📰 Dev.to · Multigrid

The discipline behind deciding which customer record is the real one, in the four architectural styles, and why AI teams meet it whether or not they wanted to.

Change data capture into a graph: idempotent upserts, out-of-order events, deletes that must not delete, and a conflict rule written down before it is needed.

Residency is four separate guarantees that get sold as one. Here is how they come apart, and what to verify for each.

Blocking, pairwise scoring, and a threshold chosen from a precision/recall table with the asymmetric cost of a bad merge worked out on the page.

The specific failure mode of every format you will be handed — PDF, HTML, DOCX, CSV, email, images — with the check that detects each one before it reaches the

A four-cell cost matrix in, an optimal threshold out, derived and then verified against total cost over 10,000 scored transactions.

Exact and near-duplicate detection with MinHash and suffix arrays, n-gram decontamination against your eval sets, and what the deduplication studies reported.

What you must pin to reproduce a result from six months ago, which of those things are not naturally versioned, and the manifest that records all of them.

Nine assertions to run between extraction and indexing, each written as code and each paired with the specific downstream failure it makes impossible.

Every point in a document pipeline where bytes move and somebody is charged, and the cost model those points combine into.

An identifier design that survives merges, splits and re-ingestion: opaque internal ids, an alias table, tombstones, and ids that never stop resolving.

The whole path in one place — ingest, clean, dedupe, chunk, index, verify — as an order of operations with the checkpoint after each step.

Uncertainty from a single dataset with no distributional assumption, worked by hand on ten measurements before the twenty-line script.

A confidence interval computed on a 200-example benchmark, showing arithmetically that a one-point win is indistinguishable from noise.

Five layers of schema context ordered by token cost, the specific wrongness each one removes, and the guardrails that make a generated query safe to run.

The exact line items in public filings that carry AI spending, the API that retrieves them, and why nothing comparable exists for the private labs.

How to get from a question to an answer you can trust, including the verification step that separates a real number from a plausible one.