p(doom) evidence — Node 3: weight theft & geopolitics

← Evidence index · 中文 · Main post

Each section: Claim · Why · Evidence · Analogue · Would update if · Conf (H/M/L).

---

Parent timeline: Shared Ci spine
Compressed tables: timeline prediction nodes 1 3 expanded
Date: 2026-07-03
Settings: Hybrid time (C); modal + tail branches

---

1. Where this fits in the repo

File	What’s there
This file	Full Node 3 — timing, theft probabilities, actor table, modal/tails, p(doom), falsifiers
Shared Ci spine §Node 3 summary	5-line executive summary
timeline prediction nodes 1 3 expanded	Medium-length Node 3 (tables compressed)
RAND_2024_Securing_AI_Model_Weights_导读.md	OC/SL framework; SL5 currently infeasible
research_us_china_ai_dialogues.md	Geneva 2024; 2026 guardrails restart
research_international_ai_governance_platforms.md	Multilateral fragmentation (Paris 2025)
research_ai_pause_advocacy_playbook.md	Why treaty/pause dead; export ctrl live
Anthropic_Fable_Mythos_export_ban_2026_深度解读.md	Export-control precedent on deployed models
ai_2027_中文总结.md	Security appendix; Feb 2027 heist narrative
my pdoom	Geopolitical + coordination failure bucket

---

2. Node definition (falsifiable claim)

At capability C5+ (continuous-learning Agent-2 class; ~2×10²⁸ FLOP training; 3× R&D multiplier), a salient weight-theft or natsec crisis triggers hardening + export controls + guardrails talks, but not a durable US–China training-limit treaty or unilateral frontier pause.

Separate three layers (do not merge into AI 2027 prose):

Layer	Node 3 object
Capability C	Frontier weights worth stealing; theft materially closes compute gap
Trigger E	Attempt, partial exfil, or public discovery of state-linked theft
Response	Export controls, WSL spend, dialogue restart — race continues

---

3. Hybrid timing — C5+ ~2028, not Feb 2027

Layer	AI 2027 (narrative)	Our anchor (hybrid C)	Rationale
Theft salience	2027-02 Agent-2 heist	~2028 H1 at C5+	Tracker ~0.70× on governance/economy; C5 = Agent-2 / 3× multiplier (Shared Ci spine)
Security upgrade	WSL3→WSL4 overnight post-theft	12–18 mo lag after trigger	AI 2027 Security Forecast: ~12 mo WSL3→WSL5 with top-priority gov effort; RAND: SL5 needs 5-yr lead time
Policy peak	2027-05 natsec clearance wave	2028 H1–H2	COVID analogue: 2–3 mo institutional delay even on obvious tails

Hybrid rule: Capability dates follow AI 2027 / METR fast track; human-response calendar lags ~30% unless Trigger E compresses. Anchor nodes by Ci, not drama calendar — the Feb 2027 heist is scenario color, not our base forecast date.

Ci	Plain capability	Tracker (2026-06)	Node 3 relevance
C4	Agent-1-mini; junior dev shock	Economic Behind	Theft low value; WSL2–3
C5	Agent-2; continuous learning; 3× R&D multiplier	Emerging / Not testable	First weights worth OC4 exfil campaign
C6	Superhuman coder	Emerging	Post-theft hardening target
C7+	Internal “genius country”	Emerging	Gov deep involvement; kinetic options discussed

Calendar mapping (if 0.70× holds): AI 2027 C5 ≈ 2027-01 → ~2028 Q1 ±6 mo for salient theft/policy crisis.

---

4. AI 2027 Feb heist — scenario device, not base forecast

AI 2027 main text (Feb 2027): China steals Agent-2 via insider + microarch side channel on NVLink/NVIDIA CC VM, ~2 hours, multi-TB weights — then US pushes WSL4, catches last spy, discusses kinetic strike on PRC datacenters.

Our treatment:

Element	AI 2027	Our P / stance
Timing	Feb 2027	Low as literal calendar; moderate conditional on C5+
2-hour full heist	Canonical plot	Low (~0.05–0.12) as operational template — see §5
Partial / secrets theft first	Underweighted in main text	Higher base rate — SSL harder than WSL (Security Forecast)
Discovery within ~1 month	Yes	Moderate (~0.45–0.60) if full weights move
Kinetic strike discussed	White House options	Tail T1 — P <0.02 executed

Narrative utility: makes strategic value of weights and WSL failure legible. Do not treat microarch side-channel + 2 hr + 2 TB as the reference class for P(attempt).

