OBSERVABILITY PLATFORM · Profile revision 1
Braintrust
An observability and evaluation platform with nested application spans, OTel ingestion, raw JSON, query export, scoring, masking, and separated data-plane options.
Why it is in the map
Braintrust contributes rich execution/evaluation records, export surfaces, and concrete hosted or customer-cloud data custody controls.
Tracevity boundary
A documented field can support reconstruction, but its presence alone does not prove completeness, authenticity, authorization, or external settlement.
Profile findings
What the documentation can—and cannot—establish
Each row distinguishes source evidence from Tracevity's bounded interpretation. “Not documented” describes the reviewed sources; it does not prove an implementation cannot record the artifact.
System identity
DocumentedTrace, root, span, parent, project, and metadata fields correlate nested activity.
Tracevity interpretation
This is strong execution identity but not a guaranteed stable agent instance or versioned agent definition.
Limit
Agent instance, definition version, and runtime version are application/integration supplied rather than mandatory.
Primary evidence (1)
- Braintrust: Advanced tracing Span IDs, root span, parents, attributes, and metadata
Principal and delegation
UnknownPrincipal-like metadata is possible but no standard verified principal or delegation model is documented.
Tracevity interpretation
Application metadata must not be treated as authenticated authority evidence without another source.
Limit
Initiator, authenticated principal, delegation chain, credential scope/expiry, and elevation are unknown.
Primary evidence (1)
- Braintrust: Advanced tracing Metadata and context
Instruction and context
DocumentedPrompts, messages, expected data, retrieval context, and other inputs can be retained.
Tracevity interpretation
Recorded fields do not prove context completeness or source authenticity.
Limit
Source identifiers/versions, context fingerprint, and explicit missing-context representation are not universal.
Primary evidence (1)
- Braintrust: Advanced tracing Input, output, expected, and metadata fields
Decision artifacts
Partially documentedEvaluator outputs and explicit assessment artifacts are representable.
Tracevity interpretation
A score or explanation is a recorded evaluation, not hidden reasoning or proof of motive.
Limit
Plan, policy result, selected alternative, calibrated confidence, and refusal reason are not universal fields.
Primary evidence (1)
- Braintrust: Examine traces Score spans and trace evaluation
Model activity
DocumentedCore model request/response, identity, token/cache, cost, latency, and error evidence is representable.
Tracevity interpretation
Exact model version and completeness depend on the emitting integration.
Limit
Provider build/version and raw content completeness are not uniformly guaranteed.
Primary evidence (1)
- Braintrust: Examine traces LLM span details
Tool and MCP activity
DocumentedTool identity, arguments/results, error, latency, and nested correlation are representable.
Tracevity interpretation
Tool telemetry does not prove credential authority or destination settlement.
Limit
MCP server, retry, credential context, and downstream provider correlation are not universal.
Primary evidence (1)
- Braintrust: Examine traces Function and tool span types
Effects
Not documentedBraintrust records action telemetry but not a universal external-effect chain.
Tracevity interpretation
A successful function/tool span cannot establish external state change.
Limit
Effect type, requested effect, acceptance, external effect state, and state delta are not canonical.
Primary evidence (1)
- Braintrust: Examine traces Tool and function span outputs
Human control
Partially documentedHuman evaluation evidence can be retained with a trace.
Tracevity interpretation
Post-run review is not necessarily pre-action authorization.
Limit
Approval, rejection, edit, interruption, takeover, rollback request, reviewer authority, and timing are not universal.
Primary evidence (1)
- Braintrust: Examine traces Scores and trace review
Outcome
Partially documentedReported success/failure and evaluation evidence can be reconstructed.
Tracevity interpretation
These records do not independently verify a downstream result.
Limit
Externally verified success, ambiguous/partial settlement, rollback, and compensation are not canonical.
Primary evidence (1)
- Braintrust: Examine traces Trace outputs, errors, metrics, and scores
Trace integrity
Not documentedBraintrust traces are mutable lifecycle-managed records, not a documented append-only ledger.
Tracevity interpretation
Trace hierarchy and timestamps do not provide tamper evidence or independent verifiability.
Limit
Clock provenance, immutability, append-only behavior, signature, tamper evidence, and independent verification are absent.
Primary evidence (1)
- Braintrust: Advanced tracing Span ID override behavior
Portability
Partially documentedNative query export and OTel ingestion provide material portability surfaces.
Tracevity interpretation
Asymmetric ingestion/export does not establish outbound OTLP or lossless round-tripping.
Limit
Stable schema version, outbound OTLP, import destinations, conversion contract, and exact semantic loss remain unknown.
Primary evidence (1)
- Braintrust: API reference BTQL export formats
Privacy
DocumentedMasking, custody choice, hosted retention, and deletion controls are documented for sensitive trace payloads.
Tracevity interpretation
Customer-hosted custody does not itself establish redaction, integrity, or minimized capture.
Limit
Mask coverage, screenshots, hashing, customer defaults, and enterprise-specific retention remain deployment dependent.
Primary evidence (1)
- Braintrust: Security Control plane and data plane
Reconstruction reading
Do not collapse these findings into one score.
This profile describes documented evidence surfaces. Whether they are sufficient depends on the reconstruction question and the other identity, authorization, tool, and destination records available.
Potentially useful evidence
- Identity: Trace, root, span, parent, project, and metadata fields correlate nested activity.
- Context: Prompts, messages, expected data, retrieval context, and other inputs can be retained.
- Decisions: Evaluator outputs and explicit assessment artifacts are representable.
- Models: Core model request/response, identity, token/cache, cost, latency, and error evidence is representable.
- Tools: Tool identity, arguments/results, error, latency, and nested correlation are representable.
Explicit gaps or uncertainty
- Principal: Initiator, authenticated principal, delegation chain, credential scope/expiry, and elevation are unknown.
- Effects: Effect type, requested effect, acceptance, external effect state, and state delta are not canonical.
- Integrity: Clock provenance, immutability, append-only behavior, signature, tamper evidence, and independent verification are absent.
Source register
7 reviewed primary sources
- Examine tracesBraintrust · OFFICIAL DOCUMENTATION · observed August 28, 2026 · currentOpen source ↗
- Advanced tracingBraintrust · OFFICIAL DOCUMENTATION · observed August 28, 2026 · currentOpen source ↗
- OpenTelemetry integrationBraintrust · OFFICIAL DOCUMENTATION · observed August 28, 2026 · currentOpen source ↗
- API referenceBraintrust · OFFICIAL DOCUMENTATION · observed August 28, 2026 · currentOpen source ↗
- SecurityBraintrust · OFFICIAL POLICY · observed August 28, 2026 · currentOpen source ↗
- Plans and limitsBraintrust · OFFICIAL POLICY · observed August 28, 2026 · currentOpen source ↗
- View logsBraintrust · OFFICIAL DOCUMENTATION · observed August 28, 2026 · currentOpen source ↗
Next question
What evidence does your use case require?
Move from documented field presence to an explicit reconstruction target, or examine selected directional format mappings without changing this system's documentation posture.
Trace Reconstruction Requirements Explore compatibility evidence