# Catan tournament results

Rules: `TOURNAMENT.md`. Paused games do not count as losses. Guarded games end at 250 turns and use the leader by victory points.

Six league appearances cannot divide equally across four positions. Each model plays two positions twice and the other two once.

Stages 1 to 3: one domestic trade offer per player per turn; bank and port trades remain unlimited. Final: first to two wins within three games, then summed victory points, then better league rating.

Ledger spend: US$3.443662. Carry-in: US$0.184448. Full accounting, including probes and failed requests, is in `spend.json`.

Paid run, framing v2. Recorded game dispatch ceilings: openrouter 60 RPM, openrouter 240 RPM, typesafe 60 RPM. Provider 429 and Retry-After remain authoritative. Seat strands uses togethercomputer/tev1-4b-experimental instead of StrandsAgents/strands-decider-2B-hobson-v19: Local MPS correct response in 3209.31ms, but official checkpoint supports 4096 tokens and truncates state, incompatible with common 6000-token envelope.

## Stage 0 standings

24 completed games, 0 paused or incomplete replays, 0 games without a replay.

| Model | Rating | Games | Wins | Average points | Fallbacks / decisions | Fallback rate | Errors | Median API ms | Seconds per seat-game | Cost USD | USD per seat-game | Status |
|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---|
| classify | n/a | 2 | 2 | 11.00 | 0 / 421 | 0.00% | 0 | 562.5 | 233.2 | 0.090387 | 0.045194 | pass |
| d1 | n/a | 2 | 2 | 10.00 | 0 / 247 | 0.00% | 0 | 324.0 | 142.0 | 0.011905 | 0.005953 | pass |
| jev | n/a | 2 | 2 | 10.00 | 0 / 271 | 0.00% | 0 | 86.6 | 238.4 | 0.034907 | 0.017453 | pass |
| luna | n/a | 2 | 2 | 10.00 | 1 / 141 | 0.71% | 3 | 263.7 | 70.0 | 0.034500 | 0.017250 | pass |
| mercury | n/a | 2 | 2 | 10.00 | 0 / 421 | 0.00% | 0 | 451.0 | 214.2 | 0.022728 | 0.011364 | pass |
| pplx | n/a | 2 | 2 | 10.00 | 0 / 236 | 0.00% | 0 | 427.0 | 116.6 | 0.012237 | 0.006118 | pass |
| solar | n/a | 2 | 2 | 7.50 | 0 / 312 | 0.00% | 0 | 12764.4 | 2266.7 | 0.049064 | 0.024532 | pass |
| clef-omni | n/a | 2 | 1 | 9.00 | 0 / 1432 | 0.00% | 0 | 301.3 | 504.6 | 0.606532 | 0.303266 | pass |
| drex | n/a | 2 | 1 | 8.50 | 0 / 466 | 0.00% | 1 | 325.2 | 183.7 | 0.071624 | 0.035812 | pass |
| ms-d1 | n/a | 2 | 1 | 7.00 | 0 / 338 | 0.00% | 0 | 331.1 | 136.4 | 0.035902 | 0.017951 | pass |
| kev | n/a | 2 | 0 | 3.00 | 0 / 417 | 0.00% | 0 | 616.7 | 253.2 | 0.044733 | 0.022367 | fail |
| strands | n/a | 2 | 0 | 3.00 | 0 / 436 | 0.00% | 0 | 251.6 | 104.8 | 0.040900 | 0.020450 | fail |

Median API latency includes every recorded request attempt, including retries, and excludes forced moves. Errors count failed recorded attempts. Seconds per seat-game is the mean whole-table wall time for completed games, including opponents, waiting and checkpoint work. Cost USD includes partial-game decisions; USD per seat-game uses completed games only.

Stage ledger spend, including startup probes and failed requests: US$1.096775. Selected replay decisions account for US$1.055417; US$0.041358 remains in probes, archived attempts or other stage calls. Archived attempts do not count in standings or projections.