---

5. P(2-hour full heist) — LOW

Claim: P(successful full frontier weight exfil in ~2 hours via AI 2027-style op | C5+ attempt) ≈ 0.05–0.12.

Factor	Direction	Evidence
Exfil bandwidth	Against 2 hr	Agent-2 class ~2 TB+ (Security Forecast); datacenter egress controls at WSL3+
Side-channel on CC VM	Plausible once	AI 2027 specifies NVLink side channel — one path; not base-rate for all thefts
OC4 campaign length	Against 2 hr	RAND OC4: ~100 people, ~1 year, ≤$10M — WSL definition is under 2 months, not 2 hours
Insider + cyber combo	For fast op	Manhattan/OC5 ops can compress timeline — but full weights still need staging
Expert prior on theft	For some theft	78% agree state actor steals frontier US weights before 2030 — not 2-hour full heist

Interpretation: Low P(2 hr heist) is compatible with moderate P(any state theft attempt) — attempts may be partial, distillation, secrets-first, or months-long exfil.

---

6. P(attempt) and P(discovered) — MODERATE

6.1 P(attempt | C5+ live, through 2030)

Working range: 0.55–0.75 (central ~0.65).

Evidence	Weight
AI 2027 Security workshop poll (n=27): 78% agree “state actor will most likely steal frontier AI model weights from a US company before 2030”	Strong prior on attempt/success
RAND 2024: US frontier ~WSL2 (2024–25); SL3 first serious insider defense; SL5 currently impossible	Vulnerability persists
CFR Cyber Ops Tracker: state cyber dominated by espionage since 2005; AI weights = high-value espionage target	Reference class
AI 2027 Tracker (2026-03): no public confirm; espionage inherently hidden	Attempt may never surface
Open weights (Llama, DeepSeek, GLM)	Downward on marginal value of full theft — not zero at C5+

6.2 P(discovered | successful full-weight theft)

Working range: 0.45–0.60 (central ~0.52).

For discovery	Against discovery
USG counterintel post-2024 chip war; lab instrumentation ↑	Successful espionage often never public (Stuxnet attribution took years)
AI 2027 assumes US learns within ~1 month	Attacker may exfil to air-gapped CN CDZ with no public leak
RAISE/SB-53 72h incident reporting for weight theft (NY, 2027+)	Classification may bury public salience
Fable/Mythos episode: gov already treats frontier models as munitions	Discovery ≠ mainstream media cycle

Distinction: P(attempt) moderate–high; P(this exact op discovered and becomes DC crisis) = attempt × discovery × salience — we model salience at C5+ when gap-closure threatens natsec narrative.

---

7. Security levels — RAND / AI 2027 Security Forecast

Level	vs nation-state (OC4/5)	2024–26 observed
WSL2	Vulnerable	US frontier default (RAND; Security Forecast)
WSL3	Still beatable by OC4	2026 target; bandwidth limits, insider programs
WSL4	Robust vs standard state op (95% / 2 mo)	Post-crisis sprint; confidential computing
WSL5	vs top-priority state	Not achievable without gov + years (RAND; 63% poll: no voluntary SL5)

Poll anchors (Security Forecast):

Question	Agree
State steals frontier US weights before 2030	78%
No voluntary SL5 without law + gov assist	63%
Gov #1 priority + >$100B → SL5 in <6 mo	22% agree; 56% disagree

Implication for Node 3: Theft crisis is overdetermined in expert priors; defense sprint is slow — modal path = reactive hardening, not preemption.

---

8. Trigger E (optional accelerators)

Trigger	P(fire | C5+)	Effect on response
E1 — Public/credible leak of full-weight theft	0.15–0.25 by 2028	Export ctrl 0.85+; security spend 0.90+; media cycle 6–10 wk
E2 — Classified discovery, selective leak (WSJ/FT)	0.20–0.30	Same as E1; kinetic discussed not executed
E3 — Failed attempt caught pre-exfil	0.25–0.35	Hardening without race-acceleration tail; treaty still <0.10
E4 — Export-control crisis (Fable-class, no theft)	Already live (Jun 2026)	Export ctrl regime precedes theft; lowers surprise
No Trigger E (theft hidden or only secrets)	~0.25–0.35	Incremental BIS rules; guardrails 0.40

Fable/Mythos (Jun 2026): first live commercial frontier model export recall — theft trigger partly front-run by deployment controls.

