decision model · League 1
Winnow-12B
FinishedPublished
Results f24d1f41, updated 16 November 2026, 21:24 UTC; scored by GNU Backgammon 1.08, 2-ply.
How we called it
- Model ID
- winnow-12b
- Provider or endpoint
- NVIDIA RTX PRO 6000 Blackwell (96 GB), Q8_0 GGUF, author's server
- Date run
- 2026-10-19 to 2026-10-26
- 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 40 options per request; more legal plays are split into equal groups, a winner per group, then among the winners; in this data 92 decisions were split into groups of 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40
- Sampling settings
- not sent (decision request)
- Notes
- a Gemma 4 12B IT fine-tune, run through its author's inference server with the author's configuration (Q8_0 GGUF); no published price
Results in League 1
- Error rate (PR)
- 49.745.9 to 53.6
- Decisions
- 2,015 (2,015 scored, 0 invalid)
- Engine's best play
- 38.9%
- Blunder rate
- 32.3%
- P95 equity loss
- 0.456
- Mean loss, worst 5%
- 0.688
- Median / P90 s per decision
- 2.46 / 5.45
- $ per 1,000 decisions
- data-v="9.615">9.615estimate · RTX PRO 6000 Blackwell
- Risk judgement, margin 0.001
- 55.2 (95% interval 52.9 to 57.2); risks found 33.7% (263 / 780), right safety 76.6% (573 / 748)
- Risk judgement, margin 0.02
- 54.5 (95% interval 52.7 to 56.2); risks found 33.6% (233 / 693), right safety 75.3% (629 / 835)
- No trade-off
- 487 (left out of every risk metric)
Risk call card
55.2risk judgement, margin 0.001
right risks263
wrong risks102
needless risks175
missed risks415
right safety573
Hit risk only: being hit. Gammon risk and priming are not in it.
Close choices
A decision is a close choice when the second most likely play has at least half the probability of the first (internally: split).
Probabilities over plays, not sampled reasoning. Method
- Decisions
- 2,015
- Eligible (full probability distribution)
- 1,705
- Left out, by reason
- asked in groups of options 81; not audited 229
- Close choices
- 40.6% (692 / 1,705)
- Ties in the top two
- 0 decisions (0 at a positive probability, 0 at 0; the two most likely plays tied: 0, the second tied with the third: 0); 0 of them close choices. Ties are broken by the order the plays were shown, never by the engine.
- Between a safe play and a riskier one
- 35.1% (243 / 692) of the close choices; 14.3% (243 / 1,705) of all eligible decisions
| Decisions | Number | p1 (most likely play) | p2 (second) | p1 + p2 |
|---|---|---|---|---|
| Close choices | 692 | 0.395median 0.388 | 0.291median 0.286 | 0.686median 0.681 |
| Other decisions | 1,013 | 0.719median 0.728 | 0.132median 0.131 | 0.851median 0.885 |
| p1 + p2 | 0.0–0.1 | 0.1–0.2 | 0.2–0.3 | 0.3–0.4 | 0.4–0.5 | 0.5–0.6 | 0.6–0.7 | 0.7–0.8 | 0.8–0.9 | 0.9–1.0 |
|---|---|---|---|---|---|---|---|---|---|---|
| Close choices | 0 | 0 | 3 | 38 | 74 | 122 | 142 | 116 | 93 | 104 |
| Other decisions | 0 | 0 | 0 | 1 | 15 | 42 | 100 | 149 | 250 | 456 |
| Decisions | Number | Hitting rolls | Newly made points | Pip counts | Longest prime | None / 1 / 2 / 3 / 4 differ |
|---|---|---|---|---|---|---|
| Close choices | 692 | 50.6%350 / 692 | 20.4%141 / 692 | 13.6%94 / 692 | 21.1%146 / 692 | 227 / 276 / 124 / 53 / 12 |
| Other decisions | 1,013 | 56.7%574 / 1,013 | 42.9%435 / 1,013 | 10.4%105 / 1,013 | 29.9%303 / 1,013 | 188 / 380 / 313 / 117 / 15 |
| Decisions | Number | Gap between the two plays' values | Most likely play: behind the engine's best | Second most likely play: behind the engine's best |
|---|---|---|---|---|
| Close choices | 692 | 0.126median 0.059 | 0.111median 0.029 | 0.119median 0.040 |
| Other decisions | 1,013 | 0.152median 0.082 | 0.088median 0.009 | 0.157median 0.077 |
| Margin | In Risk pays positions | In Safety pays positions |
|---|---|---|
| margin 0.001 | 50.4%132 / 262 | 47.0%111 / 236 |
| margin 0.02 | 48.9%109 / 223 | 48.7%134 / 275 |
| Margin | Risk pays | Safety pays | No trade-off | Risk pays, small stake | Risk pays, medium stake | Risk pays, big stake |
|---|---|---|---|---|---|---|
| margin 0.001 | 40.0%262 / 655 | 38.1%236 / 620 | 45.1%194 / 430 | 53.4%39 / 73, under 200 decisions | 40.5%85 / 210 | 37.1%138 / 372 |
| margin 0.02 | 38.3%223 / 582 | 39.7%275 / 693 | 45.1%194 / 430 | N/A0 / 0 | 40.5%85 / 210 | 37.1%138 / 372 |
At the 0.02 margin a Risk pays position has a stake of at least 0.02, so the small band is empty. A big stake means playing safe costs more by the engine's value, not that the decision was harder.
| Margin | Close choices: missed risks | Other decisions: missed risks (weighted to the close choices' mix) | Difference, close minus other | Crude difference |
|---|---|---|---|---|
| margin 0.001374 decisions | 62.0%crude 62.1% (82 / 132), under 200 decisions | 67.4%crude 66.9% (162 / 242) | −5.4 points95% interval −15.4 to +5.6 | −4.8 points95% interval −14.8 to +5.4 |
| margin 0.02330 decisions | 64.8%crude 65.1% (71 / 109), under 200 decisions | 65.2%crude 66.1% (146 / 221) | −0.4 points95% interval −10.8 to +11.9 | −0.9 points95% interval −10.9 to +9.1 |
By game phase
| Phase | Error rate (PR) | Decisions |
|---|---|---|
| Opening | 37.8 | 298 |
| Contact | 61.0 | 1,283 |
| Holding game | Too few decisions | 67 |
| Race | 20.2 | 265 |
| Bear-off | Too few decisions | 102 |