Avatar fallback quality ladder: degrade presence without losing the task

Define modes by capability, not glamour

Write an ordered ladder such as full motion and speech; reduced motion with captions; static portrait with text; text-only conversation; and handoff or unavailable state. Each mode must specify what remains: transcript, input, stop, confirmation, status, and action controls. Do not use “low quality” as a vague catchall. A lower-fidelity mode can be fully functional if the service facts remain visible and operable.

The ladder is not a performance benchmark. It is an interaction contract for when a client cannot sustain a presentation feature. The active mode should be visible in a concise, non-alarming way and should never imply that a task has completed merely because an animation finishes.

Preserve task state across a change

Presentation mode is not conversation state. Keep turn identifiers, current draft, action proposal version, and transcript independent of renderer selection. When motion playback fails, the system can retain the current response as text and continue the same turn; it must not regenerate the answer or resubmit a tool call just to recover visual presentation.

Worked example, hypothetical: a visitor is reviewing a booking proposal when avatar audio becomes unavailable. The system announces “Audio is unavailable; continuing in text,” displays captions as ordinary transcript text, leaves the confirmation disabled until current facts are reviewed, and keeps the same edit links. On audio restoration it does not replay confidential earlier content automatically.

Respect preferences and alternatives

Honor prefers-reduced-motion from the start, rather than treating it as an error state. Provide controls to choose text or captions even when full motion works. Captions and transcripts should convey service content, not a decorative approximation of facial expression. Audio-only output has no visual fallback unless the product provides one; a text transcript is the dependable common representation.

The counterexample is a fallback that removes the avatar but also removes the stop button, live status, or keyboard route. Another is a restoration that begins autoplay speech without user request. Accessibility does not begin after degradation; it defines which elements are essential in every rung.

Test escalation and restoration races

Simulate media decode failure, microphone denial, network loss, reduced-motion preference, CPU pressure, renderer reload, audio-route change, and partial recovery. Assert that the system chooses a known lower mode, records a bounded diagnostic for operators, and preserves one transcript entry per turn. Test recovery during generation, during a confirmation, and after a user pressed Stop.

Acceptance checks: every mode has keyboard-accessible controls; task state and draft survive a mode switch; one visible status tells the user what changed; restoration requires no repeated consent; and fallback cannot duplicate a consequential request. Test on target devices rather than declaring the ladder proven from desktop emulation alone.

Mode check. Expose a non-production test control that forces each rung. Teams cannot verify a graceful fallback that only exists as an unobserved error handler. Test a lower mode before adding animation polish, and keep the test control unavailable to public users.

Keep the limits honest

CSS media queries can expose a user preference for reduced non-essential motion, and WebVTT defines timed text-track behavior for media. Neither source guarantees that a browser can detect all GPU, network, or audio failures. Capability signals can be incomplete, and quality decisions need conservative defaults.

This guide intentionally goes beyond a browser performance budget: budget work asks whether a target is efficient, while a quality ladder says what the service becomes when it is not. It is original system-design guidance, not a claim that any avatar vendor or rendering stack provides these modes automatically.

If a response contains an editable proposal, the text-only mode must expose the same fields, sources, and confirmation state as the animated view. Do not make a person return to motion or audio merely to inspect an amount, date, warning, or cancellation route. Test a mode switch after a user edits one field and before they confirm; the stale-version rule must remain intact. This is where a fallback ladder joins action safety: visual richness may disappear, but the evidence needed for a decision cannot.

Document whether a lower mode changes data use, such as disabling camera processing or local speech capture. Explain the change before it happens where feasible. A text fallback should not silently activate a new collection channel simply because the audiovisual route failed.

Assign restoration ownership

Decide which subsystem may request an upgrade and which conditions make it safe. A renderer that recovers must not independently replay speech, reopen a camera, or change an action.

Telemetry should describe mode changes without storing conversation content unnecessarily. Improve the lowest essential rung first: the ladder works when text-only remains a service, not when the top mode looks impressive.