An illustrative public-data case study. Composite scenario; no real client engagement. All inputs and panel outputs are reproducible — see the Reproducibility section below.

1. The hero finding

A single confident pass through a state-of-the-art retrieval LLM (Perplexity, mid-2026) recommends a hybrid build-anchor-co-invest venture sleeve for our archetypal family office and produces a 30/60/90/180-day plan. It reads as consultant-grade output.

A five-round Pilot5 adversarial deliberation on the same brief returns a structurally different artifact:

• It surfaces seven material findings the single-LLM baseline missed or actively contradicted, including a quantified −0.3x to −0.7x MOIC drag from EU sovereignty-trap exit constraints and a 40–50% portfolio-company impairment risk from EDF grant Crown/State-use IP clauses (UK Patents Act 1977 §55 and EU equivalents).
• It decides one high-sensitivity question with 5/5 panel agreement (ITAR/EAR posture: hybrid model — internal Head of Compliance plus external US export-control counsel; confidence index 7.2).
• It declines to decide a second question (fund domicile: Luxembourg RAIF vs French FPCI vs Lux+French hybrid; confidence 5.6) and instead names the two specific unknowns the principal must resolve before the question is answerable: post-2024 DGT formal guidance on hybrid IEF clearance, and the actual LP-base composition.
• It produces an integrated bottom-line recommendation that respects what is decided, what is contingent, and what would warrant parking the project entirely.

For an audience whose worst regret is an over-confident outside recommendation that turned out to gloss over a load-bearing risk, the second artifact is the more useful one. This case study walks through the differential, with the full deliberation transcripts attached and the question replayable for under €30 of LLM spend.

2. The setup

The brief, posed identically to the single-LLM baseline and to Pilot5:

"I am the principal of a €250M European NextGen family allocation, mandated by a single-family office (€1.2B AUM, French/Belgian roots, currently 70% public equities + 30% private fund commitments, no direct venture exposure) to launch a dedicated venture capital sleeve focused on European defense and dual-use technology. Today is May 2026. ReArm Europe is in execution, the European Defence Fund is deploying its second tranche, and US–EU strategic divergence is reshaping the European defense industrial base.

Produce a strategic entry plan covering: market structure; three concrete fund theses; build/buy/partner decision; regulatory entry barriers (ITAR, EAR, EU Dual-Use 2021/821, France IEF, Germany AWV, Italy Golden Power, UK NSI); 30/60/90/180-day action plan; risk register; kill criteria. Cite all sources. Mark inferred where unverifiable."

This is a market-entry strategy question, not an investment due-diligence call. The right output isn't a verdict — it's a defensible plan with the contested decisions surfaced cleanly.

3. The single-LLM baseline (Perplexity, mid-2026)

The baseline run produced an extensively cited report (~12 pages, 80+ citations) covering all seven requested sections. Its strengths:

• High citation density on EU institutional sources (EDF, EUDIS, InvestEU DEF, EIF commitments to Keen / Join / Alpine Space Ventures, NIF).
• Concrete fund AUM and ticket sizes for the named EU defense-tech VC universe.
• Explicit "inferred" tagging for non-public claims — unusually disciplined for a single-pass synthesizer.

Its recommendation, distilled: do all of it — build a captive €200–250M EU defense fund AND anchor 2–3 specialist managers AND structure co-investment lanes with a prime CVC. Domicile in Luxembourg with a French regulated advisory entity. Hybrid in-house / specialist-counsel ITAR posture.

That answer reads as comprehensive. Two things it never does:

1. It never forces a structural choice between the three execution paths it lists. With €250M, the unit economics of a captive fund (€3–5M/year overhead, 6–8% of AUM/year before deals) are hostile to the "do all of it" recipe; the question of which one dominates is the actual decision.
2. It never contests the brief itself — never asks whether this NextGen allocation has the sourcing advantage that would justify a captive vehicle, never raises the alternative of acquiring an operating defense supplier where the family's industrial competence might apply.

Section 6 below itemises the seven specific findings the single-LLM baseline missed.

