Distance

Argued. Not a forecast. Issued in publication v0.7. Open forecasts: 0.

The curve and the bound are both on the site. They do not share a unit, so this page does not publish a single gap, a percent complete, or a date.

What is measured

SW-M-001: in February–March 2026 the public frontier on Time Horizon 1.1 was about 12 hours at 50 percent, interval 5 hours to 61 hours, and about 1.5 hours at 80 percent. SW-M-002: the suite is unreliable above 16 hours. The 50 percent interval crosses that line. The point estimate does not. SW-M-003: the series page is no longer actively updated. No software-horizon forecast is open.

What is argued

On the Bound page, further CMOS efficiency is about a 200-fold gap from an H100 near 1.4 × 10¹² FLOP/J to a 16-bit mean near 2.9 × 10¹⁴ FLOP/J. A separate engineering wall says a three-month training run past about 2 × 10²⁸ FLOP is infeasible at good utilization with current techniques. The second is not the ceiling of a later stack.

The distance

Twelve hours of skilled-human software work is not a fraction of a FLOP/J ceiling, and not a fraction of a training-run FLOP budget. Placing the horizon “on the way” to either bound would invent a conversion. No such conversion is published here.

The nearer constraint depends on the question. For one more training run at today’s utilization, data movement binds before the CMOS efficiency ceiling. For how long a software task an agent finishes, neither bound is the instrument. That instrument stopped, and the suite is already unreliable above 16 hours. The empty domain lines are a third distance: no named series converts a gain in software into discovery, manipulation, adoption, or self-improvement.

Withdrawn if a named source publishes a conversion from skilled-human task time to FLOP/J, or to a training-run FLOP budget, with the assumption stated and a result that would retire it. Until then the gap stays unnumbered.

Rows: SW-M-001, SW-M-002, SW-M-003, SW-F-001. Ceilings: Bound.