What is an institutional AI simulation?
An institutional AI simulation reads a draft disclosure the way a room of institutional analysts would, then surfaces where they converge, where they diverge, and the questions the filing will trigger, before it is filed.
One score is not enough
A single legibility number tells you the average reader's experience. It hides the desk that will misread the filing worst. An institutional simulation spins up a panel of distinct archetypes and reads the text through each lens, so blind spots become visible.
The panel spans a macro or sovereign allocator, a multi-strategy quant, a sector active manager, a credit or risk analyst, a retail reader, and a regulatory reader. Each weights different legibility factors according to what it cares about.
Consensus, deviations, and predicted questions
The simulation computes three things. Consensus comprehension is the panel average. Deviations are the members who understand the text materially worse than consensus, which are the blind spots a specific desk would hit. Predicted questions are the questions the filing will trigger, ranked by how many panel members ask them.
That turns a draft into an actionable pre-file checklist: fix the blind spots and pre-empt the questions.
How LyraMind implements it
The personas are synthetic deterministic archetypes, not any real institution's AI, and LyraMind never labels one as a firm it isn't. Each archetype's comprehension is computed by penalizing the legibility factors it weights, and it asks its templated question when its lens is not cleanly satisfied. A top macro driver for the allocator's question is pulled from the Context Graph.
When frontier models are configured (off by default), they can literally parse the draft and join the panel under the same contract, honestly labeled. Their spread of comprehension is reported as empirical misread risk. This measures the clarity of public text and does not predict how any specific named model will respond.
Related concepts
The simulation reads off the legibility scan's factors and pulls drivers from the knowledge graph. It is the panel inside the Corporate AI Readiness Report, contributing per-tier readiness and the predicted questions.
- A panel of six analyst archetypes reads the same draft, each through its own lens
- Returns consensus comprehension, deviations (blind spots), and predicted questions
- Personas are synthetic deterministic archetypes, never a real firm's AI
- The allocator's question pulls a top macro driver from the Context Graph
- Configured frontier models can join the panel, honestly labeled and off by default
- Measures clarity of public text; never predicts a named model's response
Questions
Are the personas real institutions' AI?
No. They are synthetic deterministic archetypes. LyraMind never labels a persona as a firm it isn't.
What is a deviation in the simulation?
A panel member whose comprehension is materially below consensus. It marks a blind spot: the kind of desk or reader most likely to misread the filing.
Can real frontier models join the panel?
Yes, when configured. They are off by default, honestly labeled, and their spread of comprehension is reported as empirical misread risk.