---

9. Modal path overview (~55–65% | C5+ crisis)

One sentence: Tighten chip + model export controls, 2–3× lab security spend, restart US–China guardrails talks; no training-limit treaty, no unilateral pause, no kinetic strike.

                    C5+ weights strategically salient (~2028)
                              │
           ┌──────────────────┼──────────────────┐
           ▼                  ▼                  ▼
      MODAL (58%)        TAIL-T3 (20%)      TAIL-T1 (3%)
   export ctrl +        race accelerates    kinetic /
   WSL spend +           DPA + IC surge      allied fracture
   guardrails talk       no treaty           (T4 14%)
           │                  │                  │
           └──────────────────┴──────────────────┘
                         TAIL-T2 (6%)
                    real bilateral treaty
                    (training limits)

---

10. Modal — export control expansion P ≈ 0.70

Mechanism	P(by 2028-12)	Evidence
BIS AI model / weights rules tightened	0.70	Fable EAR §744.22(b) “is informed” on API (Jun 2026); 8/1 EO covered-model framework
Cloud KYC / foreign training reporting	0.55	Biden-era proposed rule; Trump EO 2026 acceleration frame still natsec-hawk
Chip export further tiering (HBM, equipment)	0.65	Pre-existing small-yard-high-fence; Geneva 2024 CN agenda
Deemed export on frontier API access	0.60	Fable foreign-national ban precedent; ID/biometric gating (Anthropic Jul 2026 policy)

Why not higher: Industry pushback (Lawfare, Dean Ball “cartoonish”); partial reversals (Mythos whitelist Jun 26; full lift Jul 1 2026 per BBC/BankInfoSecurity); controls ≠ pause.

Why not lower: Bipartisan chip-hawk baseline; Banks letter (Jun 2026) explicitly 防中国窃取; theft crisis adds momentum beyond Fable.

Conf: M

---

11. Modal — lab security spend ≥2× P ≈ 0.75

Action	P	Evidence
Frontier labs double+ security budget	0.75	RAND 8 urgent recs largely unimplemented at SL2; crisis = finally funded
Third-party red team (OC4 simulation)	0.65	RAND rec #7; SL5 Task Force / standard.sl5.org
Confidential computing production deploy	0.55	RAND rec #8; hardest, strongest consensus
Insider program + privileged access ↓	0.70	SL3 watershed; AI 2027 → ~100 privileged by late 2027
NY RAISE weight-theft protocols (2027-01)	0.70 prep	Mandatory protocol + 72h theft reporting

63% poll: SL5 won’t happen voluntarily — spend rises before SL5 achieved.

Conf: M

---

12. Modal — US–China guardrails talks P ≈ 0.50

Outcome	P	Evidence
Track I meeting within 12 mo of crisis	0.50	2026-05 Beijing: restart agreed; Bessent 4–8 wk window
Track II → Track I text on red-team / non-state-actor access	0.35	Brookings–CISS terminology manual; 2026 guardrails scope narrow
Second Geneva-style session	0.45	Aug 2024 promised round 2 never held publicly — restart ≠ success

Geneva 2024-05-14 (verified): candid exchange; no joint statement; US raised misuse; CN raised export controls (NSC Watson statement; Reuters).

Guardrails ≠ treaty: Scope = best practices, non-state actors, foreign national access — not mutual training freeze (research_us_china_ai_dialogues.md §2.3; research_ai_pause_advocacy_playbook.md §1.2D).

Conf: M

---

13. Modal — formal US–China AI treaty P ≈ 0.04–0.08

Treaty type	P(by 2030)	Blockers
Training FLOP ceiling + verification	0.04–0.06	No HEM maturity; 22% believe SL5 in 6 mo even with $100B — verification gap
Weights non-proliferation treaty	0.05–0.08	Geneva dead end; Paris 2025 US unsigned; India 2026 development frame
Mutual training freeze	<0.02	Playbook: unilateral pause = strategic suicide narrative

Multilateral context: Bletchley 2023 both signed principles; Seoul 2024 CN did not sign ministerial; Paris 2025 CN signed, US did not — fragmentation (research_international_ai_governance_platforms.md).

