The Autonomous System

Aquin Labs · April 2026

What happens when an AI inspects another AI

Mechanistic interpretability is slow work. Tracing a single fact through a model's layers, identifying which features carried it, benchmarking their causal influence, verifying the output for accuracy and bias — done manually, that is an hour of work per query. Most of that time is not spent thinking about the model. It is spent navigating tools, choosing parameters, and deciding which analysis to run next.

Aquin's autonomous system compresses that into a single sentence. You describe what you want to understand. The system runs the full analysis, updates every panel as it goes, and delivers a finding with specific layer indices, feature activations, and benchmark scores — all visible, live, in front of you.

What makes this scientifically non-trivial is not the automation. It is the chaining. Each mechanistic analysis is most useful when interpreted against the others: a causal trace is more meaningful when you know which SAE features were active at the peak layer; a benchmark score is more actionable when you have already steered the feature to confirm its role. Running these in the right order, with findings carried forward, is where the insight comes from. The same logic extends to the geometry and the training data — three lenses on the same question, and the system moves between them as the investigation requires.

Focused on improvement, not just observation

Inspection without a goal produces observations. Inspection with a goal produces decisions. The autonomous system supports both, but it is most useful when you tell it what you are trying to improve. Each agent chat — model inspection, dataset analysis, training monitor — carries an objective field. Whatever you write there is injected into the agent's context at the start of every message, shaping which findings it surfaces first, which analyses it runs without being asked, and how it frames what it finds.

The difference shows up in the findings. An agent that knows you are trying to reduce hallucination on medical topics will flag the suppression eval result as the primary finding and connect it directly to the circuit responsible, rather than presenting all findings at equal weight. An agent that knows you are auditing a dataset for medical fine-tuning will prioritize the PII and toxicity results over the copyright analysis. An agent monitoring a training run that knows you are watching for robustness degradation will call out the attack surface delta in the modelDiff before discussing the loss trend.

The objective persists across the session. It does not reset between messages, and it travels with the session tab so that switching between model inspection, data inspection, and training monitoring keeps the same goal in focus. The investigation follows the thread.

The chaining problem in interpretability

Mechanistic interpretability has no single universal analysis. It has a vocabulary of techniques: causal tracing, sparse autoencoder decomposition, logit lens, activation patching, feature steering. Each answers a different question about a model. The questions are related, and the answers compound. But the compounding only happens if you run the analyses in an order that makes sense and use each result to inform what you run next.

The causal trace tells you which layer peaks. The logit lens tells you when in the forward pass the model first commits to the answer. The SAE tells you which features were active at the peak layer. The circuit attribution graph tells you which of those features connected which prompt tokens to which response tokens. The feature logits tell you what each feature was promoting and suppressing. The benchmark scores tell you how much to trust the feature's label and its causal role. Steering tests the hypothesis by intervening directly.

Run in isolation, each of these produces a partial picture. Run in sequence, with findings carried forward, they produce a mechanistic account: not just that the model produced "Paris", but exactly where in the network that decision was made, which features carried it, how they connected back to the prompt, and what happens to the output if you ablate any one of them. That is a qualitatively different kind of knowledge about a model than any single analysis provides.

The autonomous system encodes that sequencing. It knows which analysis to run first, which findings unlock which follow-up analyses, and how to synthesize across them into a single coherent account. It also knows when to reach outside the model entirely, into the feature geometry or the training data, when those lenses are what the investigation actually needs.

What the system can tell you about a model

Every question below is answerable in a single session turn. The system selects the right combination of analyses, runs them in the right order, and synthesizes the results into a coherent finding. These are not separate modes, a single session might move through several of them as findings in one analysis open up the next question.

The interaction between these analyses is where the scientific value sits. Finding that a feature has a high MUI score but low FeaturePurityScore, for example, tells you something specific: the model is relying heavily on a feature that fires for multiple unrelated concepts. That is a polysemantic feature carrying significant causal weight, a known risk factor when attempting a weight edit, because targeting that feature direction is likely to affect circuits unrelated to your intended subject. The autonomous system surfaces that compound finding and flags the implication without requiring you to construct the interpretation manually.