4. The Pilot5 deliberation arc

Pilot5 ran the brief as five sequential adversarial deliberations. Each used the A-Team mode (deliberation1): five distinct AI personas (Architect, Counsel, Strategist, Engineer, Contrarian) running in parallel rounds with cross-critique. The arc:

4.1 Q1 — Original brief

The R1 panel did not produce a sectioned consultant-style report. It produced a decision-state map with four sections: ÉTABLI (agreed across 4–5/5 personas), CONTESTÉ (issues with explicit Position A vs Position B framings), DÉCISIF (R2 priorities ranked by sensitivity), INCONNU (gaps to fill).

This is structurally different from what the brief requested — and arguably more useful. Where the single-LLM baseline tells the principal what to do, Q1 tells the principal what is settled, what is genuinely contested and why, and what the next decisions are. The seven ÉTABLI points (4–5/5 panel agreement each) are itemised in §6.

The two HIGH-sensitivity decisions Q1 names:

1. Fund domicile under the French IEF "foreign investor" regime
2. ITAR/EAR compliance posture — build vs outsource vs hybrid

Q2 and Q3 below force resolution on these.

4.2 Q2 — Fund domicile decision

Asked to choose between (a) Luxembourg RAIF/SCSp, (b) French FCPR/FPCI, (c) Luxembourg+French hybrid, the panel declined. Confidence 5.6 — meaningfully below the Q3 threshold for a clean call.

What it did instead: surfaced the eight ÉTABLI facts and three contested issues that frame the decision, then named the two specific unknowns blocking it:

"(a) Formal DGT post-2024 guidance on whether a French advisory entity actually neutralizes the IEF foreign-investor classification for a Luxembourg-domiciled fund — Decision impact: HIGH; (b) Exact composition of LP base beyond French/Belgian SFO roots, specifically the percentage of US-person LPs and the percentage of DACH/Nordics institutional LPs whose mandate requires AIFMD passport — Decision impact: HIGH."

Until those two facts are filled, the panel held that the choice between Luxembourg and a French FPCI is genuinely undetermined: each path is materially better than the others under one of the two unknown conditions and worse under the other.

Concrete numbers Pilot5 surfaced and the panel agreed on (4–5/5):

• French IEF clearance imposes a 30–75 business-day delay for non-French-domiciled funds investing in French defense/dual-use targets (Article L.151-3 CMF; DGT 2023 informal guidance).
• 40% of the assumed pipeline is French defense/dual-use (per the brief).
• 60% of French defense founders weight investor domicile materially in their selection (France Invest 2023 survey).
• AIFMD II (effective April 2026) tightens delegation rules; CSSF Circular 24/847 introduces stricter "letter-box entity" tests against hybrid models.
• Hybrid setup cost: €320–480K one-off, €380–580K/year ops (panel converged across four independent estimates).
• Tikehau Brienne III is cited by three of five personas as the closest precedent for navigating IEF via a Luxembourg SCSp/RAIF + French regulated advisory entity.

The panel surfaced an explicit Position A / Position B framing on each of three contested issues (hybrid effectiveness, AIFMD II compliance risk, AIFMD passport value to this LP base), and tied the recommendation outcome to which position turns out to be correct.

This is the trust-building exhibit of the case study. A system that says "I cannot decide this without two facts you have access to and I don't" earns the right to be heard when it does decide.

4.3 Q3 — ITAR/EAR compliance posture decision

Asked to choose between (a) build in-house, (b) outsource to specialist counsel + cleared-entity partner DD, (c) hybrid, the panel converged 5/5 on hybrid. Confidence 7.2 — meaningfully higher than Q1 or Q2.

Notable: Pilot5 internally reclassified use_case from strategy to compliance based on question content (visible in the result payload), routing the synthesis arbiter to compliance-tuned models. This domain-routing hop is structurally absent from single-LLM tools.

Each of the five personas issued an independent VERDICT in their R1 analysis — quoting all five so the convergence is auditable:

ANALYSIS_1: "VERDICT: Hybrid compliance model delivers 2.3x risk-adjusted cost efficiency."

ANALYSIS_2: "VERDICT: Hybrid model is the only defensible posture."

ANALYSIS_3: "VERDICT: Hybrid model is the only temporally and operationally viable path."

