Human verdicts decide provisional claims.
A claim is not confirmed until a human says so.
Inspect, run lenses, ground, present, record.
The lifecycle is fixed. The lenses do not score, average, or recommend; they return what they found and stop. A human reads the findings and records a verdict. The verdict is written against the claim, and the underlying claim is never silently changed to match it.
01 — Inspect
Review starts from the thing a human or agent proposed — a diff, a captured belief, a provisional claim. The subject is fixed before any lens runs.
02 — Lenses
Security, correctness, tests, documentation, accessibility, performance. Each lens is a fixed check, not an opinion generator, and each runs only where it has something to look at.
03 — Ground
A finding names the decision, the belief, the test, or the document it came from — cited at a revision. A finding with no citation is not published as a finding.
04 — Present
The lenses return what they found and stop. They do not score, average, or recommend. A human reads the findings and decides.
05 — Record
Confirm, reject, or leave provisional. The verdict is written against the claim; the underlying claim is never silently changed to match the verdict.
Illustrative example — not live data
New or contested claims require a human verdict.
Elapsed time never promotes a provisional claim to confirmed. Only a human moves it from provisional into confirmed — the gate fails until a recorded verdict exists.
Is the subject inside the admitted review scope?
passed
Does every blocking finding carry a revision-bound citation?
passed
Has a human reviewer confirmed this contested change?
failed
Each rung clears or blocks. The top rung decides.
A review is not a single gate but a runged escalation. Each lens check is a rung that may pass or block, and the human verdict is the top rung that decides.
- 01scope
Is the subject inside the admitted review scope?
passed
- 02evidence
Does every blocking finding carry a revision-bound citation?
passed
- 03security
Does the security lens find a blocking contradiction?
blocked
- 04verdict
Has a human reviewer confirmed this contested change?
pending
The security finding blocks. The verdict is pending until a human decides.
One provisional claim, six lenses, one human verdict.
An illustrative claim and its findings. Each lens cites what it found and stops; the claim stays provisional until a human decides.
Claim — rev_01HQ8M4T2R7XKC
A new endpoint caches session validity for sixty seconds in the auth middleware.
proposed by agent — Claude Code — sonnet — 2026-03-11T08:52:14Z
Verdict
Rejected on the security finding. The claim stays provisional.
human — reviewer — 2026-03-11T09:14:02Z. The verdict is recorded against the claim, not applied to it silently.
Security
A reviewed decision requires revocation to take effect within one request. The cache relaxes that guarantee and is named with the decision it contradicts.
blocked
Correctness
No existing test asserts the one-request revocation guarantee. The lens surfaces the gap rather than scoring the change against a missing spec.
unreviewed
Tests
The added path is covered; the lens does not infer a pass rate and does not claim the suite is green where it has no evidence.
verified
Documentation
docs/security/session-lifecycle.md describes the guarantee this change relaxes. The lens cites it at its recorded revision.
verified
Accessibility
No UI surface changed. The lens records no finding rather than manufacturing one to fill the column.
verified
Performance
The cache improves lookup latency and is not contested on performance grounds. The lens reports what it measured and nothing else.
verified
Illustrative example — not live data
From provisional claim to human verdict.
The escalation rung shows the path a review takes: inspect the subject, run the lenses, ground every finding in a citation, create the evidence capsule, and present for human verdict. No step is skipped; no timer promotes.
Inspect
Fix the subject before any lens runs.
Lenses
Run security, correctness, tests, docs, a11y, perf.
Ground
Cite every finding at a revision.
Capsule
Record the evidence capsule.
Verdict
A human decides. No timer promotes.
Illustrative example — not live data
Lenses, citations, and the human verdict workflow.
The review engine and the lens foundations are built and evolving. zephr_review, the MCP tool that exposes review to agents, is shipped and is one of Zephr’s 29 tool surfaces. The platform review ledger is in development. Every card below carries its own honest status.
zephr reviewHuman verdict workflow
EvolvingA claim enters provisional, a human confirms or rejects it, and the verdict is recorded against the claim without rewriting it. New and contested claims require a human verdict.
review engine - ADR-023
Security lens
EvolvingChecks a change against reviewed decisions and recorded beliefs that govern the changed surface. Reports a finding with a citation, or records none — never a confidence score.
review engine - lens foundations
Correctness lens
EvolvingSurfaces whether an existing reviewed spec governs the change, and names the gap when no spec exists. Does not infer correctness from a missing test.
review engine - lens foundations
Tests - Documentation - Accessibility - Performance lenses
EvolvingEach lens runs only where it has something to inspect, cites what it finds, and records no finding where it has no evidence. No lens publishes a pass rate.
review engine - lens foundations
zephr_review MCP access
EvolvingThe MCP tool that exposes review to agents, one of Zephr’s 29 tool surfaces. An agent can request a review and read the findings; it cannot confirm its own claim. Hosted reviews run only once an organization review policy is provisioned; without one the tool abstains.
mcp-server tool catalog
Citations and evidence grounding
EvolvingEvery published finding names its source at a revision. A finding the lenses cannot ground is withheld rather than printed as a plausible paragraph.
evidence graph - ADR-023
Review ledger / platform history
PlannedThe durable, queryable history of who decided what, at which revision, against which claim. The substrate that makes a review auditable across time.
platform ledger - in development
Managed or advanced judgment
PlannedHosted review that answers on pull requests from the cloud plane, and any higher-order judgment assistance. Designed, not built, and gated on the hosted plane.
managed review - designed, not built
Review proposes. A human disposes.
Run the Code Review sample to see a hermetic code-change review with cited findings, coverage, and an evidence capsule. Guard is the separate governance surface, and its T3 AI judge is deferred by design.