Conversion analytics

karani funnel                   # conversion, segmented, with intervals
karani coach --window-days 7    # the derived weekly review
karani variant list             # named strategy variants

The funnel

application -> response -> screen -> onsite -> offer

One durable row per application, written when you record a submission. That is why the write-back matters: karani submitted is not bookkeeping, it is the only thing that creates a label (ADR 0022).

karani funnel splits the rates by positioning angle, strategy variant, company, comp band, warm-versus-cold, and the posting’s age when you applied. Plus an autopsy: response rate by seniority and remote status, and keyword coverage for applications that got a response against those that did not.

Every rate ships with its interval

Not a style preference. Three applications and one response is not a 33% response rate — it is a number between about 2% and 80%, and reporting the point estimate alone invites a decision the data cannot support.

So every rate carries a 95% interval, and the intervals are what you compare rather than the percentages.

Fit bands are not pooled across scorers

A fit score is only comparable within the model and prompt version that produced it. Measured over 598 rows:

Those are not the same scale, and averaging them produces a number that describes nothing. So karani funnel pools fit bands only when one scorer produced every score, and otherwise reports per-scale with a warning.

This is also how the autopilot’s draft floor was found to be broken: it had been tuned against the third scorer and matched almost nothing under the first.

Variants, and the rule that refuses

A variant is a named strategy — a positioning angle, a cover-letter structure, a resume ordering — that applications get tagged with so you can find out whether it works.

karani variant create data-platform-first --dimension positioning \
  --description "Lead with platform scale, not causal ML"
karani variant assign 1482 data-platform-first
karani variant status

Promotion needs both of:

• a minimum sample on each of two arms, and
• non-overlapping 95% intervals.

Short of that, there is no winner, and karani says there is no winner. A test pins that three applications do not produce one, because the temptation to read a winner out of three applications is exactly what this rule exists to resist.

Time to first response

Tracked per application and per company. It is the most actionable number in the funnel, because it tells you when to follow up rather than whether to — a company whose median first response is three days has not replied in twelve for a reason.

Fast lanes

A role with high fit, posted in the last three days, is flagged in karani actions as a fast lane. Applying on day one is measurably different from applying on day nine, and the funnel reports response rate by posting age at application, so this is a measured effect rather than folklore.

The coach review

karani coach --window-days 7
karani coach --slack --save

A derived review over the same ledger: what converted this week, what did not, which variant is ahead and whether it is ahead by enough to say so. Derived means it adds no new data — if the funnel cannot support a claim, neither can the review.

What it costs you to get this

The labels come from three commands:

karani submitted 1482 --cold      # the application exists
karani status 1482 screen         # it progressed
karani outcome 1482 rejection     # it ended

The watcher does the second and third for you when it is confident. The first is always yours, because only you know you pressed submit.

An application karani does not know you sent contributes nothing to any number on this page.

Next

• The inbox watcher — how the later labels arrive without you typing them.
• ADR 0022 — the full argument, including what was measured and when.