Read this first. These are experimental results, not a validated list.
Every number below is measured. There are no outreach metrics anywhere on this page because no
message has been sent. Rows are labelled by verification state: confirmed means an independent
reviewer, instructed to refute it, found external evidence that it holds; weak means it
survived but with a real argument against it; unverified means the model scored it and nobody
has checked it yet.
Executive overview — every chart on this page in one read
60 seconds
Next step: human inspection. The agent has taken this as far as
automated evidence allows. What refines it now is a person reading each row and saying why it is
right or wrong. Open any prospect, mark it Yes, No or Borderline, add the
reason, then export your review as a CSV. Those labels become the next calibration round, exactly
the way the Krishna rounds were run.
Download
CSV exports. The audit file carries every
evaluated person including the rejects and the reason each one was rejected.
Discovery funnel
Everything that happened between a raw company list and forty delivered prospects. Nothing is hidden by truncation.
What already existed — the work done before Akshay left
Read from the shared Drive folder, read-only. This is why the segments never converted, and it corroborates the agent's own conclusion from a completely different direction.
Seed profiles saved by persona
Round 1 lists, scored under the current agent
ICP score distribution
All 918 evaluated people, by band. A model that cannot spread scores cannot rank inside a band.
Company fit against person fit
Both gates sit at 7.0. A great company with the wrong person is rejected, and so is the reverse.
recommended
borderline
rejected
Why prospects were rejected
650 rejections, attributed to the rule that fired first.
Adversarial verification
73 delivered rows were handed to independent reviewers told to refute them and to default to refuted under uncertainty.
Experiment history
Five versions. Recommendation rate is the share of evaluated people the model would send; score dispersion is its ability to rank inside a band.