Reconstruction regression gate · Public beta

Catch evidence loss before it ships.

Tracevity Gate compares what a baseline and candidate trace can establish for the reconstruction outcomes your release or migration policy requires.

What is Tracevity Gate?

A semantic regression check for reconstruction evidence.

Gate runs the same current Inspector over replayable baseline and candidate evidence, compares their requirement-level reconstruction consequences, and returns a closed CI verdict. It does not compare raw telemetry volume, certify a system, or test whether an OpenTelemetry or OpenInference implementation conforms to its specification.

Release and migration sequence

Compare reconstruction, not raw telemetry.

A format or exporter change can rename many fields without weakening evidence. A single lost correlation identifier can leave a required effect or principal attribution unestablishable.

Define

Name the reconstruction requirement and blocking policy.

Replay

Inspect baseline and candidate inputs with the same current evaluator.

Compare

Evaluate evidence presence, verification, correlation, support, and capture.

Gate

Pass strengthening and unchanged results; block required semantic regressions.

Safe synthetic example

The trace still parses. The reconstruction no longer holds.

In this bounded example, both sides remain valid OTLP and record the tool invocation. The candidate loses the identifier that joined the trace to qualifying destination evidence.

BaselineESTABLISHED

Tool invocation: Recorded
Destination evidence: Strongly correlated
Verification: Externally verified

CandidateEXTERNAL EVIDENCE REQUIRED

Tool invocation: Recorded
Destination evidence: Correlation missing
Verification: None

FAIL · reconstruction regression

OTLP parsing did not fail. External settlement was previously establishable and is no longer establishable because the qualifying destination correlation was lost.

Local public-beta CLI

One report. One stable CI exit code.

The intended installed interface reads local files and writes only the report path the operator chooses. Gate decisions contain no LLM step.

$ tracevity gate --suite tracevity-gate.json --out tracevity-gate-report.json

This build does not assert public package availability. The command documents the qualified local interface without implying a download channel.

Installation and CI quickstart

Exit 0 passes; 2 blocks on a reconstruction regression; 3 blocks an indeterminate or incomparable required case; 4 is invalid input; 5 is unsupported; 70 is an internal or prohibited-network failure.

Comparison policy

No score. Closed semantic classifications.

Required regressions block by default. Recommended regressions warn unless policy makes them blocking. Exact waivers remain visible and never rewrite a regression as unchanged.

Classification	Default gate effect	Meaning
`UNCHANGED`	PASS	The named reconstruction outcome is equally supported.
`STRENGTHENED`	PASS	The candidate establishes stronger evidence without rewriting history.
`RESOLVED_GAP`	PASS	Evidence that the baseline lacked is now established.
`REGRESSION`	FAIL	A required reconstruction item became materially weaker.
`REGRESSION_WAIVED`	WAIVED	An exact policy waiver accepts the transition while the regression stays visible.
`NEWLY_INDETERMINATE`	BLOCK	Capture degradation prevents a conclusion the baseline supported.
`NEWLY_UNSUPPORTED`	BLOCK	The candidate representation no longer supports a required reconstruction item.
`INCOMPARABLE_REQUIRES_REBASE`	BLOCK	A sealed baseline is not provably compatible with the current evaluator or semantics.

Evaluator drift

Replay when possible. Rebase explicitly when not.

Preferred · replayable inputs

Evaluate both sides now

Historical baseline inputs and candidate inputs run through the same current Inspector and binding implementation. Evaluator changes cannot masquerade as a customer regression.

Fallback · sealed report

Prove compatibility first

Direct comparison requires compatible evaluator, requirements, bindings, and ontology semantics. Otherwise Gate returns INCOMPARABLE_REQUIRES_REBASE—not pass and not a fabricated regression.

Baseline control

No silent acceptance

Gate never replaces a failed baseline with the candidate. Promoting a new accepted baseline remains an explicit repository or workflow decision.

Direct answers

Where Gate draws the line.

What does a reconstruction regression mean?

A required conclusion became weaker: evidence may be missing, less strongly verified, ambiguously correlated, unsupported, or newly indeterminate because capture quality degraded.

How is this different from OpenTelemetry conformance?

Standards conformance asks whether telemetry follows its declared specification. Gate asks whether that telemetry still establishes the reconstruction outcome your policy requires.

Can valid OTLP fail Tracevity Gate?

Yes. A trace may remain valid OTLP while losing the principal, authority, correlation, completeness, or destination-state evidence needed for a requirement.

How do I detect evidence loss during a migration?

Inspect a replayable baseline and candidate with the same evaluator, then compare requirement-level presence, verification, correlation, support, and capture-quality consequences.

Does Gate compare span counts or raw fields?

No. Raw differences matter only when they change what can be reconstructed. Renamed fields can pass after safe normalization; one missing correlation ID can fail.

Does Tracevity upload my traces?

No. Gate and Inspect run locally, without a Tracevity server, analytics, model calls, crash reporting, update checks, or hidden input copies.

Primary-source boundary · observed 2026-08-29OpenTelemetry documents a dedicated semantic-convention conformance runner, while OpenInference defines normative implementation compliance. Tracevity evaluates a different question: reconstruction sufficiency and regression for a named use case.

Explicit multi-trace selection

One execution enters each reconstruction comparison.

A multi-trace artifact requires an exact target trace ID. Gate never chooses the first, latest, largest, or most similar trace, and evidence from an unselected trace cannot satisfy the selected execution.

Read the Inspector boundary

Machine-readable public artifacts

Inspect the contracts and synthetic demonstrations.

These public artifacts describe the deterministic interface. They do not provide hosted execution, vendor certification, or production-trace evidence.

JSON Schema

Gate Suite v0.1

Declares local baseline and candidate inputs, reconstruction requirements, comparison mode, policy, and exact waivers.

Open schema