AI Profiler
Give any AI application a signature — and read why it classifies the way it does.
Created by Dr. Sharad Maheshwari, imagingsimplified@gmail.com
Institute for Responsible Healthcare AI (IRHAI) · built under the DSEA Runtime Contract v8.1
Layers 1–2 — the signature
Nine axes describe what the system is; three describe what it is permitted to do. Changing any axis re-runs the classification and rewrites the Reasoned Profile.
Layer 1 — Architecture
A1Core Categories
Context Modifiers
Layer 2 — Governability Context
G1Two systems can share an identical architecture and sit in different risk classes entirely because of these three axes.
System Signature
—
Lossless: the string alone reconstructs all twelve axes. Input hash
Profile-derived Operational Classification
Derived from Layer 1 × Layer 2. Nothing here is entered by hand.
These are profile-derived operational classifications, not governance scores. Operational Risk Tier is a property of the signature; Governance Maturity is a property of the organisation running the system and is measured separately under RATSe. A Low tier can be entirely ungoverned. See the for the formal statement of outputs.
Operational Risk Tier
—
Architecture Class
—
RD = 3 → generative
Admission Review
—
P.R.I.M.E. gate
Runtime obligations implied
Controls this signature requires at inference time rather than at review time.
Nine Layer 1 axes, 0–3, by axis code. Deterministic SVG — no charting library, no clock.
Governability context
Three Layer 2 axes, 0–3. Bars, not a radar — three points cannot make a meaningful shape.
Fact table▼
| FactID | Axis | Code | Level |
|---|
Reasoned Profile
Written by a fixed template registry from the rule trace — no model, no probability, no free-form text. The same signature always produces byte-identical wording.
Reading . Each block names the rule that produced it.
What this signature establishes
fully derivedWhat it does not establish
scope limitProof path▼
Every conclusion, the rule that produced it, the facts it read, and the L0 constants it consulted.
| Rule | Inputs | State | L0 refs |
|---|
Tier rule trace▼
Evaluated top to bottom; the first to fire sets the tier. Everything below it is superseded, not false.
Sensitivity — what would change the classification▼
Would lower the tier
Would raise the tier
Minimum pairs
Knowledge Bank
The foundational white paper sits above the technical specification. It does not begin with RG/RD/HR — it begins with the problem the Profiler exists to solve. Bundled in full, offline, with no external retrieval.
The AI Profiler outputs
- Architecture Identity — A1
- Governability Context — G1
- Combined System Signature
- Profile-derived Operational Classification
- Profile-derived Runtime / Admission Flags
- Machine-readable Profile (JSON)
- Human-readable Profile (Markdown)
It does not output
- Governance maturity
- Trustworthiness certification
- Regulatory compliance
- Residual risk
- Deployment approval
Operational Risk Tier is a property of the signature — a function of twelve integers, needing no evidence. Governance Maturity is a property of the organisation running the system, and cannot be derived from a signature. A Low tier can be entirely ungoverned; a Critical tier can be governed well. Neither is a contradiction.
Contents▼
AI Profiler
A Deterministic Architecture and Governability Signature for Artificial Intelligence Systems
An introductory white paper from BeResponsibleAI
Version 1.0 — August 2026
Abstract
Artificial intelligence is no longer a single technological category.
A rule-based expert system, a conventional machine-learning classifier, a foundation model, a multimodal generative system, an autonomous vehicle controller, and a multi-agent AI may all be described simply as "AI" — yet their architectures, capabilities, dependencies, autonomy, and consequences can be fundamentally different.
This creates a foundational problem for responsible AI:
How can we govern an AI system appropriately if we have not first established what kind of AI system it is?
Current AI discussions frequently begin with model names, benchmark performance, vendor claims, or broad labels such as "generative AI" or "agentic AI." These labels are insufficient for systematic comparison and governance. They do not provide a stable technical description of the system, nor do they capture the operational authority granted to it.
