What the evidence can—and cannot—say

Use separate axes of evaluation.

Replay integrity asks whether the stored instructions reconstruct the expected state. Image reconstruction asks how closely an output matches an admitted reference. Source support asks whether a cited passage supports a particular interpretation. Aesthetic review and artist-style validity ask different questions again.

A study can pass replay checks and still produce an unattractive painting. It can improve an image metric while weakening a source-scoped property. A concept inventory can grow without demonstrating explanatory coverage or an artist-wide style model.

Bind each result to what was actually executed.

Useful records include the engine revision, program and input hashes, dependency and numeric conventions, budgets, seeds, output identities and verification receipts. Lossless pixel comparisons should use the declared decoding and quantization rules rather than screenshots or compressed films.

Authoring and replay have different dependencies. Some paintings, model caches and large experiment outputs exist outside Git. A source checkout or small summary does not establish that a new machine has the artifacts needed to reproduce the study.

Publish a result with its limits.

A technical publication should identify its question, method, exact cohort and comparator, then report failures and abstentions alongside successes. Extra compute or paint should be stated, rather than presented as a matched-budget improvement.

The current research does not establish independently accepted artist-wide style, recovered historical gestures or calibrated physical pigment behavior. These remain research questions. The Academy technical notes summarize inspected repository evidence; they are not peer-reviewed papers or fresh replications of the experiments.

What the records carry