Where tacit cognition
becomes executable infrastructure.
Specialist funds have methodology that lives in partner heads — fifteen years of pattern-recognition, founder-coaching instinct, evidence-rigor calibration. None of it written down. We encode it. The result is queryable, replayable, owned by you (L2), and survives partner transitions.
Three layers. One canonical hierarchy.
The architecture is canonically defined by three layers. The relationship between them is what the manifesto names. See /substrate §05 for the L0 / L1 / L2 IP boundary that maps onto these.
FIG. 01 ── METHODOLOGY · SUBSTRATE · MODEL ── L0 · L1 · L2
Methodology is the invariant. Models are replaceable execution layers. L0 doctrine · L1 framework · L2 instance — three IP boundaries that make the architecture work.
Your YAML. Your scoring. Your voice.
FIG. 02 ── L0+L1 ARCHITECTURE · OVERLAY INJECTION AT BOOTSTRAP
LEFT · ARCHITECTURE
The Generic Architecture (L0 + L1) is fund-agnostic — the eleven agents, the audit chain, the L8 floors, the consent gateways. It's segment-agnostic by design.
RIGHT · OVERLAY
The methodology overlay (L2) — customers/[fund].yaml — is what makes it yours. It loads at system bootstrap. Every agent references it. Every audit chain entry tags its version.
Phase 0 is when we extract the overlay from partner heads. Two 90-minute methodology workshops produce the first draft. The async questionnaire fills the structured fields. The PoC validates it works against synthetic data. By Day 10, the overlay is v0.1 — your methodology in machine-readable form, owned by you (L2), versioned in git.
Load-bearing decisions are solved twice.
For any decision where being wrong is expensive — IC recommendations, regulatory claims, LP letter framings, override moments — the substrate solves the problem twice via independent routes. Different reasoning paths, different evidence sets, different agents. Then compares.
FIG. 03 ── ROUTE A · ROUTE B · COMPARATOR · SURFACE-ON-DISAGREEMENT
Hidden disagreement is the most expensive failure mode in AI deployment. We refuse to allow it.
Every output labeled by trust character.
Every claim the substrate produces carries an explicit evidence mode. When a partner reviews a Hexframe score, they don't just see the number — they see what kind of trust produced it. Calibrated trust is bounded uncertainty.
FIG. 04 ── EVIDENCE-MODE BREAKDOWN · CONFIDENCE BAND · κ ≥ 0.85 FLOOR
# Example: Hexframe score with evidence mode breakdown
hexframe_composite: 8.17
evidence_modes:
[verified]:
weight: 0.62
sources: [pubmed:34521098, clinicaltrials:NCT04812345]
[inferred]:
weight: 0.22
method: cross_portfolio_pattern_match (k=11)
[model_derived]:
weight: 0.12
method: lr_synthesis_of_founder_narrative
[partner_override]:
weight: 0.04
note: "P_2 sub-dimension adjusted -0.4 — founder signal weak"
flagged:
- counter_evidence: pubmed:36101234 (resolution_required)
confidence_band: [7.91, 8.42]
kappa_last_verified: 0.87 # ≥ 0.85 floor maintainedThree modes label every output: [known] from system memory, [verified] just-now via tool, [inferred] from reasoning. Calibrated. Bounded. Inspectable.
The Hexframe-class primitive.
Every specialist fund has some version of a multi-axis scoring framework. We refer to ours abstractly as a Hexframe-class primitive — three top-level axes, six sub-dimensions, configurable per segment. The Impact overlay realizes this as one set of axes. A Climate overlay realizes it differently. A Healthtech overlay differently still. The pattern survives; the realization adapts.
FIG. 05 ── 3 AXES · 6 SUB-DIMENSIONS · 3 THRESHOLDS · v1
scoring:
framework: hexframe_v1
axes:
- { name: AXIS_1, sub: [SUB_1A, SUB_1B] }
- { name: AXIS_2, sub: [SUB_2A, SUB_2B] }
- { name: AXIS_3, sub: [SUB_3A, SUB_3B] }
weighting:
AXIS_1: 0.40 # configurable per overlay
AXIS_2: 0.30
AXIS_3: 0.30
thresholds:
thesis_fit: 7.0
diligence_pursue: 7.5
ic_recommend: 8.0
override_conditions:
- { trigger: founder_signal_strong, max_delta: 1.0 }
- { trigger: contested_evidence, max_delta: -1.5 }
kappa_verification:
floor: 0.85
cadence: weekly
drift_response: halt_scorer_emit_blockerThe pattern, captured.
Founder coaching is the hardest methodology to encode. Most specialist funds have a partner whose coaching style is the single most valuable thing the fund delivers to portcos. We capture that pattern as workflow, not text.
FIG. 06 ── FOUNDER · SYSTEM · PARTNER ── L8 F2 (DIRECT PATH ABSENT)
Lead time before crisis.
Most fund failures aren't surprises. They have lead-time signals — founder communication patterns shift, calendar density spikes, sentiment proxies decline, decision velocity slows. The methodology overlay encodes which signals matter for your fund, at what thresholds, with what lead time.
≤7 days
WELLBEING SIGNAL LEAD TIME
target across all engagements
FIG. 07 ── LEAD-TIME SIGNALS · ≤7 DAY INTERVENTION WINDOW · CRISIS
Calibrated trust under model uncertainty.
The sophisticated buyer's actual fear is not “the AI will hallucinate” — that's been solved enough times. The structural fear is: “How do I maintain stable confidence levels as my models get upgraded?” That's epistemic trust under model volatility. The substrate addresses it directly. LP and regulator inspection rights (D7 §12, see /substrate §13) make epistemic trust verifiable.
Every replay improves calibration.
FIG. 08 ── DECAY (TYPICAL DEPLOYMENT) · COMPOUNDING (SUBSTRATE)
LEFT · DECAY
Most AI deployments decay. Prompts age. Models churn. The marginal value of an AI tool today is lower than it was 18 months ago.
RIGHT · COMPOUNDING
The substrate compounds. Every replay test sharpens schema migration safety. Every κ ≥ 0.85 verification refines inter-rater calibration. Every Tier promotion validates the methodology overlay against observed outcomes. Every L2 → L1 pattern migration improves the framework for the whole cohort.
Every decision deepens the substrate. What you build never decays.
The model changes.
The substrate remembers.
── FUND AI OS
Methodology is the invariant. Models are replaceable execution layers. The substrate carries continuity across model volatility. That's the architecture.
§ END ── ENGAGE