Evidence flows into the sandbox freely. Nothing leaves without paperwork.

The asymmetry

The two directions carry different risks, so they get different permissions.

Consider what each direction costs when it goes wrong. Production evidence flowing into the sandbox too freely costs nothing — the worst case is a researcher spending time on a sample that turned out to be uninteresting. Sandbox conclusions flowing into production too freely costs a rule change made on untested evidence, applied to live capital, with the previous state possibly not recorded. Those are not comparable exposures, so they do not get comparable gates. The inbound side is open by default. The outbound side has a defined workflow, a persistence window and a written verdict, and there is no informal route around it.

Why inbound is open

A sandbox fed on old data researches a market that has stopped existing.

The temptation with any isolation boundary is to make it symmetric, because symmetry looks like rigour. It would be a mistake here. Research value depends almost entirely on the evidence being current: a coefficient study run on last year's volatility regime, a branch-weight comparison run before a structural shift, an execution analysis conducted on a sample that predates a change in spreads — each of these produces a clean, well-formed, useless answer. Keeping the inbound channel unobstructed is what makes the sandbox worth having, and it costs nothing because a research conclusion has no authority until it has been promoted.

Why outbound is gated

The dangerous property of a sandbox finding is not that it is wrong. It is that it is persuasive.

An experiment produces a number, and a number carries authority that its provenance does not. By the time a finding has been discussed twice it has usually shed the filters it was computed under, the sample size it rested on, and the fact that three adjacent configurations were tested and this was the one that looked best. What survives is the conclusion, stated confidently, with none of the fragility attached. The outbound gate exists to keep the assumptions travelling with the finding — which is why anything leaving carries its evidence window, its sensitivity ranges and the conditions under which it held.

The one-way rule in practice

The production surfaces have no field a sandbox result can be entered into.

A boundary maintained by remembering to respect it does not survive an exciting result. This one holds structurally: the Legend constants, the branch weights and the rule set are edited deliberately and rarely, and there is no path by which a sandbox output writes to them automatically. The gap is the enforcement. An operator who has just found something promising cannot act on it in the next cycle even if they want to, because acting on it requires an edit that is itself an event — dated, deliberate, and visible in the next review.

• Sandbox outputs are sensitivity and comparison, never evidence of a live change.
• Nothing writes to production automatically; every promotion is a deliberate edit.
• A rejected candidate stays in the library with its verdict — negative results are results.

What this costs

The workflow is slow on purpose, and the cost lands on the ideas that were right.

It is worth being honest about the price. A good idea discovered on a Tuesday does not reach production that month; it waits out a persistence window, and during that window the operator watches a rule they now believe is inferior continue to govern their trading. That is a real cost and it is paid disproportionately by correct ideas, since wrong ones would have failed the window anyway. The reason to pay it is that the operator cannot distinguish, on Tuesday, between the ideas that will survive and the ones that will not — and a workflow applied only to the ideas one is unsure about is a workflow applied to nothing.

The key idea

Isolation is only meaningful in the direction where the damage travels.

The general form is worth carrying beyond this layer. Any boundary between a research surface and a production surface should be examined for which way harm actually flows, and gated in that direction only. Gating both directions produces research nobody can do; gating neither produces production nobody can trust. The sandbox is permeable where permeability is free and rigid where it is not, and the rigidity has to be structural rather than remembered, because the moment it is most needed is the moment the operator is most convinced it can be skipped.