Nascens

Architecture

Built on the Atlas spine.

The architecture behind Nascens is proven across four operational domains — private equity monitoring, insurance operations, procurement compliance, and clinical decision support. This is the venture-formation instantiation: the same episodic intelligence model, applied to the formation funnel.

Formation intelligence is a mentoring tool — not a founder dashboard, not a self-assessment platform. The architecture reflects that. What mentors see and what founders see are deliberately different. That asymmetry is not cosmetic. It is structural.

The Atlas Venture Spine (AVS)

Seven stages. Every interaction — a pitch session, a milestone review, a pivot declaration, a mentor observation — triggers a full run of the spine. State persists; the next event picks up exactly where the last left off.

01
Intake
Raw text: pitch decks, session notes, milestone updates, advisor observations
02
Normalize
LLM parses free text into the 6-dimension signal vector with confidence labels
03
Opportunity State
Current venture state classified from signal vector: forming, validating, progressing…
04
Stage Transition
Transition triggers evaluated against population history; threshold evidence assessed
05
Trajectory Estimate
Forward projection derived from population analogs; confidence tier applied
06
Guidance
Actionable output surfaced: what to prioritize, what assumptions to stress-test, mentor match signal
07
Track
Episodic record committed to population database; state persists until next event

No normalization, no downstream operation. The LLM normalization step at Stage 02 is a gate, not a filter. If the intake text cannot be confidently parsed into the signal vector, the run halts and surfaces the gap rather than propagating low-quality signal.

From free text to structured signal.

The founding constraint of venture formation intelligence is that almost all meaningful data exists as free text — pitch narratives, session transcripts, mentor observations, cohort notes. No program operates with structured intake at the field level. Nascens does not require them to.

The LLM normalization layer, powered by Anthropic claude-sonnet-5, converts any free-text submission into a 6-dimension signal vector. Each dimension is scored, a confidence label is assigned, and the labeled vector — not the raw text — is what flows downstream.

The six dimensions of the signal vector:

Problem

Specificity, evidence of customer pain, framing of the gap being addressed

Solution

Differentiation, feasibility, alignment to the stated problem

Market

Addressable scope, segmentation clarity, demand evidence

Team

Domain credibility, coverage of key roles, founder–problem fit

Evidence

Validation artifacts: surveys, pilots, LOIs, revenue, usage data

Assumptions

Explicit dependencies not yet validated; a first-class signal, not a deficiency

The 7th dimension: Coachability — mentor view only. Whether a founder integrates feedback, follows through on commitments, and adapts when evidence contradicts their framing is the most predictive single indicator of venture trajectory that experienced mentors track. It is tracked as a 7th signal dimension in Nascens. Mentors see it in the formation view. Founders see only guidance. This asymmetry is architectural — when founders can see their own coachability assessment, they optimize for the signal rather than the substance, and the signal degrades. The design prevents that.

Confidence labeling

Every field in the normalized vector carries a confidence tier. Downstream operations — guidance, trajectory estimates, population matching — inherit and display this label. Guidance is never presented without its epistemic provenance.

Novel
LLM-derived from this venture's text alone. No population analog confirmed. Appropriate for early-stage ventures and first intake events.
Requires: intake text only
Comparable
Matched to population analogs. Guidance draws on observed patterns from similar ventures in the database.
Requires: ~100+ venture interactions in population
Established
Empirically grounded. Signal patterns and transition outcomes are statistically supported by population data.
Requires: ~500+ venture interactions, longitudinal data

The database is the product.

Every normalized intake event — anonymized, stripped of identifying information — is committed to a shared population database. The database compounds: each new venture interaction makes every subsequent guidance event more precise. This is the core thesis behind the Nascens business model. The program is not just consuming a software tool; it is contributing to a compounding intelligence asset.

Population-grounded guidance

When a venture's signal profile matches prior population analogs, guidance draws on observed outcomes — not generic frameworks. The mentor knows what actually resolved similar assumption gaps in similar ventures.

Compounds with every interaction

Intake events, milestone reviews, pivot declarations, and terminal outcomes all feed the database. The signal patterns that indicate stalling, the assumptions that correlate with successful transitions — these emerge from the population over time.

Cross-program intelligence

Participating programs contribute to and draw from the shared population. A university cohort's venture outcomes inform accelerator guidance. The population is cross-institutional and anonymized by design.

Longitudinal by default

The episodic model means the database captures ventures across their full formation arc — from first pitch to progress, pivot, or exit. Longitudinal patterns are first-class data, not reconstructions.

Event-triggered, not continuous.

Nascens does not poll. It does not require ventures to maintain dashboards, check in weekly, or generate continuous telemetry. Ventures — especially at the formation stage — do not produce continuous signals. They produce episodes: a pitch, a customer interview, a pivot conversation, a milestone achieved or missed.

The episodic model matches how programs actually operate. A mentor session happens. An intake form is submitted. A cohort review occurs. Each of these is a discrete event that triggers the AVS pipeline. Between events, venture state persists unchanged. The next event picks up from that persisted state, not from scratch.

This design choice has a practical consequence: the intelligence is available at the moment it is needed — when the mentor is preparing for a session, when the program director is reviewing a cohort — without requiring ongoing administrative burden from founders who are already resource-constrained.

Opportunity states

Each venture is at one of six opportunity states at any given time. State transitions are event-driven and explicit — not inferred from stale data.

Forming Validating Progressing Stalling Pivoting Terminal

State transitions require threshold evidence in the signal vector. A venture cannot self-report into "Progressing." The evidence dimension must support it.

Weeks between events — by design

A venture may go 6–8 weeks between sessions. The persisted state model means the next mentor touchpoint picks up full context: what changed, what the outstanding assumptions were, what the trajectory estimate indicated at last update.

Anonymization is first-class.

The population database contains venture signal patterns, not venture identities. No cross-venture comparison surfaces individual data to another venture or another program. Anonymization is architectural — it is not a policy applied to a database that also holds identifying information. Identifying information and signal records are maintained in separate, non-linked stores.

Signal records are anonymous by construction

When a normalized signal vector is committed to the population database, it carries no venture name, founder name, institution, or program identifiers. The association exists only in the program's own isolated record store.

No cross-venture leakage

Population matching surfaces pattern counts and aggregate outcomes — "ventures with this assumption profile typically resolved it via X" — not individual venture data. No participant can reconstruct another venture's record from guidance output.

Program isolation

Each participating program's ventures are visible only to that program's authorized users. Population intelligence is aggregate and anonymized. A program's individual venture data never leaves its isolated partition.

The architecture is available for technical review.

If you run a university entrepreneurship program, accelerator, or mentor network and want to understand the full technical specification, reach out directly.

Request Access About PCI Technology