Run an experiment inside one allowance

Use this example to allocate one allowance across actor and grader calls in two experiment arms. Run it from a Harn checkout with OPENAI_API_KEY exported and harn on PATH:

harn run --standalone --no-sandbox examples/conservative-experiment.harn -- --live

The command makes four direct OpenAI text requests. It disables the process sandbox so the child Harn processes can reach the provider. Use it only with the checked-in example. Set HARN_BIN to an absolute executable path when testing a particular build; the parent uses that executable for every child.

The initial allowance is $2.40. Each arm receives $0.60 for its actor and $0.60 for its grader. The actor repairs a Python addition function; the grader reads the actor's answer. Each allocation starts one independent process with budget.admission: "conservative" before its first provider call. The parent checks the allocation sum in integer millionths of a dollar before launching any child. Unused allowance stays with that allocation and isn't reassigned.

Read the terminal JSON report. A successful run contains four runs, each with a distinct arm/role pair, one measured provider call, complete actual usage, and a conservative ceiling of "0.6". Every child must have zero in-flight and uncertain reservation, no denied attempts, and settlement at or below its allocation. cost_usd reports actual priced usage; settled_upper_usd reports the conservative charge. Keep them separate.

If a child fails or returns incomplete usage, the parent stops without launching the remaining allocations. Don't restart the whole experiment against the same $2.40 allowance: each new process starts a fresh ledger. Account for the earlier run before allocating a new attempt. Use a machine spend quota when the ceiling must survive process restarts.

To check the refusal path without a provider request, give a single actor only $0.01:

harn run --standalone --no-sandbox examples/conservative-experiment.harn -- --call actor 10000 'Return hello'

The report must contain ok: false, insufficient_allowance, one denied attempt, zero provider calls, and zero settled or uncertain allowance. The supported request's full-context bound exceeds $0.01 even though its prompt is short. A successful four-call run is the positive control for that refusal.

See conservative admission for supported billing shapes and settlement rules. These calls demonstrate allocation and accounting, without establishing actor or grader accuracy.