A prediction that can't be checked is a mood. This page lists every bet this newsletter has made in public, what would prove each one wrong, and how often I look. I score them in the open: right or wrong, the result goes here, and a scorecard post goes out every quarter. The first one is due in December 2026.
How to read it: each entry says what I bet, what would count against it, where I made it, and where it stands today. Statuses are open (no decisive evidence yet), waiting (the test hasn't been run or the clock hasn't started), supported or against (with the evidence linked), and settled.
Factories and the agent loop
P1 · Any step that can't sit inside an agent loop gets squeezed out, fast. A step must be callable by a program, able to hand over state, and checkable by something independent of the agent. Revise if: into 2027 a major vendor's agent still works only inside its own IDE; a mainstream VLA enters a ±0.3 mm process station with Cpk data; virtual PLCs haven't reached mainstream production lines within three years; or two years on, agent calls still stop at the engineering desk instead of production runtime. From: Is Manufacturing at Dusk or Dawn?, Aug 25, 2026. Status: open; Anthropic's MHS standard for callable lab equipment (Aug 2026) counts as early support.
P2 · Organizational shape follows coordination cost. Wrong if, five years after mainstream industrial agent products reach general availability, plants that adopted them show no statistically visible change in management depth or crew size. From: The Shape of Organizations, Sep 1, 2026. Status: waiting (clock starts at GA).
P3 · Coordination comes before muscle. Wrong if products that replace muscle head-on (humanoids swapping whole lines of people) reach meaningful revenue at scale before coordination-layer products do. From: The Shape of Organizations, Sep 1, 2026. Status: open, no counterexample yet.
P4 · The sixty-year line hasn't moved. Wrong if wire-harness installation gets fully automated in volume production. From: The Shape of Organizations, Sep 1, 2026. Status: open.
P15 · Precision stations need scaffolding, not just a smarter model. Wrong if a ±0.3 mm process enters production on a stronger model alone, with no simulation, no dry run and no in-line re-inspection engineered around it. From: Engineering the Map, Sep 22, 2026. Status: open.
Lab and equipment interfaces
P6 · State handover is the gap. Without a standard place to hand over task state, MHS-style agent control stays limited to short, single-scope tasks in manufacturing, and that gap becomes the opening for competitors and extensions. Checked by: whether the open-source schema defines state handover. From: MHS: Three Questions, One and a Half Answered, Aug 27, 2026. Status: waiting.
P7 · Someone will build the independent check. Between device drivers and the cell's safety model, there is no standard layer that checks what an agent tells physical equipment to do; a dedicated product or standard will appear to fill it. Checked by: a cell-level validation product or draft standard. From: MHS: Three Questions, One and a Half Answered, Aug 27, 2026. Status: open.
P8 · On the shop floor, agents start with the slow loop. If MHS-style control reaches factories, its first contact-task jobs will be batch-to-batch drift compensation (read the correction, write the parameter), not the fast control loops. Checked by: the task mix of the first shop-floor cases. From: MHS: Three Questions, One and a Half Answered, Aug 27, 2026. Status: open.
Robot learning
P9 · The endgame is hybrid. Learned prior plus test-time search or verification: about 75%. Pure end-to-end policy: about 15%. Explicit-model planning at the core: under 5%. I move toward pure policy if search gains vanish once data grows tenfold, or zero-shot success without search passes 95%. From: Amortize or Defer, Sep 15, 2026. Status: open.
P10 · A certificate layer becomes standard. An explicit constraint or certificate layer becomes a standard component of industrial robots before 2030: about 55% (lower for consumer robots), built by the deployers who pay for failures. From: Amortize or Defer, Sep 15, 2026. Status: open.
P11 · Constraint-novelty curve. As obstacle or fixture geometry moves away from the training set, a pure policy's failure rate rises steadily while a hybrid stays roughly flat. Checked by: the first controlled experiment that plots it. From: Amortize or Defer. Status: waiting for an experiment.
P12 · Latency crossover. With required reaction time over solve time on the x-axis, hybrid ≈ policy below 1 and hybrid ≫ policy above 1; cheaper inference moves the crossover left. From: Amortize or Defer. Status: waiting for an experiment.
P13 · Silent failures. A pure policy's constraint violations go unflagged; the share of a hybrid's violations that get flagged roughly equals its solver's completeness. From: Amortize or Defer. Status: waiting for an experiment.
P14 · Parameter extrapolation. On path-dependent tasks with a window length set at run time, a policy with memory fails outside its training range; a planner holds for any value. From: Amortize or Defer. Status: waiting for an experiment.
AI and work
P16 · The open side stays human for now. AI will flatten work whose end point is defined in advance and cheap to judge, but posing problems, choosing routes and defining what counts as understanding stay human-led. Wrong if, by Sep 2027, an AI poses a new problem or theory, with no human supplying the statement or route, and mathematicians accept it as important. From: The Millennium Problem Was the Easy Part, Sep 29, 2026. Status: open.
Markets and supply chains
P5 · Solar won't clear fast, or never. Overcapacity neither clears immediately nor stays cheap forever; the road back to cost is far longer than the textbook says, and in sub-industries running near 60% utilization the supplier list gets reshuffled within two years. Checked by: capacity exits after the 2027 standards take effect; supplier lists before 2028. From: Two Ledgers, One Price, Sep 8, 2026. Status: open.
General-purpose robots (check in October 2027)
P22 · A humanoid post with no human in the loop. The first hotel or retail chain runs a humanoid robot post 24/7 with no human intervention. From: The Brain Has Arrived, Oct 15, 2026. Status: open.
P23 · Decision models become standard. Jev-style decision models become standard in service robots. I concede the architecture bet if mainstream service robots instead run on pure end-to-end VLA. From: The Brain Has Arrived, Oct 15, 2026. Status: open.
P24 · No more programs per task. A maker publicly announces its new service robots no longer need a program written for each task. From: The Brain Has Arrived, Oct 15, 2026. Status: open.
Software after free (check at the end of 2027)
P17 · Seat licenses die. More than half of the main SKUs of mainstream industrial and creative software move to pricing by usage or outcome. From: SaaS Is Dead. Free Killed It., Oct 20, 2026. Status: open.
P18 · A dumbbell market. At least three mid-sized industrial software vendors are acquired or shut down; the giants hold on through ecosystem and certification, while a thousand AI-native small tools grow underneath. From: SaaS Is Dead. Free Killed It., Oct 20, 2026. Status: open.
P19 · Five-person CAD. A team of five or fewer ships professional-grade CAD or CAE with real paying production users. From: SaaS Is Dead. Free Killed It., Oct 20, 2026. Status: open.
P20 · MCP becomes the new operating system. An OS vendor absorbs an agent-tool protocol at the system level, or a second mainstream protocol appears. From: SaaS Is Dead. Free Killed It., Oct 20, 2026. Status: open.
P21 · Trust is the moat. At least one company that insures or certifies AI-generated engineering files raises institutional funding. From: SaaS Is Dead. Free Killed It., Oct 20, 2026. Status: open.
Scorecards
None yet. The first one goes out in December 2026, and every quarter after that. When a bet settles, it moves here with the evidence.