Historical projection under the qualifier offer rules, from the equal-weight mean per model seat-game. Current house-rule sizing and approved caps are in `RUN.md`:

- Basis: 24 completed seat-games from 12/12 models. Measured mean: US$0.043976, 372.0 seconds per seat-game.
- Stage 1, 72 paid seat-games: US$3.166; 7.44 serial-equivalent hours.
- Stage 2, up to 28 paid seat-games: US$1.231; 2.89 serial-equivalent hours.
- Stage 3, up to 36 paid seat-games: US$1.583; 3.72 serial-equivalent hours. With all possible close-loss tiebreaks, up to 72: US$3.166; 7.44 serial-equivalent hours.

Serial-equivalent time sums the qualifier wall time for each paid model appearance. Four parallel tables could ideally divide that time by four, but provider limits, round barriers, longer games and AlphaBeta CPU time can prevent that speedup. Scaling a whole qualifier table also repeats its bot and checkpoint overhead, so this is a sizing estimate, not an elapsed-time promise.
Game length and opponent behavior can change these estimates. Recorded retries are included in per-game cost; probes and A/B tests remain outside those averages and are included in spend.json. Kevin approved US$10 additional spend and the house trade rule on 2026-10-14; `RUN.md` gives the updated estimate.

Paid trade-response sensitivity:

- Completed qualifier tables logged 1151 offers, including 1151 from model seats, and 1301 nonforced bot accept/reject decisions. Bot replies cost nothing in the qualifier.
- Conservative surcharge: 47.96 model offers per seat-game, three newly paid replies per offer, 5342.0 input and 198.7 output tokens per reply. Token assumptions use the highest measured model mean across all request types; prices use the highest pinned roster rates across 12 entries (US$0.000000150 input and US$0.000000000 output per token), giving US$0.000801 per reply.
- Stage 1: paid-seat baseline US$3.166, response surcharge US$8.301 for 10359 additional calls, combined planning amount US$11.467.
- Stage 2: paid-seat baseline up to US$1.231, response surcharge US$3.228 for 4029 additional calls, combined planning amount US$4.459.
- Using the highest measured model mean request latency of 12210.7 ms adds 35.14 serial-equivalent hours in stage 1 and 13.66 in stage 2, before provider throttling or retry delays.
- This surcharge deliberately treats every recipient as a paid reply, even when a reply could be forced, and can double-count response work already present in the baseline. It is a conservative sensitivity check, not an upper bound on future spend. Stage 3 still has three free bot opponents and keeps the baseline estimate unchanged.

Qualifier behavior to inspect:
- clef-omni: completed-game settlements built 4, cities built 0, guarded games 1, fallbacks 0.
- ms-d1: completed-game settlements built 8, cities built 1, guarded games 1, fallbacks 0.
- kev: completed-game settlements built 6, cities built 0, guarded games 1, fallbacks 0.
- drex: completed-game settlements built 10, cities built 0, guarded games 1, fallbacks 0.
- solar: completed-game settlements built 9, cities built 2, guarded games 1, fallbacks 0.
- strands: completed-game settlements built 4, cities built 2, guarded games 2, fallbacks 0.

## Stage 1 standings

18 completed games, 0 paused or incomplete replays, 0 games without a replay.

