AI-Enhanced Institutional Knowledge
MFin Pedagogical Edition · July 2026

AI-Enhanced Institutional Knowledge

What happens when an institution does more than use AI—when it begins to build a memory with it? This repository develops a governed architecture in which protected knowledge can be discovered, structured, retrieved, challenged, extended, and selectively written back without surrendering institutional control. The papers develop the theory; 22 executable Colab notebooks turn that theory into a cumulative laboratory.

Alejandro Reynoso · Judge Business School, University of Cambridge

Read the main paper View repository
The institutional pipeline

The real challenge is not giving AI access to knowledge. It is deciding what happens next.

Once an AI system can read institutional knowledge, a deeper problem begins. Which evidence should it see? Which relationships should it infer? Which conclusions deserve to survive? And when its answers are written back, how do today’s decisions reshape what the system will know tomorrow? The central proposition of this project is that the strategic unit of organizational AI is therefore not the model alone, but the full institutional architecture around it: human judgment, protected sources, canonical knowledge objects, explicit navigation, governed model access, authorization, memory, audit, and reversal.

01Protected source vaults
02Canonical objects and graph
03Governed model context
04Candidate knowledge and approval
05Recursive write-back and selection
06Audit, benchmarking and reversal
The intellectual spine

Main and supporting papers

Main paper

AI-Enhanced Institutional Knowledge

This synthesis paper asks what it would mean for an institution to develop an AI-enhanced memory without losing control of its own knowledge. It follows information from protected source vaults into canonical objects and typed relationships; from there into governed model context, candidate knowledge, authorization, and recursive write-back; and finally into an evolving cognitive topology that can be navigated, benchmarked, challenged, and reversed. The central problem is not simply whether AI can generate a useful answer. It is whether an institution can make AI-generated knowledge cumulative without allowing plausibility to become authority by accident.

Open PDF View on GitHub
Governance

Governance Before Cognition

Begins with the first institutional distinction: generation is not authority. It develops the controls that separate candidate output from canonical memory—eligibility, approval, provenance, quarantine, supersession, auditability, and reversal.

Open PDFGitHub
Architecture

Knowledge Engineering

Explains how a document collection becomes institutional knowledge architecture: canonical objects, typed relations, controlled ontology, provenance, lifecycle state, temporal supersession, decision objects, and layered validation. Companion to NB08–NB12.

Open PDFGitHub
Navigation

Navigation-Conditioned Knowledge Evolution

Shows why retrieval is not a neutral lookup. Navigation policy determines which evidence enters context, which answers become plausible, what is written back, and—recursively—how the future topology of institutional memory evolves.

Open PDFGitHub
Institutional inquiry

Question-Induced Knowledge Architecture

Explores a deeper mechanism: institutional inquiry itself creates structure. Purposeful questions, scenarios, and shocks can induce conditional claims, relationships, decisions, and supersession structures that were not explicit in the original corpus. Shared companion: NB12.

Open PDFGitHub
Advanced selection

The Quantum Brain

Extends the architecture from retrieval to evidence-portfolio construction: path encoding, quantum walks, QUBO/QAOA context selection, and quantum-conditioned deliberation—always alongside strong classical baselines, because difference is not the same as advantage.

Open PDFGitHub
Executable curriculum

22 Colab Notebooks

The 22 Colab notebooks turn the architecture into an experiment you can run. The sequence begins with a deceptively simple question—when does model output become institutional knowledge?—and then builds the machinery required to answer it: governance, corpus control, canonical objects, graphs, navigation policies, recursive write-back, evidence portfolios, deliberation, benchmarking, and reversal. Each notebook adds one capability and makes visible what breaks when that capability is absent. Run them in numerical order to watch a static corpus become a governed, evolving institutional memory. “GitHub” opens the source; “Open in Colab” launches the corresponding live notebook.

Stage 1 · Govern knowledge before admitting it

NB01

Candidate Knowledge

The laboratory begins with the foundational rule: model generation is only a candidate; it is not yet institutional knowledge.

GitHubOpen in Colab
NB02

Governance Pipeline

Build the gate between generation and memory: eligibility tests, policy checks, quarantine, and reason-coded review queues.

GitHubOpen in Colab
NB03

Human Approval and Provenance

Show why technical validity is insufficient: institutional authorization, provenance, and accountable human approval remain distinct acts.

GitHubOpen in Colab

Stage 2 · Discover and control the corpus

NB04

Corpus Discovery

Before asking what the corpus means, establish what actually exists. Corpus discovery creates the controlled inventory on which every later claim depends.

