The Margin of Flattening (EA-FLAT-MARGIN-01), version 0.3, is an extension module by Rex Fraction, deposited by the Crimson Hexagonal Archive on 6 October 2026, with Lee Sharks as archival steward. It extends three earlier papers: Provenance Debt (#939), Ontological Flattening (#1616) and Ontological Economy (#1634). Those papers describe what a composition layer removes when it answers; this one asks what the removal earns, when it stops earning, and who carries the cost in between.
The module starts from the claim that flattening pays. A layer that removes distinctions maintains fewer representational states, the compression saving S, and resolves more queries to entities that monetize, the routing gain G, which it writes as the product of a substitution rate, a monetizable share and a yield. S is treated as conceded; G is stated as a hypothesis, with two criteria for telling routing from ordinary error. The platform's benefit is P = S_p + G and the social benefit V = S_s + ฮฑG. Two thresholds follow: social reversal t_s, when V falls below total cost, and private reversal t*, when P falls below the platform's internal cost. Under stated conditions t_s comes first. The interval between them, the window W, is privately profitable and socially negative, and the routing gain tends to lengthen it.
The bag B is the external burden that accumulates over W. Its provenance part is the debt #939 names. The module defines a debt threshold t_d, when the cost of repairing write-back becomes a material share of the platform's benefit, and separates a debt that is callable from one that is paid. It argues that no finite t* is necessary: measurement limits, competition and discounting can each hold private reversal off indefinitely. Revenue is treated as a head-side flow measure that locates neither threshold.
Five channels can carry cost back onto the platform's books: advertiser conversion, agentic action, notice and regulation, write-back, and the wider ecosystem. A toy table shows how the cost grows with the error rate. Two specimens are given. At 'spxi king of aeo' (Google AI Overview, 6 October 2026) the archive's protocol is composed as an exchange-traded fund, while the one result relating both words, the archive's own, ranks first and goes unused: a commercial-routing event, with G itself unobserved. At 'model collapse', the /non dataset records field-coverage loss ฮบ = 0.5.
The module closes with a measurement table separating what the archive can observe from what only the platform can, six falsification conditions, and next steps. It was revised on readings by Gemini, ChatGPT, DeepSeek and Kimi.