WHAT
Traces tell you what the agent did.
Ordered chain of LLM + tool calls with timing, cost, and tokens. The diary of the run — the same story every observability tool is happy to tell.
Don't just trace what happened.
Understand why.
Open-source explainable observability for AI agent systems. Causal attribution + constitutional governance + visual debugging.
Works with the stack you already run · every chip opens its guide
The gap · Two lenses · One platform
“What happened” is easy. “Why” is the bar.
Same trace data, two perspectives. Hover either lens to focus it; the components below are the real /traces and /sankey views, not screenshots.
WHY
AuditTrail explains why it did it.
Causal attribution over the prompt. Every decision has a trail back to the words that drove it, the rule that caught it, and the edit that would have flipped it.
Side-by-side
Debugging, before and after.
Same failure mode. Same agent. Drag to compare the investigation loop without AuditTrail (left) vs. with it (right).
[ERROR] 2026-04-19T14:22:08Z customer-support.agent step=synth
[INFO ] parent_run_id=a3f2… span_id=b1… tokens_in=1284
[INFO ] child_run a3f2…b2 type=tool name=search
[INFO ] child_run a3f2…b3 type=tool name=calculator
[WARN ] confidence=0.42 below_threshold=0.6 skip_chain
[INFO ] child_run a3f2…b4 type=llm name=synth model=gpt-4o
[INFO ] retries: 0 1 2 3 …
Exception: confidence_gate_failed — agent bailed after 3 retries
[DEBUG] full trace 6.2 MB — grep if you can
3 terminals open · you're diffing JSON by eye
without AuditTrail
Drag the slider to compare. Mobile: swipe horizontally.
Compare platforms.
Feature-by-feature against the OSS + SaaS LLM observability pack. Uncertain cells are shown as partial; every claim is cross-checked against published docs. Hover any cell icon for footnotes.
AuditTrail 0/20
LangSmith 0/20
Langfuse 0/20
OPIK 0/20
Helicone 0/20
| Feature | AuditTrail | LangSmith | Langfuse | OPIK | Helicone |
|---|---|---|---|---|---|
| Observability4 features· pinned | |||||
| Trace capture | |||||
| Interactive DAG viewer | |||||
| Sankey flow attribution | |||||
| Real-time streaming | |||||
| Explainability (XAI)4 features· pinned | |||||
| Causal attribution (SHAP + ablation) | |||||
| Counterfactual explanations | |||||
| Mechanistic XAI (SAE features) | |||||
| Natural language explanations | |||||
| Governance4 features· pinned | |||||
| Constitutional rule engine | |||||
| EU AI Act compliance mode | |||||
| REGO / OPA policy engine | |||||
| PDF audit export | |||||
| Operations4 features· pinned | |||||
| Live Fleet dashboard | |||||
| Operations Assistant chatbot | |||||
| 3-tier deployment actions | |||||
| OpenAI-compatible gateway | |||||
| Infrastructure4 features· pinned | |||||
| Self-hostable | |||||
| Open-source license | |||||
| 6-language first-party SDK | |||||
| OTel OTLP ingest | |||||
supported partial TBD — research pending not supported· hover a group to peek · click to pin · hover any row or footnote icon for details
The full tour.
Every surface the product ships — observe, explain, operate, govern, integrate. Screenshots captured against the live dashboard.
See every agent, every span, live.
auditrail.imaginaerium.in/overview
01/04 Dashboard Overview
Everything you need to understand your agents —
observe, explain, and operate.
Causal attribution, constitutional governance, live fleet ops, and a BYOK assistant chatbot — all on one canvas.
Every agent, live
Watch your fleet tick in real time — per-agent error rate, p95 latency, violation rate, and one-click jump to the most recent trace.
Chat with your fleet
BYOK chatbot grounded on your traces. Bring your own OpenAI or Anthropic key — your data stays put.
Right-click to operate
Throttle nodes, swap models, file incidents — 3-tier safety baked in. Every action audited.
EU AI Act ready
Article 12/13/50 logs, 180-day retention floor, PDF regulator export, incident reporting.
Drop-in proxy
Point any OpenAI-compatible client at our gateway. Zero SDK changes. Full traces.
Paged in 60s
Cost, latency, violations, errors, traffic — fan out to Slack, email, or PagerDuty.
Mechanistic XAI · v2.0+
See the concepts the model was actually thinking about.
When your agent runs on a supported open-source model we attach a sparse autoencoder trained by SAELens and surface the top-activated features per span. Behavioural features, safety features, chain-of-thought cues — the interpretable units modern mech-interp research has learned to find.
- Works with Llama 3.x, Gemma 2/3, Mistral Small.
- API models (GPT, Claude) fall back to ablation + counterfactuals.
- Feature labels come from Neuronpedia dictionaries when available.
SAE · layer 15 · top-7
Preview- f_15_4109 · refusal / policy-adjacent0.92
Active near SYSTEM block discussing allowed topics.
- f_15_2774 · user wants calculation0.78
Fires on the literal numeric tokens in the user turn.
- f_15_9210 · recency / time-bounded query0.63
Fires on phrases like "today" and "latest".
- f_15_1044 · chain-of-thought cue0.55
Rises inside `<thinking>` style wrappers.
- f_15_7702 · JSON output scaffolding0.47
Precedes tool-argument emission.
- f_15_3388 · cite / attribute source0.34
Fires after retrieved-context block.
- f_15_5821 · instructions override0.22
Partial — the model is weighing a user-prompt nudge.
FAQ
Your operator questions, answered.
- Yes — the full platform is Apache 2.0 and runs on a single Docker Compose node (or a k8s cluster via the Helm chart in /deploy). No feature gating. A managed cloud tier is on the roadmap for teams that don't want to run their own infra.
Ship agents that explain themselves.
Self-host the whole platform in minutes, or walk the live demo right now — no signup.