Engineering Investigations

I investigate how AI systems fail, how to measure them, and how to make them trustworthy. The Code Archaeology series reads real code to find where design, implementation, and runtime diverge.

Featured Investigation

Code Archaeology · Part 10

"I Went Looking for a Speculative Abstraction. I Found a Correct Architecture Boundary."

"Code Archaeology Part 10 — The Abstraction That Earned Its Keep: after nine investigations into declared-but-unexecuted features, I went looking for one more — a single-implementation strategy ABC that looked like speculative abstraction. The evidence said otherwise: the interface lives in the domain layer with zero framework dependencies, every consumer depends on the facade, the injection point exists, and the second implementation is not a fantasy — CoVe and NLI are real, active research. This is the investigation where the pattern broke: the abstraction was right."

Finding

"A single-implementation strategy ABC turned out to be a correct dependency-inversion boundary — the series' first positive finding."

Read Investigation

Code Archaeology

Forensic investigations into AI system failures — where architecture, implementation, and runtime diverge.

Code Archaeology · Part 9

HIGH

"The Handover Said 'Not Wired.' The Code Was Wired — to the Wrong Embedding."

"The handover doc said the Review Agent wasn't wired to grounding — it was wired from the start, but to the wrong embedding model."

Code Archaeology

Grounding

Evidence

Review

Embedding

Engineering

"2026-08-02" · "15 min"Read more

Code Archaeology · Part 8

CRITICAL

"The Index Was 1024 Dimensions. The Queries Were 1536. The Search Never Worked."

"The knowledge base was indexed in 1024 dimensions; API queries ran in 1536 — two vector spaces in one system, and search could never work coherently."

Code Archaeology

Embedding

Vector Search

RAG

PGVector

Engineering

"2026-08-02" · "15 min"Read more

Code Archaeology · Part 7

CRITICAL

"The API That Said Yes But Never Wrote Anything: A 201 That Was a Lie"

"POST /articles returned 201 Created, GET returned 404 — the entire API write path had zero commit() calls, and every 'successful' write vanished on session close."

Code Archaeology

Persistence

Transactions

FastAPI

SQLAlchemy

Engineering

"2026-08-02" · "15 min"Read more

Code Archaeology · Part 6

HIGH

"Sixteen Event Types, Zero Subscribers: The Real-Time Architecture Nobody Connected"

"Sixteen typed event types fire into an empty listener list — the WebSocket endpoint has no clients, and the promised audit log is never written."

Code Archaeology

Events

WebSocket

Real-time

Observability

Engineering

"2026-08-01" · "15 min"Read more

Code Archaeology · Part 5

CRITICAL

"Three Audit Trails, Zero Truth: How an AI Audit System Lost Its Own Memory"

"PostgreSQL trace stores exist with zero callers; the audit record lives in a Python list that evaporates on restart."

Code Archaeology

Observability

Audit

Trace

Engineering

"2026-07-20" · "15 min"Read more

Code Archaeology · Part 4

MEDIUM

"The Third Approval Button Never Existed: How MODIFY Died on the Way From Design to UI"

"The three-way approval decision survives only in the enum — the frontend renders two buttons, the API can't carry modifications."

Code Archaeology

HITL

Approval

Engineering

Agents

"2026-07-12" · "15 min"Read more

Code Archaeology · Part 3

HIGH

"The README Said BM25. The Code Had Never Heard of It."

"A complete hybrid retrieval subsystem (RRF + keyword + reranker) has zero production callers; the README claims BM25 that doesn't exist."

Code Archaeology

RAG

Hybrid Search

BM25

Engineering

"2026-07-05" · "15 min"Read more

Code Archaeology · Part 2

HIGH

"I Expected a next_action Chain. The Workflow Engine Had a DAG —?and a Condition Nobody Evaluates."

"Edge.condition is defined and documented but never evaluated — the DAG executes every node unconditionally."

Code Archaeology

AI Agents

Workflow

DAG

Engineering

"2026-06-27" · "15 min"Read more

Code Archaeology · Part 1

CRITICAL

"I Thought My AI Was Evidence-Driven. The Code Proved Me Wrong."

"The Evidence Agent emitted citations pointing at document_id='evidence_source' — a hardcoded string, not a real source."

Code Archaeology

AI Agents

Evidence

Engineering

RAG

"2026-06-20" · "15 min"Read more