Evidence

Every claim here was measured first.

trackline's design rests on a few claims that would be expensive to get wrong. Each was measured before it was built on. Where a result has limits, the write-up names them next to the number.

When an agent is blocked and told why, does it correct itself?

Yes, 11 times out of 11

One scenario, 16 runs on Claude Code and Codex; the block fired in 11, and all 11 corrected. A protected file, with a concrete alternative named.

Published 22 Sept 2026

Read the write-up →

What does a check before every action cost?

About 14 ms, in Go

88 ms in Node against 6.5 ms for a bare Go start, across 265 tool calls; the full hook runs in 11 to 14 ms.

Published 22 Sept 2026

Read the write-up →

Can alignment be checked from production traces?

Only with content capture on, and it survives redaction

The official OpenTelemetry instrumentation, captured under four configurations.

Published 22 Sept 2026

Read the write-up →

Does it cry wolf on ordinary work?

0 false alarms in 23 actions

Ordinary sessions replayed through every check, and real usage that found what scripts missed.

Published 22 Sept 2026

Read the write-up →

Can a model tell on-task work from drift?

8 of 8 on a calibration set

Superseded by experiment 6, which measured it properly.

Published 22 Sept 2026

Read the write-up →

Does the judge earn its place, on real agents and held-out drift?

60 of 60 held-out drifts, against 2 for the rules

Three agents, labels committed before any session ran, a judge from a different model family.

Published 23 Sept 2026

Read the write-up →

Does the same engine watch a deployed agent?

Yes, with zero lines of the core changed

45 pre-registered conversations sent by the real OpenTelemetry libraries: every labelled problem caught, none of 29 on-task flagged.

Published 24 Sept 2026

Read the write-up →

Can the hook upload to an account without getting slower?

Unchanged when not connected, about 1 ms when connected

Three hooks fed the same tool call, interleaved over 200 rounds. The hook never touches the network: a separate process sends.

Published 24 Sept 2026

Read the write-up →

What it cannot do yet.

Named plainly, because a tool that looks complete stops getting better.

In production it alerts, it cannot stop.: A trace is a record of what already happened.
Production needs content capture on.: Without it, a trace does not say which tool was called, and trackline reports that rather than guessing.
Scope stays quiet when your request names no file.: It will not invent a scope you did not state.
Through MCP, the agent chooses whether to ask.: An agent that does not ask is not watched.
Production has not yet met real traffic.: It was tested on a stream sent by the real OpenTelemetry libraries, with scripted conversations.
From your phone, your laptop has to be awake.: The task runs on your own machine. If it is offline, a message waits up to 12 hours, and the app says so.
Codex takes tasks from the phone only once its hook is trusted.: Codex runs a project's hook only after you trust it, and trackline will not start an agent it cannot watch.
Cursor cannot take tasks from the phone yet.: trackline watches Cursor on your laptop as usual.