The Instructor creator on why validation beats vibes for reliability.
High Agency Pydantic > VC Backed Frameworks, with Jason Liu
Latent Space 2024
Open latent.space →Models hallucinate because they are trained to always answer; your job is to constrain them, not to hope. The biggest wins in order: ground answers in your own retrieved data with citations (RAG), force structured outputs so responses are validated before display, let the model say "I don't know", and add guardrails plus human review on high-stakes actions. Air Canada and Cursor both learned publicly that you own whatever your bot says, so treat reliability as a launch requirement, not a polish item.
17 resources.
The Instructor creator on why validation beats vibes for reliability.
Latent Space 2024
Open latent.space →The clearest end-to-end explainer of why hallucinations happen and the layered defences that work.
Lakera 2026
Open lakera.ai →Separates factuality from faithfulness errors and matches each to its fix.
Zep 2025
Open getzep.com →A prioritized playbook: grounding and abstention first, verification where errors are expensive.
Future AGI 2026
Open futureagi.com →Production-focused checklist including the stat that layered guardrails cut hallucination rates 71-89%.
Airbyte 2025
Open airbyte.com →Schema-guaranteed responses kill an entire category of made-up fields and invalid values.
OpenAI 2025
Open developers.openai.com →Connects structured output techniques across providers in one practical reference.
Agenta 2025
Open agenta.ai →How the two main open-source guardrail approaches compose into one defence stack.
Guardrails AI 2024
Open guardrailsai.com →The legal analysis of the ruling, in plain enough language for a founder.
McCarthy Tetrault 2024
Open mccarthy.ca →A hot AI startup's own bot invented a policy and triggered cancellations; the cautionary tale of 2025.
The Register Apr 2025
Open theregister.com →The business-impact view of the Cursor incident, including the labelling fix that followed.
Fortune Apr 2025
Open fortune.com →A neutral incident record, and a database worth browsing before you ship anything customer-facing.
AI Incident Database Apr 2025
Open incidentdatabase.ai →Breaks the failure into design decisions you can audit your own bot against.
Envive 2024
Open envive.ai →Short, shareable summary for convincing your team that reliability is a legal issue too.
AI Business 2024
Open aibusiness.com →Validate and auto-retry model outputs against a schema; 6M downloads a month for a reason.
Jason Liu (567 Labs) 2025
Open github.com →Open-source programmable rails: topic control, fact-checking, and jailbreak detection.
NVIDIA 2025
Open github.com →A hub of pre-built validators that intercept bad inputs and outputs before users see them.
Guardrails AI 2025
Open github.com →The same ground, over in Build the product, our Starting Up track.