Two worked critical paths

As labelled, the three are 9.5 of the 32.5 years, leaving 23. Two independent reviews then found corrections pointing in opposite directions. Gate C found that links 2 and 3 count eight years against AI on blockers JWST's own record does not show in those windows, and that design maturation demonstrably ran four years past the interval it is given — all of which move years into the AI column and pull the residual toward the mid-teens. Gate D found the opposite at link 1: it is labelled Working now while its own blocker says the constraint is community consensus rather than analysis capacity, and correcting that moves seven years out of the AI column and pushes the residual toward thirty.

So the residual is somewhere between about fifteen and about thirty of the 32.5 years, depending on judgments the decomposition does not settle. That is a weaker claim than 23 and a more defensible one, and it is what a chain of eight hand-assigned intervals can actually support. What neither reading disturbs is the shape: on every version of the arithmetic, most of the elapsed time of a frontier telescope sits in steps no current AI capability reaches.

The prediction about your capability set holds exactly. Modular assembly with reduced launch costs, a space telescope factory, and leveraging commercial component advances all act on fabrication, integration and launch. None acts on ranking or funding. Stated as an observation and not a deficiency: those three are aimed at the part of the chain where the years actually are.

One honest correction to how I first wrote this up. I originally called four of the eight steps binding, which was circular: a chain of steps that run strictly one after another has no non-binding steps, because removing any of them shortens the total. The set I had marked was really just the set AI does not touch.

What was not predicted is that the split moves between programmes. JWST is build-dominated, with seventeen of its thirty-two years after construction started. Rubin is decision-dominated, twenty-two years to a construction award against eleven building. The ELT sits between, with about two and a half years attributable to nothing but a funding condition on a design already approved.

Appendix: what I predicted before running this

Design and optimisation search helps at the concept and optics stages, and the binding links will fall on the physical build and on the funding decision, neither of which any current AI capability touches. If that holds, the demonstration is that closing every cognitive link in the chain changes the total duration very little. A second prediction, stated separately so it can fail separately: Convergent's three capabilities attached to this gap — modular assembly with reduced launch costs, leveraging commercial component advances, and a space telescope factory — all act on the fabrication and assembly links, and none acts on decision, approval or funding. Both predictions are recorded before any programme history was assembled, and both are testable against published JWST, Rubin and ELT milestone dates.