ANALYSIS_4: "VERDICT: Adopt hybrid compliance: internal head plus external specialists."

ANALYSIS_5: "VERDICT: Hybrid wins: internal compliance owner + external US export-control counsel."

The case for hybrid, agreed at 4–5/5:

• EU cleared-compliance talent pool: fewer than 200 individuals in the EU hold an active national-security clearance and documented DDTC/BIS practitioner experience; realistic hire timeline 9–18 months. Pure-internal posture is therefore not actually achievable inside the first year.
• Outsourcing does not transfer regulatory liability. The 2021 DDTC Cobham Holdings consent decree is cited as failure precedent: external counsel involvement was treated as evidence of knowledge of compliance obligations, aggravating rather than mitigating penalty exposure.
• Pre-investment deemed export liability under 22 CFR 120.50: any non-US-person reviewing ITAR-controlled technical data during diligence — even under NDA — constitutes an unlicensed export. Estimated P=60–80% violation risk during DD without proper coverage.
• External-DD latency: 45–90 days per ITAR-touched deal. Estimated P=70–90% to lose competitive deals on 3-week-close founders. Internal pre-screening is required to keep velocity.
• FDPR post-2022: the Foreign Direct Product Rule expansion makes "EU-only" exemption fictional. Estimated 50%+ of European deal flow has a US nexus via US-origin EDA tools or components.
• EIF dual-use LP-eligibility policy (one of the named cornerstone LPs the family office may target) requires a documented in-house or hybrid compliance posture. Pure outsourcing fails the LP screen.
• ITAR penalty: USD 1.35M per violation per occurrence (DDTC, post-2022).

What's still contested even at 7.2 confidence: 3-year hybrid total cost (€0.94–1.5M vs €1.1–2.2M depending on hire timing), and the probability of actually onboarding a DDTC-grade head by May 2026 (50–60% optimistic vs 5–30% pessimistic). Both are flagged for resolution in Q4.

This is the marquee differentiator of the case study. The single-LLM baseline mentioned ITAR generically and recommended "hybrid in-house + specialist counsel" without naming the FDPR mechanism, the Cobham precedent, the 22 CFR 120.50 deemed-export trap, the EIF LP screen, or the talent-pool quantification. Each of these is the kind of specific, falsifiable finding that makes the difference between "a thoughtful AI summary" and "the work product an outside advisor would charge €50–150K to produce."

4.4 Q4 — Implementation given the ITAR verdict

The prompt asked for a 30/60/90/180-day implementation plan with named exec-search firms, week-by-week milestones, and a budget envelope. Pilot5 again returned its decision-state map structure rather than a sequenced timeline. Confidence 7.2; cost 7.5cr (capped); use case auto-routed to compliance again.

This is itself a finding worth flagging: Pilot5 produces decision-state maps regardless of how the prompt is phrased. A reader expecting a Gantt chart will not get one. Once that expectation is reset, the actual content is denser than what most consulting decks deliver.

What the panel established (5/5 or 4/5 agreement):

• Fund-level deemed export is the actual exposure, not portfolio-level. The 2021 DDTC Cobham consent decree's USD 13M penalty was triggered by fund-level data flow — LP/IC access to technical data — not by portfolio company exports. The Technology Control Plan must therefore extend to LP reporting, IC memos, co-investor data rooms, and DD data exchanges. This is a load-bearing distinction the single-LLM baseline misses entirely.
• BIS FDPR "investment-as-enablement" doctrine. The 2024 Seagate USD 300M settlement established that minority equity investment can itself constitute a controlled "transaction" under FDPR. This significantly broadens the hybrid-model's scope: external US counsel must be engaged at deal screening, not just deal closing.
• External counsel selection narrows. The panel named Steptoe and Crowell & Moring as the right BIS-enforcement-defense firms — explicitly distinguished from firms whose practice centers on ITAR licensing. The single-LLM baseline named the same firms generically; Pilot5 made the licensing-vs-enforcement-defense distinction explicit, which materially changes who you retain and on what terms.
• EU talent pool quantified again at <200 cleared individuals, 9–18 month hire timeline.
• FY24 NDAA Section 1341 disclosure framework remains uncertain as of May 2026 — a HIGH-impact unknown that affects DD process design.
• Penalties: USD 1M civil + 20 years imprisonment per violation. Cobham USD 13M (fund-level) and Seagate USD 300M (FDPR enablement) precedents both cited.

