Degradation

Meridian Claims Assistant -- degradation design.

Fallback hierarchy

Rung	Condition	Behaviour	Quality
1	Primary model responds within 8s	Full structured recommendation with per-clause citations.	Best available.
2	Primary model slow (8-20s) or rate-limited	Falls to a smaller, faster model, same structured citation format, flagged "fast mode."	Slightly lower accuracy, still cited.
3	Both models unavailable, this exact claim was recently assessed	Serves the most recent draft for this claim, explicitly labelled "draft from [timestamp], not re-verified."	Stale but labelled, never silent.
4	No model available, no recent draft	Deterministic fallback: surfaces the raw policy section matching the claim's declared damage type, no generated recommendation at all -- Dana assesses manually with the relevant policy text pre-located.	Lowest, but honest and never wrong.
5	Even the raw policy-section lookup is unavailable	Explicit "Assistant unavailable -- assess this claim manually; your queue is unaffected" banner, input disabled but visible.	No recommendation offered.

%% Color Palette: Blue #0173B2, Orange #DE8F05, Teal #029E73, Purple #CC78BC, Brown #CA9161
graph TD
    R1["Rung 1: Full recommendation<br/>trigger: primary under 8s"]:::teal
    R2["Rung 2: Fast-mode fallback<br/>trigger: primary 8-20s"]:::blue
    R3["Rung 3: Recent draft<br/>trigger: no models, cache"]:::orange
    R4["Rung 4: Raw policy link<br/>trigger: no model, no draft"]:::purple
    R5["Rung 5: Unavailable<br/>trigger: raw lookup down"]:::brown
 
    R1 -->|degrades to| R2
    R2 -->|degrades to| R3
    R3 -->|degrades to| R4
    R4 -->|degrades to| R5
 
    classDef teal fill:#029E73,stroke:#000000,color:#FFFFFF,stroke-width:2px
    classDef blue fill:#0173B2,stroke:#000000,color:#FFFFFF,stroke-width:2px
    classDef orange fill:#DE8F05,stroke:#000000,color:#FFFFFF,stroke-width:2px
    classDef purple fill:#CC78BC,stroke:#000000,color:#FFFFFF,stroke-width:2px
    classDef brown fill:#CA9161,stroke:#000000,color:#FFFFFF,stroke-width:2px

Diagram: five rungs, each with an observable trigger condition, falling one at a time from a full recommendation to an explicit unavailable state -- never silently skipping a level.

Visibility check

Every rung carries its own explicit label (a "fast mode" flag, a "draft from [timestamp]" note, the raw-lookup's absence of any generated text, or the unavailable banner) so Dana can always tell which rung the assistant is currently operating on -- no rung is rendered identically to rung 1's full recommendation.

Verify: every rung states a concrete, observable trigger condition (co-18); the diagram matches the table exactly and terminates at an explicit unavailable state (co-16); no rung is visually indistinguishable from a fully healthy recommendation (co-19).

Key takeaway: The fallback hierarchy gives Dana five distinct, honestly labelled states instead of one binary "working or broken" state -- and rung 4's raw-policy-section fallback means even a total model outage leaves her with a genuine time-saving head start rather than nothing at all.