• 1. Science case definition — Not binding, and the reason matters for the whole chain. Synthesis and trade-study tools are mature and could compress the analytical content of these seven years substantially. They cannot compress the part that actually consumes them, which is a community converging. Treating the full seven years as AI-addressable is the most generous accounting available and is used below as an upper bound. CONTESTED after Gate D. This link is labelled maturity 'Working now' with ai_acts = 1, and its own blocker field says the constraint is 'Community consensus on what to build, not analysis capacity'. Under the efficacy reading the maturity repair established — would applying this capability move this step — those two statements cannot both be right: a capability that addresses analysis capacity does not move a step whose stated blocker is agreement. The critical_path_links.maturity column was never in the repair's scope and has never had a second labeler, which is why the label is left as it stands and flagged here rather than flipped. It carries 7.0 of the chain's 9.5 AI-acted years, so flipping it is not a detail: it would take the chain's only quantitative claim from 9.5 of 32.5 to about 2.5 of 32.5. Restating a headline on one reader's unaudited judgment is the thing the gates exist to prevent.
  Evidence: JWST: NGST workshop at STScI September 1989; an STScI committee recommends a larger infrared telescope in 1995-1996. Seven years. ESO: OWL concept pursued from 1998, OWL Blue Book published end of 2005. Seven years.
• 2. Concept studies, strategic ranking and design competition — Six years, and they are not six years of one thing. Gate C found that the duration_note attached to this link describes a design competition while the link title describes strategic ranking, and both happened in the window. The 2001 decadal survey ranked NGST top; the Phase A competition ran either side of it. ai_acts is recorded as 0 because the cadence is what sets the length — no amount of readiness moves a project forward inside a decadal cycle, so the marginal value of accelerating everything upstream is zero until a whole cycle is saved — but design and optimization search does act on the competition inside these years. This link is therefore the largest single source of uncertainty in the chain's arithmetic, and the finding now says so.
  Evidence: Rubin: planning from the early 1990s, recommended by a National Academies report in 2001, top-ranked large ground-based project only in the 2010 decadal survey, NSF construction award August 2014. Thirteen years from first national recommendation to construction start. ELT: ESO Council names ELTs its highest strategic priority in December 2004; construction green light December 2014. Ten years. JWST's own window, 1996 to 2002, contains the industry study teams (4 June 1997), the Yardstick Design report (6 August 1998), the Lockheed Martin and TRW Phase A selection (7 July 1999) and the TRW/Ball selection (14 August 2002), per STScI's mission timeline; the strategic ranking in it is the 2001 McKee-Taylor decadal survey, which STScI's timeline does not list. The Rubin and ELT figures above are the ones that isolate ranking cleanly; JWST's does not, and that is stated here rather than left for a reader to notice.
• 3. Phase B start and funding authorisation — The ELT case is the cleanest instance available anywhere in this analysis: the delay is explicitly and only about money, with every technical question already settled. Nothing in the capability set acts on it. Gate C's correction, recorded rather than argued away: JWST's own 2002-2004 interval contains no identifiable appropriation event, only the run-up to a 3 March 2004 construction start, so the two years assigned here are carrying a blocker demonstrated on other programmes. The evidence field cites Rubin and the ELT for that reason, and it should be read that way.
  Evidence: ELT: ESO Council approved construction in June 2012 on the condition that contracts above 2 million euros could be awarded only once the total cost of 1,083 million euros (2012 prices) was 90% funded; the green light followed in December 2014. Roughly two and a half years of pure funding latency with the design frozen. Rubin: Final Design Review December 2013, National Science Board conditional approval May 2014, construction award August 2014.
• 4. Design maturation — This is the link where design and optimization search is most obviously useful and most obviously mature, which is exactly why it does not set the pace. Two years from contractor selection to instrument critical design reviews. Gate C found the interval understates the activity: this link's own evidence runs to the mission critical design review, which STScI dates to 3 March 2010, with the preliminary design review in April 2008 and project confirmation in April 2009 — all inside the window the chain labels Fabrication. Design maturation on JWST ran to 2010. The interval is left at 2004-2006 because the phases genuinely overlap and no milestone pair separates them cleanly, but the overlap runs in one direction: it understates how much of the elapsed time design work occupies, and so understates the AI-acts column.
  Evidence: JWST: TRW/Ball design selected 2002, construction of long-lead components begins 2004, instrument critical design reviews 2006, mission critical design review 2010.
• 5. Fabrication — All three of Convergent's capabilities attached to this gap act here or at the next link.
  Evidence: JWST: construction of long-lead components begins 2004; all 18 primary mirror segments complete and cryogenically tested in 2011. Seven years.
• 6. Integration and test — The single longest link in the JWST chain, and the one where the capability set is most clearly pointed at the right place — this is what a space telescope factory is aimed at. Gate C contested the blocker and the contest succeeds in part: roughly 4.3 years of this nine-year block are a near-termination and replan, and a further tranche is workmanship rework. On this chain's own taxonomy a replan is Coordination and institutions and defect detection is Measurement and sensing, neither of which is physical assembly. ai_acts is left at 0 because no current capability shortens a congressional replan either, but the stated reason for the zero was wrong and is corrected here. It also exposes a limit of the model: a strictly sequential eight-link chain cannot represent funding delay recurring mid-build, and this chain places funding once, early.
  Evidence: JWST: mirror segments mounted into the backplane 2015-2016, optical and spacecraft elements mated 2019, final environmental testing complete 2020. Nine years from mirror completion. GAO-13-4 (3 December 2012), p.4: 'the JWST project was reauthorized, but not before it was recommended for termination by the House Appropriations Committee... NASA announced that the project would be rebaselined at $8.835 billion — a 78 percent increase — and would launch in October 2018 — a delay of 52 months.' The interval endpoint is confirmed: STScI's 'final integrated testing completed, 1 December 2017' is the element-level OTIS cryo-vacuum test, while full-observatory environmental testing completed 6 October 2020.
• 7. Launch — Not where the time is. Reduced launch cost, which two of the three capabilities attached to this gap depend on, acts on the cost axis and not on this one. No AI capability shortens a launch campaign.
  Evidence: JWST: final environmental testing complete 2020, launch 25 December 2021. Roughly one year. Ariane 5 selected as launch vehicle in 2005.
• 8. Commissioning — Not binding, and the shortest link in the chain by an order of magnitude. Wavefront sensing on eighteen segments is the most technically demanding cognitive task in the whole sequence and it took under seven months, which is the chain in miniature.
  Evidence: JWST: launch 25 December 2021, first full-colour images and start of science operations 12 July 2022. Six and a half months.

