Skip to content

CaptainMig/nerva-v10

Repository files navigation

NERVA v10 · Decision Integrity Platform

A runtime decision-gating layer for AI agentic systems and human operators.

NERVA sits between reasoning and high-stakes action. Returns one of five structured verdicts — COMMIT · HOLD · WAIT · CONSULT · TOXIC — against five published axes, with full provenance. Every gated decision is written to an auditable ledger.

This repo is the live cockpit platform. The current primary surface is NERVA v11 · Mixed-State Decision Integrity Cockpit — the instrument that exposes the full audit cluster and serves as the operational deployment for high-stakes governance.

→ Live platform: nerva-v10.vercel.app → Primary instrument: nerva-v10.vercel.app/v11.html


What it is

The NERVA platform is the operator-facing surface for the decision-gating kernel. Built for two distinct buyers:

  • AI infrastructure teams deploying agentic systems where decisions need to be gated before action — the cockpit sits one layer above operational observability and writes a separate decision-integrity trace
  • Human operators running high-stakes decisions (capital allocation, vendor selection, hiring, strategy commitments) who want a structured brake with provenance

The cockpit is the visualization. The kernel is the math. The ledger is the memory. The discipline is the product.


NERVA v11 · The primary cockpit

v11 is the production instrument. The earlier view modes (A through D) shipped during the development of the kernel's visual vocabulary; v11 is what the platform was building toward and is the surface that should be evaluated when assessing NERVA as a governance instrument.

What v11 adds

v11 is the mixed-state cockpit — the full implementation of the Bloch-vector representation, the von Neumann entropy calculation, and the adaptive threshold logic, surfaced through an operator-facing instrument cluster.

The cluster exposes five sub-panels in Audit mode:

Panel What it surfaces
<QLT> Quality Metrics Purity, entropy, margin, brake gap, reversibility, homogeneity — six kernel diagnostics, all computable from the values the kernel already produces
<MAP> Phase Map The (C, clarity) coordinate plotted against the four non-TOXIC verdict regions — instant visual read of where the decision sits in state space
<FAC> Factors Signed bars showing cᵢ − 0.5 for each axis — makes the structural shape of the verdict legible at a glance
<SEN> Sensitivity Perturbation analysis: how much a +0.1 change to each axis moves aggregate confidence, and whether the verdict flips — names the binding constraint and the operational lever
<LOG> Agent Log Execution trace of the kernel's eight-step pipeline — parse, extract, brake check, Bloch projection, compute, verdict, audit, route

What v11 supports

  • Summary / Audit mode toggle — daily-use view collapses to the verdict, metrics, and reasoning; audit view exposes the full instrument cluster for high-stakes deliberation
  • The One-Way Brake — high-stakes irreversible decisions with insufficient margin are routed to WAIT regardless of other conditions, with the brake state surfaced on the masthead
  • TOXIC refusal — five published refuse categories (non-consenting third-party harm, undisclosed MNPI, decisions made for others who should be making them themselves, active-crisis signals, deceptive theses) — NERVA returns no verdict for these
  • Receipt provenance — every scored decision produces a signed receipt with a unique hash for ledger ingestion

Why v11 over the other modes

The four earlier modes share the same kernel but expose different visual surfaces. They remain available on the platform for design reference and for users who prefer the stripped or alternative readouts. v11 is the surface the kernel was designed to fully express, and is the surface a pilot deployment would run.


The full platform surface

Mode Status What it shows
v11 · Mixed-State Cockpit Production Full instrument cluster, Summary/Audit toggle, primary deployment surface
A · Cockpit Reference Original verdict card layout — masthead, verdict word, metrics, Bloch projection, reasoning, provenance
B · Terminal Reference Monospace operator view for keyboard-driven decision intake
C · Phase Space Reference Phase-map visualization isolated from full card chrome
D · Brutalist Reference Stripped instrument-readout for high-density display environments

All modes run the same NERVA kernel. The math does not change between views.


Applications running on the platform

  • Plumb — three-layer buyside audit for media investment. Scores Optimization, Authenticity, and Pricing independently. Designed to catch the ad buy that looks great on distribution metrics and fails on cultural authenticity. Runs the NERVA kernel against media-domain-specific axis weights.

  • Data Field — operator data view for live decision feeds. Demonstrates how the platform handles streaming decision intake at scale.


The kernel

The NERVA kernel maps decisions onto a Bloch vector representation and computes confidence as a mixed-state purity measure. Five axes with default weights:

c_evidence        × 0.30
c_differentiation × 0.25
c_falsifiability  × 0.20
c_timing          × 0.15
c_alignment       × 0.10

Aggregate confidence C = Σ wᵢcᵢ. Density matrix ρ = ½(I + r_eff · σ). Von Neumann entropy S(ρ) measures the spread between directional conviction and evidence structure. Adaptive threshold τ = τ₀ + α·S + β·σ_stakes + γ·irreversibility scales with stakes and reversibility.

A decision clears the threshold when C ≥ τ. Verdict state is then determined by entropy (COMMIT vs HOLD), brake conditions (the One-Way Brake on high-stakes irreversible commits), and TOXIC refusal categories.

The math is published in full. The reference implementation lives at CaptainMig/Nerva_Prompt_Universal and runs deterministically in Python for verification when the verdict matters.


Two deployment paths

Cockpit platform (this repo) — operator-facing visualization for human decision workflows. Static HTML, no server, browser-local state. Suitable for personal use, pilot deployments, and demonstration of the methodology.

Agentic runtime layer (in pilot scoping) — SDK and API for AI agentic systems where decisions need to be gated programmatically before action commits. NERVA sits between agent reasoning and the action layer, scores the agent's stated reasoning structure against the same five axes, and routes by verdict state. COMMIT decisions flow normally; HOLD, CONSULT, and TOXIC are routed to operator review or refused. Every gated decision is written to a decision-integrity audit ledger separate from operational telemetry.

The cockpit is what you can see today. v11 is the surface that demonstrates what the runtime layer enforces programmatically. The runtime layer is what enterprises buy.


What this is not

NERVA does not predict outcomes. NERVA evaluates the structural integrity of stated reasoning before commitment. The kernel scores claims, not credentials; structure, not narrative.

The lift score — the empirical hit-rate of decisions that followed NERVA versus decisions that overrode it — is computed only from resolved outcomes in the ledger. It is never extrapolated forward. The willingness to be falsified is the product.


Status

  • v11 cockpit: Production, deployed, primary surface
  • Reference modes A–D: Stable, available for design reference
  • Plumb application: Deployed, pilot proposal v1.0 published
  • Reference kernel (Python): Deterministic implementation available in companion repo
  • Agentic runtime layer: In pilot scoping, not yet generally available
  • Pilot inquiries: Welcome — see Starpoint LLC contact below

Built by

Starpoint LLC · Solo technical founder building a decision-integrity layer for AI agentic systems and human operators.

Related repos:


NERVA v10 platform · v11 primary instrument · Starpoint LLC · patent pending

About

NERVA — decision-integrity kernel for agentic AI. Five-axis scoring, adaptive threshold τ, One-Way Brake, auditable receipts. Live v10 cockpit + mixed-state v11.

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors