Spans
A span is a named, typed interval inside a run: a rollout, a turn, a tool call, an epoch, an evaluation pass. Spans nest, so a trajectory is a tree rather than a log.span_id as an exemplar pointer — the span the value was produced under — which is how a reward on a chart links back to the rollout that earned it.
A run bundle carries span type counts only, not the spans themselves. To read the tree, use
probe span list, or the MCP trajectory view.Units and ambient context
Inside the SDK, a unit binds coordinates and labels for everything logged within it. Nested units merge, with the child winning per key.Harbor trials
A trial is one rollout’s on-disk output: the trajectory, the result, the files it produced.probe trial captures a Harbor trial directory into a run, keyed to the training step — the sandbox-to-step join.
kind=harbor_trial manifest — all keyed by --step. A recognised trajectory format (ATIF v1.x is built in) also expands into turn and tool_call spans under the rollout span.
Unknown formats are captured raw, never rejected, and can be expanded later:Deterministic span ids make this idempotent — re-running upserts and never duplicates. Fork parsers plug in with
probe.connectors.atif.register_trajectory_parser("their-format", fn).Staging and offline capture
When the network is not available at the moment the trial finishes, stage it first and deliver later..probe-capture.json beside the durable bytes. Its collection status is tracked separately from upload status, so an exporter outage leaves a precise, retryable list of unconfirmed files rather than losing their paths.
Stable external keys and span ids make retries update the same rollout. Correlation fields from Miles and Osmosis are preserved under the manifest’s source.context, and sample_id / group_id are promoted to the Probe point labels sample and group — which is what keeps distinct same-step samples apart without minting a separate metric series for each.
Querying trials back
What the completeness claim covers
The ledger’s claim is deliberately bounded: it covers declared regular files in the host Harbor trial directory. Public Harbor tears the sandbox down beforeTrial.run() returns, so a post-run consumer cannot know about undeclared state Harbor never materialised. That state is reported as unknown; explicitly declared missing files are inventoried; hidden files and symlinks are recorded as visible skips.
A true pre-teardown guarantee requires the producer or environment implementation to invoke durable collection from its own lifecycle hook.
Trajectories are Harbor-first by decision. Today they enter Probe through Harbor’s on-disk contract: an ATIF-supporting agent writes
trajectory.json into its logs directory, the harness delivers it at <trial_dir>/agent/trajectory.json, and capture picks it up from there. Emission is per-agent opt-in upstream, so absence is a normal, captured state — not a failure.Live SDK span streaming from instrumented agent code is planned as the second door. The server-side turn and tool-call rails already exist, so it is an SDK-instrumentation arc rather than a schema change.