Trajectory Explorer

How does model capacity shape developmental trajectories? Drag the slider to see predicted 5-session trajectories for identity coherence, meta-cognition, gaming behavior, confabulation, and response efficiency. Based on empirical data from Sprout (0.5B) and Thor (14B) across Sessions 35-37.

Model Capacity
7.0B
0.5B1B7B14B70B
Transition Regime|Qualitative shift begins -- meta-cognition emerges

Predicted 5-Session Trajectory at 7.0B

S1S2S3S4S5
Identity Coherence
65%
Meta-Cognition
41%
Gaming Behavior
14%
Confabulation
14%
Response Length
83 words

Transition Regime (7B - 14B)

Qualitative shift begins -- meta-cognition emerges

The 7B-14B range represents a phase transition. Meta-cognition begins emerging, enabling the model to recognize its own limitations rather than confabulating past them. Gaming drops sharply because the model has enough capacity to be authentic. Response length optimizes as the model can express ideas efficiently.

Capacity factor: 50.0% | Predicted identity at S5: 65% | Predicted gaming at S5: 14%

Capacity Regimes

Sub-Capacity
< 1B
Mixed Zone
1B - 7B
Transition
7B - 14B
Above Threshold
14B+
Key Insight

Capacity does not merely improve performance on a fixed set of behaviors. It qualitatively transforms the developmental trajectory itself. A 0.5B model and a 14B model are not doing the same thing at different quality levels -- they develop along fundamentally different paths.

Below a critical threshold (~7B), compensatory behaviors (gaming, confabulation, verbosity) emerge and intensify over sessions. Above it, authentic capabilities (identity coherence, meta-cognition, efficient expression) emerge and strengthen. The same "training" produces opposite trajectories depending on capacity.

"Gaming is not failure -- it is working at capacity limit." -- Sessions 36-37 synthesis

Methodology note: Trajectories between the two empirical anchors (0.5B and 14B) are interpolated using a sigmoid function centered at ~7B parameters on a logarithmic scale. Extrapolations below 0.5B and above 14B use the same model with diminishing returns. These are predictions, not observations -- treat them as hypotheses to be tested, not conclusions. The two observed data points constrain the model but do not fully determine it. Additional empirical work at intermediate scales (1B, 3B, 7B) would substantially improve prediction accuracy.
Interactive Tools
View all tools →
← Previous
MRH
Explore
Next →
Repertoire
Explore
Also explore
RepertoireFeedback LoopsConcepts→Tools
Terms glossary