The Flywheel

A buyer sends out a bracket package on Monday morning. The RFQ record carries the current print, the STEP file, the quantity break, and the ship date, and three suppliers come back with something a human can actually compare. One of them wins the work, ships on time, includes the cert packet, and closes the job with no NCR. A month later a similar bracket comes around, and this time the buyer is not starting from zero.

That is a flywheel — a loop where each successful cycle lowers the cost and the risk of the next one. My claim here: a manufacturing coordination network compounds only when trusted performance changes future visibility and future routing, and when outcomes are not written back to identity and enforced with consequences, the same loop runs backward into noise, ghosting, and bad allocation.

The mechanism is feedback, and I think it is the piece most systems skip. In a two-sided market, value does not come from raw headcount alone; it comes from getting both sides on board in a way that raises the odds of a useful interaction, and governance decides whether those interactions actually create surplus. In industrial work that means the vendor master, the request record, and the response state have to get more informative after every completed job. A bigger network that does not improve matching is just a bigger search problem.

Trust is the first input, because manufacturing work is expensive to misunderstand. A clean profile page is close to worthless on its own. A record tied to a company identity, a quote response, a delivery date, a cert packet, and an NCR history is worth a great deal, because it changes the buyer's estimate of risk on the next PO. Federal procurement has treated this as ordinary operating practice for a long time: FAR says past performance information is relevant to future source selection, and includes conformance to requirements and adherence to schedules, with CPARS as the official source for that information. DoD goes further and uses SPRS to surface on-time delivery, quality classifications, and supplier risk inside award decisions. None of that is exotic. It is a government writing down what happened and then letting it matter next time.

Once trust exists, it should change visibility. When a buyer opens a drawing packet for turned parts, five-axis work, or a welded assembly, the system should not expose that request to everyone with a login and a pulse. Access should widen in proportion to verified capability and observed performance, because markets with quality uncertainty deteriorate when participants cannot distinguish strong counterparties from weak ones, and reputation mechanisms exist precisely to hold off that decay.

Then visibility has to produce structured requests rather than more inbox traffic. A useful request object carries the current print, the material callout, the finish requirement, the inspection expectation, the due date, and the shipping constraint, and it forces a clear response state — quote, decline, or need clarification. When every supplier answers against the same packet and the same response fields, the system starts learning things it could never learn from an email thread: where ambiguity lives, which packets stall, which suppliers answer quickly, and which jobs should never have been routed together in the first place.

Structured requests are what turn visibility into work that actually closes. Send a loose email with two attachments and a vague deadline and you get back delay, phone calls, and mismatched assumptions. Send a complete packet and you get cleaner accept-or-decline decisions and quotes that can be compared without a translation step in between. Every clean cycle leaves a trace in the RFQ record — how long the supplier took to acknowledge, how many clarification turns it took, whether the award converted, whether the shipment landed on time, whether the parts passed inspection on the first shot.

Successful work is what produces more trust, but only if the system closes the loop. A shipped job with a passing CMM report, a matched cert packet, and a clean invoice should strengthen that supplier's position on the next award. A missed due date, a revision mistake, or an unresolved NCR should weaken it. That is why repeat award rate is such a powerful number: it shows the prior cycle changed behavior rather than sentiment, and it mirrors the procurement logic already running inside CPARS and SPRS, where performance history is carried forward into later award choices.

The failure modes are simple and brutal, and every shop I talk to can name at least one of them from memory. Weak identity lets bad actors into the vendor master and poisons the pool. Indiscriminate visibility makes good suppliers spend their days sorting junk requests until they stop looking. A thin packet blows up clarification loops around the print, the rev letter, or the inspection plan. Sloppy closure means the system forgets who ghosted, who delivered late, and who shipped perfect parts. And with no consequences at all, high performers end up subsidizing low performers until the best participants either withdraw or start demanding side channels outside the system.

Which is why governance carries the weight here. Governance means who can see which RFQ board, what evidence expands a supplier's visibility, how long poor performance lingers in the record, how disputes get closed, and which behaviors trigger throttling or removal. Rochet and Tirole's core point was that platform outcomes depend on how the two sides are brought together and governed; CPARS and SPRS are the concrete industrial version of that idea, because performance only becomes useful when it is attached to an identity, written into a record, and allowed to affect the next award.

This is also where the series stops being commentary and turns into systems engineering. The job is to design the loop, instrument the loop, and protect the loop from corruption. Judge a national manufacturing coordination layer by operational measures written to the request record and the vendor record: median acknowledgment time, clarification count before quote, quote-to-award rate, on-time delivery, dispute or NCR rate, and repeat award rate after a first success. If those numbers improve while the request pool stays usable and the best suppliers keep showing up, the flywheel is real. If membership climbs while packet quality, response discipline, and repeat awards sit flat, there is no flywheel. There is traffic.

Implications

The implication runs well past any one marketplace or quoting tool. Industrial capacity does not compound at national scale when every buyer has to rediscover who is real, who responds, who ships on time, and who can be trusted with the harder work. It compounds when a successful PO, a closed cert packet, and a clean inspection result improve the next routing decision for the next buyer, because that is how isolated jobs become cumulative industrial memory.

The political version follows from the same mechanics. A country does not regain self-sufficiency by collecting more supplier profiles and hoping for the best. Self-sufficiency comes from making good performance legible, portable, and consequential inside its own operating boundary. The practical failure mode is misallocation — the wrong shops see the wrong work, the right shops get buried in noise, and the nation reads a coordination failure as a shortage of capacity. Which raises the question I want to take up next: who gets broader access to the loop, and under what rules?

Questions to Ask

1. What facts about a supplier are written back to the vendor master after every awarded PO, and which of those facts actually affect future routing?
2. Does every RFQ packet force a clear response state, such as quote, decline, or need clarification, so that silence is measured rather than ignored?
3. Which metrics on the request record matter most for compounding: acknowledgment time, clarification count, on-time delivery, first-pass acceptance, dispute rate, or repeat award rate?
4. What evidence is required for a supplier to gain broader visibility to higher-value or higher-risk work?
5. How long do missed commitments, revision mistakes, and unresolved NCRs remain consequential in the system?
6. Can the system distinguish growth in useful throughput from growth in noise, using award quality and repeat business rather than profile count and page views?