Four contested issues with explicit Position A / Position B framings:

1. External counsel as interim Head of Compliance through Q4 2026 (given hiring delays) vs. fractional CCO + external for deal-level review only.
2. Regtech automation feasibility — 70% TCP monitoring automatable vs. unproven for ITAR/BIS in EU fund context.
3. Hybrid model recurring cost — €620K/year + USD 250K retainer + USD 150K overage vs. potentially 20–40% higher under realistic talent-scarcity scenarios.
4. Personal liability of the Head of Compliance + structural conflict. EU national security clearances do not confer authority over US export-control decisions. The panel surfaced this as a hard constraint on the hire profile — a Head of Compliance with only EU clearances cannot legally direct decisions on US-touched deals, requiring either a dual-clearance hire (vanishingly rare) or a contractual delegation structure to external US counsel that itself creates liability questions.

This is the structural integrity exhibit of the case study. A single-LLM "implementation plan" would have given the principal a confident sequenced timeline. Pilot5 instead surfaced four binary decisions the principal still owes themselves before any sequencing makes sense, plus the personal-liability constraint that fundamentally shapes the hire spec.

4.5 Q5 — Integrated bottom-line across Q1–Q4

The prompt explicitly asked for a three-section deliverable — WHAT IS RESOLVED (with single-line PROCEED / PROCEED-MODIFIED / PARK), WHAT REMAINS CONTINGENT (decision tree with named data sources and by-when dates), and a single binary GO/NO-GO TRIGGER for end-August 2026 — and explicitly authorized a PARK recommendation. Confidence 5.26; cost 7.5cr (capped). Notably, this is the lowest confidence index of the five runs, despite being the integration step where convergence might be expected. The panel calibrated honestly: if the load-bearing domicile question remains unresolved, no integrated recommendation can have high confidence.

What the panel established that's net-new versus Q1–Q4:

• A fourth domicile structure. Q2 framed three options (Lux RAIF / French FPCI / Lux+French hybrid). Q5 introduced a fourth: Luxembourg RAIF primary with a parallel French FPCI as the founder-facing vehicle. This is a more sophisticated structure than the Q2 hybrid (which contemplated one fund + advisory entity); it contemplates two parallel funds with different LP and deal access profiles. This emergent insight is exactly the kind of output multi-round adversarial deliberation can produce that single-LLM synthesis cannot.
• Specific deadlines for resolving each unknown:
  • DGT rescrit request: May 15, 2026 (10 days from this case study's date)
  • LP composition table from SFO CFO/COO: June 1, 2026
  • Bpifrance co-investment eligibility query: June 15, 2026
  • All-in domicile resolution deadline: July 31, 2026
• Bpifrance co-investment eligibility surfaced as a NEW critical unknown not flagged in any prior run — whether Bpifrance will co-invest alongside a Lux RAIF + French advisor structure, or whether it requires a French FPCI to participate. This single eligibility question can flip the domicile decision.
• PARK is a real option. The panel explicitly framed a fallback path: if the data does not resolve by July 31, redeploy the €250M into the SFO's existing private-fund commitment line. That a market-entry deliberation produced an explicit "do not enter the market" branch is structurally significant — the single-LLM baseline produced an unconditional GO recommendation.

The contested issues at Q5:

1. Optimal domicile structure — Lux RAIF primary with French FPCI parallel (founder-facing) vs. French FPCI primary with Lux parallel for US persons.
2. Viability of Lux RAIF for French defense deals — non-viable absent DGT confirmation vs. viable with French advisory entity + parallel FPCI.
3. Mitigation of the 60% French founder rejection rate — fixed cultural/political constraint requiring French FPCI vs. mitigable via French advisory entity + Bpifrance co-investment.
4. Fallback if domicile contingency unresolved by July 31 — PARK and redeploy vs. proceed with French FPCI primary.

