🤖 AI

Auditable Decision Models with Learned Abstention and Real-Time Steering

This paper introduces EvaluatorDPT, a bounded decision-control model that learns to abstain (TBD) rather than forcing predictions, enabling auditable, policy-governed AI decisions with explicit uncertainty routing and strong empirical performance on a large test set.

Original authors: Sankaranarayanan Palamadai Chandrasekaran

Published 2026-05-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sankaranarayanan Palamadai Chandrasekaran

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the manager of a busy factory. Every day, workers (the AI) bring you a stack of requests to approve or reject. In the past, you forced every worker to stamp every single request with either a big green "GO" or a big red "STOP," even if they were confused, the paperwork was missing, or the instructions were vague. This led to mistakes: sometimes you approved a dangerous machine because the worker guessed "GO," and sometimes you blocked a perfectly good idea because the worker guessed "STOP."

This paper introduces a new kind of worker and a new rulebook for your factory.

The New Worker: The "Three-Stamp" AI

Instead of forcing a "GO" or "STOP" on everything, this new AI is trained to use three stamps:

YES (Go): "I have enough proof to approve this."
NO (Stop): "I have enough proof to reject this."
TBD (Hold for Review): "I'm not sure. The evidence is weak, conflicting, or missing. I need a human to look at this."

The big innovation here is that the AI learns when to say "TBD." It doesn't just guess "YES" or "NO" and then hope a computer later says, "Wait, that guess was shaky." Instead, the AI is taught from the start that "I don't know" is a valid, important answer. It treats uncertainty as a specific category, just like "Yes" or "No."

The Factory Floor: How It Works

Think of the AI as a highly trained traffic controller at a busy intersection.

The Input: Cars (data) arrive with their destination signs.
The Decision: The controller looks at the sign.
- If the sign is clear and the road is open, they wave it through (YES).
- If the sign is clear but the road is blocked, they wave it back (NO).
- If the sign is blurry, the road is foggy, or the driver is hesitating, the controller doesn't guess. They raise a flag and say, "Hold there, let a supervisor check this" (TBD).

The paper shows that this "TBD" stamp is learned during the AI's training, not just added as a safety net afterward. This means the AI gets better at spotting the "foggy" situations naturally.

The "Steering Wheel" and the "Dashboard"

The paper also explains how humans can control this AI without rewriting its brain.

The Dashboard (Auditability): The AI doesn't just give an answer; it gives a full report card. It shows how confident it is, how it got confused in the past, and exactly how it was tested. This is like having a black box flight recorder for the AI, so if something goes wrong, you can see exactly what happened.
The Steering Wheel (Real-Time Control): Imagine you are the factory manager. If it's a rainy day (high risk), you can turn a dial to make the AI more cautious. You can tell it, "If you are even 10% unsure, send it to the 'TBD' pile." If it's a sunny day (low risk), you can turn the dial to be more aggressive. You can change these rules instantly without retraining the AI. The AI provides the options; you provide the policy.

The Results: How Good Is It?

The researchers tested this system on a massive set of practice problems (about 44,600 examples).

The Score: The AI got about 82.6% correct overall.
The Breakdown:
- It was very good at saying "No" when it should (84.9% accuracy).
- It was good at saying "Yes" (83.1% accuracy).
- It was decent at saying "I need help" (79.6% accuracy).
The Comparison: If they forced the AI to only say "Yes" or "No" (removing the "TBD" option), its performance would crash to 49.5%. This proves that having the "I don't know" option is crucial for making smart decisions.

What This Paper Does Not Claim

It is important to stick to what the paper actually says:

It does not claim this AI can fix every problem in the world or replace human judgment entirely.
It does not claim the AI is perfect at understanding emotions (the "emotion" part of the system was turned off for this test).
It does not claim this works immediately for every specific industry (like hospitals or banks) without extra testing.
It does not promise that the AI will never make a mistake.

The Bottom Line

This paper presents a tool for making AI decisions safer and more transparent. Instead of forcing an AI to guess when it's unsure, it gives the AI permission to say, "I need a human to check this." It also gives the humans a clear dashboard to see how the AI is thinking and a steering wheel to adjust how cautious the AI should be, all while keeping a detailed record of every decision for review. It's about moving from "blind guessing" to "controlled, auditable decision-making."

Technical Summary: Auditable Decision Models with Learned Abstention and Real-Time Steering

Problem Formulation

Production AI systems frequently encounter inputs characterized by incomplete, conflicting, or insufficient evidence. Traditional approaches often force these ambiguous cases into binary action labels (e.g., YES/NO), thereby concealing uncertainty, or rely on generative outputs that lack the structure required for auditable execution decisions. The central problem addressed is the need for an operational decision control interface where uncertainty is explicitly routable, policy-governed, and auditable, rather than hidden within forced predictions or free-form generation.

The paper defines a bounded decision space $Y_d = \{YES, NO, TBD\}$ .

YES: Sufficient support for an action.
NO: Sufficient support for rejection or blocking.
TBD: Insufficient certainty for direct action or rejection, representing a learned deferral outcome.

