OBSERVABILITY PLATFORM · Profile revision 2
LangSmith
An observability and evaluation platform with an explicit run/trace/thread model, rich execution payloads, export APIs, distributed context, and capture controls.
Why it is in the map
LangSmith documents one of the clearest vendor run schemas and makes its separation from LangGraph runtime persistence visible.
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
DocumentedRun, trace, project, thread/session, and parent-child fields correlate nested executions.
Tracevity interpretation
Execution correlation is strong, but application and agent-definition version are not mandatory top-level identity fields.
Limit
Stable agent instance and agent-definition version depend on application metadata and integration behavior.
Primary evidence (1)
- LangChain: Run data format Run fields and trace hierarchy
Principal and delegation
Partially documentedAn application can record principal-like metadata, but LangSmith does not standardize a verified principal or delegation chain.
Tracevity interpretation
A user value in metadata is attribution supplied by the application, not proof of authentication or authority.
Limit
Authenticated principal, delegated identity, credential type/scope/expiry, and elevation are not canonical.
Primary evidence (1)
- LangChain: Run data format Run metadata
Instruction and context
DocumentedInputs, messages, retriever runs, outputs, events, and metadata can preserve instruction and context evidence.
Tracevity interpretation
Field availability does not prove complete capture or authentic source/version provenance.
Limit
A universal context fingerprint, retrieval-source version, and explicit missing-context representation are not defined.
Primary evidence (1)
- LangChain: Run data format Inputs, outputs, events, and run types
Decision artifacts
Partially documentedDocumented generic fields can retain evaluator output or application-defined plans and decisions.
Tracevity interpretation
Stored text or evaluator output is an observable artifact, not hidden reasoning or objective motivation.
Limit
Plan, selected alternative, policy result, calibrated confidence, and refusal reason are not universal typed fields.
Primary evidence (1)
- LangChain: Run data format Run types, events, and extra data
Model activity
DocumentedModel requests/responses, usage, cost, latency, status, and errors are representable in run data.
Tracevity interpretation
Actual completeness and exact provider model version depend on the integration that emits the run.
Limit
Provider/model-version and cache fields are not guaranteed uniformly across integrations.
Primary evidence (1)
- LangChain: Run data format LLM run data, token and timing fields
Tool and MCP activity
DocumentedTool identity, arguments/results, error, timing, and parent-child correlation can be recorded.
Tracevity interpretation
A tool run does not establish the tool's authority or the downstream system's resulting state.
Limit
MCP server identity, credential/authority, retries, and downstream provider correlation are not universal.
Primary evidence (1)
- LangChain: Run data format Tool run type
Effects
Not documentedLangSmith records tool execution evidence but not a universal staged external-effect model.
Tracevity interpretation
Tool success cannot be promoted to provider acceptance, settlement, or independently verified state.
Limit
Effect type, requested effect, provider acceptance, external effect state, and state delta are not canonical.
Primary evidence (1)
- LangChain: Run data format Tool run type and status fields
Human control
Partially documentedHuman review can be recorded as feedback or annotation.
Tracevity interpretation
Post-run review is not equivalent to runtime authorization or intervention.
Limit
Approval, rejection, edit, interruption, takeover, rollback request, reviewer identity, and authority are not universal fields.
Primary evidence (1)
- LangChain: Observability concepts Feedback and annotation workflows
Outcome
Partially documentedExecution-reported success/failure and evaluated output can be reconstructed from retained runs.
Tracevity interpretation
Run status is not externally verified destination success.
Limit
Partial/ambiguous settlement, rollback, compensation, and independently verified success are not canonical outcomes.
Primary evidence (1)
- LangChain: Run data format Status, error, outputs, and events
Trace integrity
Not documentedStored run data is mutable lifecycle-managed telemetry, not a documented immutable evidence ledger.
Tracevity interpretation
Timestamps and trace hierarchy provide correlation, not integrity guarantees.
Limit
Clock provenance, immutability, append-only behavior, signature, tamper evidence, and independent verification are absent.
Primary evidence (1)
- LangChain: Data purging and compliance Purge and deletion controls
Portability
DocumentedNative export and OTel/hybrid capture provide material portability surfaces.
Tracevity interpretation
Export availability does not establish lossless conversion of every vendor run type and field.
Limit
Schema version, import destinations, documented conversion, and exact semantic loss remain incomplete or unknown.
Primary evidence (1)
- LangChain: Bulk export traces Bulk trace export fields
Privacy
ConflictingCurrent official documentation still conflicts between a blanket 400-day SaaS statement and tiered 14-day base / 400-day extended retention.
Tracevity interpretation
Tracevity preserves the conflict. Plan, trace tier, promotion, expiry, and manual deletion workflows must be resolved for the concrete deployment.
Limit
The reviewed pages do not provide one safely universal SaaS retention duration; manual-deletion scheduling does not resolve the base-versus-extended conflict.
Primary evidence (4)
- LangChain: LangSmith observability concepts Data retention
- LangChain: LangSmith administration overview Data retention
- LangChain: LangSmith billing and retention tiers Trace tiers and retention
- LangChain: LangSmith data purging and compliance Retention expiry and manual deletion
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: Run, trace, project, thread/session, and parent-child fields correlate nested executions.
- Principal: An application can record principal-like metadata, but LangSmith does not standardize a verified principal or delegation chain.
- Context: Inputs, messages, retriever runs, outputs, events, and metadata can preserve instruction and context evidence.
- Decisions: Documented generic fields can retain evaluator output or application-defined plans and decisions.
- Models: Model requests/responses, usage, cost, latency, status, and errors are representable in run data.
Explicit gaps or uncertainty
- Effects: Effect type, requested effect, provider 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.
- Privacy: The reviewed pages do not provide one safely universal SaaS retention duration; manual-deletion scheduling does not resolve the base-versus-extended conflict.
Source register
12 reviewed primary sources
- Run data formatLangChain · OFFICIAL DOCUMENTATION · observed August 28, 2026 · currentOpen source ↗
- Observability conceptsLangChain · OFFICIAL DOCUMENTATION · observed August 28, 2026 · disputedOpen source ↗
- Bulk export tracesLangChain · OFFICIAL DOCUMENTATION · observed August 28, 2026 · currentOpen source ↗
- Distributed tracingLangChain · OFFICIAL DOCUMENTATION · observed August 28, 2026 · currentOpen source ↗
- Log traces to a projectLangChain · OFFICIAL DOCUMENTATION · observed August 28, 2026 · currentOpen source ↗
- Mask sensitive dataLangChain · OFFICIAL DOCUMENTATION · observed August 28, 2026 · currentOpen source ↗
- Administration overviewLangChain · OFFICIAL POLICY · observed August 28, 2026 · disputedOpen source ↗
- Data purging and complianceLangChain · OFFICIAL POLICY · observed August 28, 2026 · currentOpen source ↗
- LangSmith observability conceptsLangChain · OFFICIAL DOCUMENTATION · observed August 29, 2026 · disputedOpen source ↗
- LangSmith administration overviewLangChain · OFFICIAL DOCUMENTATION · observed August 29, 2026 · disputedOpen source ↗
- LangSmith billing and retention tiersLangChain · OFFICIAL DOCUMENTATION · observed August 29, 2026 · disputedOpen source ↗
- LangSmith data purging and complianceLangChain · OFFICIAL DOCUMENTATION · observed August 29, 2026 · disputedOpen 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