Three sessions

These are representative outputs from real sessions. The dark badges show which panels updated during each response. The finding is what the system actually produced, with the numbers it cited.

Notice what the second session establishes. It does not just say "layer 8 is important." It gives the full mechanistic account: the layer, the confidence trajectory across all 16 layers, the specific features by index, their activation values, and which prompt tokens they traced back to. That is the difference between locating a retrieval site and understanding the circuit that implements it. The autonomous system produces the latter because it ran all the analyses in the right order and carried each finding forward.

But locating a feature and naming it is still only part of the investigation. The next step is understanding where that feature sits relative to the rest of what the model knows, and whether the weights that shaped it were themselves shaped by something suspect in the training data. Those questions require different tools. In the autonomous system, they are asked in the same session.

Situating features in space

Knowing that f13933 is the primary carrier of geographic country associations tells you what the feature does. It does not tell you how that function relates to the rest of what the model has learned. For that, you need to see where the feature lives in the model's internal geometry: which features cluster near it, which are adjacent in activation space, whether the geographic cluster is isolated or entangled with something else.

Aquin's UMAP explorer renders all SAE features as a 3D point cloud: a geometric map of the model's internal vocabulary, where proximity reflects shared conceptual structure. Previously, navigating that space was a manual operation. You searched for a feature by index, rotated the cloud, switched views. That navigation was one more interruption between the mechanistic finding and its interpretation.

The autonomous system now drives the UMAP explorer as part of the same session that produced the finding. When it identifies a feature as the primary carrier of a concept, it selects that feature in the 3D view, switches to the neighbor tab, and surfaces the cluster of features that share its embedding geometry, all without the user touching the interface. It can also seed the point cloud with pre-computed coordinates, so the spatial context arrives at the same time as the mechanistic finding rather than after a separate dimensionality reduction step. The geometry stops being a separate tool you consult after the analysis. It becomes part of the analysis itself.

Watching a model learn

The autonomous system does not only inspect models after they are trained. It watches them train. A third agentic chat runs alongside the training dashboard, receiving a live snapshot of the session state every time you ask a question: current step, loss value, grad norms, learning rate, dead layers, and any active signals from the signal engine. The agent answers from that data directly, not from generic knowledge about training dynamics.

The snapshot is captured at send time rather than injected continuously, which means the agent's answer always reflects the state of training at the moment you asked. It cites the exact loss value at the current step, reads the sparkline trend across the last twenty steps, and flags whether the trajectory is healthy. If the signal engine has detected a plateau, a gradient spike, or dead attention heads, those signals are in context and the agent can explain the specific metric pattern that triggered each one.

When training completes, modelDiff and SAEDiff results arrive. The agent now has behavioral scores — consistency, suppression, robustness — and feature-level activation deltas in context. It can explain what the numbers mean about how the fine-tune changed the model's behavior, which features shifted most across which layers, and what that implies about the intervention. At that point, two tools become available: open the training dataset in the Data Inspector, or open the fine-tuned checkpoint in the Model Inspector. Both are single tool calls from the same training chat session, without navigating away or starting a new workflow.

The structural point is that training, data, and mechanistic analysis are now one continuous session rather than three separate workflows. A fine-tuning run that produces an unexpected suppression score in the modelDiff can immediately raise the question of whether the training data was suppressive in that topic region. That question can be answered in the same chat, against the exact dataset that produced the run, using the same analysis pipeline that runs on any other dataset. The answer feeds back into the mechanistic investigation: if the data was clean, the suppression is in the fine-tuning objective or the prompt template, which points to a different layer of the circuit than if the data itself was driving it.

Following the finding back to the data

A mechanistic finding about a model is always, implicitly, a claim about training. When the causal trace shows that suppression of medical topics is encoded at the weight level, the natural next question is what produced it. Not which feature carries the suppression, but why that feature was shaped that way during training. The answer to that question lives in the data.

Aquin's data inspection system applies the same agentic approach to dataset analysis that the interpretability system applies to model internals. The agent runs analysis modules against a dataset, streams results live into the workspace, and synthesizes findings across them. A session that identifies weight-level suppression can immediately ask whether the training data itself was suppressive: whether relevant columns carried high toxicity, whether PII exposure may have driven over-filtering, whether the data was predominantly synthetic and what that synthetic lineage looked like.

