.7 to 77.7 / 781
First run
45.941.4 to 50.3 / 66.756.6 to 77.9 / 818 (risk judgement / error rate (PR) / decisions)
Run 1 and Run 2, same pairs
Run 1 (first run)Run 2 (repeat)
Risk judgement45.941.4 to 50.344.839.5 to 50.1
Error rate (PR)66.756.6 to 77.964.853.7 to 77.7
Decisions818781
Invalid decisions00
Billed cost$0.088$0.081
Status
Complete · 9 of 9 matches (league #51)
Run id
repeat-2026-11-13-glide-2b-r1
Setup
its own players row: glide · local (glide.server) · bf16; held: in no standings, table or other setup's numbers
Settings
as the original's, changed: name, note; League 1 rules (prompt bg-v2, 3 rounds, 300 turns, the original's option-order seed, mirror games on); each pair with its original match's seed and seats; one match at a time
First run's pairs
late entry: 9 of 19 pairs played, 2026-11-12; not played: Clef (not required (Ümit #6410: only the pairs the original played); local setup, not played by the original (0 of 8 locals: queued for a GPU)); not played: Cygnet (not required (Ümit #6410: only the pairs the original played); local setup, not played by the original (0 of 8 locals: queued for a GPU)); not played: Perplexity Decider Flash v1.2 (not required (Ümit #6410: only the pairs the original played); local setup, not played by the original (0 of 8 locals: queued for a GPU)); not played: Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark (not required (Ümit #6410: only the pairs the original played); local setup, not played by the original (0 of 8 locals: queued for a GPU)); not played: Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell (not required (Ümit #6410: only the pairs the original played); local setup, not played by the original (0 of 8 locals: queued for a GPU)); not played: GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell (not required (Ümit #6410: only the pairs the original played); local setup, not played by the original (0 of 8 locals: queued for a GPU)); GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (not required (Ümit #6410: only the pairs the original played); never scheduled: OpenAI Decisions out of credit (start.sh: deferred)); Mercury Decide: outside the requested scope (Ümit, 12 Nov); not played (not required (Ümit #6410: only the pairs the original played); #1200 aborted at its warm-up, no game: no OpenRouter endpoint for inception/mercury-decide); not played: Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell (not required (Ümit #6410: only the pairs the original played); local setup, not played by the original (0 of 8 locals: queued for a GPU)); not played: Winnow-12B (not required (Ümit #6410: only the pairs the original played); local setup, not played by the original (0 of 8 locals: queued for a GPU)); its numbers here: its own decisions in the 9 pairs the repeat finished (matches #1191, #1192, #1193, #1194, #1195, #1196, #1197, #1198, #1199)
Dates
prepared 2026-11-13; started 2026-11-14 00:22 UTC; finished 2026-11-14 00:33 UTC
Pairs today
9 of 19 playable (first run: 9 of 19 played)
Cost basis
$0.09 estimated (the original's billed paid-opponent costs); cap $5 per repeat
Differences from the original protocol
  • Name: '<original> [repeat 1]', a new players row; settings otherwise as the original's (the changed fields are listed under Settings); one match at a time, as the original (players.settings max_matches 1, parallel 1).
  • Runner: the original ran e4ce0e9 (late-tiny8, target/tiny8 (clean; origin/main's runner line 18cc78d with late-tiny8's decision clients); the repeat runs 6578a73 (branch small-repeat-runner): the repeats' runner 0e66836 (origin/late-reax line: decision clients reax, baselabs, drex-max, drex-local, nace and a pre-send budget ledger for decision services and openai-compatible chat, none used by this field; the pinned-pairs change 4ca4f08; admit) with late-tiny8-small3 8ebad40's decision clients and late-smallchat-on a083fb4's reasoning_prefill / reasoning_field keys (unused here). Every request of the original's 9 matches replays byte for byte: GLiDE 2B 820 of 820 requests, its opponents 558 of 558 (mailbox/small-repeat-review/byte-identity). Same prompt version bg-v2, option-order seed and rules.
  • Server: logic's GLiDE 2B server on spark2, pinned and checked as in the original (pins.sh healthy and pinned before any write and every minute, 3 failures in a row stop); the original's guard checked pinned once at its start.
44.839.5 to 50.1 / 64.853.7 to 77.7 / 78145.941.4 to 50.3 / 66.756.6 to 77.9 / 8189 of 9Complete · 9 of 9 matches

Learning

In progress12/20 verdicts · 1 running · 6 paused

Do models take better risks after notes from their own games? Every model, five series each.

Pilot

Completed

One model, finished.

Clear risk

In progress

Decisions where a risk or safety is clearly better, per setup.

Nackgammon

Completed

Does a model still play well from a different starting position?