Equal Counts Do Not Prove Two Stores Agree
Why a matching row count is the weakest possible integrity check, and what a ladder of comparisons on identity, content and provenance actually rules out.
Notes from the work
Writing
Notes on the engineering behind the projects on this site: how to make re-running a data load safe, how to check a system against something other than itself, how to tell when a model is confident and wrong, and when it should decline to answer. Each piece carries working code.
Start here
If you read three
Hybrid Retrieval Without Silent Failure
Dense and sparse retrieval fused server-side, embedding contracts that are checked rather than assumed, and refusing to index rather than storing a truncated vector.
Idempotent Ingestion Across Heterogeneous Stores
Deriving identifiers so that a redelivered event converges three stores on the same state, without a distributed transaction and without a deduplication table.
Browse
By subject, or by how much you already know
Topics
Recurring themes
- Verification
- Testing Strategy
- Model Evaluation
- Calibration
- Event-Driven Systems
- Idempotency
- Operations
- Data Integrity
- Provenance
- Reproducibility
- Retrieval
- Embeddings
- RAG
- Data Architecture
- MLOps
- System Design
- Reliability
- Lifecycle Design
- API Design
- Document Processing
- LLM Engineering
- OCR
- Interpretability
- Computer Vision
- Selective Prediction
The library grows deliberately. A piece is published when the underlying work is finished and its limitations are known, not on a schedule.
Published
More from the library
Building an Independent Verification Oracle
A checker that shares its subject's assumptions cannot falsify anything. How to build a second implementation independent enough to disagree, and how to prove it.
Calibration Is Not Classification Accuracy
A model can be accurate and badly calibrated, or well calibrated and useless. What each property means, how to measure both, and why thresholds are a decision.
Designing Replay as a Normal Path
Offset handling, publish-then-commit ordering, and broker confirmation before HTTP success: the decisions that make re-running a stream routine, not an incident.
Evidence Provenance for Knowledge Systems
What a stored generation has to carry so a claim about it can be re-checked later: captured bytes, ruleset digests, and the difference between recorded and installed.
Promotion Gates for Model Release
Promotion is not the next box along a diagram. It is a conjunction of independent checks, and one short check stops the candidate where it stands.
Production AI Systems Need Refusal Paths
Every stage that can degrade should be able to decline. Where refusal belongs across ingestion, retrieval, extraction and inference, and what omitting it costs.
03Elsewhere
Recent work by others in these areas
Listings from arXiv, shown as published metadata only. These papers are not mine and nothing here is a summary of them — follow a link to read the authors' own abstract.
Last updatedList may be out of date
DICS: Data-Informed Centroid Splitting for Decision Tree Classifiers
Transfer Learning in Nonparametric Regression with Deep ReLU Networks
Exact Algebraic Computation of Learning Coefficients for Two-Dimensional Singular Models
Discrete Diffusion Inference-Time Control with Nested Sequential Monte Carlo
Information on trajectories: martingales and random times
$TCP_α$: Margin-Controlled Confidence estimation for reliable Music Information Retrieval
A comparison between ceiling-mounted FMCW, IR-UWB and Wi-Fi radar for in-bedroom human activity monitoring and sleep interruption detection
AI4AI-Bench: Benchmarking LLM Agents in Algorithmic Design for Recursive Self-Improvement
Projecting BrowseComp-Plus onto ClimbMix: Toward More Realistic Corpora for Agentic Search
Thank you to arXiv for use of its open access interoperability. arxiv.org