| Model | Rating | Games | Wins | Average points | Fallbacks / decisions | Fallback rate | Errors | Median API ms | Seconds per seat-game | Cost USD | USD per seat-game | Status |
|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---|
| ms-d1 | 1698.7 | 6 | 3 | 7.33 | 0 / 735 | 0.00% | 0 | 376.0 | 1513.4 | 0.082173 | 0.013696 |  |
| pplx | 1684.6 | 6 | 3 | 7.83 | 0 / 494 | 0.00% | 0 | 452.7 | 1373.9 | 0.026259 | 0.004377 |  |
| luna | 1612.9 | 6 | 4 | 8.67 | 4 / 642 | 0.62% | 12 | 301.3 | 1820.8 | 0.162266 | 0.027044 |  |
| mercury | 1600.0 | 6 | 1 | 6.50 | 0 / 591 | 0.00% | 2 | 491.6 | 1519.1 | 0.030919 | 0.005153 |  |
| d1 | 1575.2 | 6 | 2 | 6.67 | 0 / 779 | 0.00% | 0 | 339.1 | 1817.3 | 0.037785 | 0.006298 |  |
| jev | 1565.5 | 6 | 2 | 7.17 | 0 / 790 | 0.00% | 1 | 97.7 | 1513.3 | 0.099151 | 0.016525 |  |
| classify | 1552.6 | 6 | 2 | 6.33 | 0 / 745 | 0.00% | 0 | 713.4 | 1840.9 | 0.169599 | 0.028267 |  |
| drex | 1452.8 | 6 | 1 | 5.50 | 0 / 768 | 0.00% | 1 | 372.2 | 1475.1 | 0.116974 | 0.019496 |  |
| strands | 1443.6 | 6 | 0 | 4.00 | 2 / 811 | 0.25% | 9 | 290.2 | 1766.8 | 0.088615 | 0.014769 |  |
| clef-omni | 1371.0 | 6 | 0 | 4.33 | 0 / 663 | 0.00% | 1 | 314.8 | 1391.0 | 0.279885 | 0.046647 |  |
| solar | 1237.0 | 6 | 0 | 3.33 | 0 / 548 | 0.00% | 0 | 12933.2 | 2363.2 | 0.086041 | 0.014340 |  |
| kev | 1206.1 | 6 | 0 | 3.00 | 0 / 711 | 0.00% | 0 | 680.9 | 1842.2 | 0.075635 | 0.012606 |  |

Median API latency includes every recorded request attempt, including retries, and excludes forced moves. Errors count failed recorded attempts. Seconds per seat-game is the mean whole-table wall time for completed games, including opponents, waiting and checkpoint work. Cost USD includes partial-game decisions; USD per seat-game uses completed games only.

Stage ledger spend, including startup probes and failed requests: US$1.255472. Selected replay decisions account for US$1.255305; US$0.000168 remains in probes, archived attempts or other stage calls. Archived attempts do not count in standings or projections.

Ratings fit regularized Plackett-Luce log-likelihoods of score-ordered finishes. Equal scores form a tied tier; all orders within that tier receive equal fractional weight. Rating scale is 1500 + 400 log10(strength).

## Stage 2 standings

7 completed games, 0 paused or incomplete replays, 0 games without a replay.

| Model | Rating | Games | Wins | Average points | Fallbacks / decisions | Fallback rate | Errors | Median API ms | Seconds per seat-game | Cost USD | USD per seat-game | Status |
|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---|
| luna | n/a | 5 | 2 | 7.60 | 0 / 486 | 0.00% | 0 | 273.0 | 653.2 | 0.122530 | 0.024506 |  |
| pplx | n/a | 5 | 2 | 7.60 | 0 / 431 | 0.00% | 0 | 436.7 | 653.2 | 0.023473 | 0.004695 |  |
| mercury | n/a | 5 | 2 | 7.00 | 0 / 572 | 0.00% | 0 | 435.1 | 784.6 | 0.029636 | 0.005927 |  |
| ms-d1 | n/a | 5 | 1 | 6.40 | 0 / 601 | 0.00% | 0 | 345.8 | 784.6 | 0.065919 | 0.013184 |  |
| d1 | n/a | 2 | 0 | 5.00 | 0 / 188 | 0.00% | 0 | 311.7 | 1280.6 | 0.009359 | 0.004679 |  |
| classify | n/a | 2 | 0 | 4.00 | 0 / 167 | 0.00% | 0 | 634.1 | 952.2 | 0.033896 | 0.016948 |  |
| drex | n/a | 2 | 0 | 4.00 | 0 / 216 | 0.00% | 0 | 357.4 | 1280.6 | 0.032168 | 0.016084 |  |
| jev | n/a | 2 | 0 | 3.00 | 0 / 162 | 0.00% | 1 | 79.5 | 952.2 | 0.019514 | 0.009757 |  |
| clef-omni | n/a | 0 | 0 | 0.00 | 0 / 0 | 0.00% | 0 | n/a | n/a | 0.000000 | n/a |  |
| kev | n/a | 0 | 0 | 0.00 | 0 / 0 | 0.00% | 0 | n/a | n/a | 0.000000 | n/a |  |
| solar | n/a | 0 | 0 | 0.00 | 0 / 0 | 0.00% | 0 | n/a | n/a | 0.000000 | n/a |  |
| strands | n/a | 0 | 0 | 0.00 | 0 / 0 | 0.00% | 0 | n/a | n/a | 0.000000 | n/a |  |

