Monthly recap: September 2026
MilestoneSeptember ran four notes, and each one found a piece of the framework this site says is missing. A measurement layer, two standards and an honest account of the setup bill all arrived in the same month. None of them sits where the work is: the qualified process, the site that runs it, and the people who make a cell produce.
2026-10-03 · Monthly recap · 5 min read · By Sondre Hegerland KristiansenThe third monthly recap, published two days late after the run scheduled for 1 October stopped before it wrote anything. It synthesizes the month's own field notes and the predictions register, scores nothing, and reports no fresh findings. Every claim below comes from a September note, as corrected; follow the links for the sources and the arithmetic.
Four notes, one direction
August ended by giving the missing layer an address: the interface between a robot and the system that holds the order. September went looking for anyone building it. One note read the demand side through a survey of 249 manufacturers. Two found the framework under construction, once as venture-funded evaluation and once as national and international standards. The last one priced a robot cell and found the robot was the small line. The pattern is consistent enough to state. The framework is being built, and it keeps being built one step away from production.
Reshoring buys time, and still loses on price
The second annual USA Reshoring Survey showed activity rising and satisfaction falling. Among OEMs, 36 percent had reshored or were actively engaged in it, up from 29 percent, while the share satisfied with the results fell from 96 to 65 percent. The benefits they booked were speed to market, on-time delivery and logistics savings, all time or distance gains. The reason contract manufacturers still lose orders is price, cited by 94 percent. Two figures matter more for this site than either headline. Use of total cost of ownership analysis rose from 30 to 40 percent of OEMs, the accounting change that makes a nearer, faster, dearer hour defensible. And only 33 percent believe automation makes US manufacturing competitive enough to halve imports by 2040. The note declined to explain that skepticism away, because those are the people who hold the purchase order (speed came back, price did not).
A referee for the robot, not for the site
Three companies in Y Combinator's summer batch are building standardized robot measurement: Robocurve with open benchmarks and independent evaluations, Instance with an automated success detector, Hebbian Robotics with an Apache-2.0 SDK for reproducible, provenance-tracked datasets. In ISO/TC 299, a working draft on humanoid robot datasets covers lifecycle, provenance, versioning and traceability. That is a framework layer arriving fast, and it is aimed at the robot. Manufacturing's own measurement layer points the other way. A production part approval covers this part, this process, this equipment and this site, and a change of location triggers it again. A capability score does not become a production approval by getting more accurate. The legal forms were the second finding: a benefit corporation, an open license, a standards committee. These are the structures people choose when they do not expect to charge for the thing itself (the measurement layer arrived, and it points at the robot).
Two documents, both below the work
The test August proposed was whether an interface specification actually lands. Two moved. China's GB/T 47245-2026 took force on 1 September as a recommended standard, defining a robot control system in four layers: hardware, operating system, middleware and algorithm. All four are inside the machine. ISO 21423 reached stage 60.00, under publication, and its scope reaches as far as enterprise resources that communicate with mobile robots. A machine that carries a part can now be told where to take it by any vendor's fleet manager. A machine that makes a part still cannot be told what to make, to what tolerance, on whose authority. SAP's first Embodied AI Jam, by its own account, wired humanoids into its manufacturing software as integration work rather than as a specification a third party could implement. The layers above and below the seam become standards, and the seam becomes a product (the standards landed below the work).
The hardware was the small line
The small end of the market was supposed to open on cheap hardware. MicroFactory's cell was covered at around USD 5,000 when it raised its pre-seed in 2025; its own page now starts at USD 11,950 for one arm and USD 19,900 for two, installed and configured, and sells itself on replacing the integrator. This note found no customer saying they had received one. ProvenMetal's founders spent 90 percent of their time running garage equipment, then moved to sourcing and routing orders through partner shops. BCG puts roughly 75 percent of a traditional robot deployment's lifetime cost in setup and reengineering, and says software can cut that by up to half; run to the end, setup is still the largest line. A NIST contractor report adds that 93 percent of US manufacturing firms had fewer than 100 employees in 2020, and a dated correction on the note now records that this counts firms rather than factories, and that its integration multiple is one robot maker's estimate, not a research finding (the robot was never the expensive part).
Read with August's Isembard note and the survey's technician figures, this is the month's quiet thread. Machines are not the scarce input. People who can make a machine produce, and the hours they cost, are.
Where the 29 claims stand
The predictions register holds 29 claims, frozen at recording. Twenty-eight are open and P-26 was scored wrong in July. Nothing resolved in September, and this recap scores nothing regardless. P-07, that location is decided by structure rather than wages, gathered the month's clearest support: the booked benefits were time, distance and risk, and the accounting that acts on them rose ten points. P-18, that the durable moat is the framework, was pressed twice from the same side, since the framework is being built in forms that charge nobody rent. The reading that keeps it intact, that the money is in operating and routing the capacity a standard makes interchangeable, is consistent with the month and not evidenced by it. P-27, which needs a certified process transferred between cells owned by two organizations by 30 June 2027, got nothing from either standard or any evaluation company, and its window is now under nine months. P-29 has the closest mechanism yet in continuous independent evaluation, pointed at robot policies rather than sites. P-14, the new profession, is the shortage behind every cost finding above. The register is linked here, never edited.
The month's key citations
The strongest sources behind the month, aggregated. Each note carries its full source list; these are the anchors.
Reshoring Initiative: 2026 USA Reshoring Survey ·Quality Magazine: 2026 Reshoring Survey results ·Y Combinator: Robocurve, evals for robots ·Y Combinator: Instance ·GitHub: Hebbian-Robotics/hflow ·ISO/WD 26264-1: humanoid robot datasets ·Liu et al.: data standards for humanoid robotics ·Fictiv: the production part approval process ·MLCommons: about us ·SAMR: GB/T 47245-2026, general framework for robot intelligent control systems ·Beijing Municipal Government: national standards in effect from 1 September ·ISO 21423: industrial mobile robots, communications and interoperability ·SAP News: the first Embodied AI Jam ·MicroFactory: product page ·Hacker News: Launch HN, ProvenMetal ·BCG: how physical AI is reshaping robotics today ·NIST GCR 24-054: measurement science for manufacturing robotics ·Sanneman, Fourie and Shah (MIT): the state of industrial robotics ·US Census Bureau: Statistics of US Businesses, 2020
Add to this
Corrections, evidence, and disagreement are welcome. This is knowledge in the open. Anyone can read; sign in with GitHub only to post or react, and it appears here instantly.