# Engineering sketches become living Napkin studies

Historical engineering drawings can seed useful contemporary Napkin studies when the drawing’s design intent is separated from the model’s numeric assumptions. The [HTML report](engineering-napkins/index.html) presents fourteen artifacts, seventy specific chat exchanges and fourteen executable scalar models. Each model asks one engineering question and states its limitations. Every numeric input is a modern illustrative assumption; none claims to reconstruct an original machine. All fourteen examples parsed and ran against current repository source, including sweeps and sensitivity. That verifies execution and arithmetic, while empirical validity still needs matching measurements.

## Artifact dates keep the chronology honest

The collection contains one manuscript, one laboratory notebook, one retrospective published illustration and eleven patent sheets. Leonardo’s [Codex Atlanticus f. 1069r](https://artsandculture.google.com/asset/atlantic-codex-codex-atlanticus-f-1069-recto-leonardo-da-vinci/oAHGKuOp3x_J4A?hl=en) depicts water machinery around 1480–1482. Bell’s [1876 notebook](https://www.loc.gov/resource/magbell.25300201/?sp=22) records an experiment. The [Watt engraving](https://commons.wikimedia.org/wiki/File:SteamEngine_Boulton%26Watt_1784.png) was published in 1878 and depicts a 1784 design. Patent dates in the report are grant dates rather than invention dates.

The substitutions change the engineering interpretation. The Wright [1906 patent](https://patents.google.com/patent/US821393A/en) centers on equilibrium and wing deformation and depicts a glider/control architecture. Goddard’s [1914 drawing](https://patents.google.com/patent/US1102653A/en) depicts staged solid-charge apparatus rather than the 1926 liquid-fuel rocket. Kilby’s [1964 patent sheet](https://patents.google.com/patent/US3138743A/en) is a substitute for his earlier notebook. These are contemporary abstractions of questions suggested by the artifacts, not reconstructed inventor conversations.

## Five exchanges make the sketch testable

The conversation first preserves a source and names one decision, then distinguishes source-supported facts, supplied measurements, assumptions and derived values. It next encodes a short unit-aware sketch with input records, expressions, checks and a scalar sweep. It asks a decision question using analytic execution and numerical sensitivity, then compares the result against matching evidence and revises justified inputs. The [technical notes](../research_notes/Engineering%20sketches%20into%20Napkin/workflows.md) and [full validated examples](../research_notes/Engineering%20sketches%20into%20Napkin/examples.json) retain the reasoning and execution traces.

The current [sketch parser](../src/core/sketch.ts) creates analytical records; it does not automatically recognize a scanned drawing as model geometry. Source metadata and any coarse linked drawing are separate drawing edits. The agent should discover capabilities, inspect the current snapshot and pass the returned revision with mutations. It should use returned record references and rerun after an edit because saved runs belong to a particular revision. Supported boundaries and numerical analysis behavior are documented in the [package guide](../docs/package-guide.md).

## Failed checks reveal the next decision

Thirteen baseline models pass their written checks. The Wright baseline deliberately fails: illustrative lift is **2116.8 N** against **2943 N** of weight at **12 m/s**. Eleven three-case sweeps contain a failing sample; Tesla, Lamarr/Antheil and Engelbart pass all selected cases. A sweep check must hold at every sample, so these failures are useful boundary exploration rather than execution errors. A passing lift check still does not answer wing-warping stability, a passing strut stress check does not establish dome buckling resistance, and a fan-power budget does not establish dust separation performance. These limits are explicit beside each model.

## Start with measurements that can challenge assumptions

Water power and wheel resolution make good first studies because flow, head, diameter and traveled distance can be measured directly. Resistance, shear capacity and pressure drop come next because empirical coefficients depend on operating conditions. Aircraft control, motor fields, turbojet matching, staging and shell stability require a wider analytical boundary and specialist evidence. The best first study is small enough that one measurement can change its conclusion. The historical image supplies an engineering question; measured evidence makes the new sketch model dependable.