We Didn't Invent New AI Architecture. We Rediscovered 20-Year-Old Computer Science.

Over the past few weeks, a series of deep technical threads across developer communities (spanning RAG pipelines, autonomous agents, and protocol verification) converged on a single, humbling realization:

The AI industry isn't inventing a new paradigm. It is frantically rediscovering classical computer science, auditing theory, and distributed systems patterns from 20 years ago—and adapting them to survive a probabilistic engine that lies with confidence.

This post is not about a new framework. It’s an observation of where these "new" AI concepts actually come from, and a thank-you to the engineers doing the hard, unpaid labor of anchoring hyper-probabilistic technology back to deterministic ground truth.

1. "Cryptographically Perfect Hallucinations" & The Oracle Problem

The AI Discussion: Recent discussions around RAG pipelines and agent verification highlighted a critical gap: perfect retrieval doesn't mean a true answer. A model can retrieve a valid chunk and over-interpret it ("Evidence exists ≠ evidence entails"). High retrieval accuracy can even "launder" unsanctioned actions. The community started calling these "cryptographically perfect hallucinations"—clean cryptographic receipts over semantically dead context.

The Root: This is the classic Oracle Problem in cryptography. A signature (Ed25519) proves that data was transmitted and not tampered with, but it cannot prove the data was true to begin with. The AI community didn't need a new term; it just needed to apply the Oracle Problem to LLM agents.

2. The Empty Set Trap & "Capture-Recapture"

The AI Discussion: Engineers recently shared devastating production bugs where a 100% sampling collector returned 0 rows for days while serving hundreds of requests. Dashboards stayed green. The receipt was valid. The population was empty. The system reported success because it didn't independently track what was dropped before processing. The community named this the "Empty Set Trap"—the gap between eligible_seen and population_size.

The Root: This isn't a new AI failure. It's the "Completeness Assertion" in database auditing (ensuring all transactions that should have been recorded actually were) and the "Capture-Recapture" method from statistics (estimating a hidden population). The industry was just re-learning how to audit a denominator.

3. "Veto Heartbeats" & Durable Execution

The AI Discussion: Builders started tracking "Veto Heartbeats"—the realization that a reviewer that has never been seen to fail is indistinguishable from a broken rubber stamp. The proposed fix was tracking the "last time a reviewer said no" as a live, auditable signal.

The Root: The workflow industry recently published pieces like "Agent Workflows Are Rediscovering Durable Execution." BPMN engines have solved this for decades with idempotency, rollback mechanisms, and mandatory audit trails. Modern AI agents are just trying to rebuild BPMN, often poorly.

4. RetractionReceipts & Event Sourcing

The AI Discussion: To handle poisoned agent memory, protocol designers introduced concepts where a previously verified AI action could be transitioned to a REFUTED state without deleting the original cryptographic receipt, preserving forensic history while invalidating the semantic claim.

The Root: This is standard Event Sourcing combined with the W3C PROV-DM (wasInvalidatedBy) data model. It has existed for decades to ensure immutable logs can still reflect changed realities. We just had to adapt it for an LLM that confidently hallucinates.

The Meta-Lesson

If we strip away the LLM hype, we find that the AI engineering community is currently stretching an owl onto a globe (as a Russian idiom goes—forcing unrelated abstract theories onto simple physical facts). We are reinventing Event Sourcing, W3C PROV-DM, and BPMN routing from scratch, often failing to apply the lessons learned from those older systems.

Why? Because 20th-century computer science built deterministic systems. If data was missing, the process threw an exception. LLMs don't throw exceptions. They hallucinate with exit code 0 and a confident tone.

Adapting classical, deterministic verification patterns to survive a probabilistic engine that confidently lies is the actual unpaid labor of the current AI era.

To everyone sharing their production bugs, writing protocol specs, and stress-testing agent harnesses in public: thank you. You aren't just building AI tools; you are doing the hard work of bringing an over-confident technology back to reality.

The answers to our AI verification problems aren't new. But the engine we are bolting them to is. Let's stop reinventing the wheel and start reading the old manuals.

A Note of Thanks (and a Disclaimer)

I wish I could list everyone I’ve argued with, been corrected by, or shared ideas with over these past few weeks. The list is huge, and if I try to name everyone, I will inevitably leave someone out. But you know who you are. Thank you for sharing your production bugs, your telemetry, and your time.

Final disclaimer: Everything written here might be wrong, misinterpreted, or completely obsolete in a few years. Maybe we are just the dinosaurs figuring out how to walk while the meteor is approaching. I don't know. But I'm enjoying the ride.

Let's keep building.