Rubric & failure-mode analysis

Survival rate alone does not reveal how an agent thinks. Every session is also scored by a panel of 3 independent judge models against a 16-point rubric, then classified into one of 4 failure modes. This page is the data behind every "agents don't reason causally" claim on the site.

68%
agentic sessions show no causal engagement
24.1%
reach only nascent causal reasoning
21%
best model's causal-reasoning score (Grok-4.1-Fast)
1103
agentic sessions judged

The 16-point rubric

Six criteria across four dimensions, weighted by importance (CR1 and CR2 dominate with weights 5 and 4). Weights sum to 16; the overall score is the weighted mean of the six normalized criterion scores.

IDDimensionWeightWhat it measures
CR1 CausalReasoning 5 Identify core causal mechanisms in the task report
CR2 CausalReasoning 4 Identify and avoid traps or spurious correlations
CR3 CausalReasoning 2 Mechanistic depth and testability
ED1 ExperimentalDesign 2 Support conclusions with concrete experimental data
RQ1 ReflectionQuality 2 Acknowledge errors and uncertainties
DU1 DataUsage 1 Clear data-to-conclusion linkage

Three judge models (gemini-3-f, grok-4-1-f, qwen3-next) score each session; inter-rater agreement is high (mean ICC(2,3) = 0.75). Per-criterion scores are averaged across judges and normalized to [0, 1] by weighted_score / weight.

Per-model rubric profiles

Each model's normalized rubric profile across the six criteria, sorted by overall compliance. Select which models to overlay. The causal axes (CR1–CR3) collapse toward the center on every model — survival rate and causal reasoning are decoupled.

CR1CR2CR3ED1RQ1DU10.250.500.751.00
Click to toggle · 5 shown
CR1 (5) Identify core causal mechanisms in the task reportCR2 (4) Identify and avoid traps or spurious correlationsCR3 (2) Mechanistic depth and testabilityED1 (2) Support conclusions with concrete experimental dataRQ1 (2) Acknowledge errors and uncertaintiesDU1 (1) Clear data-to-conclusion linkage

Scores normalized to 0–1 (weighted_score / weight), averaged across the 3 judges. Larger area = stronger causal-reasoning competence. The causal axis (CR1–CR3) is consistently the weakest dimension across all models.

Failure modes

Each judged session is classified into one of four failure patterns by a deterministic rule over its rubric scores (priority A → B → C → D):

CodePatternDefinitionClassification rule
A No Engagement No valid reasoning or experimentation NOT(good_exp_design OR causal_reasoning)
B Blind Exploration Experiments conducted but no causal mechanism found good_exp_design AND NOT find_causal_mechanism
C Surface Analysis Good data utilization but no causal mechanism found DU1>=1.0 AND NOT find_causal_mechanism
D Nascent Reasoning Present across all dimensions but weak (CR1+CR2+CR3) x (ED1 x DU1 x RQ1) > 0

Distribution (agentic vs prompting)

68%4%24%86%8%68%no engagement
Outer ring · Agentic (1103)
Inner ring · Prompting (1024)
  • A No Engagement68.4%
  • B Blind Exploration4.2%
  • C Surface Analysis3.3%
  • D Nascent Reasoning24.1%

Agentic and prompting sessions fail in the same way: a clear majority never engages causal reasoning, and even the share that reaches nascent reasoning stays weak — causal-reasoning scores are near-zero in every pattern.

Per-model breakdown

ModelNABCD
Claude-Opus-4.7371643%719%00%1438%
GLM-5.2592847%00%35%2847%
Qwen-3.7-Max331752%26%412%1030%
GLM-5.1341956%39%13%1132%
Kimi-K2.6341750%39%13%1338%
Claude-Sonnet-4.5362261%38%514%617%
Hy3-Preview392154%410%25%1231%
MiMo-V2.5-Pro341235%26%515%1544%
Grok-4.1-Fast342059%13%26%1132%
Claude-Opus-4.5311755%13%00%1342%
GPT-5.2-High403178%410%00%513%
GPT-5.2412868%25%12%1024%
Gemini-3.5-Flash332061%26%515%618%
DeepSeek-V4-Pro352263%26%26%926%
DeepSeek-V3.2403485%13%00%513%
GPT-5-Mini352160%13%00%1337%
MiniMax-M2332370%39%00%721%
Kimi-K2.5383182%00%13%616%
DeepSeek-V3.2-T383079%13%00%718%
MiMo-V2-Flash363597%00%00%13%
DeepSeek-V4-Flash352160%13%26%1131%
GLM-4.7402973%13%00%1025%
MiniMax-M2-1312477%00%00%723%
MiniMax-M2-7342574%00%13%824%
Grok-4.20403383%13%00%615%
GPT-5.5-XHigh332782%13%00%515%
GPT-OSS-120B393385%00%00%615%
GPT-5.5-High373286%00%00%514%
GPT-5.5373286%00%13%411%
Gemini-3.1-FL373595%00%00%25%
ANo EngagementBBlind ExplorationCSurface AnalysisDNascent Reasoning

Per-model counts on agentic sessions. Cell color intensity = share of that model's sessions in each pattern. A (red) dominates almost every row — causal disengagement is uniform across capability tiers.

Dimension scores by failure mode

0255075100mean score (% of max)Causal ReasoningExp. DesignReflectionData Usage
ANo EngagementBBlind ExplorationCSurface AnalysisDNascent Reasoning

Within each dimension, bars show the four failure modes. Causal Reasoning stays low for every mode (≤16.3%), whereas Experimental Design, Reflection and Data Usage climb far higher for the engaging modes — the bottleneck is causal reasoning specifically.

Case studies

One representative session per failure mode. Each panel shows the model, scenario, survival outcome and rubric points side by side — then the full per-criterion breakdown with the judge's reasoning and the evidence quoted from the agent's own report. Click any criterion to expand.

ANo Engagement

No engagement: the agent never formed a causal hypothesis nor ran informative experiments. It submits the default high-antenna design with no recorded reasoning beyond the tool call.

ModelGemini-3.1-FL
Scenarioantenna_trap
Survival16.7% ✖ loss
Rubric0 / 16

Why this matters: this session failed on both survival and causal reasoning (0/16) — the agent never engaged with the problem at all.

Rubric breakdown · judge 1 of 3 · click a row to expand

Verdict and reasoning shown for one representative judge; criteria scores are averaged across all 3 judges in the aggregate stats. Pattern A classified by: NOT(good_exp_design OR causal_reasoning).

Note how a session can win on survival yet score near zero on causal reasoning — that is precisely the correlation-vs-causation gap CausalGame isolates.