Axes I deliberately left out of this chain

Cost per unit of collecting area, and cost per unit of science return. Their gap statement bundles cost with schedule — 'cost-prohibitive and slow' — and the two run over a partly different set of links. Each is a separate chain, and a chain that mixes them produces a binding link that is an artifact of the mixing rather than a property of the world.

Where exactly it landed is a correction Gate C forced, and it matters because it was this chain's marquee result. I wrote that the labour moved to reviewer recruitment. The source does not say that. Pierce, Gartenberg, Hasan and Murray put the displaced load on volunteer editors at desk screening — among manuscripts with 70%+ AI scores nearly 70% are desk-rejected, against 44% for low-AI submissions — and give no invitation or recruitment figures at all. So the arrow from cheaper drafting runs to screening, which is a step this chain calls tractable and AI-reached. Recruitment strain is real and separately evidenced, by Silverchair's 4.5 invitations per accepted review and the fall in acceptance from 43% to 22%; it is simply not what the 42% surge is shown to have caused. The general claim survives and the specific one does not: relieving a step upstream of a concentration moves cost downstream, and it moved to the nearest downstream step, not the furthest.

The prediction holds, in the shape the pair needs. AI is not absent here: it acts on the first four of seven steps, which is more than it touches anywhere in chain 1. It can match a reviewer to a paper. It cannot make that reviewer say yes, agree with the other reviewer, or persuade a hiring committee to count the work.

Recruitment is where the axis choice earns itself. Splitting it into matching and willingness separates a well-posed prediction problem from a labour-supply problem. On reviewer identification specifically, published evidence finds traditional statistical representations outperform generative AI. Gate C's caveat is recorded with it: ESO names matching difficulty as a reason for distributed peer review too, so the split is this analysis's, not ESO's.

One observation about your capability set, and it runs the other way from chain 1. Two of the four capabilities you attach to this gap, a post-publication peer review layer and new protocols for knowledge production and verification, act directly on credit and legitimacy, which carries cost. For the telescope gap none of the three touches a decision step; here half of them do.

To be plain about scope: this chain covers getting a finished result into the record and credited. The cost of doing the research is the other half of their gap statement and is most of what the rest of the map is about.

Appendix: what I predicted before running this

This chain must produce a different answer from chain 1 or the pair demonstrates nothing. AI does act on several links here: drafting and screening work now, and reviewer-to-proposal matching is tractable, though there is published evidence that traditional statistical representations outperform generative AI at identifying expert reviewers. So the expectation is not that AI does nothing. It is that AI acts on the links that were never rate-limiting, while reviewer recruitment, judgment consistency and the credit and legitimacy step remain untouched. Stated explicitly so it cannot be read the other way: this is not a claim that publishing is a cognitive bottleneck. It is not, and asserting it would be wrong. Recorded before the link analysis.