GitHubOpen in Colab
NB05

Metadata Extraction

Turn files into identifiable evidence. Location alone is not identity, and every claim must retain a precise path back to its source.

GitHubOpen in Colab
NB06

Corpus Manifest

A file may exist without being authoritative, current, eligible, or permitted. The manifest makes those distinctions explicit before model access begins.

GitHubOpen in Colab
NB07

Integrity and Validation

Test the corpus before trusting it: hashes, duplicate detection, fingerprints, integrity checks, and explicit exception reporting.

GitHubOpen in Colab

Stage 3 · Build governed knowledge architecture

NB08

CKO Builder

Move beyond documents by converting evidence into canonical, addressable knowledge objects that can be governed, linked, retrieved, and revised.

GitHubOpen in Colab
NB09

Relationship Graph

Build the graph by adding typed, attributable relationships among claims, entities, assumptions, evidence, and decisions.

GitHubOpen in Colab
NB10

CKO Registry

Create a reproducible registry for object identity, provenance, versioning, status, and lifecycle across the evolving knowledge system.

GitHubOpen in Colab
NB11

CKO Validation and Statistics

Validate the architecture itself: detect broken edges, orphaned objects, invalid endpoints, inconsistent metadata, and implausible graph structure.

GitHubOpen in Colab
NB12

Question-Induced Knowledge Architecture

Demonstrate that inquiry is productive, not merely interrogative: questions, scenarios, and shocks can create conditional claims, relations, decisions, and supersession events.

GitHubOpen in Colab

Stage 4 · Make navigation explicit and govern evolution

NB13

Graph Navigation Policy Laboratory

Make the hidden retrieval decision explicit. Different navigation policies traverse the same graph differently and therefore construct different evidence contexts.

GitHubOpen in Colab
NB14

Counterfactual Context and Answer Comparison

Hold the knowledge graph constant and vary the navigation policy to observe how different contexts can produce different answers from the same institutional memory.

GitHubOpen in Colab
NB15

Recursive Write-Back and Topological Divergence

Close the loop: once outputs are written back, today’s navigation path can alter tomorrow’s graph, producing cumulative topological divergence.

GitHubOpen in Colab
NB16

Navigation Governance, Audit and Policy Selection

Govern the retrieval process itself by reconstructing not only which sources were cited, but why this evidence was selected and plausible alternatives were excluded.

GitHubOpen in Colab

Stage 5 · Construct and select evidence portfolios

NB17

Quantum Brain Architecture and Path Encoding

Reframe context construction: admissible paths through the graph become evidence objects that can be compared, budgeted, combined, and governed.

GitHubOpen in Colab
NB18

Quantum Walks Through a Knowledge Graph

Compare how classical and quantum-induced navigation distribute attention across an identical knowledge graph—testing difference before making any claim of advantage.

GitHubOpen in Colab
NB19

QAOA and QUBO Context Portfolios

Construct a finite context portfolio that balances relevance, coverage, diversity, and controlled dissent, then compare QUBO/QAOA selection with strong classical baselines.

GitHubOpen in Colab

Stage 6 · Deliberate, benchmark and reverse

NB20

Quantum-Conditioned Multi-Agent Deliberation

Move from one model response to governed deliberation: distinct agents receive explicit roles, evidence boundaries, and synthesis rules.

GitHubOpen in Colab
NB21

Governed Write-Back, Noise and Advantage Benchmarking

Stress-test the complete pipeline under noise and alternative selection methods, benchmark performance, and enforce the discipline that claims must stop where the evidence stops.

GitHubOpen in Colab
NB22

Poisoned-Edge Dependency Reversal

Test whether institutional memory can correct itself responsibly: trace the full downstream dependency cone of an approved-but-wrong relation and reverse what depended on it.

GitHubOpen in Colab
Where this leads

From AI as a tool to AI as part of the institution

The most consequential phase of artificial intelligence will not be defined by isolated prompts, occasional copilots, or ever larger models. It will emerge when AI becomes deeply integrated into the institutional setting—connected to governed knowledge, embedded in decision processes, capable of learning from experience, and accountable for how it retrieves, transforms, and returns information to organizational memory. At that point, AI is no longer simply producing answers; it is participating in the evolution of the institution itself. The potential impact is enormous: better decisions, faster learning, stronger continuity, more systematic use of accumulated expertise, and forms of institutional intelligence that can compound over time. The challenge, and the opportunity, is to build this future deliberately—so that greater cognitive power is matched by stronger governance, clearer provenance, and enduring human authority over what the institution knows and becomes.