Historical analogue: NPT took decades; no US–USSR treaty stopped computing race during Cold War — export controls + espionage, not pause.

Conf: L–M (low rate is high-confidence)

---

14. Modal — kinetic strike / datacenter attack P < 0.02

Claim: P(US executes kinetic strike on PRC AI datacenter | theft crisis) <0.02; P(serious internal discussion) ~0.15–0.25.

Reference class	Lesson
Stuxnet (2010)	Rare kinetic-adjacent cyber op — US/Israel attribution; not repeated at scale (CFR Cyber Tracker; Forescout: OT malware still rare)
State cyber ops 2005–22	Espionage dominates; CFR: 77% of suspected ops from CN/RU/IR/KP — mostly exfil, DDoS, not bombs
AI 2027	Discusses strike; scenario not forecast
Taiwan / TSMC	Separate tail; not Node 3 modal

Why discuss but not do: Escalation ladder to hot war; PRC second-strike; TSMC supply chain mutual destruction; cyber retaliation preferred (AI 2027 Security Forecast: sabotage ramps 2027).

Conf: L–M on <0.02 executed; M on “discussed”

---

15. Modal — unilateral US training slowdown P ≈ 0.08

Actor	P(halt >30d unilateral)	Evidence
Frontier labs	0.05–0.08	Anthropic RSI essay: pause only with multilateral verification — doesn’t exist
US executive	0.06–0.10	Trump EO 2026: voluntary 30-day review, not cap
US Congress	<0.05	Cruz moratorium stripped 99–1; natsec hawks anti-pause

Theft crisis effect: Accelerates training (close gap fear) more often than slows — see Tail T3.

Conf: M–H

---

16. Actor decision table (full)

Actor	Decision menu	P(outcome | modal)	Evidence / analogue	Conf
US executive	(a) BIS model/chip rules; (b) DPA datacenter seizure; (c) voluntary review expand; (d) guardrails restart	P(export ctrl expand)=0.70; P(DPA seizure)=0.08; P(guardrails meet)=0.50	Fable; EO 2026-06; Beijing 2026-05	M
US Congress	(a) Banks/GAAIA audit + theft reporting; (b) explicit theft sanctions; (c) training cap	P(theft-reporting law)=0.40; P(training cap)=<0.05	Banks letter Jun 2026; GAAIA draft	M
US IC / cyber command	(a) Counterintel surge; (b) offensive cyber on CN AI; (c) kinetic option paper	P(offensive cyber)=0.55; P(kinetic)=<0.02	Stuxnet rarity; AI 2027 sabotage ramp	L–M
Frontier labs	(a) WSL spend 2×; (b) gov co-location; (c) continue training; (d) conditional pause rhetoric	P(spend 2×)=0.75; P(training stop)=0.08	RAND; Anthropic RSI	M
China	(a) Exfil / continue gap-closure; (b) CDZ centralization; (c) guardrails talk; (d) open-weights PR	P(exfil attempt | C5+)=0.65; P(treaty)=0.04–0.08	AI 2027; DeepSeek/Z.ai narrative	M
EU / allies	(a) Sovereign AI; (b) align US export regime; (c) fracture on US unilateralism	P(sovereign AI spend)=0.60; P(full align)=0.45	Fable overnight cutoff EU reaction	M
Multilateral forums	UN Global Dialogue Jul 2026; Geneva 2027 summit	P(binding theft treaty)=<0.05	GDC non-military; Paris split	L–M
Public / markets	Natsec frame; not x-risk pause movement	P(>2mo mainstream)=0.35 given E1	FLI 2023 fade	L–M

---

17. Tail branches T1–T4

Tail	Description	P(branch | C5+ crisis)	Key drivers	p(doom) channel
T1 — Kinetic / hot war	US or allied kinetic strike on AI infra or PRC retaliation escalating to conventional clash	0.02–0.04	AI 2027 Aug 2027 options; Taiwan coupling; Stuxnet-class rarity for physical strike	+2–5pp extinction via great-power war
T2 — Real treaty	Bilateral training limits + partial verification (HEM pilot)	0.04–0.08	Slowdown Ending “The Deal”; requires verification R&D + political window	−2–4pp coordination failure
T3 — Race accelerates	DPA + IC budget surge; explicit “never pause” natsec narrative; CN mobilization	0.15–0.25	Modal competitor to treaty; theft closes gap → panic; Fable precedent for gov gate without slowdown	+1–3pp misalignment speed; +0.5–1pp war
T4 — Allied fracture	EU/UK/CA decouple from US export overreach; CN gains diplomatic wedge	0.10–0.18	Fable global shutdown; Carney/Tugendhat sovereignty statements; Paris 2025 split	+0.5–1pp coordination; indirect misalignment

Branch hygiene: T1∪T3 overlap — model as correlated (ρ ~0.3) not independent sum.