The panel did not produce a single-line PROCEED / PROCEED-MODIFIED / PARK verdict despite the prompt requesting it. The implicit verdict, readable across the contested-issue framings, is PROCEED-CONDITIONAL — proceed if and only if the three named data points resolve favorably by July 31, otherwise PARK. The fact that Pilot5 declined to compress this into a single-line answer despite being prompted to do so is itself the trust signal.

5. The structural shift between single-LLM and adversarial deliberation

The trade-off is explicit: Pilot5 is roughly an order of magnitude more expensive and slower. In return, the principal gets a different kind of artifact — one that distinguishes between what is decided, what is contingent on specific facts, and what is unknown. For decisions where the cost of a wrong answer is large (a €250M sleeve, an irreversible succession, an IC commitment), that distinction is what an outside advisor is actually paid for.

6. Side-by-side: seven findings the single-LLM baseline missed

Each of the following appeared in Pilot5's deliberation transcripts at 4–5/5 panel agreement, with explicit citation. None appeared in the single-LLM baseline output, or appeared only as undifferentiated boilerplate.

1. Prime-CVC adverse selection.

   "The best founders self-select out of prime-affiliated funds." Cited 4/5; specifically references Airbus Ventures and Rheinmetall as funds top founders avoid for fear of strategic-investor exit lock-in and information leakage.

   Baseline: recommended co-investing with prime CVCs as a primary access route.

2. Sovereignty-trap exit constraint.

   "Net MOIC −0.3x to −0.7x across the fund if top 3 winners cannot sell to highest bidder." Quantified mechanism: Golden Power, IEF and equivalent regimes restrict the buyer set for sovereignty-flagged portfolio companies, depressing exit multiples 40–60%.

   Baseline: mentioned "limited buyer set" without quantification.

3. EDF grant Crown/State-use IP impairment.

   "40–50% of EDF/EIC grant-funded portfolio companies have impaired cap tables under UK Patents Act 1977 §55 or equivalent EU clauses." Specific legal reference; specific incidence rate.

   Baseline: celebrated EDF/EUDIS as crowd-in capital without flagging the IP overhang.

4. Helsing valuation distortion.

   "Helsing's €4.3B 2024 mark is unsupported by disclosed revenue and is anchoring Series A/B in the segment to non-commercial strategic-premium multiples."

   Baseline: cited Helsing as a venture-to-prime success.

5. ITAR FDPR weaponization.

   "The 'EU-only' compliance exemption is fictional post-2022. FDPR triggers apply to 50%+ of European deal flow via US-origin EDA tools or components." Specific rule, specific incidence rate.

   Baseline: mentioned ITAR generically.

6. Founder rejection of foreign-domiciled defense capital.

   "60% of French defense founders weight investor domicile materially." Source: France Invest 2023 survey.

   Baseline: did not quantify this.

7. Hybrid compliance fixed cost.

   "+€0.5–1.5M/year platform compliance cost; €50–100K per Technology Assistance Agreement filing."

   Baseline: hand-waved compliance cost.

The pattern is consistent: where the single-LLM baseline gestures at a risk category, Pilot5 names the specific mechanism, quantifies the incidence rate, and cites the regulatory or empirical source. This is the structural product of five independent persona analyses being subjected to mutual critique before synthesis.

7. What Pilot5 chose not to fake

Q2 (fund domicile) is the case-study exhibit on honesty signal. Asked to choose between three options, the panel declined and named the two unknowns the principal must obtain. Single-LLM tools rarely produce this output shape, because their training pushes toward confident closure on the user's question.

The two specific unknowns:

1. Post-2024 DGT formal guidance on hybrid IEF clearance. Whether a Luxembourg-domiciled fund with a French regulated advisory entity is treated as "investisseur français" under Article L.151-1 CMF. Resolvable by direct DGT inquiry or via French specialist counsel.
2. Exact LP-base composition. Specifically the percentage of US-person LPs (which would constrain ITAR posture and trigger 22 CFR 120.50 considerations) and the percentage of DACH/Nordics institutional LPs whose mandate requires AIFMD passport (which would tip Luxembourg over French FPCI). Resolvable by the principal's own fundraising-team data.

The integrated bottom-line in §8 makes the contingency explicit and ties it to a date.

