Self-Checking Answers + Pick How Hard Alfrada Works

June 28, 2026

Alfrada can now grade its own work before you ever see it. On bigger tasks, the assistant writes a task-specific rubric, scores its own draft against it, and revises until it passes — so you get the polished answer, not the first attempt. And you're in control of how much effort goes in: a new performance slider lets you pin Fast, Medium, or Beast, or leave it on Auto.

What you can do

• Let Alfrada review and fix its own answers. After the model writes a final answer to a substantial request, it builds a rubric tailored to that exact task, grades the draft against it, and loops back with concrete revision notes until the rubric passes — all before the answer reaches you.
• Get a single clean revision on Medium, or relentless polishing on Beast. Medium does one review-and-revise pass on genuinely large deliverables. Beast keeps revising until the rubric clears (bounded so it never loops forever).
• Have generated images checked, not just text. On Beast, produced charts and images get a vision pass that flags unlabeled axes, missing legends, unreadable or cut-off text, and rendering errors. Visual and substance verdicts merge — but a visual nitpick never blocks an otherwise-good answer.
• Pin the effort level per session. A performance slider in the model picker lets you choose Fast (best budget model, no grading), Medium (workhorse + a single self-review), Beast (top-tier + full revision loop and image checks), or Auto (Alfrada decides per turn).
• See the mode it actually ran in. A badge shows the resolved mode based on the model that really handled the turn — not just what was requested — so it's honest about what you got.
• Set it once and forget it. Your pinned mode sticks for the session and is restored when you reload it.

Trivial turns are left alone: a one-line reply to "hi" or "what model are you?" is never graded, and short answers skip the loop entirely. Grading uses a cheap local critic model and runs quietly in the background, with a small "grading…" indicator while it works.

Where this shows up

• You ask for a detailed competitive analysis or a long report. You used to skim it and spot a missing section or a thin argument yourself. Now Alfrada catches that against its own rubric and fixes it before sending.
• You're on Beast and ask for a chart. The first render has an unlabeled Y-axis — Alfrada notices in the vision pass and regenerates it labeled, instead of handing you a chart you'd have to ask it to redo.
• You've got a quick back-and-forth where you just want speed. You pin Fast, and Alfrada stops grading and routing for depth — it just answers.

Try it

• "Pin Beast mode, then write me a full go-to-market brief for a B2B SaaS launch — sections for positioning, pricing, channels, and a 90-day plan."
• "Build a bar chart of these quarterly numbers and make sure every axis and label is clean."
• "Switch to Fast — I just want quick answers for the next few questions."

Heads up

• The self-review uses a grader model, and that usage is billed to you like any other model call. It only kicks in on Medium and Beast, and only on substantial outputs — Fast never grades.
• Auto won't quietly drop to a cheaper tier mid-session. The only way to lower effort is to pin a mode yourself; left on Auto, routing stays consistent and only climbs when the task needs it.