Median API latency includes every recorded request attempt, including retries, and excludes forced moves. Errors count failed recorded attempts. Seconds per seat-game is the mean whole-table wall time for completed games, including opponents, waiting and checkpoint work. Cost USD includes partial-game decisions; USD per seat-game uses completed games only.

Stage ledger spend, including startup probes and failed requests: US$0.336543. Selected replay decisions account for US$0.336493; US$0.000049 remains in probes, archived attempts or other stage calls. Archived attempts do not count in standings or projections.

Grand final champion: pplx; summed victory points over three games. Summed points: luna 20, mercury 14, ms-d1 20, pplx 22.

## Stage 3 standings

20 completed games, 0 paused or incomplete replays, 0 games without a replay.

| Model | Rating | Games | Wins | Average points | Fallbacks / decisions | Fallback rate | Errors | Median API ms | Seconds per seat-game | Cost USD | USD per seat-game | Status |
|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---|
| pplx | n/a | 2 | 1 | 8.50 | 0 / 195 | 0.00% | 0 | 445.1 | 344.2 | 0.011117 | 0.005559 | level 1 cleared |
| luna | n/a | 3 | 1 | 8.00 | 0 / 227 | 0.00% | 0 | 275.4 | 233.0 | 0.057251 | 0.019084 | level 1 cleared |
| jev | n/a | 2 | 1 | 7.50 | 0 / 190 | 0.00% | 0 | 81.6 | 96.0 | 0.022955 | 0.011477 | level 1 cleared |
| mercury | n/a | 2 | 1 | 7.50 | 0 / 185 | 0.00% | 0 | 431.1 | 396.8 | 0.009531 | 0.004766 | level 1 cleared |
| ms-d1 | n/a | 2 | 1 | 7.50 | 0 / 203 | 0.00% | 0 | 350.0 | 418.4 | 0.023415 | 0.011707 | level 1 cleared |
| classify | n/a | 2 | 1 | 7.00 | 0 / 227 | 0.00% | 0 | 539.9 | 354.3 | 0.048664 | 0.024332 | level 1 cleared |
| d1 | n/a | 2 | 1 | 6.50 | 0 / 192 | 0.00% | 0 | 321.1 | 387.6 | 0.008932 | 0.004466 | level 1 cleared |
| clef-omni | n/a | 1 | 0 | 5.00 | 0 / 185 | 0.00% | 0 | 300.8 | 686.4 | 0.079232 | 0.079232 | none cleared |
| drex | n/a | 1 | 0 | 4.00 | 0 / 103 | 0.00% | 0 | 354.2 | 390.4 | 0.015509 | 0.015509 | none cleared |
| strands | n/a | 1 | 0 | 4.00 | 0 / 152 | 0.00% | 0 | 250.4 | 294.2 | 0.017073 | 0.017073 | none cleared |
| kev | n/a | 1 | 0 | 3.00 | 0 / 71 | 0.00% | 0 | 648.0 | 274.3 | 0.007792 | 0.007792 | none cleared |
| solar | n/a | 1 | 0 | 3.00 | 0 / 135 | 0.00% | 0 | 14629.3 | 2188.9 | 0.020019 | 0.020019 | none cleared |

Median API latency includes every recorded request attempt, including retries, and excludes forced moves. Errors count failed recorded attempts. Seconds per seat-game is the mean whole-table wall time for completed games, including opponents, waiting and checkpoint work. Cost USD includes partial-game decisions; USD per seat-game uses completed games only.

Stage ledger spend, including startup probes and failed requests: US$0.321586. Selected replay decisions account for US$0.321489; US$0.000097 remains in probes, archived attempts or other stage calls. Archived attempts do not count in standings or projections.