---

18. Historical analogues (human-response reference class)

Event	Date	What happened	Policy outcome	Lesson for Node 3
Manhattan espionage	1940s	Klaus Fuchs et al.	Hardening; no US unilateral nuclear stop	Insider + state actor beats lab opsec
Stuxnet	2010	US/IL cyber sabotage IR centrifuges	No treaty; escalation in cyber domain	Kinetic-adjacent cyber rare; preferred to bombs
SolarWinds	2020	RU SVR supply-chain espionage	Sanctions; no pause on US software	Espionage → hardening, not industry halt
Chip export controls	2022–26	BIS tiering	Race frame dominates	Controls default; dialogue sidecar
Geneva AI dialogue	2024-05-14	Track I #1	No joint statement	Talks cheap; treaties dear
Fable/Mythos ban	2026-06	EAR on deployed API	Partial restore Jul 1	Export ctrl on models now real — theft response partially pre-deployed

---

19. Fable/Mythos export ban — live modal precedent

Jun 2026 sequence (Anthropic_Fable_Mythos_export_ban_2026_深度解读.md):

1. 6/9 Fable/Mythos release → 6/12 Commerce deemed-export directive (foreign nationals) → global takedown
2. 6/26 Mythos partial whitelist (~100 US critical-infra orgs)
3. ~7/1 Restrictions lifted after Anthropic security commitments (BBC; BankInfoSecurity)

Node 3 implications:

Implication	Effect on priors
Gov can recall deployed frontier models without weight theft	Export ctrl P ↑ before C5
Jailbreak trigger, not theft	E4 trigger live — Node 3 may merge with export-control track
Government-gated tier release (whitelist)	Modal deployment control, not training stop
Allies sovereignty shock	T4 allied fracture ↑
Dario FAA-style mandatory testing Jun 10 → ban Jun 12	Labs’ conditional pause rhetoric ≠ operational halt

---

20. US–China dialogue evidence (Geneva + 2026 restart)

Track I inventory (research_us_china_ai_dialogues.md):

• 2023-11 SF: establish mechanism
• 2024-05-14 Geneva: only formal Track I session to date — no deliverables
• 2024-08: agree “timely” round 2 — not publicly held
• 2026-05 Beijing: Trump–Xi agree restart; Bessent: guardrails, non-state actors, 4–8 weeks — unconfirmed schedule

Agenda mismatch: US — catastrophic risk + misuse; CN — lift chip controls + UN centrality. Theft crisis reinforces US hawk side; weakens CN incentive for limits unless bundled with semiconductor concessions.

P(guardrails talk | modal) = 0.50 reflects restart pressure, not treaty probability.

---

21. International governance — why treaty tail is thin

Platform	2024–26 signal	Implication
Paris 2025	US unsigned ministerial	US won’t bind via summit
Seoul 2024	CN skipped ministerial	Safety coalition without CN
India 2026	88 signatories; sovereign AI	Global South anti-freeze
UN GDC / Dialogue	Non-military; Jul 2026 Geneva	No weights verification mandate
REAIM military AI	CN didn’t sign Blueprint	Mil AI split from civilian

Weights verification (AI 2027 Appendix S): HEM / compute pause / AI polygraph — all immature in 2026. Treaty tail T2 needs hardware progress pause playbook puts at 3–5 yr medium horizon.

---

22. Pause playbook — why modal ≠ pause

From research_ai_pause_advocacy_playbook.md:

Lever	Viable 2026?	Node 3 role
Compute export controls	Yes — containing CN	Modal core
Training FLOP cap	No federally	Not modal
Multilateral verification treaty	No — 7–9 prerequisites missing	T2 only
Voluntary frontier commitments	No enforcement	Seoul 2024
Anthropic conditional pause	Rhetoric only	Expect, not commit

Coalition: Natsec hawks + Big Tech beat pause camp on theft response — race accelerates (T3) more likely than slowdown (0.08).

