The Compression Paradox (kunnas.com)

People will hear "compression" and reach for information theory. That is a different machine.

Shannon's 1948 paper in the Bell System Technical Journal opens with the job: reproduce at one point a message selected at another. Messages often have meaning. Then the cut: "These semantic aspects of communication are irrelevant to the engineering problem." A good Shannon compression is one you can invert. Unpack the code, get the bits back — or a specified approximation, in the later rate-distortion work.

Kolmogorov's 1965 note, "Three Approaches to the Quantitative Definition of Information" (Problems of Information Transmission 1:1), measures the information in one object as the length of the shortest program that prints it. Grünwald and Vitányi (arXiv cs/0410002) put the two side by side: Shannon entropy versus Kolmogorov complexity, rate-distortion versus Kolmogorov's structure function. The structure function's "meaningful information" is still the regular part of a string you can split from noise. Success is still: the short description regenerates the object.

The essay's paradox runs on the opposite test. "Survival of the fittest" is short. Unpacking it does not recover Darwin. It recovers strongest-wins. The portable phrase works as a substitute because it is not invertible to the mechanism. Shannon set meaning aside on purpose. Mapping the word "compression" onto Gentner's two operations is itself taking the surface of the word and dropping the working parts.

Wells Fargo's 2004 annual report called cross-selling "our most important customer-related measure" and set the goal at eight products per customer — "Going for Gr-Eight." That is a named dashboard, not a slogan in a textbook.

The Independent Directors' Sales Practices Investigation Report (10 April 2017, Shearman & Sterling) describes the live ranking tool. Daily and monthly Motivator reports carried monthly, quarterly, and year-to-date sales goals and ranked people and districts against one another. Witnesses said some managers "lived and died by" the Motivator results. The reports were dropped in 2014 after Leadership Summits where regional leaders asked to kill them because of the shaming. Retail scorecards, put in by Carrie Tolstedt, were updated daily against the sales plan.

They already had the residue sitting next to the count. Rolling Funding Rate tracked whether a new checking or savings account received more than a de minimis deposit. Tolstedt's own 2004 email said that looking at one metric alone asks for "low value, unfunded bad cross sell." The count still governed. The Board learned from the September 2016 CFPB, OCC, and Los Angeles settlements that about 5,300 employees had been fired for sales-practice violations. Product sales goals were eliminated on 13 September 2016.

The compression mapped how many products a household looked like it had. What dropped was whether the customer asked for them, needed them, and put money in. That is the essay's attribute substitution with a named load-bearing remainder, on a ranking report.

A competing account of why the bad phrase travels: receivers rebuild, they do not unpack.

Dan Sperber's Explaining Culture (1996) treats cultural items as reconstructed at each step, not copied. Claidière, Scott-Phillips, and Sperber (Phil. Trans. B 369:20130368, 2014) call the clustering cultural attraction. The pumpkin coach in Cinderella survives oral retelling because it is easy to rebuild from a hint, not because someone compressed the tale well. Bartlett's serial reproduction method in Remembering (1932) is the old lab version of the same drift.

The essay's operator is the writer's mapping kind: write an invitation, not a replacement, and the relation can travel. Attraction says even a good incomplete phrase will be retold as the salient picture — strongest-wins, markets-run-themselves — because the listener is building a version.

They disagree on the repair. If mapping kind is doing the work, "differential reproduction under constraint," left incomplete, should unpack to Darwin more often than "survival of the fittest." If attraction is doing the work, retellings of both drift toward the fight. Discriminator: a Bartlett chain of the two Darwin compressions, scored on whether unpacking still recovers variation, heredity, and differential reproduction after N retellings. Same cut on Wells: the relational metric was already in the building. If the count still ranked people, the selector beat the better compression that was already on the page.

The four-feature test in the design section is already a checklist. Run it on two objects from the same bank, not on Darwin again.

Motivator / products-per-household. Does the number invite the model or replace it? Unpacking it, do you get need, consent, and funded use, or "more products is more relationship"? Does it predict something that can fail — unfunded accounts, unauthorized accounts? Does "eight" stay inside retail banking, or travel because it rhymes with great?

Rolling Funding Rate. Same four questions. FTI's analysis in the 2017 report already showed January funding rates below the monthly average when the "Jump into January" campaign was on — the quality number moved when the pressure did.

The file is public: 2004 annual report plus the board report. If funding rate scores as the map and Motivator as the postcard, you do not need a new theory. You need to say which number you are willing to rank people on.

One question would change what I do with this.

When a relational compression and an attribute compression of the same mechanism are both on the desk, which one is allowed to rank? If the answer is "whichever unpacks to the working parts," Motivator was a writing error and the 2004 report could have been fixed by putting funding, consent, and need on the same page as the count. If the answer is "whichever is more rankable," the invitation hierarchy is still a real writing rule and not the load-bearing repair. That is the split between rewriting the slogan and changing what the scorecard may show.