Skip to content

Claim-Boundary Positioning

Status: reference positioning note. Last updated: 2026-05-28. This document does not open a new experiment arc, define a schema, or promote any assay.experiment.* artifact to a product API.

Position

Assay is not an observability replacement, trace viewer, vendor ranking, or general agent dashboard. Assay's strongest post-arc position is a claim-boundary and evidence-fidelity layer for agent systems.

The closed Runner-vs-OTel overhead arc and agent-observability fidelity arc support one shared statement:

Assay helps determine what an agent run proves, not just what it emitted.

The useful boundary is not "Runner versus OTel" or "OpenInference versus OTel GenAI." It is the boundary between reported intent, measured effect, calibration health, join strength, and the claim a reviewer may safely make.

Methodology Anchor

Future agent-observability work should start from the pattern proven by the two closed arcs:

  1. Calibrate before timing. Requested signals must be compared with observed signals before throughput, timing, or absence claims are interpreted.
  2. Respect evidence boundaries. In-process traces, OpenInference or OTel vocabularies, Runner archives, policy evidence, and kernel effects carry different claim surfaces.
  3. Classify claims. A trace/archive mismatch is a measured divergence, not automatically malicious behavior, policy failure, or root cause.
  4. Carry bounded evidence. A portable carrier should make review easier without strengthening the underlying evidence.
  5. Use delegated gates sparingly. Real infrastructure should verify a specific publication gate, not broaden a synthetic experiment by accident.

This is the product-facing form of the experiment lifecycle documented in ../experiments/arc-lifecycle-guide.md.

Public Boundary

This document records public positioning and selection discipline. It does not publish adjacent-whitespace shortlists, outreach targets, comment drafts, competitor analysis, or private sequencing notes.

The public rule is simple: arc closure is a stop condition, not permission to open every adjacent question. A new arc requires a named consumer, upstream response, stable contract ask, or concrete delegated publication gate.

Selection Rule

Open a new arc only when the question can be stated as:

Which claim can we not safely make today because trace, protocol, policy, identity, or runtime evidence is not yet bound tightly enough?

Prefer future arcs that:

  • create or validate a bounded claim class;
  • bind two or more evidence layers that currently drift apart;
  • make absence, partial coverage, clipping, or weak joins first-class;
  • produce a carrier a reviewer can inspect;
  • have a delegated baseline or a concrete downstream consumer.

Defer arcs that:

  • mainly rank tools, vendors, vocabularies, or frameworks;
  • add another dashboard without changing claim strength;
  • expand a synthetic matrix without a new claim boundary;
  • require schema promotion before a consumer exists;
  • duplicate OTel/OpenInference trace viewing instead of binding traces to measured effects.

Source Anchors

Non-Claims

  • This document does not open or prioritize adjacent experiment arcs.
  • This document does not promote evidence packs, calibration verdicts, interop rows, semantic-gap verdicts, or future binding evidence to product APIs.
  • This document does not publish outreach plans, target lists, comment drafts, competitive analysis, or private sequencing notes.
  • This document does not rank trace vocabularies, protocols, Runner, or Assay as products.
  • This document does not claim that every protocol, trace, or trace/kernel mismatch is a security issue.