You publish the thesis that mental health can be measured objectively. The portcos operationalize it. The EU AI Act will require it of them in 79 days. Your fund needs the same primitive — two layers, one substrate.
86 portcos · 12+ AI-native · two layers
The discipline you back inside your portcos is the discipline the AI Act will require from them — and the discipline your Fund III narrative needs from you.
"Making mental health as objectively measurable as physical health."
"AI in health tech remains a major focus … massive potential to help clinicians spend more time with patients by automating administrative tasks."
86 investments · €50M AuM · 9 new in last 12 months. AI-native portcos across mental health (Upheal · Thymia · Meela), chronic care (Oska · Vermut), preventive screening (Circle · Aeon), clinical workflows (Aiomics).
Each portco runs a different measurement regime — video biomarker · NLP · matching algorithm · audio capture · activity log · self-report. The narrative stays clean. The evidence cadence beneath each portco drifts unmeasured at fund level.
Notified-body assessments already cite AI Act compliance now.
Portco fails AI Act technical file → CE stalls → carry exposed. Fund cannot reconstruct evidence chain behind 2025–26 decisions → Fund III narrative softens with corporate LPs (CSS · Schwabe · Raiffeisenbank).
The thesis is the public layer. The portco evidence regimes are the recursive layer. The substrate is the same primitive at both levels.
You back Thymia for biomarker assessment, Upheal for session intelligence, Meela for therapist–patient matching, Aiomics for record structuring. Each one ships measurement-grade evidence of something previously subjective. The discipline they ship to clinicians is what the AI Act now requires from them — and what your fund needs for LPs and Fund III. One primitive. Two layers.
- AUDIT CHAIN
- every Rx reading per portco hash-chained, append-only, never-deletable, with evidence mode tagged at emission [verified] [model-derived] [self-reported] [stale] [contradictory]. Runs at both layers — portco-side under AI Act review, fund-side under LP defensibility.
- METHODOLOGY OVERLAY · L2 YAML
- each portco's Rx, Tx, wx encoded as inspectable YAML. What Thymia operationalizes for clinicians is the same artifact the fund operates on internally — versioned, portable, replay-tested.
FIG. ── ONE SUBSTRATE · TWO LAYERS · PORTCO-SIDE FOR AI ACT · FUND-SIDE FOR LP DEFENSIBILITY
FIG. ── AUDIT CHAIN · HASH-LINKED · APPEND-ONLY · NEVER-DELETABLE
built for AI Act Article 6(1) replay
Pick one AI-native portco. We rebuild its evidence chain in synthetic. You see whether it survives AI Act Article 6(1) replay or doesn't.
Phase 0 is not a pilot. Two-week structural diagnostic against one portco of your choice — synthetic data only, no production access required. The question it answers is narrow and load-bearing: under AI Act high-risk replay across two methodology versions, does this portco's evidence regime hold — and does your investment thesis evidence chain hold beneath it? Three artifacts come out the other side. You keep all three whether you continue or not.
ONE AI-NATIVE PORTCO · TWO METHODOLOGY VERSIONS · SYNTHETIC EVIDENCE · ARTICLE 6(1) REPLAY MADE VISIBLE
THE DIAGNOSTIC · WHAT TWO WEEKS PRODUCES
- 01Methodology overlay v0.1
The portco's R_x, T_x, w_x encoded as L2 YAML — evidence regime made inspectable. Versioned. Portable. Runs in your environment without us.
- 02Audit chain replay
The chain populated against synthetic Rₓ readings across two methodology versions — Article 6(1) replay simulated before notified-body assessment starts citing it. Drift visible or not.
- 03Cost model
Framed against AI Act technical-file timing, Fund III LP-defensibility, and corporate LP follow-on cadence (CSS · Schwabe · Raiffeisenbank). Risks named. What Phase 1 costs. Where this approach breaks.
Pressure-test the substrate. Ask where it breaks.
The substrate answers in its own voice — the same primitive we'd install on one of your portcos. Not a demo. The actual reasoning surface.
STREAMING · κ ≥ 0.85 · ANONYMITY HOLDS · CITATIONS BY DEFAULT
If the pattern reads true, the next step is to test whether one AI-native portco is the right Phase 0 target.
If relevant, I can send a short async questionnaire — no call needed — and build a no-charge synthetic-data Phase 0 PoC "Calm/Storm AI OS". Within two weeks, replay one AI-native portco's evidence chain against Article 6(1) before the notified bodies start citing it.
Youth mental health · evidence-mode tagging across pre-clinical signal and lived experience.
Youth flourishing · outcomes-tied diligence · methodology overlay encoded at fund formation.
Healthtech first-time fund · evidence-mode discipline encoded before first close.