Trust
Trust makes a belief's standing a first-class, enforceable property rather than a rendering hint. It owns the review state machine, the provenance chain, the verification ladder, tamper evidence, and the gating that other capabilities consume — every capability consumes trust state, and none of them may set it. The admission-scope and Firewall walkthrough lives in Concepts: Trust; this page covers the capability itself. Externally verifiable attestations are designed, not shipped (see Status honesty).
Availability
| Surface | Tier | Residency | Notes |
|---|---|---|---|
zephr_verify (MCP) | search | deferred | Version-chain integrity, provenance anchors, review state |
zephr_rules (MCP) | review | deferred | Trust Firewall policy per worktree |
zephr_status (MCP) | core | resident | Runtime state and service health |
zephr verify, verify-claim, preflight (CLI) | — | — | T1 + T2 verification and plan-time advisory checks |
Review state machine
Every belief carries exactly one review state — UNREVIEWED (the default),
CONFIRMED, STALE, DISPUTED, or REJECTED. The field is non-null from the
moment a candidate becomes a belief and is never collapsed into a numeric score:
- Agents can never self-confirm. Confirmation is scoped (scope is not transitive upward) and records who, when, and against what.
- Automated transitions are restricted to demotion. Code drift, reviewer departure, and contradiction demote rather than resolve; a demoted belief keeps its content and history. Drift- and expiry-demotion paths are keyed on attestation records, which no live production path produces yet (see Status honesty).
- Rejection is tombstoned and blocks relapse. A rejected claim cannot quietly return as a new belief.
- DISPUTED beliefs are disclosed with their dispute, never suppressed.
Core-eligible recall states are confirmed + valid; the Firewall can only narrow that set, never broaden it.
Verification ladder
Three tiers with declared limits: T1 deterministic checks assert only
mechanical facts (structure, chain integrity); T2 semantic contrast checks
against confirmed memory only; T3 budgeted judgment (rate-limited,
self-untrusted) is specified but not shipped — today zephr_verify runs T1 + T2.
It returns a verificationProfile plus warnings and reports verified only
for well-formed, provenanced, confirmed beliefs — only well-formed chain
(current-version pointer, contiguous revisions, non-overlapping transaction
intervals, one open version) passes. Escalation to T3 and the human handoff
at budget exhaustion are design, not a shipped path.
Status honesty
Shipped and registry-visible: the review-state field and its transitions, the
Trust Firewall (zephr_rules), verification (zephr_verify, CLI verify),
provenance disclosure (zephr_why), and recall gating. Specified in the trust
spec but not shipped today: signed human attestation (keyless OIDC
binding, offline-verifiable bundles, SLSA source-review slot), one-click
renewal, hash-chained revisions, and Merkle-batched tamper proofs. No live
confirm path produces an attestation record — design without assuming one
exists behind a confirmation.
Persistence, security, observability
Review state and provenance persist alongside the belief base — attestation
records do not, because none are produced on a live confirm path today
(storage topology in Architecture; local data
not encrypted by Zephr at rest). Provenance is
content-addressed and lazily deep — it spans observation to usage, survives
consolidation and revision, and is assigned by a deterministic classifier
(Concepts: Provenance). Observability: the why-card
states the reason for the current state inline; zephr_verify returns its
verification profile and warnings. Tamper evidence is structural today (chain
integrity checks); hash-chained revisions and Merkle-batched proofs are
specified, not built (see Status honesty).
Limits and failures
- No code path can create a belief in any state other than
UNREVIEWED. - Gating thresholds are configured per scope and are not lowerable from below.
- High numeric confidence never substitutes for a categorical review state —
UNREVIEWEDis disclosed verbatim. - A hand-tampered or edited artifact fails verification; signature mismatch surfaces as a trust-level error (see Continuity for the packet-resume variant).
- Demotion is recoverable by the normal human review flow, not by rewriting state — trust state is never hand-set.
For verification failures and review-flow symptoms, see Troubleshooting.