A Model Audit That Verified the Wrong Representation

An independent plugin opened a coordination model without calling the generator that built it, and without saving a copy. It read four native search sets, three clash tests and all thirteen results, matched 147 held identities one for one, and confirmed two canonical hashes unchanged to the last hex digit [TARGET]. Empty clash identifiers counted to zero. The same note then wrote the limit in one sentence: the check proves that the content is re-readable and selection-closed. It does not prove that concept-proxy geometry has reached construction clash-zero.

Twenty hours of development closed on a split rather than a verdict: keep about 80 per cent of the data-governance and verification assets, rewrite 70 to 90 per cent of the visible drawing expression and the publication layer that carries it [MODELED]. The same sentence says those percentages are an estimate of technical disposition scope, not a figure for how much of the project is done. The split is the finding. It records that the effort was real and that it was aimed at something other than the object under question.

Kept Assets Are Machine-Decidable; Drawings Are Not

Everything on the kept side is decidable by a machine. Object registers, port and connection semantics, digests and sidecars, fail-closed states, isolation of rejected attempts and independent read-back each have a written contract and a pass condition that needs no person. The current execution memory states the same ban in operational form: object counts, hashes, author-written comments, proxy ports or a machine pass do not prove topology, shop drawings or engineering correctness.

Everything on the rewritten side is decidable only by a person reading a sheet: character height at full zoom, line weight, effective viewport area, information density, and whether the symbols carry engineering meaning. The review states the rule in one line. A machine gate proves that the defined machine contract was satisfied, and where that contract does not measure readable character height, line weight, effective viewport area, sheet information density and human semantics, a pass carries no drawing value.

Four evidence layers are written so that they cannot be collapsed. Engineering fact is not a closed data set. A closed data set is not a concept implementation. A concept implementation is not an external release. The thirteen-clash read-back sits in the second layer. Construction clash-zero would sit in the fourth.

Produce a Sketch an Engineer Understands Before Objectifying

What the review proposes is not a stricter checker. It is a sequence: produce one single-line sketch, have an engineer understand it, and only then decide how to objectify and batch it. The actual sequence ran the other way often enough for the archive to carry the evidence in its own directory names, three of which record a machine pass with a visual failure, a layout attempt that failed visually, and a segment-point attempt that failed visually. Directories named after their outcome are why this is a record and not a recollection.

The review is precise about why the order matters rather than the effort. Object keys, port registers and page counts were pursued to completion first, and visual quality control was reached last, so every intermediate pass was measured against a contract that no engineer had yet agreed described a drawing. Terminal proxies later closed one hundred exact connector references. That is single-connector proxy native closure, not a supplier family.

Five Drawing Practices Are Struck Off, Not Improved

Five practices are struck off rather than improved [TARGET]:

• no further visual adjustment of the existing schedule layout;
• no declaring a drawing set complete on sheet count, row count or digest pass;
• no calling generic placeholder geometry, proxy structural members or terminal proxies an engineering level of detail or a supplier family;
• no filling unconfirmed physical parameters with one default value across a sheet;
• no expansion beyond the two pilot drawings until both have passed human review.

Two paths are costed against each other, and that is the decision this record lands on. Continuing to patch the existing twenty-four-sheet publisher is estimated at a further 8 to 20 hours with improved layout as the likely ceiling; restarting the expression layer is estimated at 20 to 36 hours to a two-drawing concept package [MODELED]. The review recommends the restart and attaches a condition that costs nothing and blocks everything: the two drawings become the only precondition gate, so no further sheet, publisher or validator is built until a person has accepted those two.

The estimate rests on a narrow basis and the review says so. It was read-only, started no heavy modelling software, changed no model, executor or existing evidence, and is not an engineer-of-record opinion, a construction release or a procurement approval. The two percentages are its own judgement with no independent recomputation behind them, and they are not two measurements of one quantity: the first counts assets retained, the second counts a layer rewritten, and nothing in the record converts one into the other.

One item sits on the wrong side of the rule the review has just written. The prototype code carrying the corrected values, four placeholder distribution objects with ratings left open and the cooling unit counts separated into mutually exclusive study branches, has not been executed and has not been read back independently. Until it is, the corrected code stands on exactly the evidence the values it corrects stood on. The review names the principal designer as the party who confirms the first gate, and stops there without assigning that run.

Limits and open items

Confirmed: a machine gate proves only that the defined machine contract was satisfied; object counts, hashes, comments, proxy ports or a machine pass do not prove topology, shop drawings or engineering correctness; the independent read-back of the coordination model proves re-readability and selection closure, not construction clash-zero. [FACT]

Disposition estimates, not completion: keep about 80 per cent of the data-governance and verification assets; rewrite 70 to 90 per cent of the visible expression layer; 8 to 20 hours to patch a twenty-four-sheet publisher; 20 to 36 hours to a two-drawing concept package. [MODELED]

Register figures: thirteen clash results with empty-identifier counts of zero; 147 held identities matched; five practices struck off; two drawings as the only precondition gate. [TARGET]

Open: a person named to accept the two drawings; an independent run of the corrected prototype code; any sheet released as a drawing. [HOLD]

Nothing here has been reviewed or sealed by an Engineer of Record. Nothing describes how any installed system behaved.

Who Accepts the Two Drawings Before Any Publisher

The useful test is not whether the hashes still match. Name the person who would accept the two drawings. If that name is missing, the gate that would stop the next publisher has nobody to arm it.

© 2026 K&K Data Service Inc. All rights reserved. Reproduction or republication is permitted only with clear attribution to K&K Data Service Inc. and a working hyperlink to the canonical URL of this article. Excerpts must preserve the technical context, maturity labels, assumptions, and limitations. No excerpt may imply project approval, field validation, certification, or endorsement that the original article does not state.

Source: K&K Data Service Inc., “A Model Audit That Verified the Wrong Representation,” https://www.kkdatasvc.com/lab/errata-and-open-questions/we-were-exactly-verifying-the-wrong-representation/.