AiNOS
§R01 · RESEARCH

Research program · AiNOS

Research for accountable AI.

We publish the definitions, architectures, and evidence behind AiNOS — starting with what it means for a single high-stakes AI decision to be auditable.

Papers02FieldExplainable AIVenueACM FAccTPartnerU of T
FIG. R01 — PAPER SPECIMEN: A CONSEQUENTIAL AI DECISION IS EXPLAINABLE WHEN ITS SOURCE-TO-DECISION RECORD IS RECONSTRUCTABLE, CONTESTABLE, GOVERNABLE, AND BOUNDED · REQUIRED PROPERTY
§R02 · PAPERS

Published work.

Conceptual definitions, architectures, and empirical evidence from the AiNOS research program. Each paper is open to read in full.

Paper 01 · Conceptual definition

Preprint · 2026

Auditable Decision Intelligence: A Decision-Centered Definition of Explainable AI for High-Stakes Decisions

A consequential AI decision is explainable when a qualified reviewer can externally reconstruct and contest its source-to-decision record — reconstructable, contestable, governable, and bounded.

In high-stakes AI, explainability is usually treated as the presence of an explanation artifact. We redefine it as an institutional property of a single consequential decision: the ability to reconstruct, contest, govern, and bound its source-to-decision record. The paper specifies the Decision Audit Contract (DAC) — the minimal record that makes one decision auditable — and shows the definition is a genuine contribution, not a relabeling of existing transparency work.

Reconstructable

The decision can be replayed from its recorded source-to-decision path.

Contestable

A reviewer can pinpoint the exact fact, rule, or judgment to challenge.

Governable

The record shows where the system may reason freely and where approval is required.

Bounded

Claims stay inside covered sources, valid context, and stated limits.

Falsifiable test

A system fully compliant with Kroll's accountable algorithms, Cobbe's reviewable automated decision-making, the EU AI Act (Art. 12 & 14), and NIST IR 8312 can still fail ADI — because none guarantees a per-decision source → claim → boundary link.

FieldExplainable AI / FAccTVenuePrepared for ACM FAccTTypeConceptual definitionLength16 pages
Read the paper (PDF) →See it in the product →

Paper 02 · Empirical study

Preprint · 2026

Semantic Fidelity Limits of Language-Model Ontology Compilation

An ontology can be internally clean, typed, and evidence-linked while material source meaning has already been lost. Structural integrity and semantic recoverability are distinct — so the immutable source stays authoritative, and the ontology becomes a revisable semantic control plane.

Language models are increasingly used to compile unstructured text into typed knowledge graphs — and the tempting next step is to let the compiled ontology replace its source as authoritative truth. We tested that claim against a repository-verified legal-domain compiler (9 statutes, 1,364 immutable source units, 3,090 objects, 5,803 assertions, 7,567 relations) and a targeted forensic audit of five failure-prone windows. Every audited relation carried explicit evidence, yet source comparison still found material losses in actors, conditions, list membership, and amendment scope — the failure boundary was the first semantic projection, not graph construction. We propose a source-first architecture: immutable sources retain authority, and generated ontologies serve as revisable semantic control planes for retrieval, navigation, and candidate generation.

Semantic completeness

Every material actor, condition, exception, and membership is recoverable from the ontology.

Canonical stability

Repeated compilations of the same source are semantically equivalent, even when wording differs.

Relational correctness

Predicates, directions, endpoints, and qualifiers preserve the source's meaning.

Authority substitution

A downstream reasoner can rely on the ontology without consulting the source.

Falsification result

A compiler with near-perfect evidence linkage — 578 of 593 audited relations carried exactly one supporting assertion — still lost actors, conditions, and amendment scope: a 13,455-character amendment unit collapsed into 33 broad claims, and a product-classification provision lost its membership matrix. A graph validator cannot detect a proposition that was never generated.

FieldOntology learning / KGVenuePrepared for arXivTypeEmpirical studyLength9 pages
Read the paper (PDF) →See it in the product →

More work is in progress across the decision-intelligence program — context benchmarks, the ontology-driven architecture, and the source-first semantic control plane. Talk to the team →