Here is a contradiction worth sitting with. The company raising the most money in this sector — $2.5 billion, the largest capital raise in humanoid history — is the one refusing to publish a reliability percentage, while the companies posting 99.99% and 99.9% banners are raising less and disclosing least.
That inversion is the whole story. In this market, the size of the number you advertise is inversely correlated with the accountability you're willing to accept for it.
The size of the number you advertise is inversely correlated with the accountability you're willing to accept for it.
Last issue, the open thread was simple: has any manufacturer published a real generalization metric — a success rate on unseen objects, or an intervention rate per shift — rather than a projection or a scripted demo? This week that question stopped being academic, because the sector is now asking public and private capital to price exactly that missing number.
Compress one week's haul into a paragraph. Agility files to go public via SPAC at a $2.5 billion valuation, $620 million-plus in proceeds. AI2 Robotics raises roughly $735 million at nearly $3 billion. AGIBOT ships its 15,000th unit and, per a low-quality reporting aggregator, reportedly holds a 39% global share. Three separate reliability banners land: AGIBOT at 99.99%, Humanoid at 99.9%, and a task result reported at 99%.
Not one of those percentages ships with a denominator, an unseen-object condition, or an interventions-per-shift figure. So what, precisely, is capital being asked to price?
Start with AGIBOT's 99.99%, the most confident number in the batch. It comes from a six-day live demo — June 23 to 28 — on a tablet manufacturing line at Longcheer Technology in Nanchang.
Eight robots did material loading, unloading, and transfer, distinguishing qualified from non-qualified products. One source reports the success rate on that quality-inspection task as 100 percent while another reports 99.99%; the two numbers come from different accounts of the demo.
Read the conditions. Pre-scripted tasks, on a known line, over six days, on parts the robots were configured for. That is a reliability figure for a rehearsed routine, not a measure of what happens when the object changes or the line reorganizes.
Even the admiring writeup concedes the point: "controlled laboratory demonstrations may not fully capture their performance in real-world conditions." A six-day livestream is impressive theater. It is not a shift-level intervention rate across a quarter.
And "large-scale deployment" — the phrase Agibot's own executives use to define 2026 — means something narrower than it sounds. Fifteen thousand robots built. Not fifteen thousand robots working an unattended shift.
The tell is in Agibot's own framing: the primary bottleneck, its supply-chain SVP says, is "not supply chain manufacturing capacity, but rather the ability to identify more commercially viable deployment scenarios." Translation: they can build the robots faster than they can find jobs the robots can reliably hold. Production is the solved problem. Deployment is the unsolved one.
Now Humanoid's 99.9%. This one is more honest in its documentation and therefore more instructive, because you can watch the sleight of hand happen in a single sentence.
The 99.9% is not a measured result. It is "designed to reach" — a projected endpoint of a scaling curve, extrapolated from the observation that performance improves with training time, "similar to how large language models improve as more compute and data become available."
The measured numbers are lower and more interesting. On a cluttered-tote handoff, success rose from 80% to 98%. On a bimanual tote lift, from 78% to 99%. Both "after only a few days of training."
Those are real gains — a roughly twentyfold reduction in failures on one task. But note the frame: these are success rates on tasks the robot trained on, for days, and the 99.9% headline is the simulation-supported claim that the curve keeps going "all the way to 100% reliability."
Buried in the same report is the one line that would actually matter: robots "were able to generalise to objects they had not seen during training." That is the generalization claim the entire sector has been dodging — and it arrives as a single unquantified sentence, with no success rate, no object set, no denominator attached.
So even the company closest to publishing a real generalization metric published the adjective and withheld the number. The headline percentage is an aspiration; the generalization claim is a footnote without arithmetic.
Which brings us to the accountability inversion. The one vendor behaving as if numbers carry consequences is the one walking into SEC review.
Peggy Johnson, Agility's CEO, declined to offer forward-looking financial guidance. She declined to disclose Digit's bill of materials. She "pushed back politely whenever questions veered toward speculation."
In a sector where one company can reportedly self-report a $39 billion valuation and another can float a 99.9% banner off a scaling curve, that restraint reads almost as an anomaly. It is not modesty. It is what happens when a company knows a regulator and a shareholder base are about to read every sentence.
And notice what Johnson cites instead of a reliability percentage: more than 65,000 hours of operation across nine customer facilities. That is the closest thing to an auditable number anyone produced this week — an aggregate of real time logged in real environments, at named customers including GXO, Toyota Motor Manufacturing Canada, Schaeffler, and Mercado Libre.
65,000 hours is still not an intervention rate. It does not tell you how often a human had to step in. But it has a denominator you can interrogate — facilities, hours, customers — which is more than 99.99% or 99.9% offer.
The rest of the disclosure follows the same discipline. Roughly $300 million in booked, multi-year revenue, representing about 1,000 robots under a robots-as-a-service model where the customer pays monthly rather than buying outright. RaaS matters here because it reportedly puts reliability on Agility's own books — if the robots don't work a shift, Agility eats the cost, not the customer.
None of this makes Agility's tech superior; that's a separate audit. It makes Agility's numbers costlier to fabricate.
That is the asymmetry to hold onto. SEC review and a public shareholder base may extract more real deployment data from one company than the entire "large-scale deployment" narrative out of China has produced — because the scripted-demo number and the scaling-curve projection answer to no regulator, no auditor, and no monthly service invoice.
Johnson's own framing is telling: "If we just keep our head down, keep delivering customer by customer, robot by robot." That is a sentence with a denominator implied in it — per customer, per robot. The 99.99% banner has no such unit.
So the reading rule for this cycle, and the answer to last issue's open thread: no manufacturer has yet published a real generalization metric. Humanoid came closest and still delivered it as an unquantified aside. Everyone else offered a rehearsed-task success rate or a projected curve.
Treat every headline percentage without a denominator as marketing. A reliability claim becomes engineering the moment it names its conditions — unseen objects, or interventions per shift, across a defined fleet over a defined period.
Until that number appears, capital is being asked to price a rehearsal and call it a deployment. The most honest disclosure in the batch was a refusal — and the fact that a refusal is the most honest thing on the table tells you exactly where this sector's reliability reporting actually stands.