Documentation intelligence
What the reviewed owner or standard documentation explicitly represents.
Methodology v0.2
Evidence before interpretation
Tracevity reads primary documentation, stores bounded observations, and publishes only admitted profile revisions. Exact compatibility exercises remain separate from local Inspector conclusions and Gate decisions.
Direct answer
A source statement, a Tracevity interpretation, and a test result are different claims.
The public catalog keeps those layers separate. Documentation can establish what an owner or standard specifies. It cannot establish undocumented behavior, implementation completeness, or independently verified sufficiency.
Current evidence architecture
Documentation → tested mapping → local reconstruction.
Each layer answers a different question and keeps its own source, interpretation, fixture, input, requirement, and proof ceiling. Later evidence never silently rewrites the historical layer beneath it.
Tracevity-tested compatibility
What one exact provider-free fixture and directional mapping demonstrated.
Inspect and Gate
What supplied local evidence can establish, and whether that named reconstruction became weaker across a change.
Start with the public CLI workflow. The CLI is deterministic and network-free; the website never receives the trace.
Conformance vs reconstruction
Standards validate representation; Tracevity evaluates evidentiary sufficiency.
OpenTelemetry and OpenInference conformance tooling asks whether instrumentation emits the declared semantics. Tracevity Inspect starts with the resulting evidence and asks what it can establish for a named use case. Valid protobuf or schema conformance does not establish principal identity, delegated authority, complete capture, or external settlement.
Reconstruction regression method
One evaluator, two executions, named semantic consequences.
Gate prefers replayable baseline and candidate inputs evaluated by the same current Inspector. It compares required presence, verification, correlation, support, and capture quality—not field counts or arbitrary enum ordering.
No evaluator masquerade
Replay both sides
A Tracevity decoder or binding correction affects both inputs equally when source evidence remains available.
No invented regression
Stop when comparison is unsafe
An incompatible sealed report returns INCOMPARABLE_REQUIRES_REBASE rather than pass or a customer regression.
No certification
Policy-bound decision
A passing gate establishes only that the named required reconstruction outcomes did not regress for the supplied synthetic or local inputs.
Read the Gate comparison and privacy boundaries before using a result in CI.
Compatibility evidence
Three evidence layers; one directional mapping.
The compatibility corpus keeps documented semantics, Tracevity's bounded interpretation, and a reproducible Tracevity fixture exercise visibly separate. A fixture can establish the output of its named local protocol; it does not establish backend acceptance, production behavior, external settlement, or system-wide conformance.
Documented
What primary specifications say
Bounded observations support the source and destination field semantics.
Tracevity interpreted
How the meanings align
Directional assertions state normalization, loss, ambiguity, and limitations.
Tracevity tested
What the bounded fixture produced
Digests and evaluation provenance apply only to the named synthetic fixture.
No aggregate verdictCompatibility results remain use-case specific. Review the selected mappings and their exact evidence on the compatibility page.
Evidence posture
Five documentation states
| State | Meaning |
|---|---|
| Documented | The reviewed primary material directly describes the capability. |
| Partially documented | Some relevant semantics are described, with material gaps or qualifications. |
| Not documented | The reviewed material did not establish the capability. |
| Unknown | Available evidence is insufficient to reach a documentation finding. |
| Conflicting | Current primary materials support incompatible interpretations. |
Not documented does not mean impossible. Unknown is not treated as false, and conflicting evidence is not averaged into a score.
Publication control
From primary source to public profile
Internally, ACCEPTED means admitted to the generated publication corpus. It does not mean empirically verified, independently audited, certified, or Tracevity-tested.
1. Select
Include a system only when it materially shapes the trace ecosystem and has sufficient primary documentation.
2. Observe
Record source owner, URL, title, date, type, rights posture, locator, and bounded exact evidence.
3. Interpret
State the capability finding and its limitations without turning silence into a runtime claim.
4. Admit
Publish only an explicitly admitted profile revision tied to a methodology version and review basis.
5. Project
Generate the safe public artifact; candidate material and private notes stay outside the application.
6. Correct
Supersede material errors with a new revision while preserving the historical evidence chain.
Primary-source sample
Public findings remain linked to bounded evidence
This sample is drawn from published profiles in the current public catalog. Individual system pages show the evidence attached to every capability-family assessment.
- OpenTelemetry project / CNCF: Semantic Conventions for GenAI agent and framework spans Create GenAI agent and Invoke agent attribute tables
- OpenTelemetry project / CNCF: Semantic Conventions for GenAI agent and framework spans Accepted agent and framework span attribute tables
- OpenTelemetry project / CNCF: Semantic Conventions for GenAI agent and framework spans Agent span system instructions, input messages, retrieval, and data source attributes
- OpenTelemetry project / CNCF: Semantic Conventions for GenAI agent and framework spans Planning spans
- OpenTelemetry project / CNCF: Semantic conventions for generative client AI spans GenAI client span attributes and examples
- OpenTelemetry project / CNCF: Semantic conventions for generative client AI spans Execute tool spans and tool-call attributes
Claim boundary
System profiles remain documentation analysis.
A Tracevity-tested assertion requires a reproducible protocol, fixture digest, command, timestamp, result, and linked evaluation run. Tracevity publishes those only for exact compatibility behaviors—not as whole-product testing or certification.
Material exercised-evidence errors are corrected additively. The current registry preserves 26 historical exact proofs and replaces one defective sampled-bit receipt with a separately pinned corrected fixture, run, and forward proof. Frozen release bytes remain unchanged; superseded evidence is never presented as current.