> ## Documentation Index
> Fetch the complete documentation index at: https://docs.cf0.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Observability

> Every Lab turn is fully traced — system prompts, tools, model I/O, rendered output. Visible to engineers and, increasingly, to the analyst in-product.

Every Lab turn is fully traced. When you ask a question, cf0 captures the system prompt, the tools called, the model inputs and outputs, the tool execution results, and what was rendered. Traces surface inside the product via the in-app Trace panel and export with the audit trail.

```mermaid theme={null}
%%{init: {"theme": "base", "themeVariables": {"primaryColor": "#1d4ed8", "primaryTextColor": "#ffffff", "primaryBorderColor": "#172554", "secondaryColor": "#ffffff", "tertiaryColor": "#fafafa", "lineColor": "#52525b", "textColor": "#18181b", "mainBkg": "#1d4ed8", "secondBkg": "#ffffff", "clusterBkg": "transparent", "clusterBorder": "#e4e4e7", "edgeLabelBackground": "#ffffff", "labelTextColor": "#18181b", "actorBkg": "#1d4ed8", "actorBorder": "#172554", "actorTextColor": "#ffffff", "actorLineColor": "#52525b", "signalColor": "#52525b", "signalTextColor": "#18181b", "labelBoxBkgColor": "#1d4ed8", "labelBoxBorderColor": "#172554", "noteBkgColor": "#fafafa", "noteTextColor": "#18181b", "noteBorderColor": "#e4e4e7", "messageTextColor": "#18181b", "sequenceNumberColor": "#ffffff", "fontFamily": "Inter Tight, -apple-system, sans-serif", "fontSize": "13px"}}}%%
sequenceDiagram
  autonumber
  participant U as Analyst
  participant L as Lab
  participant M as AI model
  participant T as Tools<br/>(filings · templates · charts)
  participant F as Trace store
  U->>L: Plain-English question
  L->>M: System prompt + tool defs
  M->>T: Tool call (filings.read, consensus.get…)
  T-->>M: Structured result
  M-->>L: Streamed answer + render calls
  L-->>U: Chart · table · cited prose
  L->>F: Full trace (prompt, I/O, latency, bytes)
```

## What's traced, per turn

* **System prompt** — the framing cf0 gives the model
* **Tool definitions** — which tools were available
* **Tool calls** — which tools the model invoked, with arguments
* **Tool results** — what each call returned, with byte size for context-bloat detection
* **Model inputs and outputs** — every token in, every token out
* **Render output** — what was streamed to the chat UI

## Per-agent metrics

cf0 tracks per-agent token usage, tool latency, and rate-limit events. Optimisation decisions are data-driven, not anecdotal.

## What a trace looks like

A single Lab turn captures the tools the model could see, the calls it made, and the bytes each call returned. Conceptually, for a DCF turn:

| Captured          | Example                                                                                      |
| ----------------- | -------------------------------------------------------------------------------------------- |
| Thread            | `nvda-q3-readthrough` · turn 4                                                               |
| Tools called      | Filings (10-K, Item 7 MD\&A) · Consensus (FY26–FY28 revenue) · DCF template (3-stage growth) |
| Output rendered   | Bear / Base / Bull fair-value cards · Sensitivity tornado                                    |
| Captured per call | Latency, byte size, input arguments, result summary                                          |
| Audit chain       | Every figure on screen walks back to one specific tool call                                  |

**Every figure that ended up on screen can be walked back to one specific tool call** — the MD\&A passage to the filings call, the forward projections to the consensus call, the fair-value math to the DCF template. That's the audit chain.

## What you can do with it

* **As an analyst** — open the Trace panel from any thread to see exactly what the model read before it answered.
* **As an org admin** — review per-member token usage and tool patterns on the [Dashboard](/workspace/dashboard).
* **In an audit** — pair traces with the [audit trail export](/security/citations-and-audit) for a turn-by-turn record of how an output was produced.