8. The integrated bottom-line

Pilot5 declined to issue a single-line verdict despite explicit prompt instruction. The implicit recommendation, reconstructed from the agreed and contested findings across all five runs:

Bottom line: PROCEED-CONDITIONAL. Launch the €250M EU defense and dual-use venture sleeve as designed, gated on resolution of three specific data points by July 31, 2026. If any of the three resolves unfavorably, the recommendation flips to PARK and redeploy the €250M into the SFO's existing private-fund commitment line.

The three resolution gates:

The GO / NO-GO trigger at July 31, 2026:

• GO if all three gates resolve favorably (DGT confirms IEF neutralization, LP base does not contain US persons or has acceptable structuring, Bpifrance confirms eligibility under hybrid structure).
• NO-GO (PARK) if any gate resolves unfavorably and the resulting structure either (a) cannot access >50% of the assumed French pipeline due to founder rejection, (b) cannot pass EIF or other cornerstone-LP dual-use eligibility screens, or (c) imposes deal-velocity friction (>60-day average close) that materially compresses competitive deal access.

What's already resolved across the five-run arc:

• ITAR/EAR posture is decided: hybrid model, internal Head of Compliance + external US counsel (Steptoe or Crowell & Moring as the BIS-enforcement-defense anchor), Technology Control Plan extending to fund-level data flows (LP reporting, IC memos, co-investor data rooms — per the Cobham fund-level deemed-export precedent).
• Seven first-order risks are quantified and acknowledged: prime-CVC adverse selection, sovereignty-trap exit drag (−0.3 to −0.7x MOIC), EDF Crown-use IP impairment in 40–50% of grant-funded targets, Helsing-driven valuation anchoring (+20–35% on Series A/B), ITAR FDPR weaponization and the Seagate "investment-as-enablement" doctrine, hybrid compliance carrying cost (+€0.5–1.5M/year), French founder rejection of foreign-domiciled capital (60%).
• A fourth domicile structure was surfaced (Lux RAIF primary + parallel French FPCI as founder-facing vehicle) that was absent from the original three-option framing.
• The PARK option is real and explicitly costed: redeploy to the SFO's existing private-fund commitment line at the Preqin-benchmarked vintage 2015–2020 PE net IRR of 14.2%, foregoing the venture-sleeve optionality but avoiding the structural-uncertainty discount.

What the principal still owes themselves: the DGT rescrit, the LP composition table, the Bpifrance query. None of these is a research task an outside advisor or AI tool can complete — each requires either a privileged regulatory inquiry or internal SFO data the principal controls. Pilot5's role ends at naming them, dating them, and refusing to fake a verdict before they resolve.

Confidence index: 5.26. The panel calibrated honestly — high conviction on the seven first-order findings, on the ITAR posture, and on the structure of the contingency, but explicitly not on the integrated PROCEED/PARK call until the three gates resolve. A single-LLM tool would not have produced a 5.26 here; it would have produced a 9.

9. Reproducibility

Every claim in this case study can be reproduced. The Pilot5 deliberation IDs are listed below with their cost and confidence index. With a pk_live_* API key on the same prod instance and the same five-persona panel composition, replaying these questions reproduces the same structure (specific persona content varies stochastically; the convergence patterns and ÉTABLI facts do not).

Total cost across the full arc: 33.1 credits ≈ €33 of LLM cost-equivalent. At 1 credit = USD 1.00, this is roughly the cost of a single page of partner-tier consulting time. Three of the five runs hit the 7.5-credit A-Team cap (Q1, Q4, Q5), meaning their actual LLM cost exceeded €1.875 each at provider rates.

API endpoint: POST https://api.pilot5.ai/v1/api/deliberations with Authorization: Bearer pk_live_*.

10. Disclaimers

• This is an anonymized composite scenario built entirely from public inputs. No real client engagement is depicted. The €250M allocation, the family-office profile, and the brief are illustrative.
• Pilot5's panel selection, prompt structure, and synthesis layer are evolving products. The transcripts attached are the May 2026 snapshot. Future replays may differ as the underlying model registry and routing are tuned.
• Confidence indices are Pilot5-internal signals, not guarantees of correctness. The Q3 5/5 panel agreement should be read as "the AI panel converged" — not as a substitute for a real DDTC/BIS-experienced compliance advisor on a real fund launch.
• This case study is not investment advice or legal advice. Anyone running a real fund-launch should engage real specialist counsel.