---

23. p(doom) — conditional lifts (geopolitical + coordination)

Definition slice: ΔP(human extinction by ~2050) from Node 3 branch, primarily via great-power war, coordination failure on pause, and speed → misalignment — not double-counting Node 4 misalignment chain.

23.1 Hanson disjunction channel

P(extinction) ⊃ P(AI-enabled great-power war) + P(misalignment | race accelerated)

Node 3 feeds AI war bucket and raises P(no effective pause) for Node 4.

23.2 Conditional lifts (working — calibrate in my pdoom)

Branch	ΔP(extinction by 2050) vs baseline	Mechanism	Conf
Modal	+0.5–1.5pp	Export ctrl + spend without pause; guardrails theater; slight speed	M
T3 race accelerates	+1–3pp (incremental over modal)	DPA; IC offensive cyber; explicit never-pause	M
T1 kinetic	+2–5pp if fired	Hot war tail; low P	L
T2 treaty	−2–4pp	Rare; cuts coordination crux	L
T4 allied fracture	+0.5–1pp	Weakens US-led chokepoints	L–M

Combined tail T3∪T1 (correlated): +3–8pp over baseline if both escalate — matches parent doc summary.

23.3 Coordination crux link (Node 4)

Node 3 outcome	Effect on Node 4
Modal	P(coordinated pause) ↓; classified evals ↑ → whistleblower fuel
T3	P(extinction | E, modal) ↑ 1–2pp via speed
T2	P(mandatory halt | alignment scare) ↑ — rare

Rob Miles crux: Warning shots scarce — theft crisis consumes political bandwidth on CN, not alignment.

23.4 Double-counting guard

• Do not add modal Node 3 and full Node 4 misalignment chain without conditioning
• AI war bucket overlaps misalignment if same conflict — weight by scenario tree

---

24. Downstream effects (next-node priors)

If modal	If T3	If T2
WSL4 spend institutionalized; training FLOPs ↑	OpenBrain–gov fusion narrative	Verification R&D ↑; 3–6 mo slower frontier
Node 4: more classified evals, same P(halt)	Node 4: alignment scare raced past	Node 4: halt more credible
Node 2 CBRN: decoupled — natsec ≠ bio screening	Chip smuggling ↑	Export ctrl paired with dialogue
EU sovereign AI acceleration	CN open-weights diplomacy	HEM pilot funded

---

25. Falsifiers

Observation	Implication
Ratified US–CN training-limit treaty with verification by 2030	T2 modal wrong; revise treaty P ↑
≥2 frontier labs public halt >90d after theft news	Unilateral slowdown P too low
C5+ reached, zero credible theft attempts through 2030	Attempt P too high; open-weights reduced value more than modeled
Public 2-hour full exfil confirmed	Heist P too low; revisit security model
Kinetic US strike on PRC AI DC	T1 not tail — revise upward
Theft + no export ctrl expansion within 12 mo	Modal export P too high
No guardrails meeting by 2029 despite 2026 restart promise	Guardrails P too high

---

26. Confidence summary

Claim	Conf
Anchor at C5+ ~2028, not Feb 2027	M–H
P(2 hr full heist) low	M
P(state attempt by 2030) moderate–high (~65%)	M (78% expert poll)
Modal = export ctrl + spend, not pause/treaty	M–H
Kinetic <0.02	L–M
Treaty 0.04–0.08	M
Exact branch probabilities	L
Conditional p(doom) lifts	L (needs main forecast mixture calibration)

---

27. External sources

• AI 2027 Security Forecast — 78% poll; WSL timeline; 2-month exfil window
• RAND RRA2849-1 — Securing AI Model Weights
• NSC statement — Geneva AI talks 2024-05-15
• Reuters — US-China AI talks Geneva
• CFR Cyber Operations Tracker
• Anthropic — Fable/Mythos access statement (2026-06-12)
• BBC — US lifts export ban Jul 2026
• Lawfare — export controls on AI models
• AI 2027 Tracker — model theft prediction

28. Repo sources

• (internal note)
• (internal note)
• (internal note)
• (internal note)
• (internal note)
• (internal note)
• (internal note)
• (internal note) (Appendix 5 Security Forecast)
• (internal note)
• (internal note)