pi-fold: Ephemeral Retrieval

An agent session pays rent on its context window: every held page is re-sent with every request, whether or not the model reads it again. pi-fold folds finished stretches of a session behind short placeholders, keeps every byte, and lets the agent read any fold back on demand. The recovered page then becomes rent of its own, because a recovery read lands in the projection and rides there until folding catches up with it. This third report measures the reading side and adds the ephemeral peek: a read whose result is visible for exactly one model generation and is then replaced by a one-line placeholder, with the underlying bytes intact and recoverable. A frontier model met the option in a tool description, with nothing in the workload or the prompts naming it, and took it on six of its seven reads at first exposure. At a withheld exam it chained three large reads back to back, where holding them durably would have put the projection within half a percent of the serving budget. The mechanism's intended cost is one substitution at the tail, and live probes on the same provider confirm the wire honours it, retaining all 3,840 measured stable-prefix cache tokens beneath a replaced suffix in every repetition. This campaign's runtime did not honour it: its projection freeze invalidated on the edit and rebuilt the whole prefix, so three of six withdrawals rewrote a cached prefix they should not have touched, 55.1 percent identical where the geometry requires 99.5. That is a defect in one runtime, not a property of the design. The report gives the root cause, the repair, the gate that fails on the unfixed runtime, and a later sealed run on the repaired runtime. Cache is served per request while a withdrawal is an event in the session, so that run's twenty-three withdrawn results are reported at both units: nineteen of twenty-three retained cache at a median share of 0.93, and on the fifteen distinct requests those results sit on, which is the unit in which the observations are independent, twelve of fifteen retained more than half at a median of 0.80, with three retaining nothing for reasons the record does not explain. On the billed ledger the folding arm ran the twelve-run workload for $144.15 against native compaction's $278.32, and the gap is mostly one boundary: folding crossed the provider's long-context price tier zero times in 888 billed requests, where compaction crossed on 333 of 630. The surcharge alone is $116.39, forty-two percent of the compacting arm's bill. Priced entirely at base rates the two arms sit about twelve percent apart, $161.93 against $144.15, so the reportable result is that in these twelve runs the observed cost gap was dominated by native calls crossing that price tier. The fold ceiling functions as a tier cap partly because of the configured budget and not by the folding mechanism alone, so every dollar figure here travels with its base-rate counterfactual. Two further results are bounded on purpose. The first reading of the exam credited the model's correct answers to retention in its encrypted reasoning; the project's own attribution instrument overturned that reading, finding one earlier durable read still visible at the answering request and carrying every value. And the campaign's twelve-run exam is retired: the successor campaign established that it scored most of its points on material a compaction summary transcribes. The report prints all four channels the exam scored, two of which favour compaction, and keeps from it what survives, which is the billing anatomy above and an archival record of how an instrument fails. No comparative recall claim is made in either direction, and the accounting above is sealed billing for one provider on one date rather than a verdict about the mechanisms. These cases motivate replication; they do not estimate population rates. The record includes a self-contained semantic HTML reading edition, three figures in SVG, PDF, and PNG forms, the outcome-carrying campaign extract, the repetition 7 and repetition 8 audits and their builders, the campaign pricing audit with its explicit run enumeration, the sealed evidence set after privacy review, exact source, and SHA-256 manifests. A verifier of over one hundred gates rebuilds both audits byte-identically from the sealed runs, regenerates every derived table and figure, recompiles the manuscript, and refuses by name the comparative phrasings this revision retired. Verification checks consistency inside the declared evidence boundary; it does not recreate the original provider interactions. This is the third paper of the pi-fold line. It is a preprint and has not been peer reviewed. The paper and author-created research artifacts are CC BY 4.0; pi-fold software in the source snapshot is MIT licensed, and the pinned curl source retains the curl license.