msgbench beta

Simulated reader roundtable

Before you publish,
let a few readers go first.

Paste a draft. Simulated readers with different backgrounds read it at their own pace, some to the end and some only to the sentence where they give up. Then they sit around one table and tell you where it held them, and where it lost them.

Open Message Bench How it works
  • Sentence by sentenceAttention, comprehension and memory advance one sentence at a time, reproducibly
  • Each at their own paceAge, situation and motivation set how fast they read and how far they get
  • Grounded in the textEvery remark quotes what that reader actually read
Core

Message Bench

Paste a draft and seat three to five readers. Watch each finish it or put it down at their own pace, hear them speak and answer one another, then take the moderator's edits in order of priority.

Go to the bench Profiles

Import and fine-tune readers

Start from the five built-in readers or import agent-profile-v1 files. Move a reading parameter and the engine re-reads the piece as that reader, live.

Open the profile library Settings

Preferences and saves

Choose the default seats and the replay speed. Every roundtable is saved on this device and replays without calling a model again.

Open settings

Method

One table, three layers

01

From demographics to parameters

Age, education, relation to the topic, reading situation and motivation set prior knowledge, topic relevance, processing capacity, involvement and reading speed. The table is a hand-set prior: versioned, and tunable value by value.

02

A deterministic model reads

msgbench splits the draft into events and updates each reader's attention, working memory, knowledge, recall and motivation one sentence at a time. The same input always gives the same trace, and every diagnosis points to a place in the text.

03

A language model only puts it into words

Each remark is written from that reader's own trace and the part they actually read. Quotes are checked word for word on the server, and any that do not match are dropped. With no model configured, remarks are assembled from trace templates and labelled as such.

What we do not claim

  • Not people.The seats hold simulated readers. Names and titles such as Ashley or Dr. Russell are fictional; they stand for no real audience and no real person.
  • Not calibrated.The demographic mapping and the exit model are priors. Scores are simulated readings, not measurements.
  • Not averaged.The differences between readers are the finding. We never fold them into a single score.
  • Traceable.Every quote comes from the text that reader read. A quote that does not match never reaches the page.
Open Message Bench