rd-report

rd-report — the number, and the decision it supports

DOCTOR passed, so you are now allowed to quote a number. Your output is FINDINGS.md, written to the same signature the published case studies use — agent/studies/marginal_scientist/FINDINGS.md is the exemplar, and it is worth reading before you write a line.

The audience is someone who has to act and was not in the room. They need the answer, the rule that uses it, and a reason to trust it. In that order.

Get the numbers

Two arms — the usual case, a lever:

julia --project=. agent/scripts/rdx.jl effect <baseline> <arm> --host=<host.jl> \
  --metric=enpv --scale=0.001 --annual=0.08 --seeds=48 --root-seed=2026

One arm — a level, a bottleneck, a distribution:

julia --project=. agent/scripts/rdx.jl ensemble <model> --host=<host.jl> \
  --metric=final:launch --seeds=48 --root-seed=2026

Member k of each arm is seeded from hash((root_seed, k)), so both arms are compared on the same seed stream — the paired comparison is what makes a modest effect detectable at all. Record the root seed in the report; it is what turns a number into a replayable number.

Is the effect real? effect reports significant_2se. If it is false, the honest headline is "we cannot distinguish this from noise at N members" — either raise --seeds and rerun, or report that. Never report a point estimate whose sign the data does not support. Sample-size honesty is cheap here: members are independent, so if 48 doesn't separate it, try 200 before concluding.

The signature

1. Headline — the number, in the title

## **+$19.0M ± $2.5M** — hire them

The number and the verdict, in units the reader thinks in, above everything else. Then one paragraph: the two levels being compared, the horizon, and how many standard errors from zero. If the reader stops here, they should already be able to act.

2. The decision rule

Not "the fifth scientist is worth $19M" but "hire iff the fully loaded cost over 48 months is below $19M." Turn the number into the comparison the user actually makes — that sentence is the deliverable and everything else is support. Where it is natural, extend it: a shadow price that falls as the constraint relaxes traces a diminishing-returns curve, so say where hiring should stop, not just that the next hire clears.