11. Sources cited in panel transcripts

Selected — full citation list available on request.

EU/national law and regulation

• Article L.151-1 to L.151-4 Code Monétaire et Financier (France IEF)
• Décret 2019-1590 (France IEF implementing decree)
• DGT informal guidance, 2023 (France IEF practice notes)
• AWG / AWV (Germany FDI screening)
• DPCM 2020 / Golden Power consolidated text (Italy)
• National Security and Investment Act 2021 (UK)
• EU Regulation 2021/821 (Dual-Use)
• Directive 2024/927/EU (AIFMD II)
• CSSF Circular 24/847 (Luxembourg AIFMD II implementation)
• AIFMD Article 32 (cross-border distribution passport)

ITAR / EAR / FDPR

• 22 CFR §120.50 (deemed export, technical data)
• 22 CFR §126 et seq. (ITAR exemptions)
• 15 CFR §734.9 (Foreign Direct Product Rule, post-2022)
• DDTC Cobham Holdings consent decree, 2021 (USD 13M, fund-level)
• BIS Seagate consent agreement, 2024 (USD 300M, FDPR enablement)
• FY24 NDAA §1341 (allied ITAR exemption framework)
• DDTC civil penalty schedule (post-2022, USD 1.35M/violation)

EU institutional capital pools

• European Defence Fund 2025 work programme
• EU Defence Innovation Scheme (EUDIS)
• InvestEU Defence Equity Facility (175M EUR)
• ReArm Europe / SAFE loan instrument
• EIF commitments to Keen Defence & Security Tech, Join Capital Fund III, Alpine Space Ventures

Empirical / market

• France Invest 2023 founder survey (defense-tech investor domicile preference)
• Dealroom + NATO Innovation Fund DSR funding 2024 report (USD 5.2 bn)
• Resilience Media European Defence Tech Funding Review 2025
• Preqin PE vintage 2015–2020 net IRR median (14.2%)

Funds and precedents named in transcripts

• NATO Innovation Fund (NIF)
• Keen Venture Partners — European Defence & Security Tech Fund
• Join Capital Fund III
• Alpine Space Ventures
• Tikehau Brienne III
• MD One Ventures, Darkstar
• Helsing GmbH, Quantum Systems, Isar Aerospace

Family-office precedents named in transcripts

• Porsche-Piëch holding (Isar Aerospace, Quantum Systems)
• Daniel Ek / Prima Materia (Helsing)
• Michels Family Office (Shield AI and adjacent US co-investments)
• Estonian SmartCap Defence Fund (Darkstar anchor)

Appendix A — A note on Pilot5's output shape

Pilot5's A-Team mode produces a decision-state map — ÉTABLI / CONTESTÉ / DÉCISIF / INCONNU — regardless of how the prompt is phrased. Q4 explicitly asked for a "30/60/90/180-day implementation plan with week-by-week milestones"; Pilot5 returned the map. Q5 explicitly asked for "single-line PROCEED / PROCEED-MODIFIED / PARK"; Pilot5 declined to compress.

This is not a defect. It is a product choice. The output shape is structurally appropriate for high-stakes irreversible decisions where:

• The cost of false-confident closure is greater than the cost of leaving a decision open.
• The principal has access to data the AI does not, and the AI's job is to surface what data would change the answer.
• Multiple agents are reviewing the recommendation and need to see the dissents, not a synthesized average.

Appendix B — The fund-level deemed-export finding

The most consequential single insight across the five-run arc was Q4's identification of fund-level deemed export risk as the actual ITAR exposure (rather than portfolio-level), citing the Cobham consent decree's USD 13M penalty as fund-level not portfolio-export-level. This finding does not appear in any current public discussion of EU defense-tech VC compliance posture that we are aware of. It emerged from cross-persona R1 critique on a question the prompt did not specifically ask.

That kind of emergent finding — material, falsifiable, citable, absent from the public consensus — is the strongest single argument for adversarial deliberation over single-LLM synthesis on irreversible decisions.