AI Operations
System operations
Investigations
1
8 steps executed
Model calls
7
0 schema retries
Tokens
39,765
32,244 in / 7,521 out
Estimated cost
US$0.0000
From a price table, not billing
Failed steps
0
100.0% success rate
Replayed steps
7
Served from recordings
Capability performance
Each reusable capability, with the work it did and what it cost.
| Capability | Runs | Failures | Retries | Avg duration | Tokens | Est. cost |
|---|---|---|---|---|---|---|
| Planning | 1 | 0 | 0 | 7ms | 3,067 | — |
| Evidence Extraction | 1 | 0 | 0 | 1.6s | 12,916 | — |
| Risk Analysis | 1 | 0 | 0 | 4ms | 4,877 | — |
| Policy Retrieval | 1 | 0 | 0 | 15ms | 6,391 | — |
| Challenge | 1 | 0 | 0 | 3ms | 4,050 | — |
| Verification | 1 | 0 | 0 | 2ms | 3,719 | — |
| Scoring | 1 | 0 | 0 | — | 0 | — |
| Synthesis | 1 | 0 | 0 | 5ms | 4,745 | — |
Traces record what each step did - capability, tool, retrieval counts, duration, tokens and cost. They deliberately exclude model chain-of-thought, which is neither stored nor displayed.
Human override behaviour
The most honest read on whether the system is useful.
- Findings produced
- 5
- Findings adjusted
- 0
- Override rate
- 0%
- Reviews submitted
- 0
- Reviews changing the rating
- 0
A rate near zero suggests analysts are accepting output without scrutiny. A very high rate suggests the model is not earning its place. Both are failure modes, and neither is visible without storing the AI recommendation alongside the human decision.
Grounding control
How often the anti-hallucination control actually fired.
- Evidence proposed
- 12
- Accepted
- 12
- Rejected
- 0
- Rejection rate
- 0.0%
- Mean grounding score
- 99%
Where findings land
Distribution of findings across the risk taxonomy.
Latest evaluation
Measured system quality, not asserted.
make eval.