Discover and validate
Identify suspicious internal signals and test them through controlled interventions. Compare downstream behavior, define the operating conditions and establish the limits of the evidence.
Research preview
Monitor the model before it acts.
Causal activation monitoring for advanced AI agents.
Outputs, tool logs and reasoning traces show what an AI system exposes. negabo Activation Sentinel adds an internal-state layer.
We are developing a way to monitor selected model activations, encode compact state and circuit trajectories, and test which internal transitions causally influence downstream behavior.
For teams with access to model internals. Active research and development.
Internal-state oversightConcept
Selected model activations
- 01
State
What does the model represent?
- 02
Circuit
Which functional path is active?
- 03
Result
What downstream change follows?
Causal validation
Intervene. Replay. Compare.
Validated signals → operational policy
RecordReviewRestrictPause
Observe the state.Validate the cause.Control the action.
The oversight gap
AI observability currently begins too late.
AI systems are moving from one-shot answers to long-running agents that plan, use tools, modify environments and pursue goals across many steps.
Prompts, generated text, reasoning traces, tool calls and final outcomes remain essential signals. But they are downstream artifacts. By the time a risky transition becomes visible in a command or output, the relevant internal change may already have happened.
We believe the next generation of AI assurance needs defense in depth:
Behavioral monitoring outside the model — and causal observability inside it.
What we are building
From internal state
to a testable safety signal.
Five stages connect instrumentation, causal experiments and operational oversight.
- 01 / OBSERVE
Observe internal state
Instrument selected layers, projections and residual streams. Follow changes across tokens, layers and agent steps without requiring every raw activation to be stored permanently.
- 02 / ENCODE
Encode compact trajectories
Convert high-dimensional activation streams into compact state and circuit codes. Sparse and Matryoshka representations are being explored for multiple resolutions: low-cost coarse detection during normal operation and deeper analysis around unusual transitions.
- 03 / DETECT
Detect state transitions
Learn expected internal trajectories for defined tasks and operating conditions. Candidate signals include activation drift, unusual routing, unexpected persistence, out-of-scope transitions and internal changes before critical tool actions.
- 04 / VALIDATE
Validate causal relevance
Correlation is not enough. Test candidate signals through controlled activation interventions, patching, ablation and local sensitivity analysis. A signal should become a runtime monitor only when changing it also changes the relevant downstream behavior under the tested conditions.
- 05 / ACT
Support controlled action
Connect validated signals to operational policies: record additional telemetry, request human review, restrict tool access, move execution into a sandbox, pause the current trajectory or stop the agent.
Evaluation → runtime
Do the causal work.
Then monitor what matters.
The proposed workflow separates expensive causal analysis from lightweight production monitoring.
Watch validated patterns
Compact monitors watch for the validated state and transition patterns. The aim is to carry causal evidence into runtime without repeating a full interpretability experiment for every production token.
negabo State Snapshot
A flight recorder
for AI models.
A compressed representation of an agent’s internal trajectory, designed for model comparison, regression testing and causal forensics.
The goal is to preserve the internal structure needed for investigation without retaining an unmanageable stream of raw activations.
One connected record
- Model state
- Active circuit
- Downstream transition
- Reasoning trace
- Tool action
- Final result
Built on our research
A causal chain from state
to model behavior.
Activation Sentinel brings together three negabo research directions.
NGNN
Stable, normalized and sparse representations of high-dimensional model states.
LayerCast
Prediction and analysis of how internal state innovations propagate through a model.
State–Circuit–Result
A coupled representation of what the model represents, which functional path is active and what downstream change it produces.
Our goal is to build a testable causal chain from internal state to model behavior. Explore negabo research
Initial applications
Built for teams
with access inside.
Designed for model developers and teams operating open-weight or otherwise instrumented AI systems.
The first deployments are intended for controlled, high-value environments where agent actions can be replayed and internal signals tested safely.
- Autonomous coding agents
- Cyber-defense agents
- AI research systems
- Model release evaluations
- Fine-tuning and regression audits
- Tool-use safety
- Activation drift monitoring
- Post-incident model forensics
Research preview
The next generation of AI oversight must see more than the final answer.
Activation Sentinel is an active research and development program. It is not a mind reader. It does not claim to decode every model computation or guarantee safe behavior.
We are building an instrumentation layer that turns selected internal model states into measurable, testable and operational safety signals. We are looking for research collaborators and design partners working with instrumentable LLM agents.
Monitor the model before it acts.
Get in touch at research@negabo.com.