rd-doctor

rd-doctor — the plausibility gate

validate proved the document is well-formed. It proved nothing about whether the model is about the user's system. A schema-clean model can be confident nonsense, and the failure mode of this entire workflow is a beautifully formatted wrong answer. This phase is the gate that stands between the two.

Your output is DOCTOR.md: every line machine-produced evidence, not an opinion. agent/studies/marginal_scientist/DOCTOR.md is the exemplar.

Run it

julia --project=. agent/scripts/rdx.jl diagnose <model> --host=<host.jl> --seed=1 \
  --expect-binds=scientist --expect-reusable=scientist --tally=launch

The --expect-* flags come from the brief's pre-registered failure modes — do not invent them here. A check whose expectation was written before the run existed is evidence; one you chose after seeing the output is decoration. If the brief didn't pre-register them, go back and fix the brief, then run.

Exit 0 = clear to report. Exit 1 = at least one failing check; you may not quote a number. Diagnose every arm, not just the baseline — an arm that changes the lever can change which resource binds.

The nine checks

Check	Severity	Asks
transitions_fire	fail	did every transition spawn at least once?
completions_exist	fail	did anything finish?
contention_exists	fail	does anything actually run out?
intended_resource_binds	fail	is the scarce thing the one the brief named?
modality_intent	fail	are the resources the brief called reusable declared conserved?
horizon_sufficient	warn	is anything still in flight at the end; are pools still trending?
bounded_pools	warn	is anything growing without limit that isn't a declared tally?
ledger_wired	warn	do reward and cost both move?
stochasticity	warn	is any outcome actually uncertain?

contention_exists is the one that matters

If nothing binds, there is no study. A model where every transition always gets what it asks for is an accounting identity with extra steps, and the marginal value it reports is just the accounting cost. When this check fails, do not tune the model until it passes — go back to the user and report the finding: four scientists are enough; the fifth is worth their loaded cost and no more. That is a real, valuable, publishable answer, and manufacturing scarcity to avoid delivering it is the worst thing you can do in this workflow.