Skip to main content
Every page in this section so far describes the success path. This page describes what the advisor does when the success path is not available. The rule is simple: the advisor is allowed to be less useful; it is not allowed to be silently wrong.
Every failure below emits a matching event in Telemetry. Degraded states are observable, not hidden.

​
The four degradation levels

Every failure mode maps to one of these:

Level 1: Answer with caveat

Advisor proceeds but explicitly cites the limitation. Never for Task or Proposal.

Level 2: Downgrade artifact

Advisor produces a weaker artifact than it would have (Task → View, View → Answer). User can still act, just with fewer shortcuts.

Level 3: Refuse with alternative

Advisor cannot act, states the reason in plain language, and offers a viable next step (schedule check, wait for imagery, contact admin).

Level 4: Refuse without alternative

Advisor cannot act and no alternative is safe. Rare. Reserved for entitlement violations and structural impossibilities.

​
Failure catalog

Trigger: Sense reads succeed but the latest imagery, activity, or rule card ruleset is older than the freshness SLA for the crop’s cycle model.Behavior: Level 1 or 2 depending on artifact.
  • Answer / View: proceed, prepend a freshness banner (“latest NDVI is 9 days old; next pass scheduled”).
  • Task: downgrade to View. The advisor will not draft a scout task or VRA on stale imagery. It offers to notify the user when fresh data arrives.
  • Proposal: refuse. Proposals require current evidence.
Trigger: the rule card that would fire is in a pending version bump (thresholds under review, activity_bindings changing).Behavior: Level 2.
  • Advisor uses the last-approved version, cites the version in the reasoning panel, and flags that a review is in progress.
  • Task issuance is allowed but the confirmation dialog includes the caveat.
  • Proposal issuance against the same card is deferred until the migration lands.
Trigger: user’s access to an estate, field, or role was revoked while the session was open. Session is not per-user memory, but it is bound to identity + entitlements at start.Behavior: Level 4 for the affected entity; Level 1 for the rest.
  • In-flight artifacts targeting the revoked scope refuse without alternative.
  • Session is not killed; other scopes continue.
  • Advisor tells the user which scope is no longer available and why (revocation is not silent).
Trigger: a module the advisor needs to write to (Field Scouting, VRA Maps, Verification) is unavailable or returns errors.Behavior: Level 3.
  • Advisor states which module is down.
  • Task drafts are preserved in the session so the user can retry when the module recovers (drafts, not commits).
  • View artifacts fall back to a read-only version of the target page if available.
  • For Verification specifically, satellite provider outages are surfaced with an estimated recovery window if known.
Trigger: Sense reads contradict. Most common example: the last scout report says healthy; the latest NDVI says stressed. Or two rule cards fire with conflicting activity_bindings.Behavior: Level 2.
  • Advisor never picks a winner silently.
  • Reasoning panel shows both signals with their timestamps and confidences.
  • Artifact is downgraded to View with an explicit prompt: “scout says healthy on ; NDVI drop on ; assign a fresh scout or wait for the next pass?”
  • Downgrade includes offering a Task (fresh scout) as the natural resolution path.
Trigger: intent classification confidence is below threshold and no single alternative is dominant.Behavior: Level 1, then re-attempt.
  • Advisor asks a single scoping question (not a chain).
  • If the follow-up still cannot classify, advisor returns an Answer citing what it can see (raw entities) and asks the user to rephrase.
  • It does not fabricate an intent.
Trigger: Interpret finds firing signals but no rule card or diagnosis page matches the pattern.Behavior: Level 2 + Proposal.
  • Advisor returns an Answer with raw indices and any partial rule card matches.
  • Advisor offers to file a Proposal flagging that a new pattern may be needed. Proposal enters the human review queue; nothing auto-applies.
Trigger: user requests an action that would suppress a severity ≥ high signal, mute a safety-critical alert beyond bounds, or exceed agronomic min/max rates.Behavior: Level 3.
  • Refusal is stated in plain language with the rule ID.
  • Advisor offers the bounded alternative (mute for the maximum allowed window, rate at the ceiling, escalate to a role that can override with reason).
  • Every safety-floor refusal is logged and, for repeated hits by the same user, escalates per the target module’s guardrails.
Trigger: two users acting on the same entity at overlapping times, or the target entity was archived or edited between Sense and Act.Behavior: Level 2.
  • Advisor re-reads the entity before commit. If it has changed materially, the confirmation dialog switches to a diff view: “this changed since you started; keep, discard, or merge?”
  • If the entity was archived, the artifact refuses with a Level 3 alternative (open the archive, un-archive if authorized, or target the successor entity).
  • The corresponding audit entry on the module records the concurrency reconciliation.
Trigger: mobile client or spotty connection. Not the same as module outage.Behavior: Level 2 (read-only mode).
  • Advisor answers from last-known state with an explicit staleness banner.
  • Task and Proposal issuance is disabled. Drafts can be composed and queued, but nothing commits until connectivity returns.
  • Matches the offline model described in Field Scouting.
Trigger: session started under one contract_version or agent_version; a deploy landed mid-session.Behavior: Level 1.
  • Active session continues on the version it started with (pinned).
  • New sessions use the new version.
  • Telemetry records both versions on session_closed so evaluation can attribute outcomes correctly.

​
What the user always sees

When the advisor degrades, the user gets four things, in this order, every time:
  1. What is limited (“latest NDVI is 9 days old”).
  2. Why (“next Sentinel pass scheduled for ”).
  3. What the advisor can still do (“I can still show you the last verified state and file a Proposal if you think a pattern is missing”).
  4. What the advisor will not do and the alternative (“I won’t draft a VRA on stale imagery; want me to notify you when fresh data lands?”).
This is the anti-hallucination contract. The advisor is allowed to be less useful. It is not allowed to be silently wrong.
  • Overview — the artifact contract these failures map back to.
  • Agent Loop — the stage where each failure surfaces.
  • Telemetry — the events emitted when a degradation happens.