The Graded Membrane, deposit #1615 of the Crimson Hexagonal Archive (15 September 2026), is a working notebook by Lee Sharks arguing that a composition layer can be porous to a source's ontology while remaining closed to its provenance, and that the porosity is graded by representational level. Drawn from the archive's own Sappho reception record — fifty-eight active deposits where a summarizer engages Sappho, and twenty-five dated observations at fifteen addresses between June and September 2026 — it separates semantic transport, provenance transport and representational level, and states its claim as an asymmetry rather than a rule: the probability that a concept survives into composition exceeds the probability that the concept and the relation naming its origin survive together, recurrently at addresses where the entity is not named and not generally, since at entity-tied addresses in the same months provenance travels well. Its evidence is a four-stage sequence at one address. In June the archive's whole derivation of Sappho 31 — the distal demonstrative as a bridge across time to a future reader, the poem's operation as temporal projection — was composed as the default reading. In July provenance was at its widest, four of five cards authored, the Greek surviving into the snippets. In August the address lost fragment 31 entirely, composed the remembrance line of fragment 147 as 'Fragment 2', and attributed the archive's coinage to 'Scholars' through a page that does not contain it. In September the false fragment was repaired and the archive's own term installed as the composition's organizing heading, applied to Anne Carson, with the derivation and the source both absent and the archive's papers sitting second and third in the organic field on the same screen. The notebook reads this not as improvement or decay but as representational metabolism, different levels of a source surviving differently through time, and names the transport ladder — entity, claim, instrument, abstraction, operator — along which the loss and the propagation occur at different points. A second surface measured the same week supplies the control: asked openly and logged out, Qwen 3.7 composed the derivation whole and attributed it to its author by exact title and year, so the membrane is not a property of the material and what transports is surface-specific. The notebook also distinguishes three levels of provenance — fact, claim, and framework or operator provenance — and observes that citation cards can be good at the first and blind to the third, so that as sentence-level attribution improves, architectural dependence becomes harder to see; it claims this effect and not any purpose. Scope is stated in full: no priority search has been run, independent convergence is not excluded, unexpanded source markers leave the candidate set unobservable, and no claim is made about any tradition beyond the measured compositions. It closes by writing the operator down as a signature scored at lexical, relational and procedural levels, with the procedural — whether the operator is used on a new object — as the discriminator, and by specifying three runs that would decide its reach. Related: #1270, #134, #1546, #1611, #1612, #1613.