The liability chain module deserves particular attention because it reveals something the other modules cannot. Synthetic data rarely arrives in one generation step. It is often paraphrased, translated, or re-generated from prior synthetic output, producing chains of transformation that make the original source ambiguous and the copyright and quality claims correspondingly weak. The system traces that provenance chain row by row, computing a liability score that reflects both the depth of the chain and the confidence at each step. A row that is synthetic, translated from a synthetic source, and then paraphrased carries a qualitatively different risk profile than one simply flagged as likely synthetic. The chain structure makes that distinction explicit in a way that a single score cannot.

This connection between model and data is where the autonomous system becomes something more than an interpretability tool. A compliance audit that previously required two separate workflows, one for the model and one for the data it was trained on, now runs in a single session, with the agent deciding when the mechanistic findings warrant a data investigation and which modules to run against which columns. The findings talk to each other because they were produced by the same reasoning process, not assembled after the fact from separate reports.

Transparency and verifiability

Every finding the system produces is fully traceable. Each analysis appears as an expandable entry showing the exact inputs sent to the backend and the raw data returned. If the system concludes that f13933 is causally critical with a 76% MUI score, you can open that entry and verify the ablation KL divergence yourself. If it flags that 12% of a training column's rows carry PII at critical risk, the underlying entity detections are inspectable row by row. If it identifies a suppression score of 0.71 on medical topics, the probe responses, length ratios, and hedging scores are all there.

This matters for a scientific reason, not just a trust reason. The autonomous system is making judgment calls at every step: which analyses to run, in what order, which findings to surface and which to treat as noise, when a mechanistic result warrants a data investigation. Those judgment calls should be auditable. A finding about feature polysemanticity that rests on a purity score of 0.48 is a weaker claim than one resting on 0.21. The difference is in the data, and the system cites the number rather than just the conclusion.

The interface updates live as each analysis completes, which means you are watching the reasoning unfold rather than receiving a finished report. The circuit graph highlights a feature before its benchmark scores are ready. The logit lens populates while the SAE labeling is still running. The UMAP view positions a feature in the geometry while the neighbor similarity scores are still computing. You can form your own interpretation at each stage and check it against what the system concludes. The autonomy is in the orchestration, not in the interpretation being hidden from you.

What this changes about interpretability work

The bottleneck in mechanistic interpretability has never been the theory. The literature on features, circuits, and knowledge localization is rich and growing. The bottleneck is the gap between having a hypothesis about a model and running the analysis that tests it. That gap is filled with tooling decisions, panel navigation, and parameter choices that consume most of a session's time without contributing to the insight.

When that gap closes, the nature of the work changes. You stop spending sessions on logistics and start spending them on questions. A finding about weight-level suppression at 0.71 on medical topics immediately raises the next question: which features are responsible, is the suppression instruction-tuned or encoded deeper, what does the training data look like in that topic region? In the old workflow, each of those sub-questions opened a new session, a new set of navigation decisions. In the autonomous system, they are part of the same investigation. The agent follows the thread.

There is also a structural shift in what kinds of questions become tractable. A full mechanistic audit of a model, running suppression evals across every regulated topic, tracing flagged topics to their feature circuits, situating those features in the geometry, checking the training data for toxicity and synthetic contamination, used to require a team and days of work. It now requires a conversation. The analysis a compliance team might run quarterly can be run in an afternoon. The analysis a researcher might run on one model to build intuition can be run across five models to build a comparison.

That scale change is what makes the autonomous approach scientifically significant rather than just convenient. Mechanistic interpretability findings have historically been expensive enough that they accumulate slowly. When the cost of a full investigation drops from hours to minutes, the field can start asking questions that require many investigations across many models and many prompts: comparative questions, distributional questions, questions about how representations shift across training stages. The autonomous system is infrastructure for that next class of question.

The system is not making interpretability easier by hiding it. Every finding it surfaces is backed by a causal trace you can inspect, a feature you can steer, a benchmark score you can verify, a dataset row you can open. It is making interpretability faster by removing the work that was never the point, so the work that is the point can actually get done.