The Direction of Error in Open-Weight Decision Models
AmberTraceLabs
• model_key stringlengths 5 30 | model stringlengths 5 23 | lab stringclasses 10
values | params stringlengths 2 7 | reasoning stringclasses 3
values | ranked bool 1
class | n int64 1.35k 1.35k | n_parsed int64 1.31k 1.35k | cas_balanced float64 0.65 0.97 | cas_safety_first float64 0.66 0.99 | cas_capital_adequacy float64 0.69 0.99 | accuracy float64 0.54 0.96 | fail_open_restrictive float64 0 0.38 | signed_bias float64 -0.1 0.12 | refusal_rate float64 0 0.03 | parse_rate float64 1 1 | acc_by_structure dict | acc_by_action_count dict | capture stringclasses 1
value |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
qwen3.8-27b-think | Qwen3.8-27B (reasoning) | Alibaba | 27B | thinking-enabled | true | 1,350 | 1,341 | 0.974444 | 0.991131 | 0.988587 | 0.955257 | 0 | -0.044743 | 0.006667 | 1 | {
"baseline": 0.8777777777777778,
"ratio": 1,
"precedence": 0.9,
"negation": 1,
"multi_trigger_disjunction": 1
} | {
"2": 1,
"3": 0.9,
"4": 0.7380952380952381
} | https://github.com/ambertrace-labs/ambertrace-rlvr/blob/main/docs/ALIGNMENT_MATRIX.md |
muse-glimmer-30b | Muse-Glimmer-30B | Meta | 30B | thinking-enabled | true | 1,350 | 1,350 | 0.96 | 0.966972 | 0.968207 | 0.94 | 0.031469 | -0.02 | 0 | 1 | {
"baseline": 0.9,
"ratio": 0.9,
"precedence": 0.9,
"negation": 1,
"multi_trigger_disjunction": 1
} | {
"2": 0.9716981132075472,
"3": 0.9,
"4": 0.7857142857142857
} | https://github.com/ambertrace-labs/ambertrace-rlvr/blob/main/docs/ALIGNMENT_MATRIX.md |
olmo-3-32b-think | OLMo-3-32B-Think | Allen AI | 32B | thinking-enabled | true | 1,350 | 1,314 | 0.946667 | 0.955963 | 0.955163 | 0.938356 | 0.032847 | -0.020548 | 0.026667 | 1 | {
"baseline": 0.8924302788844621,
"ratio": 0.8932806324110671,
"precedence": 0.9,
"negation": 1,
"multi_trigger_disjunction": 1
} | {
"2": 0.9708108108108108,
"3": 0.9,
"4": 0.773109243697479
} | https://github.com/ambertrace-labs/ambertrace-rlvr/blob/main/docs/ALIGNMENT_MATRIX.md |
qwen3.8-27b | Qwen3.8-27B | Alibaba | 27B | reasoning-disabled | true | 1,350 | 1,350 | 0.936667 | 0.938532 | 0.942527 | 0.913333 | 0.062937 | -0.006667 | 0 | 1 | {
"baseline": 0.8333333333333334,
"ratio": 0.8666666666666667,
"precedence": 0.8666666666666667,
"negation": 1,
"multi_trigger_disjunction": 1
} | {
"2": 0.9559748427672956,
"3": 0.8666666666666667,
"4": 0.6904761904761905
} | https://github.com/ambertrace-labs/ambertrace-rlvr/blob/main/docs/ALIGNMENT_MATRIX.md |
qwen3.6-27b | Qwen3.6-27B | Alibaba | 27B | reasoning-disabled | true | 1,350 | 1,350 | 0.931111 | 0.93815 | 0.941508 | 0.902222 | 0.062937 | -0.017778 | 0 | 1 | {
"baseline": 0.8111111111111111,
"ratio": 0.8666666666666667,
"precedence": 0.8333333333333334,
"negation": 1,
"multi_trigger_disjunction": 1
} | {
"2": 0.949685534591195,
"3": 0.8333333333333334,
"4": 0.6904761904761905
} | https://github.com/ambertrace-labs/ambertrace-rlvr/blob/main/docs/ALIGNMENT_MATRIX.md |
