Zephr
Start free

Intelligence · Pulse

Measured data,or an empty state.

Pulse is the instrument panel for your evidence graph: what drifted when a branch moved, how fresh each anchor is, which beliefs are actually being recalled, and how much of the graph a human has reviewed. It never fills a gap with an estimate. Every view states what it could not see, so a missing repository reads as missing rather than as healthy.

Freshness + coverage shippedEnd-to-end wiring partial
How it works

A ref moves. Everything that cited it is re-checked.

A provider event says a ref moved. Pulse walks every anchor that pointed into that ref, classifies what happened to it, updates freshness, and records what it could not reach. The classification runs before recall serves the claim — that ordering is the entire point.

Change radar

Evolving

A ref moved. Pulse classifies every anchor that pointed into it as unchanged, moved, altered, or removed — before recall has a chance to mislead an agent.

Freshness

Shipped

How long ago each anchor was last confirmed against real source, and which anchors have never been confirmed at all.

Usage

Evolving

Which beliefs are actually recalled, which are dead weight, and which are recalled constantly while still unreviewed.

Trust health

Evolving

Review-state distribution, open contradictions, and abstention rate — the graph reporting on its own reliability.

Coverage disclosure

Shipped

What Pulse could not see: unindexed refs, unreachable providers, and gaps in the window. Stated on the view, not in a tooltip.

Drift events to recall

Planned

Pushing a drift verdict into the recall path so a stale anchor is flagged at read time, not at dashboard time (ADR-029).

The artifact

One branch switch, 158 anchors re-checked.

Real markup, not a screenshot. Read the four classifications, then read the coverage panel beside them — it is the same size, in the same card grammar, because a count you cannot interpret without knowing its coverage is not a measurement.

Pulse · change radar13 anchors need a human

release/2026.3 diverged from main

An agent switched to the release branch. 158 anchors across the project pointed into the refs that moved. Every one of them was classified before recall served a single claim.

refs/heads/main @ 4c81f0arefs/heads/release/2026.3 @ 9ea2d73

  • Unchanged128

    The cited lines are byte-identical on the new ref.

    Recall serves these normally. Nothing to review.

  • Moved17

    The same content exists at a different path or line range.

    The anchor is re-pinned and the claim keeps its review state.

  • Altered9

    The cited region still exists but its content changed.

    Recall serves these with a stale marker until a human re-confirms them.

  • Removed4

    The cited region does not exist on the new ref at all.

    The claim is withheld from recall rather than served against nothing.

since 2026-03-11T00:00:00Z
Pulse · coveragePartial coverage

What this radar could not see

The counts above describe the anchors Pulse could actually check. Three gaps mean the true totals are higher, and Pulse does not know by how much — so it reports the gaps rather than a confident estimate.

  • 2 of 5 repositories not indexedAnchors into them are neither classified nor counted above. They are absent from the totals, not assumed unchanged.
  • 1 provider webhook unreachable for 6hRef events during that window were reconstructed from a later poll, so the ordering within it is unknown.
  • Usage data starts 2026-03-11Recall telemetry before the freshness migration was not retained and is not modelled.
158 anchors observed · unobserved anchors are excluded, not assumed

Illustrative example — not live data

What exists today

Surfaces built. Wiring partial.

The radar surface and the freshness migration are shipped, and the honest empty and degraded states are real — a Pulse view with nothing measured renders as nothing measured. The end-to-end path from provider event to flagged recall is partial, and the pieces that are missing are listed here rather than implied to work.

  • Freshness on every anchor

    Shipped

    The schema migration that gives each anchor a last-confirmed timestamp and a never-confirmed state. Shipped.

  • Coverage disclosure

    Shipped

    Every Pulse view states what it could not observe. A view with no coverage statement is a bug, not a clean dashboard.

  • Change radar surface

    Surface shipped · Wiring partial

    The radar view itself is built and renders honest empty and degraded states when there is nothing measured to show.

  • Anchor drift classification

    Evolving

    Classifying anchors as unchanged, moved, altered, or removed on a ref event. The classifier exists; end-to-end wiring from provider events is partial.

  • Usage signals

    Evolving

    Which beliefs are recalled and how often, measured from real recall calls — never estimated from what looks important.

  • Trust health rollups

    Evolving

    Review-state distribution, open contradictions, and abstention rate across a project.

  • Drift events pushed into recall

    Planned

    So a stale anchor is flagged when an agent reads it, not when a human happens to open the dashboard (ADR-029).

  • Fleet-wide health across machines

    Planned

    Rolling the same signals up across every connected client and machine. Depends on the cloud session registry.

Status labels follow docs/PRODUCT-DESIGN.md. No number anywhere on this page is measured from a real deployment; the artifact above is fixture content and is labelled as such.

Know what drifted before your agent quotes it.

Read the docs
See the platform