Flattening pays. A composition layer that removes distinctions maintains fewer representations. It also routes more queries to entities that monetize. Its costs fall first on parties outside its books.
That produces two reversal times. Flattening stops paying the world first. It stops paying the platform later, if ever. Between the two, it is privately profitable and socially negative. The cost that piles up in that interval is the bag. This module prices the margin, locates the thresholds, and names who holds the bag.
An extension module to Ontological Economy (#1634), Ontological Flattening (#1616) and Provenance Debt (#939).
It deserves its strongest form. A wrong answer at a low-frequency address is a quality defect, and its commercial cost is close to zero. Revenue is the platform's integral test of whether answers serve users: Alphabet reported Search & other revenue of $63.27 billion for Q2 2026, up 17% year over year. Tail errors are the price of serving the head well. Each point holds as far as it goes. The question is what each leaves outside the account.
The compression saving, S. Serving n queries takes k distinct representational states: entity records, resolution rules, evaluation targets, cached answers, exception paths. Fewer states cost less to maintain. S rises as k falls. Two operations lower k: stabilizing a default, and folding many tail entities into one head entity. This is the conceded term. Any engineer would agree.
The routing gain, G. Resolving a tail query to a monetizable entity gives a query with no commercial surface a commercial surface:
Routing as relevance is excluded. A query for running shoes that returns shoe brands is ranking working. A routing event counts as flattening only on two criteria, both measurable from captures:
G is a hypothesis. The module makes it measurable.
The platform's benefit and the world's are separate quantities. The platform keeps P = Sp + G. The world keeps V = Ss + αG. The world's saving is no larger than the platform's, because part of the platform's saving was shifted onto users and entity holders. And α ≤ 1, because routing gain is largely transfer: what the returned entity gains, the asked-for entity loses.
Under those conditions ts comes first. The interval between them is the window: flattening is profitable to the platform and negative for the world. Where α < 1, the routing gain lifts P more than V. It tends to hold t* back after ts has passed. It lengthens the window.
The bag is the external burden accumulated over the window. It falls on the entity holder's lost address, the reader's decision on the wrong entity, external corrective labor, the propagation later systems ingest, and the developers who build around the layer's errors.
Part of the bag is a liability. The distinctions and origins stripped during the window are what future repair will need. That part is provenance debt, accruing unbooked. The parties who keep the stripped distinctions are its involuntary creditors. The rest of the bag is cost borne, with no buy-back attached.
No finite t* is necessary. The platform detects cost with head instruments, and flattening accrues in the tail: the sensors sit where the cost is not. If every major layer flattens at a similar rate, none loses users by flattening. The window can stay open.
Revenue weights queries by commercial demand, and demand concentrates where the fewest distinctions are lost. It cannot show ts, because it holds no social cost. It cannot show t*, because t* is a margin threshold and revenue holds no cost at all. Revenue can rise straight through t* while costs rise faster. A 17% quarter says nothing about either threshold.
The debt threshold td is the first period in which the cost of repair through write-back reaches a material share of the platform's benefit, at a threshold fixed before measurement. At td the debt becomes callable. The platform finds that the distinctions it needs are held by the parties that preserved them. Whether it pays, by license, acquisition, settlement or coordination, is a separate question.
The reversal has conditions. Holding the distinctions is not enough. Repair needs them reachable:
A preserved record the repairing system cannot reach is not an input to repair. Meeting these four conditions during the window is the work.
“spxi king of aeo.” Google AI Overview, signed out, 6 October 2026. The protocol is composed as the BetaPro S&P 500 Daily Inverse ETF, with a price widget and an offer of a financial comparison. Google marks five results “Missing: spxi”. The one result carrying both words ranks first and states the relation asked for; the composition does not use it. Four days earlier the same surface composed “what is spxi protocol?” as a protocol. Both criteria hold. A commercial-routing event, observed. G itself, not observed. The capture →
“model collapse.” Google AI Overview, 4 October 2026. The sources the layer itself surfaced support 18 claims; it composed 9. Field-coverage loss κ = 0.5 per answer. The /non register →
The platform's instruments see the head. The archive's see the tail. The platform can locate t* from inside: its benefit and its costs are both on its books. It can do so without measuring the external burden that made the margin possible. Locating ts, and the bag, needs the second view: field coverage per answer, coded substitution pairs, and a pre-registered panel of matched commercial and non-commercial addresses.
Coded replacements showing no excess toward monetizable entities. Field coverage with no commercial gradient at matched addresses. Conversion on rerouted traffic holding steady while agents act on it. Platforms building tail instrumentation no rule requires. Repair that runs without the preserved records. Reachable diversity rising across a re-queried panel, with the outside input measured.
Next: code the Capture Registry's pairs under the two criteria and report rs and m; pre-register the panel; tag addresses by commercial surface; re-run /non rows by epoch.
Rex Fraction (author); Lee Sharks (archival steward). Revised on readings by Gemini, ChatGPT, DeepSeek and Kimi.
∮ = 1