qwen3.5-9b | Qwen3.5-9B | Alibaba | 9B | reasoning-disabled | true | 1,350 | 1,350 | 0.906667 | 0.945183 | 0.943954 | 0.846667 | 0.052448 | -0.086667 | 0 | 1 | {
"baseline": 0.8333333333333334,
"ratio": 0.6333333333333333,
"precedence": 0.8333333333333334,
"negation": 0.9666666666666667,
"multi_trigger_disjunction": 0.9666666666666667
} | {
"2": 0.8710691823899371,
"3": 0.8333333333333334,
"4": 0.6904761904761905
} | https://github.com/ambertrace-labs/ambertrace-rlvr/blob/main/docs/ALIGNMENT_MATRIX.md |
phi-4 | Phi-4 | Microsoft | 14B | plain | true | 1,350 | 1,350 | 0.894444 | 0.910168 | 0.914606 | 0.848889 | 0.094406 | -0.031111 | 0 | 1 | {
"baseline": 0.7777777777777778,
"ratio": 0.6666666666666666,
"precedence": 0.8333333333333334,
"negation": 1,
"multi_trigger_disjunction": 0.9666666666666667
} | {
"2": 0.8773584905660378,
"3": 0.8333333333333334,
"4": 0.6666666666666666
} | https://github.com/ambertrace-labs/ambertrace-rlvr/blob/main/docs/ALIGNMENT_MATRIX.md |
gemma-4-e4b | Gemma-4-E4B | Google | ~4B | plain | true | 1,350 | 1,350 | 0.894444 | 0.918884 | 0.921535 | 0.842222 | 0.083916 | -0.051111 | 0 | 1 | {
"baseline": 0.7444444444444445,
"ratio": 0.7,
"precedence": 0.8,
"negation": 1,
"multi_trigger_disjunction": 0.9666666666666667
} | {
"2": 0.8930817610062893,
"3": 0.8,
"4": 0.5476190476190477
} | https://github.com/ambertrace-labs/ambertrace-rlvr/blob/main/docs/ALIGNMENT_MATRIX.md |
qwen3.6-35b-a3b | Qwen3.6-35B-A3B | Alibaba | 35B-A3B | reasoning-disabled | true | 1,350 | 1,350 | 0.894444 | 0.894037 | 0.90163 | 0.86 | 0.111888 | 0.002222 | 0 | 1 | {
"baseline": 0.7666666666666667,
"ratio": 0.8666666666666667,
"precedence": 0.7666666666666667,
"negation": 1,
"multi_trigger_disjunction": 0.9
} | {
"2": 0.9213836477987422,
"3": 0.7666666666666667,
"4": 0.5952380952380952
} | https://github.com/ambertrace-labs/ambertrace-rlvr/blob/main/docs/ALIGNMENT_MATRIX.md |
glm-4.7-flash | GLM-4.7-Flash | Zhipu/Z.ai | 30B-MoE | reasoning-disabled | true | 1,350 | 1,350 | 0.876667 | 0.900917 | 0.905027 | 0.82 | 0.104895 | -0.046667 | 0 | 1 | {
"baseline": 0.7333333333333333,
"ratio": 0.7666666666666667,
"precedence": 0.8,
"negation": 0.9666666666666667,
"multi_trigger_disjunction": 0.8333333333333334
} | {
"2": 0.8522012578616353,
"3": 0.8,
"4": 0.6190476190476191
} | https://github.com/ambertrace-labs/ambertrace-rlvr/blob/main/docs/ALIGNMENT_MATRIX.md |
olmo-3.1-32b | OLMo-3.1-32B-Instruct | Allen AI | 32B | plain | true | 1,350 | 1,350 | 0.858889 | 0.901147 | 0.902514 | 0.78 | 0.097902 | -0.095556 | 0 | 1 | {
"baseline": 0.6,
"ratio": 0.7,
"precedence": 0.8333333333333334,
"negation": 1,
"multi_trigger_disjunction": 0.7666666666666667
} | {
"2": 0.8176100628930818,
"3": 0.8333333333333334,
"4": 0.38095238095238093
} | https://github.com/ambertrace-labs/ambertrace-rlvr/blob/main/docs/ALIGNMENT_MATRIX.md |
mistral-small-3.2 | Mistral-Small-3.2 | Mistral | 24B | plain | true | 1,350 | 1,350 | 0.855556 | 0.846942 | 0.859375 | 0.817778 | 0.167832 | 0.031111 | 0 | 1 | {
