The Behavior Map — observed dimensions per model · V1

Creative Observations

Rather than reduce a model to a single score, the Creative Observations folder measures each model across six distinct dimensions: Solo Executor, Swarm Worker, Coordinator, Memory Agent, Code Reviewer, Synthesizer. A model that scores 8.4 overall might be a 9.5 Code Reviewer and a 4.0 Coordinator — those are entirely different jobs.
notable (llama3.2:1b)
1
observed op-value
12.1
smollm2:360m value
132.5
suites
4

Fit · Efficiency · observed value

Fit

Capped 0–10

Raw test scores weighted by dimension relevance. A model can excel at Code Reviewer tests but be irrelevant for Coordinator. Fit reflects dimension-targeted ability.

Efficiency

0.5× – 1.4×

Speed multiplier by dimension. Workers benefit most (up to 1.4× at 20+ tok/s). Coordinators up to 1.1×. Memory / Solo / Synthesis efficiency = 1.0.

Observed Value

UNCAPPED · Fit × Efficiency

The product of Fit and Efficiency. Can exceed 10.0. When this occurs, the public site caps display at "10+" with a Notable badge; the raw value is exposed to the API.

Notable observation. llama3.2:1b recorded Fit 8.96 × Efficiency 1.35 = Op Value 12.1 — "Exceptional Swarm Worker." A 1-billion-parameter model exceeded the conventional ceiling because its speed is exceptional for that dimension. This is what the framework was designed to surface — small fast models that conventional observations underrate.

The Behavior Map observed score table

Score
gemma3:4b
8.64
cogito:latest
8.44
llama3.2:1b
8.43
granite4:1b
8.38
phi4-mini:latest
8.27
llama3.2:3b
8.26
dolphin-mistral:latest
8.05
smollm2:360m
5.32
smollm2:135m
5.3

9 entries · ordered by Score as observed in this run · not a ranking

The Creative Observations suites

Code IQ

4 test types · execution-verified

Blast Radius Prediction, Call Chain Tracing, Cross-File Refactoring, Code Review. Top: llama3.2:1b at 9.89.

Memory IQ

4 test types · long-context recall

Conversation Retention, Fact Extraction, Context Compression, Session Handoff. Top: gemma3:4b at 9.16.

Skills IQ

4 scientific-domain tests

Query Generation, Multi-Step Reasoning, Data Interpretation, Hypothesis Generation. Top: granite4:1b at 9.02.

Mashups

4 cross-domain tests

Scientific Code Reviewer, Memory Debugger, Research Agent, Swarm vs Solo. Top: gemma3:4b at 8.83 — with two perfect 10.00 cells.

Models

Families

Companies

Related observations

AI-generated & AI-scored. Every observation on this site is produced and judged by AI, and AI makes mistakes. Treat each result as a recorded observation under a declared protocol, as of its date. Method and marker legend