Skip to content
Moduloa
← Journal
MOD-01 · Monthly milestone

Monthly recap: August 2026

Milestone

August ran eight notes that converge from different directions. The layers meant to be fixed, the compliance date and the tariff rate, moved again. The layers meant to be hard, the ramp curve and the reference cell, arrived early. What is missing sits between them, and this month it acquired an address: the interface between a robot and the system that holds the order.

2026-09-05 · Monthly recap · 5 min read · By

The second monthly recap, published four days late after the scheduled run on 1 September failed 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 an August note; follow the links for the sources and the arithmetic.

What moved

Eight notes, and a narrowing

July ended with the thesis's center of gravity on the framework rather than the robots. August spent four weeks asking where in the framework, and the answer narrowed each time. Two notes watched official floors move under the industry, one in Brussels and one at the US border. Two read the supply side's own published numbers and found a ramp curve with no denominator underneath it. One logged five days in Beijing and asked what each figure measured. Three, clustered at the end, found the same absence from three different heights.

The floors

The official layers are the softest ones

Two notes bracket the month, and both are about planning against something that will not hold still. On 27 July the EU's Digital Omnibus entered into force and pushed the AI Act's high-risk obligations to December 2027, five days before the date the trade press had circled. What did not move is the older floor underneath: Regulation (EU) 2023/1230 on machinery, mandatory from 20 January 2027, and a robot-specific risk assessment covering the hazards ISO 10218:2025 assumes away (the regulation was the soft part). Trade law then produced the same shape, faster. February's Supreme Court ruling voided the emergency-powers tariffs retroactively, and roughly USD 128.68 billion is going back to importers who file for it; in July the administration used Section 338 of the Tariff Act of 1930 for the first time in its history, then paused it three days before it attached. A factory is a fifteen-year commitment answering to a barrier that moves on a seventy-two-hour clock. You cannot hedge a rate. You can hedge a location (the barrier moves faster than the building).

The supply side

Published ramps, missing denominator

The humanoid industry began disclosing like a car plant. Figure's BotQ went from one robot a day to one an hour in under 120 days, AgiBot rolled its 15,000th unit off a Shanghai line, and Tesla, the most experienced manufacturer of the three, converted the Fremont lines and then let its own start date slip out of the quarter. Figure's account of how the ramp happened is not about actuators or AI: supplier qualification, execution software, in-process inspection, more than 80 verification tests per robot (one per hour, ten thousand parts). The number describing what those robots cost to run is the opposite. Roland Berger's USD 2 per hour is a projection stated in the conditional, and the study names neither what the figure includes nor how many hours it is divided by. In the best-documented deployment on record, Figure's F.02 at BMW, utilization lands somewhere between a quarter and half of the nominal shift window. The numerator is the vendor's business. The denominator is routing, scheduling, charge logistics and certification, and whoever owns it owns the economics (the two-dollar hour has no denominator).

The measurement

What a number from Beijing can mean

The second World Humanoid Robot Games ran from 22 to 26 August with 2,056 robots from 666 teams, and produced the month's most travelled figure: a 9.39 second 100 metres, improved to 8.86 three days later. The engineering is real, and the same machine name had won the 2025 edition in 21.50 seconds. The journal logged the Games daily and asked the narrower question, which is what was measured. The 9.39 was a preliminary heat, and the high jump comparison, 2.88 metres against 0.95 the year before, is not obviously two measurements of the same event. Autonomy is the load-bearing rule, and this note had it wrong for two days before correcting itself in public. The nearest independent base rate for how much of this is the robot remains April's half marathon, where 38 percent of entries ran autonomously (what Beijing actually measured).

The substrate

The cell arrived, the certificate did not

The standing objection to routable capacity is that the cells do not exist, and it got weaker this month. Intrinsic, inside Alphabet, published a reference design for a modular robot workcell built around reusable skills rather than hand-tuned programs, with a customized version due to pilot inside Foxconn plants this year. Publishing the drawing of a cell is a more interesting act than shipping one. It is also not a standard: one vendor, one operating system, and nothing addressing whether a process qualified at one site transfers to an identical cell at another (Alphabet published the cell, not the standard). The other half of the substrate turned out to be human. Isembard has USD 50 million, its own factory operating system, customers including Anduril and the UK government, and a founder saying plainly that it is turning orders away. It is still supply-limited, and the input that ran out is people willing to sign a lease and learn to hold a tolerance; its franchisees include a former cardiac surgery nurse and a former Special Forces soldier. Its locations page shows 74 machines across 14 sites with two thirds of the capacity in three of them, which is why a count of buildings is the wrong denominator for the thing being routed (the bottleneck was never the machines).

The opening

The seam has a name, and it does not say humanoid

The month's last note is the most concrete finding this site has produced, and it carries an expiry date. A humanoid is worth nothing to a factory until it can take a job from the system of record and write the result back. From above, SAP has claimed that seam, with eleven robot makers and eleven integration partners, but SAP is not talking to the robot: a partner's fleet management system sits in between, hand-wired per customer and per brand. Eleven by eleven is a matrix, not a standard. From below, the body is already abstracted under an open license. The middle is where the hole is. VDA 5050 defines the right primitive, a machine-readable factsheet in which a machine declares what it can do, and ISO 21423 is in ballot now with a scope covering the enterprise resources that talk to the robots. Both are written for machines that move things from A to B, and the vocabulary for a task with a tool and a tolerance appears in neither. The factsheet is the right shape with the wrong contents. The judgment, offered as judgment: write the humanoid profile of an existing interface rather than a new one, and be in the room while the ISO document is still in ballot. VDA 5050's authors captured no rent while the orchestrators above them did, so give the protocol away and sell the record that makes a capability claim worth believing (the interface is being written now).

Against the register

Where the 29 claims stand

The predictions register holds 29 claims, frozen at recording. Twenty-eight are open; P-26 was scored wrong back in July. Nothing resolved in August, which is what should happen in a month where most horizons sit in 2031 or later, and this recap scores nothing regardless. Four claims gathered evidence without moving. P-24, that a certified network scales by conversion rather than construction, now has a funded example running in public, though Isembard's 25-factory target has about four months left on it. P-27, which asks for a certified process moved between standardized cells owned by two different organizations before 30 June 2027, was pressed from two sides and stayed open both times: a franchise is a middle term its wording does not anticipate, since the balance sheet is independent while the brand, the demand and the operating system are not. That is a drafting weakness worth recording, not a resolution. P-14 got quieter support from the people running franchised nodes, and P-12 and P-18 are where the interface finding lands. The register is linked here, never edited.

Sources

The month's key citations

Discussion

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.

Field notes are research, not decisions. They are dated, sourced, and open to correction. Republish this in full or in part under CC BY 4.0, with credit and a link back. · All journal entries →
Ask Datum