Review intelligence, progress & Pulse
These three capabilities close the observability-product surface: review-intel is the
LLM-augmented review layer that produces findings flowing through the Trust
state machine; progress is the evidence-anchored change graph and confirmation queue over
recorded activity; pulse is the product observability surface that renders signals produced by
the measurement plane without computing them. The ownership split is deliberate:
trust owns verdict state and provenance; review-intel owns the LLM that produces findings
into that state.
Capability status
| Capability | Tier | Status | Evidence |
|---|---|---|---|
review-intel | Differentiator (Tier 2) | Evolving | Governed findings, citations, lens set, abstention, TrustBench eval |
progress | Tier 3 | Evolving | Evidence-anchored graph + confirmation queue; depends on surfaces/memory |
pulse | Observability UX | Shipped | Product surface; consumes measurement, does not produce it |
review-intel
review-intel sits at Tier 2 — it depends on Tier 0
(economics, observability), Tier 1 (surfaces, memory), and Tier 2 (trust).
It ships the LLM-augmented review layer: governed findings, citations, a lens set, abstention
at budget exhaustion, and TrustBench review-quality evaluation. trust owns verdict state;
review-intel never writes it directly — findings flow through the review state machine and a
hand-tampered or edited artifact fails verification (see Trust).
progress
progress ships as the evidence-anchored change/progress graph and confirmation queue answering
"what changed, what task did it complete, and is that trustworthy". A progress claim carries
scope identity, evidence anchors, a content digest, and a reviewState of unreviewed,
provisional, confirmed, rejected, or revoked. This reviewState governs progress items
specifically; it is independent of, and does not feed, the belief reviewState in trust —
the two share value names but apply to different entities. No claim becomes durable confirmed
without a human-accountability path.
pulse
pulse ships as the product observability surface under /pulse/*. The shipped redirect is
limited to /pulse/jobs → /runtime/jobs; no blanket /pulse/* alias is claimed. It consumes signals produced by observability, trust,
code-intelligence, economics, memory, and platform and renders health, usage, trust,
drift, storage, and diagnostics — without inventing, re-specifying, or computing those signals.
If the trust capability reports a metric unevaluated, pulse renders not_evaluated rather
than a number.
Availability
| Surface | Tier | Residency | Notes |
|---|---|---|---|
zephr_review_usage (MCP) | review | deferred | Review-intelligence findings and TrustBench quality |
zephr_status (MCP) | core | resident | Runtime state and service health (consumed by Pulse) |
/pulse/* | — | cloud | Product observability surface |
/runtime/jobs | — | cloud | Canonical target for the shipped jobs redirect |
Limits and failures
review-intelabstains at budget exhaustion rather than guessing; it never self-confirms a finding into thetrustverdict state.progresscannot promote a claim toconfirmedwithout a human-confirmed path; an auto-derived claim is never directly confirmed.pulserenders only what it consumes — a missing upstream signal surfaces asnot_evaluated, never a fabricated value.
Related capabilities
governance/defect provide the accountability band this surface reports on; observability
owns the measurement plane pulse consumes; trust owns the review state review-intel
writes through.