Probabilistic computation needs deterministic continuity.
That's the architectural problem we solve.

Irreproducible cognition + institutional memory decay.

The hidden crisis in venture is not data fragmentation. Fragmentation is the symptom. The crisis is that institutional reasoning is irreproducible — decisions made in 2022 cannot be reconstructed in 2026. The IC's evidence weighting lived in a partner's head. The thesis nuance lived in three Slack threads. The override rationale lived in one email. None of it survives partner departures or model upgrades.

SaaS fragmented institutional reasoning. Substrate recombines it.

Five disciplines. Architectural, not declarative.

Five non-negotiable disciplines define what makes the substrate substrate-grade. Generic AI tools have none of these. Substrate-grade architecture has all five, architecturally enforced — not asserted as policy.

Every enduring infrastructure company solves an entropy problem.

Feature companies sell capability. Infrastructure companies solve entropy. The strongest companies in the history of software are entropy-solvers — they exist because some critical institutional resource accumulates disorder under usage, and they impose deterministic order. Reasoning under probabilistic computation is the next entropy frontier.

Three layers. Three IP boundaries. Stable across all surfaces.

The architecture is canonically defined by three layers. The IP boundary is canonically defined by three levels. The relationship between them is the source of the manifesto.

Observe → Explain → Constrain → Govern.

Infrastructure systems evolve through four stages. Most current AI deployments are at stage 1 or 2. The substrate operates at stage 4.

Four tiers. Each agent occupies one.

Every agent in the substrate occupies one of four tiers. Tier ceilings are declared per agent and architecturally enforced — for L8-floor agents, the higher-tier code path does not exist.

FIG. 07 ── CONCENTRIC TIERS · STEWARD ON THE OUTER FLOOR

LLMs operate as compilers over typed specifications. Not as autonomous deciders.

The architectural choice that makes everything else possible: the frontier model is not given free-form authority over institutional decisions. It compiles structured intent — defined in the customer overlay — into structured outputs that flow through the audit chain. The model is execution, not judgment.

Observe → Diagnose → Propose → Approve → Execute → Measure → Recalibrate.

Seven steps. Every step traceable. Every action graded against reality. Loop latency under 5 minutes for standard workflows. The substrate doesn't run open-loop — outputs are continuously validated against observed outcomes, and the methodology overlay gets updated when calibration drifts.

RECALIBRATE feeds back into the methodology overlay when measured drift exceeds threshold. The fund's reasoning improves under load, not in spite of it.

Load-bearing decisions are solved twice.

# Example: IC recommendation diverge-and-reconcile
decision_type: IC_RECOMMENDATION

route_A:
  method: hexframe_composite_score
  evidence_set: [pubmed:34521098, clinicaltrials:NCT04812345, cochrane:CD013999]
  output: 8.17  ·  recommend_invest

route_B:
  method: counter_evidence_search + cohort_pattern_match
  evidence_set: [internal:portco_cohort_DPUK, pubmed:36101234]
  output: 7.84  ·  recommend_pursue_diligence

reconcile:
  agreement: false
  delta: 0.33
  action: SURFACE_TO_PARTNERS
  status: BLOCKING — IC scheduled requires resolution

Disagreement is a feature, not a failure. The substrate raises hidden contradictions before they become 18-month write-offs.

Trust is distributed, not scalar.

Confidence is not “0.87 confident.” Confidence is a topology — distributed across evidence types, each with its own trust character. A score of 8.17 means nothing without knowing what's underneath it. The substrate decomposes every output into the trust-state distribution that produced it.

Every abstract claim ships with operational tether.

We have a discipline that prevents drift into systems philosophy. Every category-naming abstraction is paired with a concrete operational mechanism. The phrase doesn't ship without the tether. Below: the doctrine table that governs every page of this site.

If we ever ship a phrase without a tether, we've drifted into theater. This doctrine prevents that.

LPs and regulators have standing inspection authorization. Five steps. Architecturally cooperative.

Every stakeholder of the substrate retains standing authorization to invoke a five-step verification protocol at any time. Reasonable notice for coordination-required steps (typically 3-5 business days); immediate for the others. The substrate is built to cooperate with verification, not to resist it.

Past decisions replay against past methodology.

The customer overlay is versioned. When you amend it — new scoring weight, new sub-thesis area, new override condition — every prior audit chain entry stays tagged with the version active at the time of the original decision. If an LP asks in 2029 why a 2026 decision was made, the substrate reconstructs the exact context using the 2026 overlay, not the 2029 one.

Patterns travel. With attribution. With compensation.

When a pattern proves cross-cutting across multiple engagements in the same L1 — when a scoring refinement, a workflow innovation, or an evidence-handling protocol developed inside one customer's L2 turns out to apply to the broader specialist VC L1 framework — that pattern can migrate L2 → L1. With explicit attribution. With compensation. With the original customer's concurrence.

The substrate compounds. AI tools decay.