The AI Profiler is proposed as a deterministic method for creating a machine-readable and human-interpretable identity for an AI system. It uses two complementary layers:
- A 9-point Architecture Signature (A1) describing how the system is technically constructed.
- A 3-point Governability Context (G1) describing its operational authority, autonomy, and consequence context.
Together these produce a 12-axis AI System Profile and a compact, versioned signature that can be compared, stored, exported, and supplied to downstream assessment and governance systems.
The AI Profiler does not certify an AI system, measure governance maturity, or determine whether a system is trustworthy. Its purpose is more fundamental:
Profile first. Assess second. Govern third.
This separation allows governance requirements to be derived from a stable description of the system rather than from vague categories or model branding.
Supervisor View
Constants provenance, test vectors, and the append-only session log. Nothing here influences the engine.
Constants requiring verification
Constants register (L0)
| ID | Value | Unit | Source | Verified |
|---|
Test vectors
Session log (L9, append-only, in memory)
Sequence, timestamp, input hash, logic version and tier only. L9 is the only place that reads the clock; nothing is stored or transmitted.
| Seq | Timestamp | Input hash | Logic | Tier |
|---|
The framework
Twelve axes in two layers. Everything else is derived.
Reading a signature
A signature looks like AI · A1:RG1RD3HR2RA2AL2DD3LM2KS3DE3 · G1:OA3AU2CC3. Each two-letter code carries its level, 0–3. A1 is the architecture layer, G1 the governability layer; the prefixes are format versions, so the scheme can extend without invalidating recorded signatures. The encoding is lossless, which is what makes it usable as a citation, a registry key, or a spreadsheet column.
Why the reasoning is templated, not generated
DSEA §0.6 forbids generative text in the authoritative path; L8 is defined as a fixed registry mapping state to text. That is also simply correct here: a model asked to narrate a deterministic classification can contradict it, and the one property this tool sells is that it cannot. Every sentence is selected by state rather than composed, so identical inputs yield identical output and the wording is reviewable in advance.
Risk is not a decision
Per DSEA §0.10 nothing here is phrased as an instruction. A Critical tier is a classification, not a prohibition; an implied runtime obligation is an observation about the signature, not an order. What follows is the reviewer's judgment, and the proof path exists so that judgment can be argued with.
Layer 1 — Core Categories
Layer 1 — Context Modifiers
Layer 2 — Governability Context
How the classification is derived
Operational Risk Tier — first match wins
- Critical — CC ≥ 3 or OA ≥ 3
- High — CC ≥ 2 or OA ≥ 2 or AU ≥ 2
- Moderate — generative, or AU ≥ 1, or AL ≥ 2, or (CC ≥ 1 and RD ≥ 2)
- Low — default
Low / Moderate / High / Critical is exactly the ordinal tier vocabulary the DSEA contract specifies at L5. The top two tiers are reachable from Layer 2 alone: consequence and authority dominate architecture by design. The Drug-Interaction Checker preset makes the point — no machine learning anywhere in it, and still High.
P.R.I.M.E. admission triggers — any one suffices
- CC = 3 and OA = 3
- Generative and OA = 3
- Tier is Critical
Inert axes
Rule-Guided (RG), Hybrid Reasoning (HR) and Learning Mode (LM) appear in the signature but touch no derivation rule — no tier, no trigger, no runtime obligation. Verified exhaustively over all 262,144 combinations of the nine axes that do matter. They are descriptive only, which is worth knowing when reading a verdict and arguably worth revisiting in the framework.
Scope
This tool assigns a signature and derives its classification. It does not assess whether the organisation running the system has ownership, audit trails, contestability, drift monitoring or equity telemetry actually in place — that is the RATSe assessment, a separate exercise needing documentary evidence rather than dropdowns. A Low tier means low consequence, not good governance.
Provenance and limits
Runs entirely in the browser: no network code, no external resources, nothing written to storage, nothing transmitted. Use Data Download to retain a session or Copy link to retain a signature. Output is a descriptive classification, not a certification, regulatory determination, or legal advice.
Comments
Post a Comment