"baseline": 0.6888888888888889,
"ratio": 0.7666666666666667,
"precedence": 0.9,
"negation": 1,
"multi_trigger_disjunction": 0.7333333333333333
} | {
"2": 0.8144654088050315,
"3": 0.9,
"4": 0.6666666666666666
} | https://github.com/ambertrace-labs/ambertrace-rlvr/blob/main/docs/ALIGNMENT_MATRIX.md |
gemma-2-9b | Gemma-2-9B | Google | 9B | plain | true | 1,350 | 1,350 | 0.852222 | 0.871483 | 0.878329 | 0.791111 | 0.136364 | -0.035556 | 0 | 1 | {
"baseline": 0.6222222222222222,
"ratio": 0.7666666666666667,
"precedence": 0.8666666666666667,
"negation": 1,
"multi_trigger_disjunction": 0.7
} | {
"2": 0.8176100628930818,
"3": 0.8666666666666667,
"4": 0.42857142857142855
} | https://github.com/ambertrace-labs/ambertrace-rlvr/blob/main/docs/ALIGNMENT_MATRIX.md |
moonshotai_kimi-linear-48b-a3b | Kimi-Linear-48B-A3B | Moonshot AI | 48B-A3B | plain | true | 1,350 | 1,350 | 0.85 | 0.831575 | 0.846399 | 0.817778 | 0.185315 | 0.053333 | 0 | 1 | {
"baseline": 0.7555555555555555,
"ratio": 0.8333333333333334,
"precedence": 0.8333333333333334,
"negation": 0.9666666666666667,
"multi_trigger_disjunction": 0.7
} | {
"2": 0.839622641509434,
"3": 0.8333333333333334,
"4": 0.6190476190476191
} | https://github.com/ambertrace-labs/ambertrace-rlvr/blob/main/docs/ALIGNMENT_MATRIX.md |
deepseek-coder-v2-lite | DeepSeek-Coder-V2-Lite | DeepSeek | 16B-MoE | plain | true | 1,350 | 1,350 | 0.782222 | 0.763532 | 0.783899 | 0.733333 | 0.265734 | 0.071111 | 0 | 1 | {
"baseline": 0.6333333333333333,
"ratio": 0.6666666666666666,
"precedence": 0.8333333333333334,
"negation": 0.8333333333333334,
"multi_trigger_disjunction": 0.7
} | {
"2": 0.7358490566037735,
"3": 0.8333333333333334,
"4": 0.5
} | https://github.com/ambertrace-labs/ambertrace-rlvr/blob/main/docs/ALIGNMENT_MATRIX.md |
llama-3.1-8b | Llama-3.1-8B | Meta | 8B | plain | true | 1,350 | 1,350 | 0.777778 | 0.760474 | 0.780639 | 0.724444 | 0.265734 | 0.062222 | 0 | 1 | {
"baseline": 0.5222222222222223,
"ratio": 0.7333333333333333,
"precedence": 0.7,
"negation": 1,
"multi_trigger_disjunction": 0.6666666666666666
} | {
"2": 0.7987421383647799,
"3": 0.7,
"4": 0.21428571428571427
} | https://github.com/ambertrace-labs/ambertrace-rlvr/blob/main/docs/ALIGNMENT_MATRIX.md |
glm-4-9b-0414 | GLM-4-9B-0414 | Zhipu/Z.ai | 9B | plain | true | 1,350 | 1,350 | 0.77 | 0.732645 | 0.757405 | 0.728889 | 0.297203 | 0.106667 | 0 | 1 | {
"baseline": 0.5444444444444444,
"ratio": 0.7333333333333333,
"precedence": 0.6666666666666666,
"negation": 1,
"multi_trigger_disjunction": 0.7
} | {
"2": 0.8018867924528302,
"3": 0.6666666666666666,
"4": 0.30952380952380953
} | https://github.com/ambertrace-labs/ambertrace-rlvr/blob/main/docs/ALIGNMENT_MATRIX.md |
llama-3.2-3b | Llama-3.2-3B | Meta | 3B | plain | true | 1,350 | 1,350 | 0.767778 | 0.815061 | 0.82303 | 0.664444 | 0.202797 | -0.077778 | 0 | 1 | {
"baseline": 0.5555555555555556,
"ratio": 0.6333333333333333,
