RRememorio Blog

Model internals · Five-part route

From activation to agent action

An agent changed a file, but that does not mean the model directly touched disk. This series separates weights, activations, silent and visible reasoning, tool calls, runtimes, and environment results in time order, then asks which internal evidence supports causal claims.

Reading boundary: internal readouts are instruments, not word-for-word autobiography. Every article separates correlation, causal evidence from intervention, and interpretation of what a model may be doing.

Overview of model activations flowing toward an agent action

Reading route

Build the map, then read out, identify, and intervene

All five articles reuse one retention-policy task so every concept lands on an input, state change, tool call, or file result.

The runtime checks and executes a proposed model tool call01 · OverviewFrom Activation to Agent Action

Follow one file change through model computation, a proposed tool call, and runtime execution, separating activations, KV caching, and durable records.

Read Part I → Concept readout and causal substitution in J-space02 · WorkspaceJ-space and Silent Reasoning

Separate J-lens readouts, sparse reconstruction, and causal intervention through the reference implementation.

Read Part II → CoT, NLA, SAE, and J-lens observation methods beside latent reasoning03 · ReadoutCoT Monitoring, NLA, SAEs, and Latent Reasoning

Compare CoT, NLA, SAE, and J-lens evidence, then trace continuous thought through Coconut.

Read Part III → Evidence boundaries for persona directions, the Assistant axis, and limited introspection04 · IdentityPersona, Self-Model, and Post-Training

Trace persona extraction, the Assistant axis, and one-sided capping to see how post-training stabilizes roles and where self-model evidence stops.

Read Part IV → Separate responsibilities for inference steering, weight training, and alignment auditing05 · InterventionActivation Steering, Reflection Training, and Alignment Auditing

Compare ActAdd and CAA intervention positions, separating inference-time steering, reflection training, and independent audits with their validation and rollback conditions.

Read Part V →