Skip to content

Core Concepts

Provon turns evidence from real agent runs into reviewable artifacts that support repair, knowledge retention, evaluation, and model improvement.

View as Markdown Open the plain-text version of this page.

The Improvement Loop#

flowchart LR
  agent["Agent"] --> capture["Gateway, OTLP, or transcript sync"]
  capture --> traces["Trace evidence"]
  traces --> findings["Findings"]
  traces --> knowledge["Knowledge"]
  traces --> datasets["Datasets"]
  findings --> repair["Repair workflow"]
  knowledge --> agent
  datasets --> evaluate["Evaluation or fine-tuning"]
  evaluate --> models["Improved model"]
  models --> gateway["AI Gateway"]
  gateway --> agent

Every derived artifact keeps a path back to the evidence that motivated it. This makes improvements reviewable instead of treating model output or automated diagnosis as unquestioned truth.

Traces#

A Trace records one causally connected execution. For an agent, that can include model calls, tool calls, retrieval, handoffs, latency, token usage, cost, errors, and the final response.

Multiple traces can belong to one conversation. Stable conversation identity lets Provon evaluate a multi-turn outcome instead of judging isolated requests.

See Tracing and the Trace model.

Findings#

A Finding is a supported, actionable problem distilled from one or more conversation occurrences. It combines a cause assessment, severity, confidence, remediation direction, and evidence references.

Findings are intended for review and repair. They can be confirmed, dismissed, resolved, or handed off to an issue tracker.

See Findings.

Knowledge#

Knowledge captures reusable facts, constraints, and preferences found in conversations. Projects define Learning Goals that tell Provon what is worth retaining.

Knowledge items remain linked to their source evidence and can be reviewed before being consumed by people, coding agents, or Dataset collection policies.

See Knowledge.

Datasets#

A Dataset owns structured Examples for evaluation and model improvement. Examples can preserve successful behavior, correct failed conversations, learn from user corrections, or teach project knowledge.

Datasets are consumption-neutral: the same reviewed examples can support evaluation, supervised fine-tuning, or export to another system.

See Datasets.

AI Gateway#

The AI Gateway is an OpenAI-compatible model boundary. It resolves model targets, applies routing and governance policy, records provider attempts, and emits trace evidence.

The Gateway is optional. Applications can continue calling providers directly and export OpenTelemetry instead.

See AI Gateway.

AI Models#

The Node deployment can import and serve open-weight models, train project-specific checkpoints from Datasets, and expose running models behind the same Gateway contract used for cloud providers.

Model hosting and fine-tuning are separate from hosting the Provon control plane. See AI Models and Self-hosting.

Next Steps#