Eleven agents. One audit chain.
A methodology overlay that travels with your fund.
Six customer-operational specialists. Five architectural agents that keep the substrate substrate-grade. One overlay that's yours. The architecture is segment-agnostic. For deeper architectural doctrine, read /substrate →
System of systems.
FIG. 01 ── FUND LAYER · CONSENT-GATED COMPOSE · PORTCO LAYER
LEFT · FUND LAYER
Eleven agents at fund tier. Six customer-operational specialists handle sourcing, diligence, IC memos, founder support, LP reporting, portfolio strategy. Five architectural agents — Audit Curator, L8-Enforcer, Registry Maintainer, Eval Suite Runner, Version Controller — keep the substrate substrate-grade. All operate within a methodology overlay encoding your specific scoring frameworks, sub-thesis weights, LP voice tiers, regulatory posture.
RIGHT · PORTCO LAYER
Optional. When a founder consents, a per-company instance runs inside their tenant — eleven agents in parallel architecture: six domain operational (Finance, Marketing, Product, Engineering, People [privacy-bounded], Strategy) + five architectural. Aggregate signals flow back to the fund layer through consent gateways at k ≥ 8 anonymity. Revocation halts data flow within 24 hours.
Four tiers. Every agent occupies one.
Every agent in the substrate has a declared tier ceiling. For L8-floor agents, the higher-tier code path does not exist. Tier ceilings are architectural, not policy.
FIG. 02 ── OBSERVER · DRAFTER · OPERATOR · STEWARD
Six specialists. Two with L8 hard caps.
FIG. 03 ── 6 SPECIALISTS · 5 STEWARDS · ONE SOLE-EMITTING CORE
Five Tier-4 Stewards. Most engagements never think about them. That's the point.
The architectural agents keep the substrate substrate-grade. They surveil L8 floors, enforce audit chain integrity, manage versioning, and run κ verification. Most customer engagements never directly interact with them — they operate as the substrate's immune system. But without them, "substrate-grade" is a claim, not a property.
The substrate's immune system. Invisible when working. Load-bearing when stressed.
Append-only. Hash-chained. Sole-emission. Replayable. Governance-grade.
Every system action — every score, every memo, every LP draft, every wellbeing inference — writes to an append-only audit chain via the Audit Curator (sole-emission authority per tier). Each entry hash-links to the prior entry. Tampering breaks continuity end-to-end and is detectable. Beyond defensibility (explaining past decisions), the audit chain enables governance (constraining future ones).
FIG. 04 ── ONE CURATOR · FIVE BLOCKS · CONTINUITY BY CONSTRUCTION
SOLE-EMISSION
Exactly one Audit Curator agent per tier holds emission authority. All other agents route emission requests through the curator. Chain integrity by construction, not policy.
REPLAY
Given an audit ID, the substrate reconstructs the exact context that produced the original output — overlay version, model version, evidence set — and reproduces it bit-for-bit.
SCHEMA MIGRATION
Overlay version-tagged on every entry. Migrations replay against last 100 entries before promotion. Past decisions stay defensible across overlay updates.
NEVER-DELETABLE
Severity:critical entries are filesystem-level append-only. No delete API exposed. Deletion attempts emit a new severity:critical entry (recursively) and themselves persist forever.
ARTICLE 9 SFDR
Every impact KPI movement timestamped, evidenced, traceable. LP letters cite audit chain IDs inline. Regulator-defensible by architecture, not by post-hoc reconstruction.
2026-Q3-W4-T0042 | agent=diligence_lead | overlay_v=1.4.2 | model=opus-4.7
input_hash: f8a3c7b9
output_hash: 2e4d1a06
evidence_set: [pubmed:34521098, clinicaltrials:NCT04812345, nice:NG185]
hexframe_composite: 8.17
partner_overrides: [{sub_dim: P_2, delta: -0.4, rationale: "founder_signal_weak"}]
txn_id: TXN-2026-Q3-W4-0017
severity: info
hash_prev: 7c4f2a8e
hash_self: 3b8d6c5f
emission_curator: audit_curator_fund_tier_v1Every field is queryable. Every entry is replayable. Every claim carries lineage.
Commitments encoded as code-path absence.
L8 architectural floors are commitments that cannot be lifted by policy. They are encoded as the absence of the lifting code in the substrate — not as runtime checks, not as authorization roles, not as configuration flags. The L8-Enforcer surveils continuously. Refusal events emit severity:critical NEVER-deletable. Self-tests run quarterly with 5/5 PASS expected.
FIG. 05 ── FIVE PILLARS · NOTCHED CAPS · ABSENCE BY DESIGN
Floors evolve only via MAJOR amendment cycle + domain-architectural redesign evidence + stakeholder concurrence. BOTH required. Either alone is refused.
Observe → Explain → Constrain → Govern.
FIG. 06 ── FOUR STAGES · ASCENDING PLATEAUS
OBSERVE
Logs, metrics, events.
Where most AI bolt-ons stop.
EXPLAIN
Replay, evidence lineage, audit chain.
Where mature audit-aware AI operates.
CONSTRAIN
Refusal under threshold · tier caps · escalation.
Where the substrate constrains its own outputs.
GOVERN
L8 architectural floors · authorization topology · schema lock · reversibility.
Where the substrate operates by design.
Defensibility explains past decisions. Governance shapes future ones.
Three layers. Each one revocable.
FIG. 07 ── CONCENTRIC BOUNDARIES · OUTWARD REVOCATION
Your methodology, in a single YAML file.
The Generic Architecture is fund-agnostic. The customer overlay is where your specific methodology lives: scoring axes, sub-dimensions, decision thresholds, LP voice samples, sub-thesis priorities, IAC composition, regulatory cadence. Every agent loads it at bootstrap. Every audit chain entry tags the version.
overlay_version: 1.0.0 # MAJOR.MINOR.PATCH
scoring:
framework: hexframe_v1
axes: [AXIS_1, AXIS_2, AXIS_3]
sub_dimensions: 6
thresholds:
thesis_fit: 7.0
diligence_pursue: 7.5
ic_recommend: 8.0
evidence_base:
sources: [pubmed, clinicaltrials, cochrane, nice]
counter_evidence_band: [0.85, 0.90]
kappa_verification:
floor: 0.85 # Cohen's κ weighted variant
cadence: weekly_primary_monthly_voice_quarterly_comprehensive
lp_voice_tiers:
- { name: tier_1, samples: 10, framing: rigor_first }
- { name: tier_2, samples: 8, framing: regulatory_cascade }
- { name: tier_3, samples: 6, framing: returns_balanced }
- { name: tier_4, samples: 4, framing: institutional }
governance:
l8_floors:
investment_principal: tier_2_hard_cap
founder_support: tier_2_hard_cap
wellbeing_pathway: default_off_opt_in_required
federation_k_floor: 8
consent_gateway_2: fail_closed_default
confidence_escalation_floor: 0.70Axes are abstract placeholders until Phase 0 workshops resolve them. The overlay is versioned in git. You own it (L2 IP).
What we don't build.
§ END ── DEEPER