# A Workflow for Full Traceability of AI Decisions Source: https://www.youtube.com/watch?v=GXKPN3NFuxM Recap page: https://rapidrecap.app/video/GXKPN3NFuxM Generated: 2025-11-18T23:04:16.556+00:00 --- ## Quick Overview The proposed workflow for full AI decision traceability, exemplified by the Fung AI model, relies on two phases: training and inference, using a cryptographically secured, tamper-proof, and auditable system called DEBOMB (Decision Bill of Material) to ensure every input, process step, and output is verifiably linked to the original, untampered components, thereby resolving accountability gaps and providing crucial evidence for regulatory compliance. **Key Points:** - The proposed workflow divides AI decision traceability into two phases: Training and Inference. - The core solution is DEBOMB (Decision Bill of Material), which creates a cryptographically secured and tamper-proof record of the entire AI process. - DEBOMB documents every component, including raw model output, intermediate activations, training data, hyperparameters, and final probability scores. - The system ensures that for high-stakes decisions, like medical diagnoses, the resulting artifact is verifiable against the original, untampered code and data. - The Fung AI model case study showed its training phase was 157 times slower than a highly optimized baseline, illustrating the performance penalty for rigorous security. - The integrity of the system is guaranteed because the hardware itself validates the process, meaning even the model owner cannot tamper with the audit trail. - The ultimate goal is to provide a granular, auditable trail that connects every decision back to its source components, satisfying regulatory demands for transparency and accountability. ![Screenshot at 02:24: The visual explicitly shows the comparison between the intentionally moved DEBOMB artifact \(representing the secure process\) and earlier efforts, highlighting the core concept of securing the decision trail against external tampering.](https://ss.rapidrecap.app/screens/GXKPN3NFuxM/00-02-24.png) **Context:** The video introduces a complex problem facing AI deployment, particularly in high-stakes fields like medical diagnosis: the lack of transparency and accountability when an AI decision proves harmful or inaccurate. The speakers discuss the urgent need for a workflow that guarantees full traceability, ensuring that decisions are not only verifiable but also demonstrably untampered, addressing the inherent trust gap between AI systems and human auditors or regulators. ## Detailed Analysis The discussion centers on establishing a rigorous, fully traceable workflow for AI decisions, necessary because current systems lack transparency, creating accountability gaps when failures occur. The proposed solution involves a two-phase process: Training and Inference, secured by a system called DEBOMB (Decision Bill of Material). DEBOMB creates a cryptographically signed, tamper-proof artifact for every decision, linking the final output back to the exact training data, hyperparameter settings, intermediate layer activations, and the specific version of the compiled model used. The authors use the Fung AI model, which classifies mushroom images, as a concrete example, noting that its rigorous, secure training process resulted in a 157x slowdown compared to a baseline model. Critically, this security is hardware-enforced, meaning the hardware validates the process, making the audit trail untamperable even by the model's owner. This comprehensive, granular record allows compliance checkers to automatically audit the decision path, satisfying regulatory demands for explainability and integrity, which is vital for high-stakes applications like medical AI. ### The Two-Phase Workflow - Training phase establishes the traceable model and its comprehensive blueprint - Inference phase generates the verifiable artifact for every decision. ### The DEBOMB Artifact - It contains raw model output, intermediate layer activations, and final probability scores - It is cryptographically linked to the original, untampered model version. ### Fung AI Case Study - Training took 157x longer than the optimized baseline due to rigorous security - The model correctly distinguished poisonous from edible mushrooms. ### Security and Integrity - Hardware validation ensures the process is tamper-proof, protecting against data manipulation or unauthorized modifications to the training process. ### Accountability and Auditing - The system allows auditors to trace every decision metric (accuracy, sensitivity, specificity) back to the exact training run and data set, ensuring compliance with regulations like the EU AI Act. ![Screenshot at 00:00: Introductory screen showing podcast setup and a call to 'Become a Member Today!'](https://ss.rapidrecap.app/screens/GXKPN3NFuxM/00-00-00.png) ![Screenshot at 00:14: Speaker introducing the problem: the urgent need for traceability for AI in high-stakes roles like medical diagnosis.](https://ss.rapidrecap.app/screens/GXKPN3NFuxM/00-00-14.png) ![Screenshot at 00:47: Speaker outlining the need for a solution that moves beyond mere good intentions to verifiable, enforceable systems.](https://ss.rapidrecap.app/screens/GXKPN3NFuxM/00-00-47.png) ![Screenshot at 01:27: Speaker detailing the two phases: Training and Inference, which form the basis of the proposed workflow.](https://ss.rapidrecap.app/screens/GXKPN3NFuxM/00-01-27.png) ![Screenshot at 02:49: Visual emphasis on the cryptographic bubble protecting the data during processing.](https://ss.rapidrecap.app/screens/GXKPN3NFuxM/00-02-49.png) ![Screenshot at 03:37: Speaker explaining the key to the whole system: cryptographic linking back to the original, untampered DEBOMB T.](https://ss.rapidrecap.app/screens/GXKPN3NFuxM/00-03-37.png) ![Screenshot at 04:51: Speaker explaining that the hardware itself holds the developer accountable by validating the process.](https://ss.rapidrecap.app/screens/GXKPN3NFuxM/00-04-51.png) ![Screenshot at 06:34: Speaker summarizing the key takeaway: the workflow proves the model was trained properly and fairly, preventing claims of using different test sets.](https://ss.rapidrecap.app/screens/GXKPN3NFuxM/00-06-34.png)