Three sessions.
One vertical.

Settled: registered apprenticeship programmes in the US produce average earnings premiums of 40–70% above control groups in the first five years post-completion (Mathematica, 2018; Urban Institute, 2021). Contested: whether earnings premiums persist beyond 10 years or whether they reflect selection effects rather than programme causality — no RCT evidence exists; all studies are quasi-experimental. Unknown: which programme design features (duration, employer ratio, credential type) drive the earnings premium.

Selection bias concern: quasi-experimental designs in this literature use comparison groups of workers who applied to apprenticeships but were not accepted. This comparison group may differ systematically from participants on unobservable motivation and employer quality dimensions. The earnings premium estimate could partially reflect employer quality selection rather than human capital development. No instrumental variable study has cleanly resolved this.

“The current programme design has a 3:1 employer ratio (3 apprentices per journeyperson mentor). The literature on mentorship quality in apprenticeship outcomes consistently identifies 1:1 or 2:1 ratios as the threshold above which outcome quality declines. The programme is operating above this threshold in 60% of registered sites. The design does not match what the literature shows works.”

Confidence: Probable that the programme produces positive earnings outcomes. Contested on magnitude due to selection bias. Gap on design feature evidence. Programme recommendation: commission a randomised expansion study with design feature variation to establish causal mechanism before the next reauthorisation cycle.

Settled: large-scale government IT programmes (>$500M) have well-documented failure rates. GAO data: 80% of large IT programmes exceed original cost estimates; 40% exceed by more than 50%. Contested: whether cloud-native migration programmes perform better than legacy modernisation — limited government-specific data; commercial sector data shows better outcomes but government regulatory and security constraints are materially different. Unknown: contractor performance at this specific scale with the proposed hybrid multi-cloud architecture.

Base rate calibration: the $2.1B programme falls in the >$1B government IT category. GAO historical data for this category: 23% delivered on time and on budget; 47% delivered with >50% cost overrun; 30% cancelled or restructured. Applying base rates, expected cost at completion: $2.8–3.4B. Expected timeline: 18–36 months beyond current projection.

“The GAO base rates pool programmes across agencies with very different procurement maturity. DoD has significantly improved IT procurement outcomes since the establishment of the JEDI/JWCC contracting vehicles. DoD-specific base rates for cloud programmes post-2020 are better than the GAO aggregate. Using the aggregate overstates the risk for a programme with this contracting structure.”

Three risk concentration points based on programme-specific analysis: (1) Requirements lock — 60% of DoD IT overruns originate from requirements changes after contract award. (2) Contractor concentration — single prime contractor model creates single point of failure. (3) Security clearance pipeline — cleared personnel bottleneck is the leading cause of schedule slippage. Address all three in programme governance before contract award.

Settled: export controls enacted 2022–2024 have measurably slowed adversary access to leading-edge GPU hardware (A100/H100 class). Contested: whether the slowdown is temporary (before domestic fab capacity comes online) or persistent. Unknown: the actual computing capacity currently deployed for adversary AI development — open-source indicators are a partial and potentially lagging signal.

Indicator reliability: the current assessment relies on three open-source indicator categories: (1) research publication citations of specific compute resources, (2) satellite imagery of data centre construction, and (3) semiconductor import partner data. Each has known limitations. Research citations undercount classified and non-published work. Satellite imagery has 6–9 month latency on construction-to-operational timelines. Import data captures official channels, not grey market.

“The assessment treats the three indicator categories as independent, but they are correlated — a grey market acquisition strategy would specifically avoid all three detection methods simultaneously. The absence of signal in all three channels is consistent with both low capability and sophisticated evasion. The assessment cannot distinguish between these hypotheses on open-source evidence alone, and should explicitly state this limitation rather than presenting a point estimate.”

Confidence: Contested. Assessment should be presented as a range with explicit uncertainty bounds rather than a point estimate. The three-indicator methodology cannot distinguish low capability from sophisticated evasion at current signal strength. Recommendation: flag this epistemic limitation explicitly in the NSC brief and recommend augmented collection to resolve the ambiguity.