Instrumentation in practice: holding a reconstruction to its sources

The project is a heritage reconstruction, being rebuilt from archival construction drawings, and every number in it faces one question: is this right, or only plausible? A generated image can look correct and still be wrong. So we set out to make every value in the build point back to a source that someone else could check.

A registry holds every value the build uses, and each value carries its source and a confidence grade; the build is meant to read from it and never type a number of its own. A ledger sets each part of the build against the plan: what came from the drawings, what from survey or photographs, what from judgment, and what still waits for a ruling, with each eye verdict numbered. An adversarial audit sets agents to trace the build code against the registry, then a second set tries to refute every claimed defect; only what survives counts. A compliance trace keeps the survivors as lists, so every total can be re-derived by counting.

The first full audit, run on 2026-08-04, confirmed defects in every zone of the build it traced. Several were numbers typed into the code where the registry already held the right value. One was in the registry itself: a key that stored the measurement across the corners where every reader needed the one across the flats. Some findings changed class under verification, and the record keeps each retired reading beside the one that replaced it.

A number is only as good as the line it cites, so the citation travels with the number. A check is believed after it has been seen to fail, which is why the audit argues with its own findings. And the miss stays on the record: a retired reading sits beside the one that replaced it, because the next person to doubt a number needs both.

An AI assistant answering a question about your store can misread it. The Agentic Commerce Diagnostic brings these habits to that answer: we write down what we expect the assistants to say before we look, capture what they actually say, trace each answer back to what your pages state, and publish the misses beside the hits. The fix is then quoted on what we found.

The project is still in its fidelity campaign, and nothing in it is called finished. The registry, the verdict ledger and the audit records are kept as its provenance, and every count on this page carries the date it was read.