Unlike standard reject-option models where abstention is a post-hoc threshold applied to a binary classifier, this work treats deferral (TBD) as a supervised class within the label space. This allows the model to internalize the semantic structure of ambiguity during training rather than treating low-confidence regions as a uniform rejection zone.

Methodology

Architecture

The proposed model, EvaluatorDPT, utilizes a transformer encoder (BERT-base-uncased) with a maximum sequence length of 128 tokens. The architecture employs a multi-head design:

Primary Decision Head: Predicts the bounded distribution over $\{YES, NO, TBD\}$ .
Auxiliary Heads: Structured channels for values (10 labels) and emotions/sentiments (28 labels).
- Note: For the specific evaluation reported in this paper, the emotion head is disabled. Empirical claims are restricted to the decision head, while auxiliary channels are presented as an architectural capability for future policy control.

Training and Data

Dataset: A curated corpus derived from Natural Language Inference (NLI) and stance-style materials (SNLI, MultiNLI, SemEval). Entailment maps to YES, contradiction to NO, and neutral/insufficient evidence to TBD.
Boundary Refinement: The model undergoes a targeted refinement epoch using a small subset (323 examples) mined from prior diagnostic errors. This subset focuses on hard semantic regions, including negation, contradiction, lexical-overlap traps, role/number swaps, and hedged/modal statements.
Loss Function: Weighted loss (1.0 for decision, 0.3 for value, 0.3 for emotion) with label smoothing (0.05).
Configuration: Cosine scheduling with warmup, unfrozen encoder layers 9–11, and a global learning rate of $5.0 \times 10^{-6}$ .

Operational Interface

The system separates the model's learned decision distribution from the deployment policy.

Inference-Time Control: Operators can adjust routing behavior by changing recorded operating thresholds (e.g., increasing the deferral rate for high-risk workflows) without retraining the model weights.
Auditability: The interface outputs a bounded distribution, confidence scores, and structured auxiliary signals. The evaluation package includes calibration data, threshold sweeps, confusion matrices, and reproducibility commands to support external review.

Key Contributions

Learned Deferral Formulation: Modeling decision routing as a bounded three-way output where TBD is a learned supervised outcome, allowing ambiguous examples to shape the representation during training.
Separation of Decision and Policy: The model emits a distribution, while deployment thresholds remain external and inspectable. This separates the representation of uncertainty from the choice of risk tolerance.
Auditable Inference-Time Control: A framework where routing policies (thresholds, fallback rules, auxiliary signal usage) are recorded as operating policy rather than embedded in model weights, enabling real-time steering without retraining.
Domain-Agnostic Interface: The output space is independent of specific downstream domains, allowing the same bounded labels to be applied to moderation, approval workflows, or compliance reviews with domain-specific calibration.
Audit-Oriented Evaluation: A comprehensive evaluation structure that includes classwise behavior, calibration evidence (ECE), threshold sweeps, multi-seed stability checks, and reproducibility materials, moving beyond simple aggregate accuracy.

Results

Performance Metrics

On the held-out test split ( $n=44,597$ ), the model achieves:

Accuracy: 0.8260
Macro F1: 0.8252
Per-Class F1:
- YES: 0.8314
- NO: 0.8486
- TBD: 0.7956

Comparative Analysis

Learned vs. Forced: Collapsing the model to a binary YES/NO decision (removing TBD) causes Macro F1 to drop significantly to 0.4945, demonstrating the necessity of the learned deferral class.
Thresholding: Applying reject thresholds to a binary classifier at 90% coverage yields higher Macro F1 (~0.863) but excludes 10% of cases. The learned TBD model maintains high performance across the full coverage set.
Calibration: The model demonstrates good calibration with an Expected Calibration Error (ECE) of 0.0338 on the validation set.
Refinement Impact: The boundary-refinement pass did not materially increase raw Macro F1 (parent: 0.8254 vs. refined: 0.8252) but reduced calibration error and high-confidence error rates (e.g., high-confidence error at threshold 0.85 dropped from 0.0558 to 0.0485).

Error Analysis

The primary residual error involves confusion between true TBD cases and direct action labels (YES/NO), particularly TBD cases misclassified as YES. This highlights the semantic difficulty of distinguishing ambiguous inputs from those with decisive evidence, a region targeted by the boundary-refinement subset.

Significance and Claims

The paper claims that learned deferral can be effectively evaluated as an auditable decision-control interface. The significance lies not in achieving maximal automation or a new language-model backbone, but in providing a controlled routing mechanism under ambiguity.

Key assertions include:

Explicit Uncertainty: Uncertainty should be an explicit, routable outcome (TBD) rather than a hidden property of a confidence score.
Policy Separability: The model distribution and deployment threshold policy can be examined and modified independently, supporting governance and risk management.
Reviewability: The evaluation evidence (calibration, stability, confusion matrices) allows external reviewers to inspect the system's behavior and reconstruct the evaluation process, a standard the authors argue is necessary for operational AI systems.

The authors remain modest regarding generalization, noting that the results are specific to the evaluated corpus and that domain transfer requires separate validation, calibration, and threshold selection. The auxiliary value and emotion channels are defined as architectural capabilities but are not empirically validated in this specific release.