Chain Design

The object

A chain is practices composed in sequence: one stage's output is the next stage's input. Design decides a spec; engineering builds to the spec; the built thing constrains whatever comes after. Each link takes what the last link produced and produces what the next link takes.

Chains of this kind are the ordinary shape of work. Requirements → build → operate. Author → editor → reader. Clinician → protocol → patient. In each, an upstream practice fixes an object — a spec, a manuscript, a protocol — and a downstream practice works inside what that object fixed. The chain is usually inherited rather than chosen. This note treats it as chosen.

Order matters

Composition carries its order: for functions in general, g∘f and f∘g are different objects — textbook mathematics; composition is non-commutative. Applied to practices: designing-then-engineering yields a different thing than engineering-then-designing. Design first and the spec constrains the build. Build first and the artifact constrains the spec — the design becomes a description of what already exists. Same two practices, two products.

So the sequence is a choice, and it is the first design decision. A chain is chosen link by link, and chosen again in its order.

Interfaces filter

A link passes to the next link only what it can state. The interface between two practices is a made object — the spec, the document, the handoff note — and it carries exactly what fits in it. What a practice can formalize crosses. What it cannot formalize stays behind.

Polanyi named the remainder: the tacit dimension, the part of a skill its holder cannot fully state (The Tacit Dimension, 1966). A tacit-dependent stage has no way to hand its tacit part forward; the interface has no field for it. A nearby note on this site works the same boundary — a practice's formalizable part runs without the practitioner, its tacit part cannot (when thought becomes artificial). Kin, argued there; used here as a premise about handoffs.

The spec crosses the handoff. The skill stays with the hand.

Links degrade

Loss per link is measured, and has been for nearly a century. Bartlett passed material along chains of people — the serial reproduction method, most famously run on the folk tale "The War of the Ghosts" — and recorded what each handoff did to it: reproductions shortened, dropped detail, and conventionalized toward the receivers' existing schemas (Remembering, Cambridge University Press, 1932). His own terms were conventionalisation and rationalisation — at every link, the unfamiliar was rewritten into the familiar. The caveat: the 1932 study was loosely controlled — small samples, no objective scoring — and the finding stands today on later controlled transmission-chain work that reproduces its direction (Mesoudi & Whiten, 2004; 2008). Per-link loss is an experimental result, and its shape is specific: chains do not merely blur what they carry, they bend it toward what the receivers already expect.

The chain's shape also prints onto the product. Conway: "organizations which design systems … are constrained to produce designs which are copies of the communication structures of these organizations" ("How Do Committees Invent?", Datamation, 1968). Brooks cited it as Conway's law (The Mythical Man-Month, 1975) and the name stuck. Its empirical standing is mixed and worth stating plainly: tests of the mirroring hypothesis find it "a prevalent pattern but not universal" — supported in industry and single-firm studies, largely unsupported in open collaborative software projects (MacCormack, Rusnak & Baldwin, 2012; Colfer & Baldwin, 2016). Prevalent, contested at the edges, and real enough to design against: who talks to whom along the chain shows up in the thing the chain makes.

Closing the loop

Control theory supplies the completing fact. An open-loop system runs without measuring its output, so its errors go uncorrected; a closed loop measures the output against what was wanted and feeds the difference back as correction (Åström & Murray, Feedback Systems, Princeton University Press, 2008). In a chain of practices the uncorrected errors are Bartlett's per-link losses, each link adding its own on top of what arrived — an open chain compounds them, link by link, with no point at which the loss appears as a signal. A loop makes the loss measurable. That is the structural difference: a longer chain adds links that lose; a loop adds a measurement that corrects.

The consequence is a design rule. Where correction matters more than throughput, close the loop rather than lengthening the chain. Length adds capacity and adds error; measurement adds neither, and removes error.

The proposal

Chain design is three choices and one judgment.

• Choose the links. The fewest practices that do the work. Every link is a place to lose.
• Choose the order. Composition is directional. Decide which practice constrains which, on purpose.
• Choose the interfaces. Decide what must survive each handoff and build the handoff to carry it — and keep tacit-dependent stages adjacent, in one person or one room, so the un-statable never has to cross a document.

The judgment: when the cost of accumulated error exceeds the cost of measurement, stop chaining and close the loop. Add the measurement, feed it back, shorten the chain.

The limit

Chains trade correction for specialization, and the trade often pays. Porter's value chain describes value created precisely by specialized activities in sequence — each link does its one thing well because it does only that thing (Competitive Advantage, Free Press, 1985). The filtering is itself part of the value: an editor removing is the value; a spec forcing statement is the value; the interface's discipline is part of what each link buys. Filtering is loss and refinement at once, and Bartlett's chains show both faces — conventionalization is degradation from the sender's side and legibility from the receiver's.

So the claim is bounded. Chains are the right structure where specialization is worth more than correction; loops are the right structure where correction is worth more than throughput. The design question is which trade a given stage is making — decided per chain, per handoff, and never in general.

Kin to When Thought Becomes Artificial — the formalizable/tacit boundary the interfaces section leans on — Tacit — the remainder itself — Engineering Therapy — the process engineered, the core held — and The Recursive Ladder — the minimum depth of self-reference.

Rests on: Bartlett, Remembering: A Study in Experimental and Social Psychology (Cambridge University Press, 1932) — serial reproduction — with its modern controlled replications in Mesoudi & Whiten, Journal of Cognition and Culture 4(1) (2004) and Phil. Trans. R. Soc. B 363 (2008); Conway, "How Do Committees Invent?", Datamation 14(4), 28–31 (1968), with its mixed empirical standing per MacCormack, Rusnak & Baldwin, Research Policy 41(8), 1309–1324 (2012) and Colfer & Baldwin, Industrial and Corporate Change 25(5), 709–738 (2016); Polanyi, The Tacit Dimension (Doubleday, 1966); Åström & Murray, Feedback Systems: An Introduction for Scientists and Engineers (Princeton University Press, 2008); Porter, Competitive Advantage (Free Press, 1985); and the non-commutativity of composition, which is textbook mathematics. All of these are established prior art. What is proposed is only the framing — a chain of practices as a designed object with three choices and one judgment — offered to be argued with.