Where inefficiency can be captured

The value of an action a to an operator o at size q is not its gross spread but what survives the operator's own costs and risks:

V(a, o, q) = G(a, o, q) − C_execution − C_operations − C_funding − E[L] − C_exit − ρ(a, o, q)

The cost terms carry the operator's index. Two operators can rationally assign different values to the same lot because their trucks, storage, capital and exit channels differ; this is comparative advantage, and it needs no forecast to exploit. The gross value G and the expected loss E[L] are estimates, and estimates are where other participants err. The size term q makes the action space combinatorial: a thin book or an ageing inventory turns one nominal deal into many economically distinct ones — which is why the Menus table carries one row per quantity option.

An efficient-markets reading (Fama, 1970) says neither source should survive. The literature since then explains why both do: information is costly to acquire, arbitrage is costly to execute, quality is hidden, attention is bounded, histories are selected, stock perishes, and size moves price. Each is a named mechanism with a named author, and each maps to a reject reason.

The two sides decay at different rates. Attention is the cheapest inefficiency to remove: once evaluation costs fall, the untested part of the menu gets tested, and Grossman–Stiglitz predicts the residual shrinks toward the new cost of information. Estimate-side edges last while histories stay private and adverse selection stays uncorrected. Cost-side edges last as long as the marginal participant must own the physical or legal means of execution; Shleifer–Vishny limits do not move when models improve. The durable position is therefore a cost advantage the operator already holds, plus a model that reads it — not a model alone.

Partial observability is a consequence, not a criterion. When the asset is consumed — eaten, installed, soldered in, retired — the only true outcome is the sale to the end user or the write-off, and it arrives after the decision. Consumption therefore produces the delayed, path-dependent label that selection bias feeds on; it is why the Sales table is an event tape and not a revenue total. It also removes supply, which is what stops open capital from competing the edge away; assets with a consumer are structurally less efficient than assets that are merely traded.

Sizing is part of the estimate. Because impact and capacity enter the value function through q, the question is never "is this deal good" but "at what size is this deal good for this operator". A model that predicts unit economics and treats quantity as an afterthought will be right on average and wrong at the margin that matters.

The falsifiable statement. For a market–operator pair where actions can be enumerated, outcomes attributed, constraints encoded and feedback repeated, a decision system fitted to the operator's own records should approach the best policy measurable from those records, and should beat the operator's prior policy mainly by declining false positives. The claim is weakened for a given market when no model beats a simple baseline after costs, when performance disappears once post-decision leakage is removed, when recommended actions rely on unsupported regions, or when the policy's own impact invalidates history faster than it can be refitted. No-free-lunch and no-arbitrage results still bind: compute cannot recover information the records never contained.