A modular architecture for AI-powered regulated investigation. 8 universal base components · 7 plug-and-play clusters · 6 deployment profiles. From interview-only setups to full financial crime platforms. Applicable to any regulated investigation in any industry or jurisdiction.
Version 1.2 · March 2026 · © 2026 Aditya Kaushal
Institutional Reasoning Engines (IRE) are a new class of AI system purpose-built for regulated investigation. Unlike AI assistants that answer queries or automation tools that execute predefined tasks, an IRE:
- Maintains a working hypothesis that evolves as evidence accumulates
- Iterates toward a defensible conclusion rather than answering a single query
- Gates every material output through human judgment before it enters the evidentiary record
- Produces a tamper-evident chain of reasoning from raw document to final finding
- Operates within a defined methodology — not in service of a prompt
This is not a chatbot. It is not automation. It is a professional investigation workflow encoded in software — with AI accelerating the analytical steps and humans accountable for every conclusion.
| File | Description |
|---|---|
releases/v1.2/IRE_v1.2_Final.pdf |
Whitepaper v1.2 — 59 pages, full architecture, clusters, deployment profiles, adversarial critic, investigator accountability schema, worked case walkthrough |
releases/v1.2/IRE_Builder_Guide_v1.2.html |
Interactive build guide — tools, workflows, static flowcharts, investigator UX screens, AI coding guidance |
docs/ire-builder-guide.html |
Hosted version of the builder guide (GitHub Pages) |
releases/v1.1/ |
Version 1.1 archive |
The IRE is structured in two tiers:
Every regulated investigation system must implement all eight. No exceptions.
| # | Component | What It Enforces |
|---|---|---|
| B1 | Case-Scoped Isolation | Cross-case contamination prevented by architecture, not policy |
| B2 | Retrieval Verifier | Every AI claim checked against case corpus — deterministic, not AI judgment |
| B3 | Evidence Grounder | Source citation attached to every verified claim |
| B4 | Human Review Gate | No AI output enters the evidentiary record without explicit investigator approval |
| B5 | Audit Event Logger | Every system action recorded |
| B6 | Hash-Chain Immutability | Any tampering anywhere in the chain is mathematically detectable |
| B7 | Model Version Pinning | No silent upgrades mid-investigation |
| B8 | Algorithmic Bias Monitoring | Active from the first case — not contingent on fine-tuning |
| Cluster | Name | Activates When |
|---|---|---|
| A | Documentary Evidence | Case corpus contains any documents |
| B | Behavioural & Interview | Case involves transcripts or witness statements |
| C | Entity & Network Intelligence | Multiple entities whose relationships are material |
| D | Advanced Reasoning | Case complexity requires iterative hypothesis formation |
| E | Privacy Gateway | Any external AI API is used for reasoning |
| F | Institutional Memory | Sufficient historical case volume for fine-tuning |
| G | Integrity Elevation | Outputs may be used in legal or regulatory proceedings |
| Profile | Components | Use Case |
|---|---|---|
| 1 — Minimal Compliant | 8 | Simple internal investigations |
| 2 — Interview-Led | 11 | HR, whistleblower, testimony-primary cases |
| 3 — Document Investigation | 20 | Legal due diligence, regulatory exam prep |
| 4 — Corporate Fraud & FCPA/ABAC | 28 | Bribery, beneficial ownership, multi-party fraud |
| 5 — Financial Crime Full | 31 | AML, STR generation, transaction monitoring |
| 6 — Enterprise Forensic | 35 | Enforcement agency, Big 4 forensic, court-bound outputs |
The interactive builder guide is available at docs/ire-builder-guide.html and hosted on GitHub Pages. It covers:
- Prerequisites — what to document before you write a line of code
- Tools & Subscriptions — every tool needed, free and paid
- Build Phases — 6-phase sequence, each delivering standalone value
- Workflows — static flowchart diagrams for every major process
- End-to-end case lifecycle
- Document ingestion pipeline
- Entity resolution (3-tier HITL)
- Recursive reasoning loop
- Privacy gateway / pseudonymisation
- Audit chain / hash-chain design
- Investigator UX — five critical decision screens documented in full
- Per-Cluster Build Guide — tools and exit criteria for Clusters A through G
- Building With AI Tools — which AI coding tools to use for which tasks, with sample prompts
- FAQ — common questions including Indian regulatory context (RBI, PMLA, SEBI)
- Adversarial Critic (Section 7.4) — active hypothesis invalidation; blocks confirmation bias by design
- Investigator Accountability Schema (Section 12.3) — 8-field approval record with rubber-stamp detection
- Investigator UX — five critical decision screens (builder guide)
- Data Messiness section — seven production blockers with detection and fixes (builder guide)
See CHANGELOG.md for the full diff.
Methodology shapes technology. The investigation framework determines what the system is built to do. The technology is in service of the methodology — never the other way around.
Human gates are architectural, not optional. The Human Review Gate (B4) is a mandatory base component. There is no finding without an approval event in the audit chain. This is enforced by architecture, not policy.
Tamper-evident by default; independently verifiable with Cluster G. The mandatory base produces institutional-grade audit integrity. Cluster G elevates to independently verifiable integrity — provable by a third party without trusting your systems.
Modularity without fragmentation. No cluster replaces a base component. Clusters add capability. The base establishes defensibility.
- Investigation leaders building AI-assisted investigation capability inside regulated institutions
- Forensic practitioners (Big 4, boutique forensic firms) evaluating AI architecture for investigation workflows
- Compliance and legal technology founders building products for regulated investigation markets
- Technical architects designing AI systems that must produce defensible, auditable outputs
- Researchers in AI governance, legal AI, and regulated AI systems
If you reference this work in research, publications, or product design:
Kaushal, A. (2026). Institutional Reasoning Engines: A Modular Architecture for
AI-Powered Regulated Investigation (Version 1.2).
https://github.com/ayuvinc/Institutional-Reasoning-Engine
This work is licensed under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0).
You are free to share and adapt this material for non-commercial purposes, provided appropriate credit is given and changes are indicated.
Commercial use, productisation, or deployment requires explicit written permission from the author.
Full license text: LICENSE
Aditya Kaushal Investigations Leader | AI-Powered Investigation Systems | CFE In collaboration with Claude (Anthropic)
- CHANGELOG.md — full version history
- ROADMAP.md — planned future work
- CONTRIBUTING.md — contribution guide
- CODE_OF_CONDUCT.md
- SECURITY.md
- CITATION.cff — citation metadata
- Hosted builder guide: GitHub Pages
- Download whitepaper (PDF):
releases/v1.2/IRE_v1.2_Final.pdf - Deployment checklists:
checklists/profiles/ - Architecture diagrams:
architecture/ - Latest release:
releases/v1.2/
© 2026 Aditya Kaushal. All Rights Reserved.