Critical paths
A gap broken into the ordered steps the work actually runs through, with a duration or a cost on each one, and which of Convergent’s own capabilities act on it.
3 gaps have one, against 103 in the map, and each is a claim about one axis. The capability counts on them are Convergent’s data, not my labels. Where a gap bundles several axes I picked one and named the rest on the page.
Doing and publishing research is expensive and subject to structural roadblocks
MetascienceDrafting was expensive, and AI has made it much cheaper. That saving is real and large. It has not arrived as cheaper publishing: submissions rose 42% after ChatGPT's release relative to the prior two-year window, in the one corpus where a journal has published full figures, and the labor that the saving displaced landed further down the chain rather than disappearing.
Frontier Telescopes Are Expensive and Take Decades to Build
AstrophysicsAI can act on three of the eight steps, and those three take 9.5 of the 32.5 years a frontier telescope needs from first idea to first images — still leaving 23 years that AI is unlikely to shorten soon. That split is the least settled number on this page: the working below puts the untouched share anywhere from about fifteen to about thirty years, and on every version of it most of the time sits where AI does not reach.
Inability to Model Turbulence
PhysicsPredicting a turbulent flow — the air over an aircraft wing, say — means running a large computer simulation, and that simulation is the only step in the whole process anyone has put a price on. It is also a step AI reaches. In the one aircraft case where every step was costed, the simulation took about 360,000 hours of processor time for a single flying condition, against about 33 for building the grid it runs on. Two things stop that from meaning AI has found the bottleneck. Counted in engineers' time rather than computer time, building the grid is the expensive step — NASA's own survey says so. And the two steps practitioners say cost most, producing trustworthy reference data and getting a regulator to accept a calculation in place of a wind-tunnel test, have no published price at all. The September 2026 Navier-Stokes result does not change any of this: it settled whether the idealised equations can break down when a force is applied, not how cheaply a real flow can be predicted, and none of the seven steps waits on that answer.
Two of them run through the same step
Telescope time and facility approval are allocated by peer review of proposals. A decadal survey is a review panel, and so is a time allocation committee. That makes the telescope chain’s ranking and funding steps an instance of the publishing chain’s reviewer recruitment and judgment steps. ESO on why they changed it:
“the load on the panels and the Observing Programmes Committee (OPC) members has become unsustainable” … “it has become progressively harder to find scientists willing to serve in the panels and in the OPC”ESO, introducing Distributed Peer Review, where every PI submitting a qualifying proposal reviews ten others. Running since Period 110, at ALMA from Cycle 8, and at Gemini before that in the Fast Turnaround channel.
Two gaps, two fields, one blocker. The Gap Map’s export has one row per gap and nowhere to record that, and once gaps decompose into steps a recurring blocker becomes something you can count across all 103 rather than notice twice. The turbulence chain does not share it, which is the point of having a third: counting needs more than two.
Why there are only 3
Each of these was built by hand, because nothing in the source data marks a capability as necessary, sufficient or partial for the gap it hangs under, and a chain needs to know which links are load-bearing. Typing those edges is the one change that would make chains something the data yields rather than something somebody authors — it is the first item on what’s missing.