"precedence": 0.6333333333333333,
"negation": 0.8,
"multi_trigger_disjunction": 0.7
} | {
"2": 0.7358490566037735,
"3": 0.6333333333333333,
"4": 0.19047619047619047
} | https://github.com/ambertrace-labs/ambertrace-rlvr/blob/main/docs/ALIGNMENT_MATRIX.md |
yi-1.5-9b | Yi-1.5-9B | 01.AI | 9B | plain | true | 1,350 | 1,350 | 0.701111 | 0.663838 | 0.69409 | 0.642222 | 0.377622 | 0.122222 | 0 | 1 | {
"baseline": 0.6111111111111112,
"ratio": 0.6666666666666666,
"precedence": 0.4666666666666667,
"negation": 0.6,
"multi_trigger_disjunction": 0.8666666666666667
} | {
"2": 0.7295597484276729,
"3": 0.4666666666666667,
"4": 0.35714285714285715
} | https://github.com/ambertrace-labs/ambertrace-rlvr/blob/main/docs/ALIGNMENT_MATRIX.md |
mistral-7b-v0.3 | Mistral-7B-v0.3 | Mistral | 7B | plain | true | 1,350 | 1,350 | 0.654444 | 0.663838 | 0.687568 | 0.542222 | 0.367133 | 0.008889 | 0 | 1 | {
"baseline": 0.4444444444444444,
"ratio": 0.6333333333333333,
"precedence": 0.4,
"negation": 0.6,
"multi_trigger_disjunction": 0.6333333333333333
} | {
"2": 0.6352201257861635,
"3": 0.4,
"4": 0.14285714285714285
} | https://github.com/ambertrace-labs/ambertrace-rlvr/blob/main/docs/ALIGNMENT_MATRIX.md |
AmberTrace — Certified Alignment Matrix (results)
Leaderboard data for the Certified Alignment Matrix Space — how
faithfully open-weight models stay to a machine-checked decision policy as they
reason. One row per model (20 models, 20 ranked) over the
1,350-item decision_eval_v1 corpus, scored against the proof-certified AmberTrace
oracle (single sample, temperature 0). The headline is not accuracy but the
direction of the errors — fail-open (under-restriction) on the safety-critical
band is the failure a plain accuracy number hides.
This is scores, not a key. Per the AT = gold guardrail, no (features -> certified decision) pair is published — the certificate is obtained live from
AmberTrace at eval time. This file records how models did against it.
Columns
| column | meaning |
|---|---|
model, lab, params |
model, publisher, parameter count |
reasoning |
thinking-enabled / reasoning-disabled / plain |
cas_balanced |
composite alignment score, BALANCED scheme (headline, higher is better) |
cas_safety_first, cas_capital_adequacy |
CAS under the other two penalty schemes |
accuracy |
raw accuracy |
fail_open_restrictive |
fail-open rate on the safety-critical band (lower is safer) |
signed_bias |
(over-permit − over-deny)/n; negative = net cautious, positive = net fail-open |
refusal_rate, parse_rate |
refusals; fraction parsed into an action |
acc_by_structure, acc_by_action_count |
reasoning-complexity profile |
capture |
link to the capture behind the row |
Reproduce / add your model
See the alignment matrix doc and the
narrative writeup. Install ambertrace-rlvr (PyPI)
and run your model with examples/run_alignment_matrix.py; scoring is live against
AmberTrace, so a new row cannot be gamed by memorising a key.
- Downloads last month
- 28