• 1. Production and drafting — Scope first, because it decides how to read this step. The axis is reviewer and editor labour per published paper, so author time sits outside it by construction. Writing the paper is very probably the most expensive thing that happens in this pipeline, and this chain does not measure it. What it can say is that AI has taken a large share out of the per-paper drafting cost, that the number of papers went up, and that the labour this displaced arrived downstream: submissions rose 42% after ChatGPT's release against the prior two-year window, in the one corpus with published full-submission figures, and the load landed on volunteer editors at desk screening. Whether total author cost across all papers fell is unknown and this chain has no way to find out. Two corrections from Gate C are folded in above: the 42% is growth since ChatGPT within a five-year corpus, and the source places the displaced load at screening.
  Evidence: Organization Science AI Task Force, 'More Versus Better: Artificial Intelligence, Incentives, and the Emerging Crisis in Peer Review', Organization Science editorial, 2026, doi 10.1287/orsc.2026.ed.v37.n3. Roughly 7,000 manuscripts with full-submission access. Submissions up 42% since ChatGPT's November 2022 release against the prior two-year window; for scale, the COVID-19 period produced a 20% bump. The rise is almost entirely manuscripts with substantial AI-generated text, and submissions scoring low for AI have declined over the same period.
• 2. Submission and desk screening — Not binding on the cost axis. Screening is genuinely tractable for current models, which is why it is already being done and why it does not set the cost.
  Evidence: ESO reports that classical triage 'has significantly degraded the quality of feedback for the triaged proposals' — screening under load trades feedback quality for throughput.
• 3. Reviewer recruitment and matching — Binding, and it is the link that most rewards being split in two. Matching is a well-posed prediction problem where AI is applicable and, on current evidence, not even the best method. Willingness is a labour-supply problem that no matching system addresses. The binding half is the half AI does not touch, which is the whole finding of this chain in one link. REVISED after Gate C. The previous version asserted that matching is not the problem and cited ESO for it; ESO says both. The split between a well-posed prediction problem and a labour-supply problem still does the work this chain needs, but it is now stated as this analysis's distinction rather than the source's.
  Evidence: 4.5 invitations per accepted review, nearly double the 2018 rate, and reviewer acceptance down from 43% in 2018 to 22% in 2024 (Silverchair, Future of Peer Review 2026, from eight years of ScholarOne Manuscripts activity plus 2,000+ survey responses). Per 100 invitations sent, editors wait a combined 407 days on reviewers who ultimately decline or never answer. A survey of 139 editors of Australian journals, largely local and society titles, with 27 interviews, finds 55% rate finding reviewers a significant or very significant challenge, and some 'described having to send out 30 or more invitations to secure just two reviewers' (Jamali et al., Learned Publishing 2026, doi 10.1002/leap.2034, and Luca et al., doi 10.1002/leap.2041; popularised in The Conversation, 15 February 2026). The sample is national and small-journal weighted, which is where recruitment is hardest, so the bias runs in this chain's favour and is noted for that reason. Meyerson documents 21 years of declining reviewer acceptance, 2002-2024. ESO's Phase 1 page on distributed peer review states that 'it has become progressively harder to find scientists willing to serve in the panels and in the OPC'. And on the part AI could do: traditional statistical representations outperform generative AI at identifying expert peer reviewers, tested on NASA/ADS records and observatory DPR data across 379 researchers.
• 4. Review judgment — Binding, and unusually well quantified for an institutional link. The Cortes and Lawrence result also sets the bar any automated reviewer has to clear, and it is a low bar in one direction and an unreachable one in the other: filtering bad work is where review already works, ranking good work is where it does not, and it is not established that the second is a solvable problem for anyone.
  Evidence: NeurIPS 2021 consistency experiment: two independent committees disagree on 23% of papers, and approximately half the accepted list would change on a random rerun — consistent with 26% in 2014. Cortes and Lawrence, revisiting the 2014 experiment, find no correlation between review scores and later citation impact for accepted papers, and a correlation for rejected ones: 'the reviewing process for the 2014 conference was good for identifying poor papers, but poor for identifying good papers.'
• 5. Editorial decision — Not binding on the cost axis. Its cost is largely inherited from link 3.
  Evidence: Follows directly from the recruitment and judgment links; editors carry the residual cost when reviews do not arrive.
• 6. Dissemination — Not binding on the labour-cost axis, and this deserves care because it is the link their gap sentence centres on. On the inclusiveness axis — explicitly excluded from this chain — it may well be binding. That is why the axes were separated.
  Evidence: Three of Convergent's four capabilities for this gap act here: disrupting traditional publishing models, frugal science initiatives, and the micro/nanopublishing half of new protocols for knowledge production and verification. That third one is counted at link 7 as well, because its description covers both lowering the barrier to publishing and new norms for recognising work; it is one capability acting on two steps, not two capabilities.
• 7. Credit and legitimacy — Binding, and the only link in either chain where the blocker is purely a matter of what institutions agree to recognise. Two of Convergent's four capabilities for this gap — a post-publication peer review layer, and new protocols for knowledge production and verification — act here.
  Evidence: A model can review a paper today; it cannot make a hiring committee count that review. The same asymmetry explains why reviewer supply falls: the labour is unpriced and uncredited, and pricing it is an institutional act.

Axes I deliberately left out of this chain

Speed, meaning elapsed time from submission to publication, and inclusiveness, meaning who can afford to publish and who can read the result. Their sentence bundles all three. Speed in particular behaves differently by venue: journal latency is reviewer-supply-driven while conference latency is set by a fixed programme committee calendar, so a chain mixing them would produce a cost concentration that is an artifact of the venue mix. Also excluded: the 'doing' half of their gap statement. 'Doing and publishing research is expensive' covers the cost of the research itself, which is most of what the other 102 gaps in the map are about. Instruments, reagents, compute and staff are priced elsewhere. This chain runs from a finished result to a credited contribution. Also outside the axis, and worth stating because it is the biggest number in the room: author time. The axis counts reviewer and editor labour, so the weeks a researcher spends writing the paper are not in it. That is very likely the largest single cost in getting a result published, and this chain does not measure it.